The Lindahl Letter

Greetings, readers and avid listeners and technology enthusiasts! You're either reading or tuned in to the audio-only podcast of The Lindahl Letter, now in its 154th week. Remember, an extra fresh and original edition lands in your inbox every Friday. Now certified with a three year proven track record. Today, we delve into an intriguing theme: “My 2024 Predictions.” Let's explore the future together.1. Generative AI's Expanding Horizons: We're on the brink of witnessing a generative AI leap forward into agency and actions. The upcoming wave is set to introduce more nuanced language models and sophisticated image generators. Imagine a world where content creation and design are revolutionized, and software development is seamlessly intuitive. A standout prediction? Micro-targeting will become a staple, with chat agents offering highly personalized experiences, finely attuned to individual preferences and interests. I think it's actually going to get uncomfortable with how far people are going to take targeting in 2024. 2. The Evolution of AI-Powered Automation: AI's influence in automation is deepening its roots across various sectors. We'll analyze potential milestones in logistics, retail, and online services. Could these innovations redefine job roles and reshape business workflows? Let's ponder the possibilities. I think people are going to jump in and start automating all sorts of things that might deserve automation and others that maybe should have waited for a more mature point in the technology development curve.3. AI Ethics and Legislative Landscape: As AI entwines more with our daily lives, the drumbeat for ethical standards and regulations grows louder. We'll reflect on how various nations might navigate the governance of AI and its impact on global AI development and cooperation. Don’t worry this won’t be a purely legislative capture point of view or a comparative political analysis.4. AI policy may abound: AI's Role in Government and Public Sector is going to increase. AI's potential in enhancing public administration, policy formulation, and citizen services is immense. The discussion will spotlight AI's applications in public safety, urban planning, and social welfare initiatives, marking a significant shift in governmental functions. I think things on this front are going to get moving at a rapid speed in 2024. 5. Quantum Computing Breakthroughs: Turning our gaze to quantum computing, I’ll speculate on its role in tackling problems that are currently beyond classical computing's reach. From cryptography to material science, the implications are vast and profound. Maybe 2024 is the year people use some of that IBM quantum computing and share the results. 6. AI in Healthcare - Next Frontiers: The healthcare sector is ripe for AI-driven innovation. We'll dive into expected advancements in personalized medicine, advanced diagnostics, and AI's emerging role in streamlining healthcare administration and enhancing patient care.What’s next for The Lindahl Letter? * Week 155: Generative AI's Expanding Horizons* Week 156: The Evolution of AI-Powered Automation* Week 157: AI Ethics and Legislative Landscape* Week 158: AI policy may abound* Week 159: Quantum Computing Breakthroughs* Week 160: AI in HealthcareIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you for joining us this week. Stay curious, stay informed, and enjoy the week ahead! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Please accept this note of gratitude for being a part of this journey. We made it. Some of you have been a part of the entire journey. Three years ago on January 26, 2021, this very Substack started rocking and rolling along a weekly journey to share research notes. My backlog of weekly items to cover is still pretty darn larger. It contains well over 100 blocks of potential writing content. A few larger projects are lurking within that backlog that could be stacked up block by block into future books, manuscripts, or larger articles. I’m feeling reflective today about the nature of and future of Stubstack as a publishing platform. A lot of writers flocked to Substack and it as a platform has certainly helped nurture independent writing. Newsletters have come a long way over the years, but for the most part they are fundamentally the same asymmetric writer to audience communication method. Within the situation we are experiencing at the moment it appears to be Substack as a platform that has changed. I believe and consider it to be true that sunlight has always been the best disinfectant. Understanding begins the path to knowledge. Substack as a platform is at a crossroads. It’s my guess that the platform that is Substack will radically change in 2024. To be honest about that change, I’d have to say I’m not entirely sure what will happen [1]. Generally, I’m going to keep writing and publishing until a move to my core WordPress domain is required [2]. Everything is all set up over at that domain just in case things have to be moved, but I’m legitimately hoping that the Substack community survives the year. My corpus of writing is well over 5 million words and while none of them are particularly spicy or super eventful they were written to be shared. You can tell here now that you made it to the third paragraph that this missive has gone back to my previous writing strategy and is not reduced into highly curated bullet points. That was something that I tried out to see if that modern communication strategy would work for my research notes. I think my preference going forward will be to write in a more long form communication structure. Bullet points have a place and are great for reducing large amounts of information into something more palatable. My writing generally has been more about being a self contained research note that brings forward a degree of understanding about something complex. That will most certainly be the standard going forward. Things that catch my attention are going to receive coverage and that will be the ongoing basis of each weekly Lindahl Letter.Links and thoughts:* I listed to Hard Fork this week “The Times Sues OpenAI + A Debate Over iMessage + Our New Year’s Tech Resolutions” * I read this paper “Mixtral of Experts” https://arxiv.org/pdf/2401.04088.pdf Footnotes:[1] I read this article from Platformer:[2] https://www.nelslindahl.com/ or here https://www.nelslindahl.com/weblog/ What’s next for The Lindahl Letter? * Week 154: My 2024 predictionsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Maybe a longer title for this post could be, “Bayesian Models and Elections: A Dive into the Dance of Uncertainty.” This is the 150th transmission of the Lindahl Letter.In the vast and often unpredictable theater of electoral forecasting, the quest for precision is a relentless pursuit. The choreography of voter behavior is a complex ballet, orchestrated by a myriad of factors—societal tremors, economic tides, the charisma of candidates, and the machinations of campaign strategies. Amidst this swirling cauldron of variables, the call for a more nuanced forecasting method is loud and clear. And what answers the call with a finesse born of probabilistic reasoning is the realm of Bayesian models. These statistical marvels stand at the confluence of data and uncertainty, offering a refined lens to dissect the electoral enigma.The essence of Bayesian statistics, a legacy of Thomas Bayes, is a narrative of evolving beliefs in the face of emerging evidence. It's a realm where estimates aren't static, but dynamic, continually reshaped by the rhythm of new data—a narrative that resonates with the pulsating heart of electoral dynamics.In the Bayesian narrative, the tale begins with initial beliefs, our prior probabilities. As the story unfolds with new data—the likelihood—our beliefs morph, culminating in updated beliefs or posterior probabilities. This dance of iterative learning is akin to the dynamism of electoral scenarios, where a single debate, policy announcement, or campaign rally could tilt the scales of public sentiment.A compelling act in the Bayesian play is its ability to weave historical election data into the forecasting fabric. It’s not just about the now, but a dialogue with the past, understanding how the ghost of incumbency, the whisper of economic indicators, or the shout of demographic shifts have choreographed electoral outcomes before.And then, there’s the magnum opus of Bayesian models—the articulation of uncertainty. Unlike the static snapshot often rendered by traditional polling, Bayesian models compose a symphony of probability distributions. They unveil a spectrum of possible electoral outcomes, each with its associated probability, painting a picture of electoral reality that's as rich as it is realistic.The spotlight often falls on case studies like the 2012 and 2016 U.S. Presidential Elections, where the Bayesian choreography, as orchestrated by platforms like Nate Silver’s FiveThirtyEight, navigated the electoral tumult with a commendable degree of accuracy. By embracing uncertainties and dancing with historical context, Bayesian models orchestrate a forecast that traditional polling methods seldom match.Yet, the narrative isn’t without its share of cliffhangers. The hurdles of data scarcity, model misspecification, and computational intricacies are challenges that beckon solutions. Despite these, the Bayesian voyage into electoral forecasting holds a promise—of rendering narratives that are not only statistically sound but resonate with intuitive clarity.As the electoral saga continues to unfold, the allure for better forecasting tools is a relentless whisper. Bayesian models, with their eloquence in narrating the dance of uncertainty, emerge as potent companions for pollsters and policymakers. They underline an electoral truism—in a realm replete with uncertainties, understanding and embracing these uncertainties isn’t just the hallmark of wisdom, but a cornerstone of robust electoral forecasting.A few scholarly articles I found interesting this week:Linzer, D. A. (2013). Dynamic Bayesian forecasting of presidential elections in the states. Journal of the American Statistical Association, 108(501), 124-134. https://www.ocf.berkeley.edu/~vsheu/Midterm%202%20Project%20Files/Linzer-prespoll-May12.pdf Lock, K., & Gelman, A. (2010). Bayesian combination of state polls and election forecasts. Political Analysis, 18(3), 337-348. https://academiccommons.columbia.edu/doi/10.7916/D88K7GV1/download Heidemanns, M., Gelman, A., & Morris, G. E. (2020). An updated dynamic Bayesian forecasting model for the US presidential election. Harvard Data Science Review, 2(4), 10-1162. https://assets.pubpub.org/wbec6d9k/9dfc3335-6d48-4f8e-bf5d-0011c7817a09.pdf Olsson, H., Bruine de Bruin, W., Galesic, M., & Prelec, D. (2021). Election polling is not dead: a Bayesian bootstrap method yields accurate forecasts. Preprint at https://osf.io/nqcgs/ If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Title: The Confluence of Agent Systems and Expert Opinion in Election SimulationsIn the ever-evolving landscape of political science and technology, curiosity often paves the way for innovative approaches and fresh perspectives. Recently, a wave of curiosity has washed over me, primarily centered around the exploration of diverse agent systems to simulate elections. This intersection of technology and electoral processes opens up new realms of possibilities, allowing us to mimic, analyze, and potentially enhance our understanding of elections in a simulated environment.Agent systems provide a powerful tool for creating algorithmic or model-based simulations. These systems can be meticulously crafted, incorporating synthetic focus groups and panels that emulate real-world election scenarios. The meticulous design allows for the creation of adversarial agents that can engage in debates or various activities that accurately mirror the complexities and dynamics of an actual election. However, my curiosity doesn’t end here. I’m also deeply intrigued by the fusion of election simulation with expert opinion systems. By appending the term ‘systems’ to ‘expert opinion’, the concept transcends beyond individual viewpoints, fostering an environment where aggregated expert opinions are diligently worked upon and analyzed. These collected data become a potent resource, providing invaluable insights that can be seamlessly integrated into the simulated election models.Imagine the immense potential unlocked by the combination of these two realms. The simulated agents, fortified with synthesized expert opinions, could operate in a nuanced manner that echoes the depth and diversity of actual election contenders and voters. These enhanced agents could engage in debates, make decisions, and navigate the election simulation with a level of sophistication that brings us closer to understanding the myriad factors influencing election outcomes.Through this amalgamation, the simulation becomes a crucible where technological prowess meets the wisdom of expertise. The interplay between algorithmically driven simulations and the rich reservoir of expert opinions can unveil unprecedented avenues for exploring electoral processes. It can deepen our comprehension, offering a clearer lens through which we may view the multifaceted realms of elections.In conclusion, the fusion of various agent systems with expert opinion systems presents a promising frontier in the world of election simulations. By harnessing the collective wisdom of experts and embedding this knowledge within algorithmic agents, we stand on the brink of developing more nuanced, realistic, and insightful election simulation models. The curiosity driving this exploration is not merely a personal quest, but rather a shared journey towards enriching our understanding and approaches to simulating and analyzing electoral processes.What’s next for The Lindahl Letter? * Week 147: Bayesian Models* Week 148: Running Auto-GPT on election models* Week 149: Sentiment Analysis* Week 150: Voter ModelsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Let’s get to work unpacking the Delphi Method & Door-to-Door Canvassing. This is the Substack deep dive you don’t want to miss.Hey there, dear weekly Substack readers!Today, let's embark on a journey through two fascinating potentially election related realms: the Delphi Method and Door-to-Door Canvassing. While these two concepts might seem worlds apart, there is a curious intersection between the two of them that's worth exploring. So, grab your favorite beverage, and let's dive in! I had two shots of espresso and they were delightful. Delphi Method: More Than Just Ancient GreeceFirst up, the Delphi Method. No, we're not time-traveling back to ancient Greece using the world’s finest Delorean, but we are diving into a method inspired by the oracle of Delphi. It's a structured communication technique designed for interactive forecasting. Picture this: a group of experts, multiple rounds of questionnaires, and a quest for consensus. The beauty of this method? It taps into collective intelligence while ensuring every expert voice gets its moment in the sun, minus the overshadowing by dominant personalities. Generally this method is unlikely to occur naturally today due in part to the overwhelming decay in civility that has occurred. People just don’t cross over into different parisian camps these days. Something has distinctly changed in our politics. Prokesch, T., Von der Gracht, H. A., & Wohlenberg, H. (2015). Integrating prediction market and Delphi methodology into a foresight support system—Insights from an online game. Technological Forecasting and Social Change, 97, 47-64 [1]. Dalkey, N. C., Brown, B. B., & Cochran, S. (1969). The Delphi method: An experimental study of group opinion (Vol. 3, p. 107). Santa Monica, CA: Rand Corporation [2].Door-to-Door Canvassing: Old School politics, But otherwise pure GoldSwitching gears, let's talk about the age-old art of door-to-door canvassing. It's personal, it's direct, and it's all about that face-to-face interaction. Whether it's political volunteers rallying support or grassroots movements gathering opinions, this method has stood the test of time. Why? Because nothing beats the authenticity of a real conversation. Sure people are less likely to want to answer the door or talk politics at the front door, but this method does still show signs of working.Green, D. P., Gerber, A. S., & Nickerson, D. W. (2003). Getting out the vote in local elections: Results from six door-to-door canvassing experiments. The Journal of Politics, 65(4), 1083-1096 [3]. The Unexpected CrossoverNow, for the fun part. How do these two methods intertwine?Imagine harnessing the Delphi Method's expert-driven insights to supercharge a door-to-door canvassing campaign. Before our canvassers even lace up their shoes, we could have a panel of experts—from veteran canvassers to communication gurus—forecasting the best strategies, pinpointing challenges, and highlighting golden opportunities.That type of Magic could happenMarrying the Delphi Method's structured insights with the grassroots authenticity of door-to-door canvassing could:* Elevate the Message: Crafting narratives that truly resonate.* Stay Two Steps Ahead: Predicting and preparing for potential challenges.* Strategize Like a Pro: Directing efforts where they count the most.Wrapping UpMerging the old with the new, the traditional with the innovative, can lead to some unexpected and powerful synergies. And isn't that what we're all about here on Substack? Exploring, questioning, and connecting the dots in unexpected ways.Stay curious, and until next time!Dr. Nels LindahlFootnotes:[1] https://www.researchgate.net/profile/Heiko-Von-Der-Gracht/publication/260755254_Integrating_prediction_market_and_Delphi_methodology_into_a_foresight_support_system_-_Insights_from_an_online_game/links/5c339e35299bf12be3b5592a/Integrating-prediction-market-and-Delphi-methodology-into-a-foresight-support-system-Insights-from-an-online-game.pdf [2] https://apps.dtic.mil/sti/trecms/pdf/AD0690498.pdf [3] https://d1wqtxts1xzle7.cloudfront.net/45996527/Getting_Out_the_Vote_in_Local_Elections_20160527-16511-1wf5rrd-libre.pdf?1464362918=&response-content-disposition=inline%3B+filename%3DGetting_Out_the_Vote_in_Local_Elections.pdf&Expires=1696684199&Signature=S6S9UNmWYRMepopnbBWQlGkCn4q4C889yqi3aoE~-47Z~DL2Hpw5TWKDz6Cq4IF9~gp-sfEPaVehWkrW7YiYQLLL0f6XEsNNtlU3WUl4NSee2JH1B2CNTWcy9glqPjVo6KBfe6oKUYr4YlCatCXgDgJEL~HtRsIiwswn4XxGpWAv~7sLT-X5M8Zc13wVlYl8MEzNF32WpOM5JaJUtUA8Z5k8G2cMgHWzRRYyB6GXf1Pr2MWovSCameHEHC~G44wcCYoK-54jdYdnP605msL6gifKpj0dp58ETNOcFnBvCVwU5hjfUcIgjCJLpDTlXV7OLokXfL3uydrisDoN~H3QZA__&Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Don’t panic, the Google Scholar searches are coming in fast and furious on this one [1]. We had a footnote in the first sentence today. Megan Tomlin writing over at neo4j had probably the best one line definition of the difference by noting that knowledge graphs are going to be in the human readable data camp and vector databases are more of a black box [2]. I actually think that eventually one super large knowledge graph will emerge and be the underpinning of all of this, but that has not happened yet given that the largest one in existence Google holds will always remain proprietary. Combining two LLMs… right now you could call them one after another, but I’m not finding an easy way to pool them into a single model. I wanted to just say to my computer, “use Baysian pooling to combine the most popular LLMs from Hugging Face,” but yeah that is not an available command at the moment. A lot of incompatible content is being generated in the vector database space. People are stacking LLMs and working in sequence or making parallel calls to multiple-models. What I was very curious about was how to go about the process of merging LLMs, combining LLMs, actual model merges, ingestion of models, or even a method to merge transformers. I know that is a tall order, but it is one that would take so much already spent computing cost and move it from sunk to additive in terms of value. A few papers exist on this, but they are not exactly solutions to this problem. Jiang, D., Ren, X., & Lin, B. Y. (2023). LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion. arXiv preprint arXiv:2306.02561. https://arxiv.org/pdf/2306.02561.pdf you can see more content related to this one here https://yuchenlin.xyz/LLM-Blender/Wu, Q., Bansal, G., Zhang, J., Wu, Y., Zhang, S., Zhu, E., ... & Wang, C. (2023). AutoGen: Enabling next-gen LLM applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155. https://arxiv.org/pdf/2308.08155.pdf Chan, C. M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., ... & Liu, Z. (2023). Chateval: Towards better llm-based evaluators through multi-agent debate. arXiv preprint arXiv:2308.07201. https://arxiv.org/pdf/2308.07201.pdf Most of the academic discussions and even the cutting edge papers like AutoGen are about orchestration of models instead of merging, combining, or ingestion of many models into one. I did find a discussion on Reddit from earlier this year about how to merge the weights of transformers [3]. It’s interesting what things end up on reddit. Sadly that subreddit is closed due to a dispute over 3rd party plugins. Exploration into merging and combining Large Language Models (LLMs) is indeed at the frontier of machine learning research. While academic papers like "LLM-Blender" and "AutoGen" offer different perspectives, they primarily focus on ensembling and orchestration rather than true model merging or ingestion. The challenge lies in the inherent complexities and potential incompatibilities when attempting to merge these highly sophisticated models.The quest for effectively pooling LLMs into a single model or merging transformers is a journey intertwined with both theoretical and practical challenges. Bridging the gap between the human-readable data realm of knowledge graphs and the more opaque vector database space, as outlined in the beginning of this podcast, highlights the broader context in which these challenges reside. It also underscores the necessity for a multidisciplinary approach, engaging both academic researchers and the online tech community, to advance the state of the art in this domain.In the upcoming weeks, we will delve deeper into the community-driven solutions, and explore the potential of open-source projects in advancing the model merging discourse. Stay tuned to The Lindahl Letter for a thorough exploration of these engaging topics.Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=knowledge+graph+vector+database&btnG= [2] https://neo4j.com/blog/knowledge-graph-vs-vectordb-for-retrieval-augmented-generation/ [3] https://www.reddit.com/r/MachineLearning/comments/122fj05/is_it_possible_to_merge_transformers_d/ What’s next for The Lindahl Letter? * Week 145: Delphi method & Door-to-door canvassing* Week 146: Election simulations & Expert opinions* Week 147: Bayesian Models* Week 148: Running Auto-GPT on election models* Week 149: Modern Sentiment AnalysisIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

After the adventures of last week, I started this writing adventure wanting to try to figure out what people were doing with LangChain and social media. People are both generating content for social media using LLMs and oddly enough repurposing content as well. We have to zoom out for just a second and consider the broader ecosystem of content. In the before-times, people who wanted to astroturf or content farm had some work to do within the content creation space. Now ChatGPT has opened the door and let the power of synthetic content creation loose. You can create personas and just have them generate endless streams of content. People can even download and run models trained for this purpose. It’s something I’m legitimately worried about for this next election cycle. Sometimes I wonder how much content within the modern social media spaces is created artificially. Measuring that is actually pretty difficult. It’s not like organically created content gets a special badge or recognition. For those of you who were interested in finding out insights on any topic with a plugin that works with the OpenAI ChatGPT system then you could take a moment and install “The Yabble ChatGPT Plugin” [1]. Fair warning on this one I had to reduce my 3 plugins down to just Yabble and be pretty explicit in the prompts within ChatGPT to make it do some work. Sadly, I could not just login to Yabble and had to book a demo with them to get access. Stay tuned on that one to get more information on how that system works. I had started by searching out plugins to have ChatGPT analyze social media. This has become easier now with the announcements that OpenAI can openly use Bing search [2]. Outside of searching using any OpenAI tooling like ChatGPT, Google was pretty clear on the reality that what I was really looking for happened to actually be marketing tools. Yeah, I went down the SEO Assistant rabbit hole and it was shocking. So much content exists in this space that is like watching a very full ant farm for the most part. Figuring out where to jump in without getting scammed is probably a questionable decision framework. Whole websites and ecosystems could be synthetically generated pretty quickly. It’s not exactly one click turn key deployments, but it is getting close to that level of content farming.I was willing to make the assumption that people who were going to the trouble of making actual plugins for ChatGPT within the OpenAI platform are probably going to be more interesting and maybe are building actual tooling. For those of you who are using ChatGPT with OpenAI and have the plus subscription you just have to open a new chat, expand the plugin area, and scroll down to the plugin store to search for new ones…I also did some searches for marketing tools. I’m still struck with the possibility that a lot of content is being created and marketed to people. It’s not the potential flooding of content that becomes so overwhelming that nobody is able to navigate the internet anymore. We are getting very close to the point where it would be entirely possible for the flooding of new content to occur in ways that simply overwhelm everybody and everything with new content. This would be like the explosion of ML/AI papers over the last 5 years, but maybe 10x or 100x even that digital content boom [3].Footnotes:[1] https://www.yabble.com/chatgpt-plugin[2] https://www.reuters.com/technology/openai-says-chatgpt-can-now-browse-internet-2023-09-27/ [3] https://towardsdatascience.com/neurips-conference-historical-data-analysis-e45f7641d232 What’s next for The Lindahl Letter? * Week 144: Knowledge graphs vs. vector databases* Week 145: Delphi method & Door-to-door canvassing* Week 146: Election simulations & Expert opinions* Week 147: Bayesian Models* Week 148: Running Auto-GPT on election modelsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

And now it’s time to pivot toward the “Learning LangChain” topic…The most fun introduction to LangChain seems to be from DeepLearning.ai with Andrew Ng and Harrison Chase [1]. You can expect to spend a couple of hours to complete the process of watching the videos and absorbing the content. Make sure you use a browser window large enough to support both the jupyter notebook and the video. You are probably going to want these items to run side by side. This course covers models, prompts, parsers, memory, chains, and agents. The part of this learning package that I was the most interested in learning more about was how people are using agents and of course what sort of plugins could that yield as use cases in the generative AI space. Going forward I think agency will be the defining characteristic of the great generative AI adventure. These applications are going to do things for you and some of those use cases are going to be extremely powerful. After that course I wanted to dig in more and decided to go ahead and learn everything I could from the LangChain AI Handbook [2]. This handbook has 6 or 7 chapters depending on how you count things. My favorite part about this learning build is that they are using Colab notebooks for hands-on development during the course of the learning adventure. That is awesome and really lets you get going quickly. A side quest spawned out of that handbook learning which involved starting to use Pinecone in general which was interesting. You can do a lot with the Pinecone including building AI agents and chatbots. I’m going to spend some time working on the udemy course “Develop LLM powered applications with LangChain” later this weekend [3]. You can also find a ton of useful information within the documentation for LangChain including a lot of content about agents [4].You might now be wondering what alternatives to LangChain exist… I started looking around at AutoChain [5], Auto-GPT [6], AgentGPT [7], BabyAGI [8], LangDock [9], GradientJ [10], Flowise AI [11], and LlamaIndex [12]. Maybe you could also consider TensorFlow to be an alternative. You can tell from the combination of companies and frameworks being built out here a lot of attention is on the space between LLMs and taking action. Getting to the point of agency or taking action is where these spaces are gaining and maintaining value. Footnotes:[1] https://learn.deeplearning.ai/langchain/lesson/1/introduction [2] https://www.pinecone.io/learn/series/langchain/ [3] https://www.udemy.com/course/langchain/[4] https://python.langchain.com/docs/modules/agents/ [5] https://github.com/Forethought-Technologies/AutoChain [6] https://github.com/Significant-Gravitas/Auto-GPT[7] https://github.com/reworkd/AgentGPT[8] https://github.com/miurla/babyagi-ui [9] https://www.langdock.com/ [10] https://gradientj.com/[11] https://flowiseai.com/[12] https://www.llamaindex.ai/What’s next for The Lindahl Letter? * Week 143: Social media analysis* Week 144: Knowledge graphs vs. vector databases* Week 145: Delphi method & Door-to-door canvassing* Week 146: Election simulations & Expert opinions* Week 147: Bayesian ModelsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

You can feel the winds of change blowing and the potential of people building out election expert opinion chatbots. Maybe you want to know what they are probably going to use to underpin that sort of effort. If you were going out to build some generative AI chatbots for you might very well use one of the 5 systems we are going to dig into today.* Voiceflow - This system may very well be the most prominent of the quick to market AI agent building platforms [1]. I have chatbots deployed to both Civic Honors and my main weblog powered by Voiceflow.* LangFlow - You are going to need to join the waitlist for this one to get going [2]. I’m still on the waitlist for this one… * Botpress - Like Voiceflow this system lets you pretty quickly jump into the building process of actual chatbot workflows [3]. To be fair with this one I was not able to build and deploy something into production within minutes, but you could do it pretty darn quickly if you had a sense of what you were trying to accomplish. I built something on Botpress and it was pretty easy to use. After login I clicked answer questions from websites to create a bot. I added both Civic Honors and my main Nels Lindahl domain. They just jumped in and advised me that the knowledge upload was complete. Publishing the bot is not as low friction as the Voiceflow embedding launch point, but it was not super hard to work with after you find the share button.* FloWiseAI - You will find this is the first system on the list that will require you to get out of your web browser, stretch a bit, and open the command line to get this one installed with a rather simple “npm install -g flowise” command [4]. I watched some YouTube videos on how to install this one and it almost got me to flip over into Ubuntu Studio. Instead of switching operating systems I elected to just follow the regular Windows installation steps.* Stack AI - With this one you are right back into the browser and you are going to see a lot of options to start building new projects [5].All of these chatbots built using a variety of generative AI models are generally working within the same theory of building. The conversation is being crafted with a user and some type of exchange with a knowledge base. For the most part the underlying LLM is being used to facilitate the conversational part of the equation while some type of knowledge base is being used to gate, control, and drive the conversation based on something deeper than what the LLM would output alone. It’s an interesting building technique and one that would not have been possible just a couple of years ago, but the times have changed and here we are in this brave new world where people can build, deploy, and be running a generative AI chatbot in a few minutes. It requires some planning about what is being built, you need some type of knowledgebase, and the willingness to learn the building parameters. None of that is a very high bar to pass. This is a low friction and somewhat high reward space for creating conversational interactions. Messing around with all these different chatbot development systems made me think a little bit more about how LangChain is being used and what the underlying technology is ultimately capable of facilitating [6]. To that end I signed up for the LangSmith beta they are building [7]. Sadly enough “LangSmith is still in closed beta” so I’m waiting on access to that one as well. During the course of this last week I have been learning more and more about how to build and deploy chatbots that take advantage of LLMs and other generative AI technologies. I’m pretty sure that the development of agency to machine learn models is going to strap rocket boosters to the next stage of technological deployment. Maybe you are thinking that is hyperbole… don’t worry or panic, but you are very soon going to be able to ask these agents to do something and they will be able to execute more and more complex actions. That is the essence of agency within the deployment of these chatbots. It’s a very big deal in terms of people doing basic task automation and it may very well introduce a distinct change to how business is conducted by radically increasing productivity. Footnotes:[1] https://www.voiceflow.com/ [2] https://www.langflow.org/ [3] https://botpress.com/ [4] https://flowiseai.com/ [5] https://www.stack-ai.com/ [6] https://www.langchain.com/ [7] https://www.langchain.com/langsmith What’s next for The Lindahl Letter? * Week 142: Learning LangChain* Week 143: Social media analysis* Week 144: Knowledge graphs vs. vector databases* Week 145: Delphi method & door to door canvasing* Week 146: Election simulationsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonu

Sometimes a simplified model of something is easier to work with. We dug into econometric models recently during week 136 and they can introduce a high degree of complexity. Even within the world of econometrics you can find information about proxy models. In this case today we are digging into proxy models for elections. My search was rather direct. I was looking for a list of proxy models being used for elections [1]. I was trying to dig into election forecasting proxy models or maybe even some basic two step models. I even zoomed in a bit to see if I could get targeted on machine learning election proxy models [2].After a little bit of searching around it seemed like a good idea to maybe consider what it takes to generate a proxy model equation to represent something. Earlier I had considered what the chalk model of election prediction would look like with using a simplified proxy of voter registration as an analog for voting prediction. I had really thought that would end up being a highly workable proxy, but it was not wholesale accurate. Here are 3 papers I looked at this week:Hare, C., & Kutsuris, M. (2022). Measuring swing voters with a supervised machine learning ensemble. Political Analysis, 1-17. https://www.cambridge.org/core/services/aop-cambridge-core/content/view/145B1D6B0B2877FC454FBF446F9F1032/S1047198722000249a.pdf/measuring_swing_voters_with_a_supervised_machine_learning_ensemble.pdf Zhou, Z., Serafino, M., Cohan, L., Caldarelli, G., & Makse, H. A. (2021). Why polls fail to predict elections. Journal of Big Data, 8(1), 1-28. https://link.springer.com/article/10.1186/s40537-021-00525-8 Jaidka, K., Ahmed, S., Skoric, M., & Hilbert, M. (2019). Predicting elections from social media: a three-country, three-method comparative study. Asian Journal of Communication, 29(3), 252-273. http://www.cse.griet.ac.in/pdfs/journals20-21/SC17.pdf I spent some time messing around with OpenAI’s GPT-4 on this topic. That effort drove down to a few proxy models that are typically used. The top 10 seemed to be the following: social media analysis, google trends, economic indicators, fundraising data, endorsement counts, voter registration data, early voting data, historical voting patterns, event-driven, and environmental factors. Combining all 10 proxy models into a single equation would result in a complex, multi-variable model. Here's a simplified representation of such a model:E=α1​(S)+α2​(G)+α3​(Ec)+α4​(F)+α5​(En)+α6​(VR)+α7​(EV)+α8​(H)+α9​(Ed)+α10​(Ef)+βWhere:* E is the predicted election outcome.* α1, α2​,...α10 are coefficients that determine the weight or importance of each proxy model. These coefficients would be determined through regression analysis or other statistical methods based on historical data.* S represents social media analysis.* G represents Google Trends data.* Ec represents economic indicators.* F represents fundraising data.* En represents endorsement count.* VR represents voter registration data.* EV represents early voting data.* H represents historical voting patterns.* Ed represents event-driven models.* Ef represents environmental factors.* β is a constant term.This equation is a linear combination of the proxy models, but in reality, the relationship might be non-linear, interactive, or hierarchical. The coefficients would need to be determined empirically, and the model would need to be validated with out-of-sample data to ensure its predictive accuracy. Additionally, the model might need to be adjusted for specific elections, regions, or time periods. It would be interesting to try to pull together the data to test that type of complex multivariable model. Maybe later on we can create a model with some agency designed to complete that task.Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=election+proxy+models&btnG=[2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=election+proxy+models+machine+learning&btnG=What’s next for The Lindahl Letter? * Week 141: Building generative AI chatbots* Week 142: Learning LangChain* Week 143: Social media analysis* Week 144: Knowledge graphs vs. vector databases* Week 145: Delphi methodIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

This might be the year that I finally finish that book about the intersection of technology and modernity. During the course of this post we will look at the intersection of machine learning and election models. That could very well be a thin slice of the intersection of technology and modernity at large, but that is the set of questions that brought us here today. It’s one of things we have been chasing along this journey. Oh yes, a bunch of papers exist related to the topic this week of machine learning and election models [1]. None of them are highly cited. A few of them are in the 20’s in terms of citation count, but that means the academic community surrounding this topic is rather limited. Maybe the papers are written, but have just not arrived yet out in the world of publication. Given that machine learning has an active preprint landscape that is unlikely. That darth of literature is not going to stop me from looking at them and sharing a few that stood out during the search. None of these papers is approaching the subject from a generative AI model side of things they are using machine learning without any degree of agency. Obviously, I was engaging in this literature review to see if I could find examples of the deployment of models with some type of agency doing analysis within this space of election prediction models. My searching over the last few weeks has not yielded anything super interesting. I was looking for somebody in the academic space doing some type of work within generative AI constitutions and election models or maybe even some work in the space of rolling sentiment analysis for targeted campaign understanding. That is probably an open area for research that will be filled at some point.Here are 4 articles:Grimmer, J., Roberts, M. E., & Stewart, B. M. (2021). Machine learning for social science: An agnostic approach. Annual Review of Political Science, 24, 395-419. https://www.annualreviews.org/doi/pdf/10.1146/annurev-polisci-053119-015921 Sucharitha, Y., Vijayalata, Y., & Prasad, V. K. (2021). Predicting election results from twitter using machine learning algorithms. Recent Advances in Computer Science and Communications (Formerly: Recent Patents on Computer Science), 14(1), 246-256. www.cse.griet.ac.in/pdfs/journals20-21/SC17.pdf Miranda, E., Aryuni, M., Hariyanto, R., & Surya, E. S. (2019, August). Sentiment Analysis using Sentiwordnet and Machine Learning Approach (Indonesia general election opinion from the twitter content). In 2019 International conference on information management and technology (ICIMTech) (Vol. 1, pp. 62-67). IEEE. https://www.researchgate.net/publication/335945861_Sentiment_Analysis_using_Sentiwordnet_and_Machine_Learning_Approach_Indonesia_general_election_opinion_from_the_twitter_content Zhang, M., Alvarez, R. M., & Levin, I. (2019). Election forensics: Using machine learning and synthetic data for possible election anomaly detection. PloS one, 14(10), e0223950. https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0223950&type=printable My guess is that we are going to see a wave of ChatGPT related articles about elections post the 2024 presidential cycle. It will probably be one of those waves of articles without any of them really standing out or making any serious contribution to the academy. The door is opening to a new world of election prediction and understanding efforts thanks to the recent changes in both model agency and generative AI models that help evaluate and summarize very complex things. It’s really about how they are applied to something going forward that will make the biggest difference in how the use cases play out. These use cases by the way are going to become very visible as the 2024 election comes into focus. The interesting part of the whole equation will be when people are bringing custom knowledge bases to the process to help fuel interactions with machine learning algorithms and generative AI. It's amazing to think how rapidly things can be built. The older models of software engineering are now more of a history lesson than a primer on building things with prompt-based AI. Andrew Ng illustrated in a recent lecture the rapidly changing build times. You have to really decide what you want to build and deploy and make it happen. Ferris Bueller once said, "Life moves pretty fast." Now code generation is starting to move even faster! You need to stop and look around at what is possible, or you just might miss out on the generative AI revolution.You can see Andrew's full video here: Links and thoughts:I really enjoyed this talk from Bill Gurly, “All-In Summit: Bill Gurley presents 2,851 Miles”Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=Machine+learning+election+models&btnG= What’s next for The Lindahl Letter? * Week 140: Proxy models for elections* Week 141: Building generative AI chatbots* Week 142: Learning LangChain* Week 143: Social media analysis* Week 144: Knowledge graphs vs. vec

We have been going down the door of digging into considering elections for a few weeks now. You knew this topic was going to show up. People love prediction markets. They are really a pooled reflection of sentiment about the likelihood of something occuring. Right now the scuttlebut of the internet is about LK-99, a potential, maybe debunked, maybe possible room temperature superconductor that people are predicting whether or not it will be replicated before 2025 [1]. You can read the 22 page preprint about LK-99 on ArXiv [2]. My favorite article about why this would be a big deal if it lands was from Dylan Matthews over at Vox [3]. Being able to advance the transmission power of electrical lines alone would make this a breakthrough. That brief example being set aside, now people can really dial into the betting markets for elections where right now are not getting nearly the same level of attention as LK-99 which is probably accurate in terms of general scale of possible impact. You can pretty quickly get to all posts that the team over at 538 have tagged for “betting markets” and that is an interesting thing to scroll through [4]. Beyond that look you could start to dig into an article from The New York Times talking about forecasting what will happen to prediction markets in the future [5].You know it was only a matter of time before we moved from popular culture coverage to the depths of Google Scholar [6].Snowberg, E., Wolfers, J., & Zitzewitz, E. (2007). Partisan impacts on the economy: evidence from prediction markets and close elections. The Quarterly Journal of Economics, 122(2), 807-829. https://www.nber.org/system/files/working_papers/w12073/w12073.pdfArrow, K. J., Forsythe, R., Gorham, M., Hahn, R., Hanson, R., Ledyard, J. O., ... & Zitzewitz, E. (2008). The promise of prediction markets. Science, 320(5878), 877-878. https://users.nber.org/~jwolfers/policy/StatementonPredictionMarkets.pdfBerg, J. E., Nelson, F. D., & Rietz, T. A. (2008). Prediction market accuracy in the long run. International Journal of Forecasting, 24(2), 285-300. https://www.biz.uiowa.edu/faculty/trietz/papers/long%20run%20accuracy.pdf Wolfers, J., & Zitzewitz, E. (2004). Prediction markets. Journal of economic perspectives, 18(2), 107-126. https://pubs.aeaweb.org/doi/pdf/10.1257/0895330041371321 Yeah, you could tell by the title that a little bit of content related to time-series analysis was coming your way. The papers being tracked within Google Scholar related election time series analysis were not highly cited and to my extreme disappointment are not openly shared as PDF documents [7]. For those of you who are regular readers you know that I try really hard to only share links to open access documents and resources that anybody can consume along their lifelong learning journey. Sharing links to paywalls and articles inside a gated academic community is not really productive for general learning. Footnotes:[1] [2] https://arxiv.org/ftp/arxiv/papers/2307/2307.12008.pdf[3] https://www.vox.com/future-perfect/23816753/superconductor-room-temperature-lk99-quantum-fusion[4] https://fivethirtyeight.com/tag/betting-markets/ [5] https://www.nytimes.com/2022/11/04/business/election-prediction-markets-midterms.html[6] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=election+prediction+markets&btnG= [7] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=election+time+series+analysis&oq=election+time+series+an What’s next for The Lindahl Letter? * Week 139: Machine learning election models* Week 140: Proxy models for elections* Week 141: Election expert opinions* Week 142: Door-to-door canvassingIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Trying to figure out how many republicans, democrats, and independents are registered in each state is actually really hard. It’s not a trivial task. Even with all our modern technology and the extreme power of the internet providing outsized connectedness between things and making content accessible to searches. Even GPT-4 from OpenAI with some decent plugins turned on will struggle to complete this task.Your best searches to get a full list by state are probably going to land you into the world of projections and surveys. One that will show up very quickly are some results from the Pew Research which contacted people (300 to 4,000 of them) from each state to find out more data about political affiliation [1]. They evaluated responses into three buckets with no lean, lean republication, or lean democrat. That allowed the results to evaluate based on sampling to get a feel for general political intentions. However, that type of intention based evaluation does not give you a sense of the number of voters within each state. It opened the door to me considering if political registration is even a good indicator of election outcomes. Sports tournaments rarely play out based on the seeding. That is the element of it that makes it exciting and puts the sport into the tournament. To that end back during week 134 I shared the chalk model to help explore a hypothesis related to registration being predictive. At the moment, I’m more interested to see how proxy models for predicting sporting events are working. Getting actual data to track changes in political registrations is an interesting process. ChatGPT, Bard, and Bing Chat are capable of providing some numbers if you prompt them properly. The OpenAI model GPT-3.5 has some older data from September 2021 and will tell you registered voters by state [2]. I started with a basic prompt, “make a table of voter registration by state.” I had to add a few encouraging prompts at some points, but overall the models all 3 spit out results [3]. The Bing Chat model really tried to direct you back to the United States Census Bureau website [4]. This is an area where setting up some type of model with a bit of agency to go out to the relevant secretary of states websites for the 30 states that provide some data might be a way to go to build a decent dataset. That would probably be the only way to really track the official data coming out by state to show the changes in registration over time. Charting that change data might be interesting as a directional view of how voters view themselves in terms of voter registration in a longitudinal way. People who participate in Kaggle have run into challenges where election result prediction is actually a competition [5]. It’s interesting and thinking about what features are most impactful during election prediction is a big part of that competition. Other teams are using linear regression and classification models to help predict election winners as well [6]. I was reading a working paper from Ebanks, Katz, and King published in May 2023 that shared an in depth discussion about picking the right models and the problems of picking the wrong ones [7][8]. To close things out here I did end up reading this Center for Politics article from 2018 that was interesting as a look back at where things were [9]. Circling back to the main question this week, I spent some time working within the OpenAI ChatGPT with plugins trying to get GPT-4 to search out and voter registration by state. I have been wondering why with a little bit of agency one of these models could not do that type of searching. Right now the models are not set up with a framework that could complete this type of tasking. Footnotes:[1] https://www.pewresearch.org/religion/religious-landscape-study/compare/party-affiliation/by/state/ [2] https://chat.openai.com/share/8a6ea5e7-6e42-4743-bc23-9e8e7c4f79c5 [3] https://g.co/bard/share/96b6f8d02e8e [4] https://www.census.gov/topics/public-sector/voting/data/tables.html [5] https://towardsdatascience.com/feature-engineering-for-election-result-prediction-python-943589d89414 [6] https://medium.com/hamoye-blogs/u-s-presidential-election-prediction-using-machine-learning-88f93e7f6f2a[7] https://news.harvard.edu/gazette/story/2023/03/researchers-come-up-with-a-better-way-to-forecast-election-results/[8] https://gking.harvard.edu/files/gking/files/10k.pdf [9] https://centerforpolitics.org/crystalball/articles/registering-by-party-where-the-democrats-and-republicans-are-ahead/ What’s next for The Lindahl Letter? * Week 138: Election prediction markets & Time-series analysis* Week 139: Machine learning election models* Week 140: Proxy models for elections* Week 141: Election expert opinions* Week 142: Door-to-door canvassingIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

It has been a few weeks here since we started by digging into a good Google Scholar search and you know this topic would be just the thing to help open that door [1]. My searches for academic articles are always about finding accessible literature that sits outside paywalls that is intended to be read and shared beyond strictly academic use. Sometimes that is easier than others when the topics lend themselves to active use cases instead of purely theoretical research. Most of the time these searches to find out what is happening at the edge of what is possible involve applied research. Yes, that type of reasoning would place me squarely in the pracademic camp of intellectual inquiry. That brief chautauqua aside, my curiosity here is how do we build out econometric election models or other model inputs to feed into large language model chat systems as prompt engineering for the purposes of training them to help either predict elections or interpret and execute the models. This could be a method for introducing extensibility or at least the application of targeted model effect to seed a potential future methodology within the prompt engineering space. As reasoning engines go it’s possible that an econometric frame could be an interesting proxy model within generative AI prompting. It’s a space worth understanding a little bit more for sure as we approach the 2024 presidential election cycle. I’m working on that type of effort here as we dig into econometric election models. My hypothesis here is that you can write out what you want to explain in a longer form as a potential input prompt to train a large language model. Maybe a more direct way of saying that is we are building a constitution for the model based on models and potentially proxy models then working toward extensibility and agency from introducing those models together. For me that is a very interesting space to begin to open up and kick the tires on in the next 6 months. Here are 6 papers from that Google Scholar search that I thought were interesting:Mullainathan, S., & Spiess, J. (2017). Machine learning: an applied econometric approach. Journal of Economic Perspectives, 31(2), 87-106. https://pubs.aeaweb.org/doi/pdfplus/10.1257/jep.31.2.87 Fair, R. C. (1996). Econometrics and presidential elections. Journal of Economic Perspectives, 10(3), 89-102. https://pubs.aeaweb.org/doi/pdfplus/10.1257/jep.10.3.89Armstrong, J. S., & Graefe, A. (2011). Predicting elections from biographical information about candidates: A test of the index method. Journal of Business Research, 64(7), 699-706. https://faculty.wharton.upenn.edu/wp-content/uploads/2012/04/PollyBio58.pdf Graefe, A., Green, K. C., & Armstrong, J. S. (2019). Accuracy gains from conservative forecasting: Tests using variations of 19 econometric models to predict 154 elections in 10 countries. Plos one, 14(1), e0209850. https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0209850&type=printableLeigh, A., & Wolfers, J. (2006). Competing approaches to forecasting elections: Economic models, opinion polling and prediction markets. Economic Record, 82(258), 325-340. https://www.nber.org/system/files/working_papers/w12053/w12053.pdf Benjamin, D. J., & Shapiro, J. M. (2009). Thin-slice forecasts of gubernatorial elections. The review of economics and statistics, 91(3), 523-536. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2860970/pdf/nihms190094.pdf Beyond those papers, I read some slides from Hal Varian on “Machine Learning and Econometrics” from January of 2014 [2]. The focus of the slide was applied to modeling human choices. Some time was spent on trying to understand the premise that the field of machine learning could benefit from econometrics. To be fair since that 2014 set of slides you don’t hear people in the machine learning space mention econometrics that often. Most people talk about Bayesian related arguments. On a totally separate note for this week I was really into running some of the Meta AI Llama models on my desktop locally [3]. You could go out and read about the new Code Llama which is an interesting model trained and focused on coding [4]. A ton of researchers got together and wrote a paper about this new model called, “Code Llama: Open Foundation Models for Code” [5]. That 47 page missive was shared back on August 24, 2023, and people have already started to build alternative models. It’s an interesting world in the wild wild west of generative AI these days. I really did install LM Studio on my Windows workstation and run the 7 billion parameter version of Code Llama to kick the tires [6]. It’s amazing that a model like that can run locally and that you can interact with it locally using your own high end graphics card.Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=econometric+election+prediction+models&btnG= [2] https://web.stanford.edu/class/ee380/Abstracts/140129-slides-Machine-Learning-and-Econometrics.pdf [3] https://ai.meta.com/llam

I read and really enjoyed the book by Nate Silver from 2012 about predictions. It’s still on my bookshelf. Strangely enough the cover has faded more than any other book on the shelf. Silver, N. (2012). The signal and the noise: Why so many predictions fail-but some don't. Penguin.That book from Nate is sitting just a few books over from Armstrong’s principles of forecasting. A book that I have referenced a number of times before. It will probably be referenced more as we move ahead as well. It’s a resource that just keeps on giving. Math it’s funny like that. Armstrong, J. S. (Ed.). (2001). Principles of forecasting: a handbook for researchers and practitioners (Vol. 30). Boston, MA: Kluwer Academic.My podcast feed for years has included the 538 podcast where I listened to Nate and Galen talk about good and bad uses of polling [1]. Sadly, it does not currently feature Nate after the recent changes over at 538. They reported on and ranked a lot of polling within the 538 ecosystem of content. Model talk and the good or bad use of polling were staples in the weekly pod journey. I really thought at some point they would take all of that knowledge about reviewing, rating, and offering critiques of polling to do some actual polling. Instead they mostly offered polling aggregation which is what we are going to talk about today. On the website they did it really well and the infographics they built are very compelling. Today setting up and running a polling organization is different from before. A single person could run a large amount of it thanks to the automation that now exists. An organization with funding to set up automation and run the polling using an IVR and some type of dialogue flow [2]. Seriously, you could build a bot setup that placed calls to people and completed a survey in a very conversational way. That still runs into the same problem that phone survey methods are going to face. I screen out all non-contact phone calls and I’m not the only person doing that. Cold calls are just not effective for business or polling in 2023 and the rise of phone assistants that can effectively block out noise are going to make the phone methodology even harder to effectively utilize.It’s hard to make a hype based drum roll on the written page. You are going to have to imagine it for me to get ready for this next sentence. Now that you are imagining that drum roll… Get ready for a year of people talking about AI and the 2024 election. It probably won’t get crypto bad in terms of the hype trane showing up to nowhere, but it will get loud. I’m going to contribute to that dialogue, but hopefully in the softest possible way. Yeah, I’m walking right into that by reflecting on the outcome of my actions while simultaneously writing about them during this missive.You can see an article from way back in November 2020 talking about how AI does show some potential to gauge voter sentiment [3]. That was before all of the generative AI and agent hype started. Things are changing rapidly in that space and I’m super curious about what can actually be accomplished in that space. I’m spending time every day learning about this and working on figuring out ways to implement this before the next major presidential election in 2024. An article from The Atlantic caught my attention as it talked about how nobody responds to polls anymore and started to dig into what AI could possibly do in that space, microtargeting, and Kennedy (1960) campaign references [4]. That was an interesting read for sure but you could veer over to VentureBeat to read about how AI fared against regular pollsters in the 2020 election [5]. That article offered a few names to watch out for and dig into a little more including KCore Analytics, expert.ai, and Polly. We will see massive numbers of groups purporting to use AI in the next election cycle. Even The Brooking Institute has started to share some thoughts on how AI will transform the next presidential election [6]. Sure you could read something from Scientific American where people are predicting that AI could take over and undermine democracy [7]. Dire predictions abound and those will probably also accelerate as the AI hype train pulls up to election station during the 2024 election cycle [8][9]. Some of that new technology is even being deployed into nonprofits to help track voters at the polls [10].Footnotes:[1] https://projects.fivethirtyeight.com/polls/ [2] https://cloud.google.com/contact-center/ccai-platform/docs/Surveys [3] https://www.wsj.com/articles/artificial-intelligence-shows-potential-to-gauge-voter-sentiment-11604704009[4] https://www.theatlantic.com/technology/archive/2023/04/polls-data-ai-chatbots-us-politics/673610/ [5] https://venturebeat.com/ai/how-ai-predictions-fared-against-pollsters-in-the-2020-u-s-election/[6] https://www.brookings.edu/articles/how-ai-will-transform-the-2024-elections/ [7] https://www.scientificamerican.com/article/how-ai-could-take-over-elections-and-undermine-dem

Last week we started to mess around with some methods of doing sentiment analysis and setting up some frameworks to work on that type of effort. This week we take a little different approach and are going to look at an election model. I’m actively working on election focused prompt based training for large language models for better predictions. Right now I have access to Bard, ChatGPT, and Llama 2 to complete that training. Completing that type of training requires feeding election models in written form as a prompt for replication. I have been including the source data and written out logic as a part of the prompt as well.Party registration drives the signal. Everything else is noise. That is what I expected to see within this model. It was the headline that could have been, but sadly could not be written. It turns out that this hypothesis could be tested. You can pretty easily try to view the results as a March Madness college basketball style bracket. Accepting that chalk happens or to be put more bluntly the higher ranked seeds normally win. Within the NCAA tournament things are more sporting and sometimes major upsets occur. Brackets are always getting busted. That is probably why they have ended up branding it as March Madness. Partisan politics are very different in terms of the chalk being a lot more consistent. Sentiment can change over time and sometimes voter registration does not accurately predict the outcome.We are going to move into the hypothesis testing part of the process. This model accepts a bi-model two party representation of political parties with an assumption that generally the other parties are irrelevant to predicting the outcome. The chalk model for predicting elections based on registration reads like this, the predicted winner = max{D,R} where D = registered democrats and R = registered republicans at the time of election. For example, the State of Colorado in December of 2020 that would equate to the max{1127654,1025921} where registered Democrats outnumber registered Republicans [1]. This equation accurately predicted the results of the State of Colorado during the 2020 presidential election. 30 states report voter statistics by party with accessible 2020 archives. Using the power of hindsight we can test the chalk model for predicting elections against the results of the 2020 presidential elections. Several internet searches were performed using Google with the search, “(state name) voter registration by party 2020.” Links to the referenced data are provided for replication and or verification of the data. Be prepared to spend a little time completing a verification effort as searching out the registered voter metric for each of the states took about 3 hours of total effort. It will go much faster if you use the links compared to redoing the search from scratch. Data from November of 2020 was selected when possible. Outside of that the best fit of the data being offered was used. * Alaska max{78664,142266}, predicted R victory accurately [2]* Arizona max{1378324,1508778}, predicted R victory in error [3]* California max{10170317,5334323}, predicted D victory accurately [5]* Colorado max{1127654,1025921}, predicted D victory accurately [6]* Connecticut max{850083,480033}, predicted D victory accurately [7]* Delaware max{353659,206526}, predicted D victory accurately [8]* Florida max{5315954,5218739}, predicted D victory in error [9] * The data here might have been lagging to actual by 2021 it would have been accurate at max{5080697,5123799}, predicting R victory* Idaho max{141842,532049}, predicted R victory accurately [10]* Iowa max{699001,719591}, predicted R victory accurately [11]* Kansas max{523317,883988}, predicted R victory accurately [12]* Kentucky max{1670574,1578612}, predicted D victory in error [13] * The data here might have been lagging to actual voter sentiment. The June 2023 numbers flipped max{1529360,1593476}* Louisiana max{1257863,1020085}, predicted D victory in error [14,15]* Maine max{405087,321935}, predicted D victory accurately [16]* Maryland max{2294757,1033832}, predicted D victory accurately [17]* Massachusetts max{1534549,476480}, predicted D victory accurately [18]* Nebraska max{370494,606759}, predicted R victory accurately [19]* Nevada max{689025,448083}, predicted D victory accurately [20]* New Hampshire max{347828,333165}, predicted D victory accurately [21]* New Jersey max{2524164,1445074}, predicted D victory accurately [22]* New Mexico max{611464,425616}, predicted D victory accurately [23]* New York max{6811659,2965451}, predicted D victory accurately [24]* North Carolina max{2627171,2237936}, predicted D victory in error [25,26]* Oklahoma max{750669,1129771}, predicted R victory accurately [27]* Oregon max{1043175,750718}, predicted D victory accurately [28]* Pennsylvania max{4228888,3543070}, predicted D victory accurately [29]* Rhode Island max{327791,105780}, predicted D victory accurately [30]* South Dakota max{158829,277788}, pred

I’m still spending some time digging into notebooks. This time around the topic as you might have guessed for that inquiry is figuring out how to automate a survey or more pointedly some sentiment analysis. People are building automated phone surveys with interactive voice response (IVR) systems. The next wave of this technology will be hard to tell if it is a person or a bot. Seriously, those systems are going to keep getting better at a rapid pace. The new wave of generative large language models are going to make outbound call surveys better and probably more plentiful. When the outbound call survey plugins roll in for ChatGPT, Bard, and anybody can build one for Llama 2 if they are willing to serve up a custom model. At the same time, I’m entirely sure (and hopeful) that people will be using more advanced technology to block those phone calls as well. All right, let’s shift away from considering phone calls and start to dig around into some of the automated sentiment analysis techniques that exist. We are starting to see frameworks where you can ask these new series of ChatGPT type services to act as an agent for you and complete some type of tasking. One of the things that would be interesting to ask that type of agent to complete would be to evaluate sentiment about something. I’m sure brands would like to have some automated brand evaluation methods. This will inevitably be used for politics as well. Right now we are not to the point where everybody has plugins that allow agency for ChatGPT or other toolings at the moment. That really is coming very soon as far as I can tell. Between lower energy costs and the solid platforms being built, those changes together may enable the compute for this type of interaction to happen in very conversational ways with a computer in the next 5 years.Right now you could start by messing around with Google Colab and use the forms options they have [1]. Completing some really solid sentiment analysis may require more than just focusing on the Google Colab environment. You may want to go out to somewhere like Hugging Face to get some information on how to do this with some python code [2]. A nice place to go along this journey to get some sentiment analysis done would be to venture out to the world of Kaggle and access one of their notebooks for sentiment analysis [3].Another notebook that I liked was from notebook dot community and it shares some of the natural language processing basic of sentiment analysis in really good chunks that make it easy to understand the mechanics of how things are happening [4]. At this point in the process you are probably ready to start to do some work on your own to complete some sentiment analysis and I found the right Google Colab notebook for you to start work on designing your own sentiment analysis tool [5].I was talking to somebody recently about the future of AI. My explanation may have not been what they expected to hear. Within the next couple of years I expect to see a lot of companies spin up and a lot of different creativity happening in the space. All of that will end up settling out into a commodified built in series of advancements. A lot of new features for applications and tooling will spin out off the great wave of AI builds that are happening now, but it will end up feeling more commonplace and build into technologies that exist now. These technologies will mostly supplement or augment things as we move forward. You will have to know how to interact with and work with the generative models that exist, but they are going to be built into the platforms and systems that end up winning out within the business world in the next couple of years.Content consumed this week:“The Impact of chatGPT talks (2023) - Keynote address by Prof. Yann LeCun (NYU/Meta)” “Llama 2: Open Foundation and Fine-Tuned Chat Models” https://arxiv.org/pdf/2307.09288.pdf “Stanford CS229 Machine Learning I Naive Bayes, Laplace Smoothing I 2022 I Lecture 6” “MASTER Auto-GPT in under 60 MINUTES | Ultimate Guide” Footnotes:[1] https://colab.research.google.com/notebooks/forms.ipynb[2] https://huggingface.co/blog/sentiment-analysis-python [3] https://www.kaggle.com/code/omarhassan1406/notebook-for-sentiment-analysis[4] https://notebook.community/n-kostadinov/sentiment-analysis/SentimentAnalysis[5] https://colab.research.google.com/github/littlecolumns/ds4j-notebooks/blob/master/investigating-sentiment-analysis/notebooks/Designing%20your%20own%20sentiment%20analysis%20tool.ipynbWhat’s next for The Lindahl Letter? * Week 134: The chalk model for predicting elections* Week 135: Polling aggregation* Week 136: Econometric models* Week 137: Time-series analysis* Week 138: Prediction marketsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to disc

Thank you for tuning in to this audio only podcast presentation. This is week 132 of The Lindahl Letter publication. A new edition arrives every Friday. This week the topic under consideration for The Lindahl Letter is, “Synthetic data notebooks.”People are totally working on this one actively which I thought was pretty interesting. I have a general interest in how to create synthetic data using notebooks as it helps to provide people with lessons on how to do it from an educational based perspective. Really solid automated testing process may include some of this in the development process. It makes automation even more amazing as a part of the process. It looks like the folks over at Towards AI released a nice guide to synthetic data that is geared at beginners in March of 2023 [1]. That guide walked through some of the concepts and a few pieces of information like how some report from a researcher at Gartner showed that half of future AI data will end up being synthetic data. Don’t worry I went out and found sourcing for that from Gartner and Alexander Linden who estimated by 2030 that synthetic data would outpace the regular data [2]. Those future considerations aside, the reason that is happening is that most people are going to be expanding their datasets with synthetic data to help them do training and work with models [3]. We are pretty far into the second paragraph and you might be wanting to access a couple of Google Colab notebooks to be able to do some of this yourself. Don’t worry that is about to happen for you. The team over at gretel AI shared a couple of notebooks that you can use for this type of effort:https://colab.research.google.com/github/gretelai/gretel-synthetics/blob/master/examples/synthetic_records.ipynbThe first notebook had all sorts of errors and would not work. https://colab.research.google.com/github/gretelai/gretel-blueprints/blob/main/docs/notebooks/create_synthetic_data_from_a_dataframe_or_csv.ipynbThe second one required a Gretel API key to get going which was a lot less fun than it could have been without that part of the equation. I went out to the website over at https://gretel.ai/ and they have some free elements. I got into the dashboard they have pretty easily and started to look around to see what they are offering [4]. I went out to YouTube and found a 12 minute video from one of the co-founders Alex Watson showing how to do this effort. They did quickly show how to get the API key for the above notebook. I really did follow those instructions to get that magic API key which totally worked in the 2nd Google Colab notebook link from above. I stepped through the entire notebook in about 15 minutes and was able to see the process of synthetic data generation from a dataframe or CSV which was exciting to watch and learn about in a notebook. The main model training took 7 minutes so don’t expect that it will happen in just a click.Maybe you wanted to see somebody else do some generation of synthetic data in Google Colab on YouTube. You can see YData work with a fabric environment and work in a notebook. They had 44 subscribers and the video had 28 views before I shared this link for your enjoyment.You could also check out this other YouTube video from The Next Phase team that shows more information about “Synthetic data generation with CTGAN” in a Google Colab notebook as well [5]. Maybe you wanted to switch gears a bit and learn a little bit about how to create 8-bit audio samples [6]. I’m going to share one more article here that walks through how to generate datasets as I thought it was actually pretty good [7]. I’m going to close this one out with a zoom out to what some people think is the future of these synthetic data driven creations which is in fact the politician of the open internet and eventual model collapse [8]. People have even recently gone as far as to say copies of the internet before all this generated content are worth more for training than the derivatives. We will see what happens soon. Oftentimes in the machine learning spaces people have used randomness, chaos, or other techniques of shifting things around to overcome blocks. I’ll be curious to see if something is developed to overcome these potential model collapse elements. Footnotes:[1] https://towardsai.net/p/machine-learning/a-beginners-guide-to-synthetic-data[2] https://www.gartner.com/en/newsroom/press-releases/2022-06-22-is-synthetic-data-the-future-of-ai[3] https://towardsdatascience.com/generating-expanding-your-datasets-with-synthetic-data-4e27716be218 [4] https://console.gretel.ai/use_cases/cards/use-case-synthetic/projects [5] https://colab.research.google.com/drive/18vavq2Kt8HqhSnZvvFxJUc-70NCPeDU_?usp=sharing [6] https://medium.com/mlearning-ai/python-machine-learning-gans-synthetic-data-and-google-colab-5bb43491a8c7[7] https://medium.com/nerd-for-tech/synthetically-generate-datasets-using-deep-learning-c1f6ee7a0990 [8] https://arxiv.org/pdf/2305.17493.pdf What’s next for The Lin

We are not to the point where some AI agent is going to be able to do this type of bulk image improvement by a verbal command. I really wanted to be able to just say, “ok computer, edit the photos in this folder to improve them and ping me when you are done.” This use case will probably show up at some point, but now is not that point in the timeline. The future of this type of computer usage is close, but not here just yet.Over the last couple of weeks, I have spent some time messing around with ways to complete some bulk image improvement via either scripting or applications. It turns out a lot of people use Adobe Lightroom to edit images. Adobe is really spending a lot of time and effort to share the AI powered innovations they have built into the product. Fundamentally, it is not that surprising that Adobe would try to use adoption and the major amount of brand equity they have as a moat to deflect away from the entry of new products into the photo editing process. A lot of room exists for a product to come in and be disruptive in this space. My use case here is really simple. I want to just have a folder of photos and via either scripting or sharing the folder to an application have those images improved via all these brand new AI tools we keep hearing about. I’m asking for a low friction solution to just have a tool or some tooling do a bit of work to improve the photos being taken. You might be thinking, but I thought Google Photos would basically do this for you if you use that service. It does not really work that way. You can go in and individually update photos with a single click, but it won’t just do the work on the photos in the background. Actually getting Google Photos to work in the background would essentially involve writing scripting to bounce each photo against a vision based API for improvement. That sounds tedious and is not a low friction solution. Within the marketplace generally a number of things have popped up to compete with the Adobe products including some interesting ones: * Fotor (100 employees, started 2012)* Skylum (150 employees, started 2008)Searching for Fotor in the Google Play store ended up directing me to something called “AI Photo Editor, Collage-Fotor” which was not what I expected [1]. The installation was clearly a Fotor owned application and it is apparently used by more than 10 million people with a 4.3 out of 5 star review rating. Before using the product I ended up backing out and just searching for “AI photo editor” in Google Play which was interesting [2]. That search produced a ton of options. Way more options that I was expecting to see. I know that a lot of people use smartphones or tablets and don’t really even open a computer to do editor or general computing work. Right now I’m writing this missive on a double monitor desktop setup going in the total other direction of things. The application based editing route is one thing, but I wanted to see if some other method existed. I had hoped that something in the Google ecosystem would make this task a little bit easier. During the course of a few more searches and looking around I did end up watching this video from Josiah Blizzard who used the Batch AI Adobe Lightroom plugin to edit 1,000 photos per minute. This seemed promising. It was a way more entertaining video that I expected. I thought it might be interesting to check out a few more videos related to Batch AI on YouTube and was able to find a few different options. Footnotes:[1] https://play.google.com/store/apps/details?id=com.everimaging.photoeffectstudio&hl=en_US&gl=US [2] https://play.google.com/store/search?q=AI%20photo%20editor&c=apps&hl=en_US&gl=US What’s next for The Lindahl Letter? * Week 132: Synthetic data notebooks* Week 133: Automated survey methods* Week 134: Make a link based news report automatically* Week 135: Saving some notebooks every day* Week 136: What if July was startup month? 31 days for 31 ideasIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

We are getting closer and closer to the world envisioned in Star Trek and countless other science fiction universes where you could just ask the computer to complete some sort of task or build for you. We are starting to see what would be like an AutoGPT or a plugin style version of ChatGPT with agents that work to complete some sort of request. It’s not like they have a full built in chron system with scheduling and routines. Really actionable virtual assistants are going to show up here very quickly. I’m pretty sure Google is not very far off from being able to allow their voice assistant to activate more things within routines and schedules with a real degree of agency. I’m actually fairly surprised they have not just launched that feature already. That ecosystem has the connectivity to be able to turn off and on lights or other home connected devices and connect to so many other things. What I am saying here is that you very well could have an agent do a variety of things very soon.You can imagine that it would be fun to just ask the ChatGPT agent to write and do something with fractals. To that end I spent some time asking by prompt both Bard and ChatGPT to produce code for a Colab notebook. Strangely enough Bard never produced any code that was executable within a Colab notebook. When advised I was wanting code that would execute in a Colab environment the ChatGPT model did spit out code that would execute properly [1]. I’m adding notebooks that work to my GitHub repository on an ongoing basis. Beyond messing around with the generative models it seemed like a fun idea to look around and see what people were doing with fractals in Google Colab notebooks. The notebooks I found were produced a while ago, but that made them no less fun to play with. Fractal Art With Python by Dr. Mike Lam of James Madison University: https://colab.research.google.com/drive/1tSBkON1Uj0NCfYEgELdXxC-PS9eOaUQG?usp=sharing#scrollTo=p02VqSeizpUNFractal Generation with L-Systems by Paul Butlerhttps://colab.research.google.com/github/paulgb/notebooks/blob/master/source/l-systems/Fractal%20Generation%20with%20L-Systems.ipynb I ended up wanting to just look at Python code for fractal building and manipulation. That is where I’m working with things right now in the process [2]. My base test for these new generative models relates to how well they can actively generate and manipulate fractals.Footnotes:[1] https://github.com/nelslindahlx [2] https://towardsdatascience.com/creating-fractals-with-python-d2b663786da6What’s next for The Lindahl Letter? * Week 131: Bulk imagine improvement scripting* Week 132: Synthetic data notebooks* Week 133: Automated survey methods* Week 134: Make a link based news report automatically* Week 135: Saving some notebooks every dayIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Working out of notebooks is an easy and lightweight way to just mess around with code. One of my all time favorite ways to go about doing that is in the Colab research space provided by the team over at Google [1]. They have an offering for a pro subscription, but you can get in and use things without a payment plan. Sure the promise of faster GPUs, more memory, and longer runtimes is enticing. If you really get deeply into the process of training models, then it might make sense for you to make arrangements to pay for some compute. A lot of people have shared notebooks and you can quickly start to work on things and see code in action. The part of it that is so appealing to me is that you can look at a very distinct block of action and run it in the notebook environment without having to worry about anything else in the stack. Breaking things into smaller understandable pieces is absolutely fantastic for learning about things. It’s also nice to just jump in and be able to execute the code. Here very soon according to the Google Blog page for developers we are going to see AI-powered coding free of charge within the Colab space [2]. You will be able to open the notebook and instead of adding code or text you will have access to another option to generate. That is where the magic will be located and you will be able to very rapidly use Colab in a generative way. I think it is going to bridge the gap between interacting with generative models to accomplish things and the current sort of limitations. Inside that argument would be the backward linkages to the Google ecosystem where instead of plugins or managing extensions the system will just have extensibility into a massive and mature ecosystem. I have been waiting since that May 17, 2023 post for the feature to show up. Literally, I just keep checking on an almost daily basis to see when we can get to the point of rock and rolling the generative creation of code from the PaLM 2 family of code models [3]. At this point, I’m going to admit that my favorite part of this whole thing is that they have called it Codey. Maybe the controversial Clippy was just ahead of its time, but I think Codey will be just right as a part of a larger ecosystem of generative coding efforts that are about to take flight. You will see that very little traffic occurred on Twitter about this news [4]. In terms of education and getting people going with generative code generation this is going to be epic. You will see my contributions to GitHub of notebooks go from sparse to daily [5]. I have started dropping some code to make fractals and do some various types of math into a new repository [6].All of the security for Google Colab notebooks is managed by those teams. You certainly have control over your own private notebook being saved to your Google drive. Access to the data and notebooks is controllable and the team over at Google is spinning up and then spinning down virtual machines to do this work. Most of the process is entirely ephemeral and to the end user is fundamentally just a browser. The FAQ page does not really spend much time addressing security [7]. It’s a place to very quickly be able to execute code from a browser. Using it for educational purposes makes total sense. It’s a great place to share code and learn about how that code works. It has enough power to do some basic machine learning and to that end it is a pretty awesome experience. You are not going to end up using this as a production runtime for any sort of tasking. One of the things I’m actually curious about in the future is what happens when people start using ChatGPT plugins or agents that go out and farm compute from environments like this type of browser based engagement.Footnotes:[1] https://colab.research.google.com/[2] https://blog.google/technology/developers/google-colab-ai-coding-features/ [3] https://9to5google.com/2023/05/17/google-colab-codey/ [4] https://twitter.com/search?q=colab%20codey&src=typed_query [5] https://github.com/nelslindahlx [6] https://github.com/nelslindahlx/fractals [7] https://research.google.com/colaboratory/faq.html What’s next for The Lindahl Letter? * Week 130: Build captain fractal using Colab* Week 131: Bulk imagine improvement scripting* Week 132: Synthetic data notebooks* Week 133: Automated survey methods* Week 134: Make a link based news report automaticallyIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

The topic for this week was a weighty one for sure. It’s one that I think will functionally happen, but I do not believe will be real in terms of actual practice. We are seeing a lot of emphasis on using these new AI systems that are popping up everywhere. People are not able to build these systems from scratch anymore. Using code from somebody else is becoming more and more a part of the process. When you look at how to democratize AI systems it is certainly about spreading usage and making these systems omnipresent. Security is structurally something that is important in this situation, but it remains something that generally happened before you got involved. In that scenario, really deeply considering how your AI system security posture is going to be invoked you have to consider both forward looking actions and the backward linkages within these products both in terms of the build and design. Within the calling of an API in this space it certainly is possible to consider what data is going out and what is coming back. You can pretty clearly understand what happens with the data after it is sent over and what is going to happen with it after the fact. My searches on Google Scholar for articles related to AI security were not super exciting [1]. That search with the brackets around the search term yielded about 2,900 items. Taking the brackets off opened up the results to over 3 million [2]. This may very well be a space where the academic content has not caught up with where the technology happens to be at the moment. A lot of academic research in the space is occurring, but since the papers are not really about or from production implementations the topics of how to secure, manage, and deploy are distinctly lacking. At this point in the story, I got super interested in reading about comparisons of protocols for distributed social networking. Things like the AT protocol vs activitypub [3]. Oddly, one of the best lists that I encountered was on Wikipedia where somebody is clearly keeping track of open projects related to this one [4]. You might be wondering if that list includes 30+ projects with tacking and information and you will find out that it totally does. However, being able to take a look at such a large list actually made me a little bit concerned about any of these protocols actually becoming dominant and taking over in practice. Let’s take this post in a different, more meta direction. Ok, here are all the spilled beans on that one, I ended up making an executive decision on how to manage my writing backlog. It was a super disruptive and spur of the moment decision at this point in the process. I went ahead and at the week 128 issue (which is this post) pasted into the backlog the new list of 49 items from a Google Keep note. That happened as a result of a really productive day that happened last week. During one of the vacation days at the beach recently, I started making a list of topics I would like to spend some time either researching or writing about. It turns out that list ended up including 49 total items. With the addition of those 49 items, my backlog is now tracking out to week 218. Switching things up like that means I’m going to struggle with updating the next five weeks of forward looking items on a bunch of blocks of content, but that is a solvable problem in terms of editing and writing. It’s just a bit annoying and somewhat time consuming vs. creating any really ongoing problematic situation. That does mean that starting in July you will start to get the benefit of this new backlog and strictly speaking you are going to end up waiting about a year to return to the previously structured program content. While we are zoomed out and considering some of the more meta implications of things I’ll note that I have kept the podcast part of this effort going. Readership numbers indicate that it is a smaller portion of things compared to the standard text based link clicking. About 20% of the ongoing traffic for this writing project seems to be related to the audio based podcast. The recording process is now pretty streamlined and I have it as a part of my overall creative weekend routine. These weekly blocks of writing won’t grow long enough to make recording the audio for them problematic in terms of a time commitment. Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=%22ai+security%22&btnG= [2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=ai+security&btnG= [3] https://www.theverge.com/2023/4/20/23689570/activitypub-protocol-standard-social-network [4] https://en.wikipedia.org/wiki/Comparison_of_software_and_protocols_for_distributed_social_networkingWhat’s next for The Lindahl Letter? * Week 129: How do you use Colab in a generative way?* Week 130: Build captain fractal using Colab* Week 131: Bulk imagine improvement scripting* Week 132: Synthetic data notebooks* Week 133: Automated survey methodsIf you enjoyed this content, then please take a moment an

Things have been happening with Google DeepMind. A lot of news has been dropping since I started crafting this post. I don’t want to write about the news of the moment on this one. The context and analysis within this post is going to try to be deeper than a base reaction to that recent news. The Verge (and a ton of other people) reported that, “Google’s big AI push will combine Brain and DeepMind into one team” [1]. Apparently, some drama exists around those two organizations being combined. We will almost certainly see some books about this from the principal players. Let’s hope they turn out as good as the 2020 memoir “Uncanny Valley” by Anna Wiener. Some of them won’t be that dynamic, but they should be. I spent some time reading content from DeepMind. They have a solid research page [2]. It highlights some interesting things including AlphaFold, WaveNet, and AlphaGo [3][4][5]. They have a lot of content out on GitHub in a variety of repositories [6]. They have over 190 repositories with a variety of code you could go take a look at. You may even notice that they have signed the “Lethal Autonomous Weapons Pledge” [7][8].Some good conversation is occurring about what happens when AI hallucinates, bias bounties, new bug bounties, and how red teams work [9]. You could also find out a little more about how the GitLab AI-based security feature helps to scan codebases for vulnerabilities [10]. You could go check out the Google Cloud Security Podcast EP52 Securing AI with DeepMind CISO Vijay Bolina from February 14, 2022 [11].You really have to consider that even the best pair programming with assistive technology could introduce security vulnerabilities [12]. We could probably spend an entire week looking at the medical artificial intelligence system or more clearly named Med-PaLM 2 that was released [13]. They also shared an interesting method to discover faster matrix multiplication algorithms which I thought was rather interesting [14]. Don’t worry I did go out and grab a couple (really 4) scholarly papers to share with you this week:Beattie, C., Leibo, J. Z., Teplyashin, D., Ward, T., Wainwright, M., Küttler, H., ... & Petersen, S. (2016). Deepmind lab. arXiv preprint arXiv:1612.03801. https://arxiv.org/pdf/1612.03801.pdfPowles, J., & Hodson, H. (2017). Google DeepMind and healthcare in an age of algorithms. Health and technology, 7(4), 351-367. https://link.springer.com/article/10.1007/s12553-017-0179-1Tassa, Y., Doron, Y., Muldal, A., Erez, T., Li, Y., Casas, D. D. L., ... & Riedmiller, M. (2018). Deepmind control suite. arXiv preprint arXiv:1801.00690. https://arxiv.org/pdf/1801.00690 Evans, R., & Gao, J. (2016). Deepmind ai reduces google data centre cooling bill by 40%. DeepMind blog, 20, 158. https://www.deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40 Footnotes:[1] https://www.theverge.com/2023/4/20/23691468/google-ai-deepmind-brain-merger [2] https://www.deepmind.com/research[3] https://www.deepmind.com/research/highlighted-research/alphafold[4] https://www.deepmind.com/research/highlighted-research/wavenet [5] https://www.deepmind.com/research/highlighted-research/alphago[6] https://github.com/deepmind [7] https://www.deepmind.com/safety-and-ethics[8] https://futureoflife.org/open-letter/lethal-autonomous-weapons-pledge/?cn-reloaded=1[9] https://www.theregister.com/2023/04/26/is_your_ai_hallucinating/[10] https://www.spiceworks.com/tech/artificial-intelligence/news/gitlab-launches-ai-security-feature/[11] https://cloud.withgoogle.com/cloudsecurity/podcast/ep52-securing-ai-with-deepmind-ciso/[12] https://www.techtarget.com/searchsoftwarequality/news/252523049/Developers-beware-AI-pair-programming-comes-with-pitfalls[13] https://www.verdict.co.uk/google-lauches-new-medtech-ai-system-amid-calls-for-greater-data-security/[14] https://venturebeat.com/ai/top-5-stories-of-the-week-deepmind-and-openai-advancements-intels-plan-for-gpus-microsofts-zero-day-flaws/What’s next for The Lindahl Letter? * Week 128: Democratizing AI system security* Week 129: Snapchat Security* Week 130: Generative model security* Week 131: Profiling Microsoft Azure security* Week 132: Engineering discipline securityIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Something is going to have to give here pretty soon for Hugging Face or they are going to get left behind by the extended model, plugin driven, and now internet connected GPT models that some of the biggest platforms are working to take mainstream. We are seeing the real time development creating extensibility into actions from companies like OpenAI and their partner Microsoft that are going to be a foundational groundwork for how connectivity works within these models. This ecosystem is going to be something so highly proprietary and interconnected between a set of foundational companies that no ability to directly open source a competitor is going to exist. Part of that will be due to the payment models that are going to get setup as a foundational services layer starts to get setup.People have been really excited about Hugging Face for some time now. Since 2016 the private company Hugging Face has worked to share machine learning content with the world. You can easily get to some of the training courses they have set up on NLP and Deep RL [1]. They are pretty good training courses. People have really dug into models and spaces from Hugging Face to help democratize NLP models. At one point, everybody was checking in with what Hugging Face was up to and they made a very large space in the AI and ML space. You can easily go out to Google Scholar and see a bunch of academic papers that reference Hugging Face either as two words or sometimes one word [2]. I spent some time looking for papers with a decent number of citations. Here are 5 that were selected:Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A. M. (2019). Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. https://arxiv.org/pdf/1910.03771Jiang, W., Synovic, N., Hyatt, M., Schorlemmer, T. R., Sethi, R., Lu, Y. H., ... & Davis, J. C. (2023). An empirical study of pre-trained model reuse in the hugging face deep learning model registry. arXiv preprint arXiv:2303.02552. https://arxiv.org/pdf/2303.02552Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A. M. (2020, October). Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations (pp. 38-45). https://aclanthology.org/2020.emnlp-demos.6.pdf Pfeiffer, J., Rücklé, A., Poth, C., Kamath, A., Vulić, I., Ruder, S., ... & Gurevych, I. (2020). Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779. https://arxiv.org/pdf/2007.07779 Zhang, Y., Sun, S., Galley, M., Chen, Y. C., Brockett, C., Gao, X., ... & Dolan, B. (2019). Dialogpt: Large-scale generative pre-training for conversational response generation. arXiv preprint arXiv:1911.00536. https://arxiv.org/pdf/1911.00536.pdf%7D. You can get a feel for what the Hugging Face community has been working on by taking a look at their GitHub repository [3]. They have transformers, datasets, diffusers, and tools for a variety of things that people find useful. I would describe Hugging Face as a very hands-on community where you can dig in and be a part of what is going on within the space. That is the distinctive difference from some of the other larger corporations that do open source things. Most of the corporate AI labs will produce and distribute things when they are ready. Iterative tool releases and a vibrant community full of contributions are popping up all over. We are also seeing a newer trend where companies like OpenAI are releasing models and other technology via API without open sourcing the content. This pay to play API model shields the intellectual property better, but in this space people are rapidly learning from innovations and working to build alternatives. Here are a couple of YouTube videos including one from Hugging Face:You can dig into the security features offered by Hugging Face [4]. It says clearly on the security page that they are SOC2 Type 1 certified with a link to the AICPA page [5]. With all sources where you might download code and use it you are going to want to understand the code, be ready to keep it current, and of course be prepared to take rapid action. They are even trying to compete with ChatGPT and share more open models for that type of effort [6]. I spent some time looking around to try to find some deep assessments of the security profile for Hugging Face and have not really found what I was looking for yet.Footnotes:[1] https://huggingface.co/learn/nlp-course/chapter1/1 [2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=%22hugging+face%22&btnG= [3] https://github.com/huggingface [4] https://huggingface.co/docs/hub/security [5] https://us.aicpa.org/interestareas/frc/assuranceadvisoryservices/aicpasoc1report.html [6] https://venturebeat.com/ai/hugging-face-launches-open-source-version-of-chatgpt-in-bid-to-battle-openai/ What’s next for The Lindahl Letter? * We

Security professionals the world over are concerned about what people like some of Samsung’s engineers might be doing with OpenAI’s ChatGPT services [1]. Let’s back up just a bit from that very real breaking news security reality and make sure to set the stage. Instead of starting with a complete corporate history about OpenAI, you knew that I would go out and search Google Scholar to find some of the key OpenAI related scholarly works [2]. Alternatively, you could surf over to the OpenAI website and find the section they have for over 100 papers and just download the PDFs from the source [3]. You could download papers like the much maligned “GPT-4 Technical Report” that was lengthy at over 100 pages, but did not go into the model mechanics people were interested in understanding [4]. Other papers are hosted on that page like the Dall-E-2 paper “Hierarchical Text-Conditional Image Generation with CLIP Latents” [5]. It is a lot of great content and you will get a sense that at one point OpenAI was sharing and building a great legacy of published content that helped people really understand where artificial intelligence was going. They also did some really awesome work building agents that competed in pretty complex video games against world class players. That part of the equation is what caught my attention and really pulled me into trying to understand OpenAI. Here are 5 interesting papers with a solid number of citations:Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). Gpts are gpts: An early look at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130. https://arxiv.org/pdf/2303.10130.pdf Berner, C., Brockman, G., Chan, B., Cheung, V., Dębiak, P., Dennison, C., ... & Zhang, S. (2019). Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680. https://arxiv.org/pdf/1912.06680.pdf Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016). Openai gym. arXiv preprint arXiv:1606.01540. https://arxiv.org/pdf/1606.01540.pdfGawłowicz, P., & Zubow, A. (2019, November). Ns-3 meets openai gym: The playground for machine learning in networking research. In Proceedings of the 22nd International ACM Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (pp. 113-120). https://www2.informatik.hu-berlin.de/~zubow/gawlowicz19_mswim.pdf Zamora, I., Lopez, N. G., Vilches, V. M., & Cordero, A. H. (2016). Extending the openai gym for robotics: a toolkit for reinforcement learning using ros and gazebo. arXiv preprint arXiv:1608.05742. https://arxiv.org/pdf/1608.05742.pdfNow that we have covered some of the decently cited scholarly content it is probably good to zoom out from that space into the broader media coverage of OpenAI. Most of it is about the risk, reward, and fear that this type of generative AI fosters. You will see that security is not at the forefront of these discussions. From TheVerge “OpenAI co-founder on company’s past approach to openly sharing research: ‘We were wrong’”. Ok. Where do you go from there…https://www.theverge.com/2023/3/15/23640180/openai-gpt-4-launch-closed-research-ilya-sutskever-interviewYou can hear Greg Brockman who was an OpenAI cofounder talk about the astonishing potential during a 30 minutes TED talk.The CEO of OpenAI Sam Altman sat down with ABC News for 20 minutes here:You can see a long form interview with Lex Fridman and Sam Altman here:They have an OpenAI YouTube channel that gets a lot of views. They used to share some pretty interesting videos.Ok at this point in our journey here, we have covered OpenAI in general and looked at what is being created. Now it’s probably best to try to explain the drama that surrounds the company as quickly as possible. Oddly, it’s not a drama focused on the technology being created that certainly has entered the public mind. Elon Musk was an initial board member of OpenAI back in 2015 [6]. It used to be a non-profit and things have changed. That is where the real drama exists about the company. Microsoft now has a multi-billion dollar investment in OpenAI [7]. You can read the Microsoft official blog where they detail that they have invested in OpenAI three times including 2019, 2021, and 2023 and is now the exclusive cloud provider [7]. You might be thinking how did Elon Musk invest or more accurately donate 100 million to OpenAI which now is clearly partnered with Microsoft [9]. It’s a confusing plot twist for sure and a lot of media coverage exists trying to explain how that exactly happened [10]. I’m sure we will learn more about what exactly happened in the next couple of years.Let’s circle back to the security realities of using things like OpenAI’s ChatGPT system. People in this case are going to use the system to input things to the prompt. Some of that input might very well be intellectual property owned distinctly by somebody. You could ask a model to create more content in an author’s style or t

It's all happening. I'm pivoting to AI + Security as the central topic for The Lindahl Letter. The next weekly modules I'm going to create are all going to be pivoting to a clear security based focus each week. Content for this year of The Lindahl Letter was packaged into posts encompassing weeks 105 to 156 or 52 total blocks of content. Think year divisible by week with production spread evenly. This is the 3rd year of that type of writing effort. This is week 124 in that sequence and I have yelled pivot and brought in 11 new topics that will follow the profiles of OpenAI, Hugging Face, and DeepMind. That content pivot pretty much involved highlighting a section of the backlog and just moving it beyond week 156. I’m going to keep on working to balance the topics together to bring forward the best possible set of content for this year. Getting the best possible blocks of content brought together as a research project is essential to my work moving forward. It’s possible that I will go back and rework some of the content from weeks 105 to 123 to include more of a security focus. Obviously, that won’t change the previously released podcast audio or published content, but it would update the content that gets put into a manuscript at the end of the year. That content gets edited and revised during the course of the publishing process so it is never exactly the same as the content that was published throughout the year in weekly installments. While I certainly strive to produce high quality prose my ability to edit to perfection on a weekly basis remains suspect. Editing for continuity and general cohesiveness is a different element within that large of a manuscript compared to producing a weekly installment. Putting content into an eBook and ultimately a hardback or paperback format means that it needs to be free of grammarian enraging distractions.During my initial cut at introducing security related topics into the backlog 11 topics jumped out at me and got included. My backlog editing will be pretty extreme for the next couple of weeks to really get the most out of this epic content based pivot. Right now we are at the precipice of some very serious questions about how AI will impact society. Seriously, AI and society are very alive within the public mind at the moment. Understanding how we secure, interact over time, and ultimately establish security within the watershed event of modern AI development will be more important than my rather casual walk into AI in general. You will find that this will be a good pivot that ends up providing more depth and context to what is happening currently. Each of the new installments will be written with a degree of perspective that helps prevent them from being passing observation installments. Writing pure reaction content is not where I’m trying as that provides less of an advisor type function and more of an ephemeral of the moment observation.You may well be aware that I’m still struggling with the idea that all the content that gets produced as a part of my weekly research notes could be replicated by ChatGPT in minutes. This is the 124th installment and OpenAI’s ChatGPT could easily provide alternative versions of the content I have created. Understanding how to bring larger themes together and diligently searching for the best academic articles is not something that ChatGPT does well at the moment. It’s entirely possible that the combination of content selection and prompt engineering could change the potential output. I’m pretty sure somebody will work to better adapt ChatGPT to the review of complex academic work. Having the right data sources selected for a model will be a key element of how things work moving forward. You could take the papers from my independent study ML syllabus and get a decent set of inputs for a model. However, would you add that as a layer to a current model like ChatGPT 3.5 or 4.0 or will you have focused layers for the models to consider in sequence or as a refinement step along the way. All of those things are up in the air and we will learn more about where they will ultimately end up. As a general editorial decision here and now I may consult and work with ChatGPT, but the output in terms of writing posts for the rest of the year will just include my words. For me the novelty of including some examples of the differences between what I would write and what OpenAI’s ChatGPT would create has lost a certain degree of interest for me. Maybe the novelty or the thrill of it is gone. It took a couple months for that to happen, but now I’m going to focus on providing the best research note I can each week. Even this note focusing my content going forward signals an important change in my research trajectory and was worth devoting a week of content creation to building and sharing. Links and thoughts:What’s next for The Lindahl Letter? * Week 125: Profiling OpenAI Security* Week 126: Profiling Hugging Face security* Week 127: Profiling Deep

Substack Week 123: We are wholesale oversubscribed on AI related contentThank you for tuning in to this audio only podcast presentation. This is week 123 of The Lindahl Letter publication. A new edition arrives every Friday. This week the topic under consideration for The Lindahl Letter is, “We are wholesale oversubscribed on AI related content.”You may have noticed that the topic for the post this week changed. I called an audible and wrote a different piece of content.All that AI content has hit a real saturation point. We are wholesale oversubscribed on AI related content and we have not even really crossed into the situation I’m concerned about where people are using the content to flood the open internet with synthetic content. Generally given the current state of the AI excitement level and some of the fear, apprehension, and concern a lot of people are writing or commenting. All that commenting has built into a very real saturation point where AI content is literally everywhere. Sam Altman of OpenAI recently noted in a Wired article that, “the Age of Giant AI Models Is Already Over” [1]. The initial release date of ChatGPT was November 30, 2022 so the cycle on this one was no more than a 6 month explosion of hype content. You can go out to Google Trends and pretty quickly get a sense of just how fast ChatGPT spun up in December of 2022 [2]. If you went out and added a comparison term within Google Trends of auto-gpt or AutoGPT, then you would get a sense of just how much more popular ChatGPT happens to be in the wild [3]. Keep in mind that since April 11, 2023 the searches for AutoGPT have taken off exponentially [4]. Major things are happening in terms of ChatGPT, AutoGPT, and people expanding how agents are able to cooperate. A paper from Park et al. that was just published on April 7, 2023 shows some very interesting use cases for simulating behavior using generative agents. Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv preprint arXiv:2304.03442. https://arxiv.org/pdf/2304.03442.pdf Rewind a little bit and I thought the key Stanford University paper would be that 214 page multi author one on foundation models.Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., ... & Liang, P. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258. https://arxiv.org/pdf/2108.07258.pdf You can get a sense from the demonstration of capability and use that occurred in that paper on generative agents just a couple years after scholars circled the wagons on foundation models just how fast the field is changing. Things are changing so fast that Sam Altman above noted that the age of giant models might have arrived and be complete between the publication of those two papers. It can make conducting research a very interesting thing to contemplate. Content at the bleeding edge of AI technology today could very well encounter a context, vibeshift, or even meaning change within months. Taking that into consideration I’m going to recognize that we are approaching the halfway point on the production of this year’s Substack content. To that end, I’m starting to ponder how to improve things and really make things better in terms of the content that is being produced. Links and thoughts:Footnotes:[1] https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/[2] https://trends.google.com/trends/explore?geo=US&q=chatgpt&hl=en [3] https://trends.google.com/trends/explore?date=now%207-d&geo=US&q=chatgpt,autogpt&hl=en [4] https://trends.google.com/trends/explore?date=2023-03-18%202023-04-18&geo=US&q=autogpt&hl=en What’s next for The Lindahl Letter? * Week 124: Profiling OpenAI * Week 125: Profiling Hugging Face (open and collaborative machine learning)* Week 126: Profiling Deep Mind* Week 127: Democratizing AI systems* Week 128: Building dreaming into AI systemsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

I spent some time over the last couple of weeks looking for scholarly articles that reference AI as a service (AIaaS) [1]. Then I realized that things are changing so quickly in the landscape of AI these days that none of those articles even consider Auto-GPT [2]. Unless you have been on vacation and hiding from AI news recently you have been hearing about the advent of this plugin, multiagent, and task targeted implementation of a GPT model called Auto-GPT. People are using it to do things. You can see that that GitHub repository already has 50,000 stars and the demo was from March 30, 2023. In terms of open source code projects this one has taken off with rocket boosters. Right now the project has been forked almost 8,000 times. I’m trying to find a version that can be run in a jupyter notebook on Colab. This space is rapidly changing and will probably be changing rapidly over the next 6 months.What did the OpenAI ChatGPT 4.0 model spit out on this one when prompted?AI can be both a platform and a service, depending on how it is implemented and used. The distinction between the two is not mutually exclusive, as AI applications often span across various domains and industries. Let's explore each concept:* AI as a Platform: An AI platform provides the infrastructure, tools, and libraries necessary for developing and deploying AI applications. These platforms enable developers and data scientists to build, train, and manage AI models without having to start from scratch. Examples of AI platforms include TensorFlow, PyTorch, and OpenAI's GPT series. In this context, AI serves as a foundational technology upon which various solutions can be built.* AI as a Service (AIaaS): AIaaS refers to the delivery of AI capabilities as a service, usually through APIs or web services. Companies that offer AIaaS provide access to pre-built AI models, which clients can use without having to develop their own AI capabilities. This allows organizations to leverage AI technology without investing in infrastructure or expertise. Examples of AIaaS include Google Cloud AI, IBM Watson, and Amazon Web Services' machine learning services.In conclusion, AI can function both as a platform and a service, depending on its implementation and usage. Many companies and developers use AI platforms to create AI services, which can then be integrated into various applications and business processes.I thought it might be fun to follow up and prompt to see who was the best at AIaaS… It is difficult to definitively say who the "best" AI-as-a-Service (AIaaS) provider is, as different providers excel in different areas and offer varying services that cater to specific use cases. However, some of the most prominent and reputable AIaaS providers include:* Google Cloud AI: Google offers a comprehensive suite of AI services, including natural language processing, computer vision, and machine learning. Its offerings, such as Google Cloud Vision API, Dialogflow, and AutoML, are widely used and respected for their accuracy and ease of use.* Amazon Web Services (AWS): AWS provides a range of AI services, including machine learning, computer vision, and natural language processing. Some popular offerings are Amazon Rekognition, Amazon Lex, and Amazon SageMaker.* IBM Watson: IBM's AIaaS platform, Watson, offers various services, including natural language processing, machine learning, and chatbot development. Watson Assistant and Watson Discovery are among its popular services.* Microsoft Azure: Azure offers a suite of AI services, including machine learning, natural language processing, and computer vision. Key offerings include Azure Cognitive Services, Azure Machine Learning, and Azure Bot Service.The "best" AIaaS provider for your specific needs will depend on several factors, such as the required AI capabilities, pricing, ease of integration, and the level of customization needed. It's essential to research and evaluate each provider based on your unique requirements to make an informed decision.Links and thoughts:Top 4 Tweets of the week:https://twitter.com/ericschmidt/status/1659619610063249409https://twitter.com/CaseyNewton/status/1659579798857871360https://twitter.com/robmay/status/1659610763005399042https://twitter.com/DynamicWebPaige/status/1659075284891738112Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=AIaaS&btnG= [2] https://github.com/Torantulino/Auto-GPT What’s next for The Lindahl Letter? * Week 123: Considering open source AI* Week 124: Profiling OpenAI * Week 125: Profiling Hugging Face (open and collaborative machine learning)* Week 126: Profiling Deep Mind* Week 127: Democratizing AI systemsIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus

My initial take on writing an independent study based syllabus for applied AI was to find the best collection of freely available scholarly papers that somebody could read as an onramp to beginning to understand the field. That I think is a solid approach to helping somebody get going within a space that is very complex and full of content. It’s a space that is perpetually adding more content than any one person could possibly read or consume. Before you take that approach it is important to understand that one definitive textbook does exist. You certainly could go out and read it. Russell, S. J. (2010). Artificial intelligence a modern approach. Pearson Education, Inc.. You could find the first edition, second edition, or third edition for sale on eBay or somewhere else if you wanted a physical copy of the book. The book is currently in a 4th edition run, but I don’t have a copy of that edition yet. It’s used by over 1,500 schools so a lot of copies exist out in the wild [1]. The authors Stewart Russell and Peter Norvig have shared a PDF of the bibliography for that weighty tome of AI insights as well [2]. Even with 35 pages of bibliography nobody with the name Lindahl made the cut. On a side note you can find the name Schmidhuber included twice if that sort of thing is important to you. Let’s reset for a second here. If you are brand new to the field of AI or want to read a textbook based introduction, then you should seriously consider buying a copy of the aforementioned textbook. That is a really great way to start which has worked for tens of thousands of people. My approach here is going to be a little bit unorthodox, but it works for me. My last run at this type of effort was, “An independent study based introduction to machine learning syllabus for 2022” and you can find it out on Google Scholar [3]. This outline will be the basis of a similar type of work that will end up getting crafted in Overleaf and shared out to the world. Searching for just pure introductions to artificial intelligence is really hit or miss. A lot of different introductions to various fields exist. In this case, I’m trying to zoom out a little more into a larger evaluation of content instead of focusing on any one field. Nothing I ran into during my search had the number of citations or impact of the Russel and Norvig textbook. I’m going to endeavor to structure and organize 70+ articles into a syllabus. To give you an idea of the kind of things that are going to get pulled together here are 5 different papers. Oke, S. A. (2008). A literature review on artificial intelligence. International journal of information and management sciences, 19(4), 535-570. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=f4dfcfe3f132b1079d054e2db29adc063fab0007 Contreras, I., & Vehi, J. (2018). Artificial intelligence for diabetes management and decision support: literature review. Journal of medical Internet research, 20(5), e10775. https://www.sciencedirect.com/science/article/pii/S0004370218305988/pdfft?md5=ec6948d3f66efe5e57d1336a54d1604d&pid=1-s2.0-S0004370218305988-main.pdf Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., & Aerts, H. J. (2018). Artificial intelligence in radiology. Nature Reviews Cancer, 18(8), 500-510. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6268174/ Ghahramani, Z. (2015). Probabilistic machine learning and artificial intelligence. Nature, 521(7553), 452-459. https://www.repository.cam.ac.uk/bitstream/handle/1810/248538/Ghahramani%25202015%2520Nature.pdf?sequence=1 Jiang, F., Jiang, Y., Zhi, H., Dong, Y., Li, H., Ma, S., ... & Wang, Y. (2017). Artificial intelligence in healthcare: past, present and future. Stroke and vascular neurology, 2(4). https://svn.bmj.com/content/svnbmj/2/4/230.full.pdf Obviously, working out of a textbook format will help provide a better structured introduction and that is why I plan on mentioning that in the paper and I have noted it here. On a scholarly basis I understand the need to dive directly into source papers and really get a sense of what is happening within the applied side of AI research in the last 10 years. That is why I want to spend some time preparing the independent study based syllabus for others as well as it being a good solid effort for me to learn more rapidly with an extreme degree of focus. What did ChatGPT 4.0 model come up with?An independent study applied AI syllabus will provide you with the necessary knowledge and skills to develop practical AI solutions. Below is a suggested 14-week syllabus to guide you through the essential topics and resources.* Week 1: Introduction to AI and Machine Learning* History of AI* Overview of machine learning, deep learning, and reinforcement learning* Common machine learning algorithms* Week 2: Python for AI and Machine Learning* Basic Python programming* NumPy, pandas, and matplotlib* Setting up a Python environment for machine learning* Week 3: Data Preprocessing and Exploration* Data cleaning and transfo

Right now at the start of 2023, I would probably highlight 3 AI trends: generative models, automation, and legislation. Before we get into those specific topics let’s zoom out for just a second and look at two different reports you could read to get a sense of what is going on right now. One of the great places to start would be with the recently released 2023 AI Index report from the Institute for Human Centered AI. Nestor Maslej, Loredana Fattorini, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Helen Ngo, Juan Carlos Niebles, Vanessa Parli, Yoav Shoham, Russell Wald, Jack Clark, and Raymond Perrault, “The AI Index 2023 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April 2023. https://aiindex.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf You can look around the website they have setup here:https://hai.stanford.edu/research/ai-index-2023If the 386 page PDF seems like a lot of content to consume, then feel free to read the slimed down 2023 state of AI in 14 charts that they also producedhttps://hai.stanford.edu/news/2023-state-ai-14-charts The other interesting report that I read recently was the Google Cloud 2023 Data and AI Trends Report which you can download by giving them your email. That step seemed annoying so I’m going to leave the link to the report below, but it might default back to the information pagehttps://services.google.com/fh/files/misc/data_and_ai_trends.pdf Those two reports aside, let’s jump into the 3 topics that will comprise the bulk of this missive. Generative models - generative models are probably the biggest thing happening at the start of 2023. We are setting chat based and image based models like GPT-4 and DALL-E 2 from OpenAI [1]. One of the things spinning out of the generative models are the low code, no code, and copilot efforts. This is one of the areas where you will find a lot of coverage in the media as the generative models are being used as search engines and the image generating efforts are so easy to use people have gone thumbnail happy using them recently [2]. Automation - We are going to see a tremendous amount of better automation out of the new technology that is available [3]. I’m tempted to mention that I consider a lot of the copilot efforts to be a part of early automation. Those copilot efforts will move from being inline assistive to probably a more prompt based delivery. That type of request and receiving copiloting is what I would consider a form of automation. Legislation - Countries are wholesale banning ChatGPT [4]. A lot of legislation will probably come into being related to trying to ensure safety [5]. I’m not going to spend a ton of time on this trend as it's going to unfold in a pretty predictable way.You might find that this week the large language model from OpenAI has produced a better post than what I crafted. To that end, let's find out now what the ChatGPT model GPT-4 produces after being prompted about this one…As of April 2023, the top trends in AI include the following:* Federated Learning: This approach allows AI models to be trained on decentralized data while maintaining privacy. It has gained prominence as a way to harness the potential of data from various sources without compromising user privacy.* AI Ethics and Regulation: As AI systems become more integrated into everyday life, there is an increasing focus on the ethical implications and regulatory measures required to ensure AI is used responsibly.* Explainable AI (XAI): With the growing complexity of AI models, there is a rising demand for models that provide interpretable and understandable explanations for their predictions or decisions.* AI in Healthcare: AI applications in healthcare are expanding rapidly, from diagnosis and treatment recommendations to drug discovery and personalized medicine.* AI for Climate Change: AI techniques are being used to tackle climate change by optimizing energy consumption, forecasting extreme weather events, and modeling climate patterns.* Edge AI: As IoT devices proliferate, there is a growing interest in running AI models on-device (at the edge) rather than relying on cloud-based processing. This approach can improve latency, privacy, and energy efficiency.* AI for Cybersecurity: AI is being deployed to enhance cybersecurity measures by detecting and preventing potential threats and attacks using advanced pattern recognition and anomaly detection.* Reinforcement Learning (RL): RL, a type of machine learning that focuses on decision-making, continues to gain traction as researchers explore its potential applications in various fields, including robotics, finance, and gaming.* Natural Language Processing (NLP): NLP techniques continue to advance, enabling more sophisticated AI applications in areas like language translation, sentiment analysis, and chatbot development.* AI and Robotics: AI is playing a key role in the dev

Flooding and astroturfing abound at the moment. Both of those things were happening before the advent of large language models (LLMs), but they have increased in frequency now that bad actors are able to just open the floodgates for content. Making large swaths of the internet that is just designed for search engine placement and self-referencial boosting has become so much easier recently. Sure all that bad data was abounding before this shift in what is now happening with OpenAI, Google, Microsoft, and Facebook recently sharing out chat services. It’s one of those things where it is hard to put words on a page about it. Working with one of the chat systems to make content seems to trivialize the writing process. My day starts with an hour of focused academic work. That time is the fulfilled promise of decades of training that included a lot of hard work to get to this point. I can focus on a topic and work toward understanding it. All of that requires my focus and attention on something for that hour. Sometimes on the weekends I spend a couple of hours doing the same thing on a very focused topic. Those chat models with their large language model backends (LLM) produce content within seconds. It’s literally like a 1:60 ratio for output. It takes me an hour to produce what it creates within that minute including the time for the user to enter the prompt. Maybe I did not expect this type of interaction to really affect me in this way. Everything has been questioned in terms of my writing output and what exactly is going to happen now. The door has been flung open to the creation of content. Central to that problem is the reality that the careful curation of content within academics and the publish first curation of the media are going to get flooded. Both systems are going to get absolutely overloaded with submissions. Something has to give based on the amount of attention that exists. They are not minting any new capacity for attention and the channels for grabbing that attention are relatively limited. The next couple of years are going to be a mad scrabble toward some sort of equilibrium between the competing forces of content curation and flooding.This really is something that I’m concerned about on an onboarding basis. Do all the books, photos, articles, and paintings in the before times just end up with a higher value weighting going forward? Will this AI revolution have cheapened the next generation of information delivery in ways we will not fully get to appreciate until the wave has passed us and we can see the aftermath of that scenario? Those questions are at the heart of what I’m concerned about. Selfishly they are questions about the value and purpose of my own current writing efforts. More broadly they are questions about the value of writing within our civil society as we work toward the curation of sharable knowledge. We all work toward that perfect possible future either with purpose or without it. Knowledge is built on the shoulders of the giants that came before us adding to collective understanding of the world around us. Anyone with access and an adventurous spirit can pick up the advancement of some very complex efforts to enhance the academy's knowledge on a topic. Maybe I’m worried that the degree of flooding with flatten information so much that the ability to move things forward will diminish. Sorting, seeking, and trying to distill value from an oversupply of newly minted information may well create that diminishing effect. We will move from intellectual overcrowding in the academy to just an overwhelming sea of derivative content marching along beyond any ability to constrain or consume. I’m going to stop with that last argument as it may be the best way to sum this up. Links and thoughts:Top 5 Tweets of the week:https://twitter.com/CaseyNewton/status/1651967468322320387https://twitter.com/Benioff/status/1651605849310130178https://twitter.com/AndrewYNg/status/1651605660382134274https://twitter.com/verge/status/1651355922751422464https://twitter.com/NateSilver538/status/1650899579234140168What’s next for The Lindahl Letter? * Week 120: That one with an obligatory AI trend’s post* Week 121: Considering an independent study applied AI syllabus* Week 122: Will AI be a platform or a service?* Week 123: Considering open source AI* Week 124: Profiling OpenAI If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

This is going to be an interesting edition of The Lindahl Letter. Please pardon the self-indulgence of working with ChatGPT to write a paper and ultimately expand that into a full book using prompts and the model. Sit back and relax as this will be a long one.I was able to give OpenAI’s ChatGPT a prompt of, “A paper on political debt as a concept vs. technical debt.” That created the following output:Political debt and technical debt are two distinct concepts that have different implications and applications in their respective fields. While technical debt refers to the long-term costs associated with the trade-offs made in the software development process, political debt refers to the exchange of favors, resources, or benefits for political influence or gain.Technical debt arises when developers make decisions that favor short-term gains over long-term benefits, such as prioritizing speed over quality or failing to maintain code cleanliness. While these decisions may be necessary in the short-term, they can result in higher maintenance costs and increased risk in the long-term.Political debt, on the other hand, arises when political actors trade favors, resources, or benefits in exchange for political influence or gain. This can include actions such as accepting campaign contributions from specific interest groups, providing special treatment to certain individuals or corporations, or ignoring public opinion in favor of personal interests.While technical debt is primarily a concern for software developers and companies, political debt has broader implications for democratic governance and the public interest. Political debt can undermine the principles of transparency, accountability, and ethical leadership in government, and can result in policies that do not fully address the needs or preferences of the broader public.Managing and reducing political debt requires a sustained commitment to transparency, accountability, and ethical leadership in political systems. This can include measures such as campaign finance reform, strengthening democratic institutions such as the judiciary and civil society organizations, and promoting public participation and engagement in the political process.In conclusion, while political debt and technical debt are two distinct concepts, they both demonstrate the challenges associated with short-term thinking and decision-making in their respective fields. By promoting the values of transparency, accountability, and ethical leadership, we can build stronger, more effective, and more representative political and technical systems that serve the interests of all stakeholders.That was not good enough so I thought it would be a good idea to write a full book working with ChatGPT on this one. Don’t begin a state of panic on this one, but the remainder of this effort is about 45 pages that include 21 chapters of content. I’m not entirely sure I’ll be able to read this output in one session for the audio version of this effort. To create this series of chapters I just kept prompting the model to produce more chapters. This is the result of a couple hours of my working with the interface to get this output. Chapter 1: IntroductionPolitical debt and technical debt are two distinct but related concepts that have significant implications for their respective fields. Political debt refers to the exchange of favors, resources, or benefits for political influence or gain, while technical debt refers to the long-term costs associated with the trade-offs made in the software development process.Political debt is a complex and multifaceted challenge that can undermine the principles of transparency, accountability, and ethical leadership in government. When political actors are indebted to specific interest groups or individuals, they may be less likely to consider policies that conflict with the interests of these groups. This can result in policies that do not fully address the needs or preferences of the broader public, and can undermine the democratic legitimacy of government.Technical debt, on the other hand, arises when developers make decisions that favor short-term gains over long-term benefits. These decisions may be necessary in the short-term to meet project deadlines or to address urgent technical issues, but they can result in higher maintenance costs and increased risk in the long-term.Despite their different origins and applications, political debt and technical debt share important similarities. Both concepts are associated with short-term thinking and decision-making, and both can result in long-term costs and negative consequences. Additionally, strategies for managing and reducing political debt and technical debt require a sustained commitment to transparency, accountability, and ethical leadership.This book will compare and contrast political debt and technical debt, examining the causes, consequences, and impact of these concepts on their respective fields. We will also e

This week I ended up trying to find some solid academic articles to read. This time around I found a few that have publicly available downloads to share. Digging around literature related to non-mail polling methods opens the door to a lot of different directions. You will get a sense from the journals listed below that a lot of literature exists and research is benign conducted about public opinion. My interests are more in the automated methods that would help collect sentiment to complete polling without active respondents. Data abounds and some of it holds the key to understanding how sentiment is being expressed. Being able to catalog and share that sentiment is an interesting thing to automate. Voss, S., Gelman, A., & King, G. (1995). The polls—A review: Preelection survey methodology: Details from eight polling organizations, 1988 and 1992. Public Opinion Quarterly, 59(1), 98-132. http://www.stat.columbia.edu/~gelman/research/published/voss.pdfProsser, C., & Mellon, J. (2018). The twilight of the polls? A review of trends in polling accuracy and the causes of polling misses. Government and Opposition, 53(4), 757-790. https://papers.ssrn.com/sol3/Delivery.cfm/SSRN_ID3184232_code2420517.pdf?abstractid=3184232&mirid=1 Martin, E. A., Traugott, M. W., & Kennedy, C. (2005). A review and proposal for a new measure of poll accuracy. Public Opinion Quarterly, 69(3), 342-369. http://www.asasrms.org/Proceedings/y2003/Files/JSM2003-000913.pdfPrice, V., & Neijens, P. (1998). Deliberative polls: Toward improved measures of “informed” public opinion?. International Journal of Public Opinion Research, 10(2), 145-176. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=3922d1362956bb55df25dbaa8aae8d4706067002 Let’s see what ChatGPT model 3.5 would generate related to this one.While mail polling remains a valuable tool for collecting data, advances in technology have led to the development of a variety of non-mail polling methods. In this literature study, we will explore the latest research on non-mail polling methodology.Firstly, online polling has become increasingly popular, particularly for collecting data quickly and cost-effectively. A study by the Pew Research Center found that online polling can be just as accurate as traditional phone or mail polling, but response rates may be lower, particularly among certain populations such as older adults or those with lower levels of education.Secondly, phone polling remains a popular method for collecting data, particularly for political polling. However, the rise of mobile phones has presented new challenges for phone polling. A study by the European Social Survey found that mobile phone users are more likely to be younger and have higher levels of education, leading to a potential bias in the results. Additionally, increasing rates of spam and robo-calls may lead to declining response rates.Thirdly, social media has presented new opportunities for pollsters, particularly for collecting data on younger populations. However, concerns have been raised about the representativeness of social media polls, as they may attract self-selected respondents and may not be representative of the broader population.Fourthly, researchers have explored the use of text message polling, particularly for collecting data quickly and with high response rates. A study by the National Opinion Research Center (NORC) found that text message polling had higher response rates than email or online polling, particularly among younger populations.Finally, researchers have also explored the use of location-based polling, which involves collecting data from individuals based on their geographic location. This method can be particularly useful for collecting data on local issues, but concerns have been raised about the potential for privacy violations.In conclusion, non-mail polling methods have become increasingly popular in recent years, particularly for collecting data quickly and cost-effectively. While online, phone, social media, text message, and location-based polling all have their advantages and disadvantages, researchers have found that no single method is universally superior. The choice of polling method should depend on the research question, the population being surveyed, and the resources available.Links and thoughts:Top Tweets of the week:https://twitter.com/CaseyNewton/status/1646886386031788032https://twitter.com/robmay/status/1646862194775130114What’s next for The Lindahl Letter? * Week 118: A paper on political debt as a concept vs. technical debt* Week 119: All that bad data abounds * Week 120: That one with an obligatory AI trend’s post* Week 121: Considering an independent study applied AI syllabus* Week 122: Will AI be a platform or a service?If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead. Th

Searching around on Google Scholar is something that I’m pretty adept at by now. This happens to be the 116th Substack post and almost all of them have involved some type of research. Trying to get a set of the best articles on mail polling methodology was actually really challenging [1]. I know that modern polling has moved away from mail polling in general. It’s expensive. The response rate has gotten worse over time. Probably the worst part about it is that compared to phone or internet based polling it is very slow. My interest here is to really understand polling methodologies. Getting to a high level of understanding about these methodologies will help me evaluate that current literature against automated sentiment analysis. It’s the combination of those two literatures that really has my attention at the moment. Google Scholar was actually recommending a related search, “design method mail and internet surveys” [2]. I’ll admit that the results on that search were a little more interesting, but none of it was really targeted at what I wanted to learn about. I ended up broadening my searches to include content related to mail surveys. That yielded a few highly cited sources. One of the problems with this area of research compared to AI or ML is that all the research is not easily accessible for download. Brown, M. L. (1965). Use of a postcard query in mail surveys. The Public Opinion Quarterly, 29(4), 635-637.Sudman, S., Sirken, M. G., & Cowan, C. D. (1988). Sampling rare and elusive populations. Science, 240(4855), 991-996.Dillman, D. A. (1991). The design and administration of mail surveys. Annual review of sociology, 17(1), 225-249.Scott, C. (1961). Research on mail surveys. Journal of the Royal Statistical Society: Series A (General), 124(2), 143-195.Kanuk, L., & Berenson, C. (1975). Mail surveys and response rates: A literature review. Journal of marketing research, 12(4), 440-453.Maybe you were wondering what ChatGPT model 3.5 would come up with today. Using the prompt, “write a substack post called, ‘A literature study of mail polling methodology’”, a post was created.A literature study of mail polling methodologyIn an era where online and mobile polling have become more popular, it can be easy to overlook the traditional method of mail polling. However, mail polling remains a valuable tool for collecting data, particularly for certain populations. In this literature study, we will explore the latest research on mail polling methodology.Firstly, studies have shown that mail polling can be particularly effective for reaching older populations, who may be less likely to use online or mobile polling methods. A study by the Pew Research Center found that mail polling had higher response rates among adults aged 65 and older than online or phone polling.Secondly, mail polling can also be useful for collecting data on sensitive or controversial topics. A study by the National Opinion Research Center (NORC) found that mail polling resulted in higher rates of disclosure on topics such as drug use and sexual behavior, possibly because respondents feel more comfortable disclosing sensitive information in writing rather than in person or over the phone.Thirdly, mail polling can be useful for collecting data from hard-to-reach populations, such as those without internet or phone access. A study by the American Association for Public Opinion Research (AAPOR) found that mail polling resulted in higher response rates among rural populations than online or phone polling.Fourthly, researchers have explored the use of incentives to increase response rates in mail polling. Studies have shown that offering incentives such as cash or gift cards can lead to higher response rates, although the effectiveness of incentives may vary depending on the population being surveyed.Finally, researchers have also explored the use of mixed-mode polling, which combines mail polling with other methods such as online or phone polling. This approach can help reach a wider range of respondents and increase response rates.In conclusion, mail polling remains a valuable tool for collecting data, particularly for certain populations. While online and mobile polling methods have become more popular, mail polling should not be overlooked, particularly for collecting data on sensitive topics or hard-to-reach populations. The use of incentives and mixed-mode polling can also help increase response rates and improve the accuracy of results.Links and thoughts:Top 5 Tweets of the week: It’s possible that these tweets will not show up as embeds and will just be links. This is due to a change in how Twitter is handling things and not a Substack issue.https://twitter.com/sama/status/1644878154870427649https://twitter.com/jdevalk/status/1644439739645194279https://twitter.com/MKBHD/status/1644562346843467776https://twitter.com/AndrewYNg/status/1644418518056861696https://twitter.com/CaseyNewton/status/1644358645084459013 Footnotes:[1] https://scholar.g

Earlier today I did end up getting ChatGPT Plus for a cost of USD $20 a month. That apparently will provide me better availability during high demand, faster response speeds, and priority access to new features [1]. I’m not sure if this subscription will be worth keeping. I’m about to open up the latest model GPT-4 and see what it is able to accomplish in terms of advanced reasoning, complex instructions, and more creativity. I have spent the last 6 weeks thinking about writing these polling related papers. A lot of that time and effort was spent trying to figure out what the trajectory of the current literature was and what that meant for the future of understanding and evaluating sentiment. To me polling is a sampling of sentiment on something. It’s an evaluation of respondent preference or attitude. I’m writing that explanation without using the word opinion. That word choice happened on purpose.A lot of opportunity exists to study breakdowns in modern polling techniques. This is the start of my series of inquiries into that potential area of study. To be really clear here upfront in this analysis I believe that respondent fatigue, general unavailability, and systemic methodology breakdowns have made modern polling problematic. You can see commentary in very public news sources about people being frustrated with polling [2][3]. I’m surprised so far that more of the literature did not openly discuss the breakdown in previously solid methodologies for polling the public. We are here working on these series of literature evaluations, because I’m principally interested in opinion polling and sentiment analysis. You could go out to the Pew Research Center to learn about polling basics [4]. They present research and polling in some pretty easy to understand ways. That visit to Pew might even send you in the direction of the American Association for Public Opinion Research (AAPOR) [5]. The outcome of that digging might point you in the direction of some methods of administering polling. * Traditional mail survey questionnaire * Phone based delivery of questionnaire* Email survey questionnaire* Web based questionnairesI have spent some time listening to Nate Silver on podcasts over the years. One of the interesting things Nate shared was that a recent batch of polls was more accurate than expected [6]. That made me wonder if the metrology being used had improved or if they were just tuning properly based on some type of modeled expectation. I will caveat here that I read Nate’s book:Silver, Nate. The signal and the noise: Why so many predictions fail-but some don't. Penguin, 2012.It was interesting, but I wanted to know more about the future of polling. One of the main questions that I ended up having was about the differences between polling and modeling and what method was better and would end up being more popular over time [7].* Public attitude extraction* Indicator based tracking* Modeled behavior* Simulated response analysis* Persona based economic modeling* Sentiment interviewingWhere did I end up at the end of this analysis? Two things came into focus. First, it was pretty clear to me after digging around that a clear framework for auditing polling methodologies needs to be developed. Second, a sea change in polling is about to occur where modeling as a method or some type of supplemental grounding activity will have to augment traditional polling methodologies. We have already seen extremely convoluted weighting and error ranges become forward reaching into the public mind. You could search Google Scholar for online polling best practices [8]. You could also take a look at direct search for “modern polling methodology” and see what shows up [9]. Only one result on the first page of that search had more than 100 citations and that was a handbook. It came in at 172 citations. Donsbach, W., & Traugott, M. W. (Eds.). (2007). The SAGE handbook of public opinion research. Sage.The PDF was not available over at ResearchGate or on Google Scholar. You can buy the hardcover from Sage for $215 dollars which seems a little bit out of hand. You can rent or buy the eBook from Sage as well which for a lifetime copy would set you back $117 [10].I did prompt ChatGPT from OpenAI using the GPT-4 model to generate a paper called, "A literature review of modern polling methodology" that turned out to be interesting. My exact prompt on this one was, “write an academic paper called, ‘A literature review of modern polling methodology’”.Oddly enough on the first attempt it produced a ton of content then had a “network error” with a more specific message of, “There was an error generating a response.” My only option to move forward was to click the “Regenerate response” box and hope for the best. It took 4 regeneration attempts to get a complete output. Title: A Literature Review of Modern Polling MethodologyAbstractThe present study aims to provide a comprehensive review of the modern polling methodologies, analyzing their stre

Imagine you move part of your workload to an environment that you don’t fully own. That used to be a datacenter and more recently it is probably a cloud environment. Sooner or later you might think about how to secure that environment in terms of hardware-based security or other ways to secure things. Most of the promotional pages on the internet covering this one try to bring forward the idea of securing code and data [1]. Seriously, the idea of confidential computing is everywhere on these cloud pages, but not something you hear about all the time generally [2][3][4]. It is a topic that receives a lot of industry specific coverage related to managing and operating cloud environments. The low number of highly cited “confidential computing” articles over at Google Scholar was probably indicative of the relative interest in this topic [5]. Very few of the available academic works are highly cited in this area. It’s actually an area where some research opportunity exists related to a survey of security and the outcomes of confidential computing in practice. A few of the more interesting ones with a higher level of citation ended up being:Mulligan, D. P., Petri, G., Spinale, N., Stockwell, G., & Vincent, H. J. (2021, September). Confidential Computing—a brave new world. In 2021 international symposium on secure and private execution environment design (SEED) (pp. 132-138). IEEE. https://dominicpm.github.io/publications/mulligan-seed2021.pdfRashid, F. Y. (2020). The rise of confidential computing: Big tech companies are adopting a new security model to protect data while it's in use-[news]. IEEE Spectrum, 57(6), 8-9. https://ieeexplore.ieee.org/iel7/6/9099910/09099920.pdfZhu, J., Hou, R., Wang, X., Wang, W., Cao, J., Zhao, B., ... & Meng, D. (2020, May). Enabling rack-scale confidential computing using heterogeneous trusted execution environment. In 2020 IEEE Symposium on Security and Privacy (SP) (pp. 1450-1465). IEEE. https://heartever.github.io/files/hetee.pdfThe central question here seems to be if you are going to have your data and run your workloads in the cloud how do you really make sure that effort is confidential. At this point in my analysis, I thought it might be interesting to see what ChatGPT would produce on this topic. OpenAI ChatGPT Prompt: “scrap that and rewrite it with ethics included”Sure, here's a rewritten version of the substack post about AI and confidential computing with a focus on ethics:The intersection of AI and confidential computing has become increasingly important as we realize the importance of privacy and security in computing. Confidential computing is a field that seeks to protect sensitive data and computations by ensuring that they remain encrypted and secure throughout the computing process. Meanwhile, AI relies heavily on large amounts of data to make accurate predictions and decisions, but this data is often sensitive and needs to be kept confidential to protect individuals’ privacy. This is where confidential computing can play an important role, as it can ensure that sensitive information remains private, even when being processed by AI algorithms.The integration of confidential computing into AI has several benefits. First, it ensures that personal data is kept secure and private, which is crucial for industries such as healthcare and finance. Second, it enables collaboration between organizations without risking the leak of sensitive information. Lastly, it can improve the accuracy and effectiveness of AI models as more data can be used for training without compromising privacy.However, there are also ethical considerations when it comes to the intersection of AI and confidential computing. For example, there is a risk that confidential computing could be used to conceal unethical or discriminatory practices. It is important to consider how confidential computing is used and to ensure that it is not used to violate ethical principles. Additionally, there is a need to ensure that confidential computing is not used to perpetuate biases that already exist in datasets.To address these ethical concerns, it is important to ensure that the development and deployment of confidential computing technologies are guided by ethical principles. Transparency and accountability should be central to the development of these technologies, and there should be a focus on ensuring that they are used in a way that promotes fairness, accountability, and transparency. This requires collaboration between computer scientists, ethicists, and policymakers to ensure that these principles are integrated into the design and use of confidential computing technologies.In conclusion, the intersection of AI and confidential computing has promising implications for privacy and security. However, it is important to ensure that these technologies are developed and used ethically. By integrating ethical considerations into the design and use of confidential computing technologies, we can ensure that th

You may remember that I listened to an audiobook during a road trip back to Kansas about AI. Kissinger, H. A., Schmidt, E., & Huttenlocher, D. (2021). The age of AI: and our human future. Hachette UK. Former CEO of Google Eric Schmidt has had a lot to say about the future of AI and a lot of that has been grounded in the implications of AI on society. It’s important to set the foundation of any introduction to AI ethics on the implications to people and society. Crowdsourcing a really solid introduction to AI ethics seems like something that would be easy enough to achieve. People have a lot of somewhat divergent thoughts on what exactly AI ethics includes. This effort is my attempt at structuring an introduction to AI ethics. I’ll take this and produce an independent study version of this using Overleaf LaTex to make it extra awesome. My last run at this was about 6 months ago during week 86 that was primarily about ML ethics centering on fairness, bias, and privacy as core topics [1]. It might be a good start to dig deeper this time around and look for more topics to include with a larger focus on AI and not staying limited to the world of ML. I cloned my previous project in Overleaf and then removed all the extra content. Not only should it be pretty easy to take this structure or outline and expand it to a longer form for publication, but also it is probably time to revisit using LaTeX before that rather academic skill set fades.What are other major topics to consider besides fairness, bias, and privacy?* Laws and legislation* Intellectual property * Academic citation* Human-centered design* Examples (applied or theoretical)* Data sourcing* Labor practices* Frameworks* Operationalizing* MisinformationYou could go out and read “The Ethics of Artificial Intelligence” by Nick Bostrom and Eliezer Yudkowsky [2]. Or maybe go out to the “Everyday Ethics for Artificial Intelligence” from the team over at IBM [3].As a little bonus for you here, instead of closing out this Substack post I’m going to provide you with a few chapters from a session of content generation I had with OpenAI’s ChatGPT. The session was seeded with a request to write an introduction to AI ethics. You will find that my prompts are very direct and highly word economic. The output here is really just based on getting the ChatGPT large language model to keep generating as the session output is limited per interaction. You will see evidence of this as it sometimes just sometimes appears to stop mid chapter or closes out abruptly.OpenAI ChatGPT Prompt: “write a summary for the book”The book is about the ethical considerations of artificial intelligence (AI) and its impact on society. It covers a wide range of topics including bias and fairness, transparency and explainability, privacy and security, the impact of AI on employment and the economy, the use of AI in sensitive domains, and the future of AI and its potential impact on society. The book emphasizes that ethical considerations surrounding AI are complex and multifaceted, and that it is crucial to stay informed about the latest developments in the field. It also highlights the importance of developing strategies to ensure that the benefits of AI are shared by all members of society, including policies such as retraining programs, universal basic income, and government oversight. The book also emphasizes that AI ethics is an interdisciplinary field that brings together experts from various domains such as computer science, philosophy, law, sociology, psychology, and policy.OpenAI ChatGPT Prompt: “why should i read this book”You should read this book because it provides a comprehensive overview of the ethical considerations surrounding AI and its impact on society. The book covers a wide range of topics that are important for understanding the current and future implications of AI on society. It highlights the potential benefits and risks of AI and provides strategies for addressing these challenges. Additionally, the book emphasizes the importance of staying informed about the latest developments in AI and the importance of developing policies to ensure that the benefits of AI are shared by all members of society.As AI is a rapidly advancing technology that has the potential to shape the future of humanity in ways that are difficult to predict, it is important to be aware of the ethical considerations surrounding AI. This book is a valuable resource for anyone who wants to understand the impact of AI on society and the ethical considerations that need to be taken into account as the technology continues to advance.OpenAI ChatGPT Prompt: “what would you title book”"AI Ethics: Navigating the Challenges and Opportunities of Artificial Intelligence"OpenAI ChatGPT Prompt: “tell me the chapters again”The book has the following chapters:Chapter 1: Introduction to AI EthicsChapter 2: Bias and Fairness in AIChapter 3: Transparency and Explainability in AIChapter 4: Privacy and Security in AIChapter

Lectures related to this topic are mostly about convolutional neural networks and visual recognition systems. My searches were explicitly for autonomous vehicles and did not directly include any of the methods being deployed. That in part was due to my interest in going into higher levels of analysis this time around. My searches of Google Scholar on this topic of autonomous vehicles returned results with a lot fewer citations than I was expecting to find [1]. Three of the papers in the first 30 results had more than 1,000 citations.Fagnant, D. J., & Kockelman, K. (2015). Preparing a nation for autonomous vehicles: opportunities, barriers and policy recommendations. Transportation Research Part A: Policy and Practice, 77, 167-181. https://www.caee.utexas.edu/prof/kockelman/public_html/TRB14EnoAVs.pdf Shah, S., Dey, D., Lovett, C., & Kapoor, A. (2018). Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference (pp. 621-635). Springer International Publishing. https://arxiv.org/pdf/1705.05065.pdf Bonnefon, J. F., Shariff, A., & Rahwan, I. (2016). The social dilemma of autonomous vehicles. Science, 352(6293), 1573-1576. https://arxiv.org/pdf/1510.03346 The rest of the scholarly articles seemed to have around a couple hundred citations. Maybe the collection of scholars interested in autonomous vehicle research is about that size. In practical terms it is probably a fairly expensive area of study. Even building simulated models would be computationally expensive. Collecting real world model data would be cost prohibitive for most researchers. Faisal, A., Kamruzzaman, M., Yigitcanlar, T., & Currie, G. (2019). Understanding autonomous vehicles. Journal of transport and land use, 12(1), 45-72. https://conservancy.umn.edu/bitstream/handle/11299/209218/JTLU_vol-12_pp45-72.pdf?sequence=1Schwarting, W., Alonso-Mora, J., & Rus, D. (2018). Planning and decision-making for autonomous vehicles. Annual Review of Control, Robotics, and Autonomous Systems, 1, 187-210. http://alonsomora.com/docs/18-schwarting-AR.pdfKato, S., Takeuchi, E., Ishiguro, Y., Ninomiya, Y., Takeda, K., & Hamada, T. (2015). An open approach to autonomous vehicles. IEEE Micro, 35(6), 60-68. http://cs.furman.edu/~tallen/csc271/source/openAppr.pdfYou will notice pretty quickly that all of those scholarly works are aging a bit compared to where we are now at the doorstep of modernity. I’ll be curious about digging into things a bit more over time to find more current state reviews of autonomous vehicles.Links and thoughts:Top 5 Tweets of the week:Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=autonomous+vehicles&btnG=&oq=autonomous+v What’s next for The Lindahl Letter? * Week 113: Structuring an introduction to AI ethics* Week 114: How does confidential computing work?* Week 115: A literature review of modern polling methodology* Week 116: A literature study of mail polling methodology* Week 117: A literature study of non-mail polling methodologyIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

We have reached the 7th installment of the new series for 2023 related to AI. In just a couple of weeks we are going to venture out of these framing topics into the heart of my journey back into polling and sentiment analysis. Toward the end of that journey it will be very clear on the relationships between AGI, AI, and sentiment analysis. Processing to the point of understanding then adding a degree of directionality in terms of sentiment may be beyond the current state of things at this moment. A lot of debate will go into what is understanding, the occurrence of it, and does it have to be sustained to be relevant. Commonly people try to sum that debate up within the context of next best alternative consideration and that is interesting. However, that debate could be summed up within the analogy of trying to argue that a really well drawn map could provide understanding. It certainly could provide the pathing and the next best routing assuming you have a general idea of where you are going and that path is covered on the map. Let’s zoom back out for a minute and look at a relatively current webinar. A recent one hour webinar from Stanford online professor Christopher Potts related to GTP-3 & Beyond was pretty interesting. It’s easy to find on YouTube and you can watch it faster than 1x if you enjoy that sort of thing.Outside of that consideration, I’m wondering if we are going to end up moving from people writing books and articles to a world where people are certifying books and articles for correctness. Some sort of model is used to produce content and then a person edits it to or just verifies its correctness. People have spent a lot of time trying to figure out how to handle natural language processing (NLP). As editors of that content people could act as the ultimate gatekeeper of knowledge and correctness within what is spit out of the systems. We are after all capable of processing natural language ourselves. It’s just a lot harder to explain or rationalize the rule systems that each individual person uses to achieve that objective.Here are 5 scholarly works related to NLP:Manning, C., & Schutze, H. (1999). Foundations of statistical natural language processing. MIT press. http://www.cs.cornell.edu/home/llee/papers/topost.pdfBird, S., Klein, E., & Loper, E. (2009). Natural language processing with Python: analyzing text with the natural language toolkit. " O'Reilly Media, Inc.".Chowdhary, K., & Chowdhary, K. R. (2020). Natural language processing. Fundamentals of artificial intelligence, 603-649. https://strathprints.strath.ac.uk/2611/1/strathprints002611.pdfHirschberg, J., & Manning, C. D. (2015). Advances in natural language processing. Science, 349(6245), 261-266. https://nlp.stanford.edu/~manning/xyzzy/Hirschberg-Manning-Science-2015.pdf Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R., Bethard, S., & McClosky, D. (2014, June). The Stanford CoreNLP natural language processing toolkit. In Proceedings of 52nd annual meeting of the association for computational linguistics: system demonstrations (pp. 55-60). https://aclanthology.org/P14-5010.pdf You could go out and watch Christopher Manning’s “Natural language processing with deep learning CS224N/Ling284” 23 video course on YouTube [1]. This one contains a lot of content. Maybe you were looking for a more media rich introduction to NLP. The team over at TensorFlow also has a collection of 6 videos on the topic that are going to be a lot more visually involved [2]. They are also considerably shorter than the 23 video course from Christopher Manning. Rounding this one out will be a video from the IBM Cloud team which was pretty decent. Links and thoughts:Top 4 Tweets of the week:Footnotes:[1] [2] What’s next for The Lindahl Letter? * Week 112: Autonomous vehicles* Week 113: Structuring an introduction to AI ethics* Week 114: How does confidential computing work?* Week 115: A literature review of modern polling methodology* Week 116: A literature study of mail polling methodologyIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

In total the largest knowledge graph ever created on the shoulders of giants happens to be the college academy of knowledge compiled as a part of our higher education traditions. That knowledge graph is inherently built on connections and the learning patterns of academics. It would have to be considered to be a very human based aggregate not a computer based one. It’s possible that could change over time, but I’m not entirely convinced that is going to happen in the short term. The totality of academics is a much more complex knowledge collection than any model. It also contains a lot of conflicting arguments and disagreement. Here is a video from IBM Technology that explains, “What is a knowledge graph?”If you wanted to really go deeper, then you could check out this video from Yanic.Wang, C., Liu, X., & Song, D. (2020). Language models are open knowledge graphs. arXiv preprint arXiv:2010.11967. https://arxiv.org/pdf/2010.11967.pdf Here are 3 academic papers that are highly cited within the context of knowledge graphs.Nickel, M., Murphy, K., Tresp, V., & Gabrilovich, E. (2015). A review of relational machine learning for knowledge graphs. Proceedings of the IEEE, 104(1), 11-33. https://ieeexplore.ieee.org/ielaam/5/7360840/7358050-aam.pdf Ji, S., Pan, S., Cambria, E., Marttinen, P., & Philip, S. Y. (2021). A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems, 33(2), 494-514. https://arxiv.org/pdf/2002.00388.pdf%E2%80%8Barxiv.org Wang, Z., Zhang, J., Feng, J., & Chen, Z. (2014, June). Knowledge graph embedding by translating on hyperplanes. In Proceedings of the AAAI conference on artificial intelligence (Vol. 28, No. 1). https://ojs.aaai.org/index.php/AAAI/article/download/8870/8729 Academic papers are really just a series of single meals within the greater academy of academic work. You have to really be committed to bringing together your overall thoughts on the field, the most relevant publications, and ultimately the totality of things happening in your field of academic study. You cannot for the most part hang your hat on any one academic paper being able to explain it all. Together with a multitude of other papers the content has more context. Sometimes people stop eating single serving meals to write introductions, textbooks, or other onramp materials to help people catch up. My fear has been that within the ML and AI spaces the flooding of content has made catching up almost impossible for people wanting to find the on ramp to getting up to speed on everything that is happening. This problem exists in different ways for researchers and scholars that are just trying to keep up with the endless stream of new content along the way. Every day, month, and year new papers are being published at rates that are mind boggling. Now that synthetic written content has gone mainstream thanks to ChatGPT kickstarting that marketplace. My honest guess here is that a lot of researchers will kickstart their writing endeavors with the rocket booster for written content creation that is ChatGPT. They can just edit it down and move along to publish more papers.That preamble is super important to trying to contextualize why understanding the knowledge graph is so important. An important factor within the larger aggregate knowledge graph we all share itself is changing. It’s really like somebody showed up to the party and invited way more people than could possibly attend. Yeah, I’m about to talk about the effects of content flooding on future knowledge graph management. When the intake protocols for the knowledge graph cannot tell the difference between synthetically created input data and the standard product that has been on the market for thousands of years. For the most part knowledge graph based algorithms have had a hard time dealing with satire and sarcasm content where the author is intentionally spinning a false narrative. You end up with two conflicting next best answers where one should have a very low confidence score. However, this new effort within ChatGPT is whole-sale different. You have content that looks and seems like it should be included in the knowledge graph, but it is not tested or curated in any way shape or form. It’s almost like a point in time has to be considered where things before ChatGPT and things after ChatGPT have different weights within the overall knowledge graph. That is a troubling idea to consider. Top 5 Tweets of the week:What’s next for The Lindahl Letter? * Week 111: Natural language processing * Week 112: Autonomous vehicles* Week 113: Structuring an introduction to AI ethics* Week 114: How does confidential computing work?* Week 115: A literature review of modern polling methodologyIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year

Right before sitting down to work on this post I recorded the audio for week 108. My voice has almost recovered all the way from catching that pandemic thing in December. It’s interesting that sustaining vocal quality would be one of the last things to stabilize from that distinctly unwanted experience. You have probably been able to hear that within the last few recorded podcasts. I just did not have that totality of vocal control. I might go back and record that audio segment again, but at this point it is entirely possible that won’t happen based on time and the weight of other commitments this week. I was also a little concerned when Audacity, the software I use for this recording, required an update before the recording started, but it was apparently a small one and I was able to start the recording process within a couple of minutes. It’s very possible that this missive would go out without an audio edition. You however know that is not the case.It turned out that I was able to sit down and record the audio version of this Substack post. Right now I’m working without my backlog due to some unavoidable things that disrupted my writing routine a bit. I’m hopeful that given a little bit of time a few weeks of backlog will build back up, but right now each post is as fresh as it will ever be as things are being written directly before being recorded as part of my weekend writing routine. Typically it is better to have the posts created within the 5 week planning and review cycle. At this point, that is not possible. Welcome to the now and being in the moment as the very freshest words are brought your way within this missive of consideration for Substack. For the most part households adopted some machines pretty quickly and in a sustained way. Depending on where you are, microwaves, refrigerators, freezers, dishwashers, washing machines, and dryers are a part of the technology footprint in households. None of those devices need to be connected to either WiFi or Bluetooth to be operational. They existed for many years without that type of connectivity. This essay happens to be about robots in the house and the aforementioned appliances are not really what people talk about in terms of modern robotics in the household. It’s something more mobile that gets a lot of attention. To be fair a lot of robot vacuum brands and companies now exist. They roam and clean, get stuck, and have to be rescued and maintained. I went out to find some papers that referenced the Roomba and they seem to have peaked between 2006 to 2007. Which I thought was a very interesting element to see within this literature review. They just sort of stopped in frequency. Here are three of them that were well referenced. Forlizzi, J., & DiSalvo, C. (2006, March). Service robots in the domestic environment: a study of the roomba vacuum in the home. In Proceedings of the 1st ACM SIGCHI/SIGART conference on Human-robot interaction (pp. 258-265). https://www.cs.cmu.edu/~kiesler/publications/2006pdfs/2006_service-robots-roomba.pdfTribelhorn, B., & Dodds, Z. (2007, April). Evaluating the Roomba: A low-cost, ubiquitous platform for robotics research and education. In Proceedings 2007 IEEE International Conference on Robotics and Automation (pp. 1393-1399). IEEE. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=15cb2263caf26ac68906858093ae8d7749ad7827 Jones, J. L. (2006). Robots at the tipping point: the road to iRobot Roomba. IEEE Robotics & Automation Magazine, 13(1), 76-78. https://www.researchgate.net/profile/Joseph-Jones-14/publication/3344755_Robots_at_the_tipping_point_the_road_to_iRobot_Roomba/links/5728aa7308ae2efbfdb7dce8/Robots-at-the-tipping-point-the-road-to-iRobot-Roomba.pdf None of these robots in the house are equipped with any conversational subroutines. Within the worlds created by science fiction writers it is not uncommon for the robots in the house to talk back and demonstrate some degree of personality. We currently have no legitimate AGI that would facilitate that type of exchange. It’s probably up next at some point. Now that both Microsoft (powered by OpenAI) and Google are trying to create chat-like interactions I’m guessing that chatting with the robots in the house will arrive as a feature. Right now in Google Scholar searches you can find almost 5,000 results for ChatGPT [1].Links and thoughts:Top 6 Tweets of the week:Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=chatgpt&btnG= What’s next for The Lindahl Letter? * Week 110: Understanding knowledge graphs* Week 111: Natural language processing * Week 112: Autonomous vehicles* Week 113: Structuring an introduction to AI ethics* Week 114: How does confidential computing work?If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead. This is a public episode. If you would li

Part of what made Twitter so interesting is the diversity of argument and the townhall nature of it being the first place things show up in the feed. I’m not sure any other company or social platform could attract the same amount of hyperactive content creation users geared at news and coverage of the moment. People are arguing and I’m sure papers will soon be arriving to describe the end of social media. This weekend I’m going to spend a bit of time reading a book by Robert Putnam of “Bowling Alone” fame called “The Upswing”.Putnam, R. D. (2000). Bowling alone: The collapse and revival of American community. Simon and schuster.Putnam, R. D. (2020). The upswing: How America came together a century ago and how we can do it again. Simon and Schuster.It’s probably somewhere in the meta analysis between social capital and social media that a compelling story about why Twitter as a company would not happen today exists. During the course of this analysis you are going to receive two different lines of inquiry. First, I’ll consider the nature of Twitter and a few books related to it and silicon valley in general. Second, we will dig into some of the AI and sentiment analysis scholarly work related to that field of study to help keep the writing trajectory for the year on track. Books have arrived to tell the stories of what happened in Silicon Valley. A lot of unlikely things happened within the borders of the space described as silicon valley. Some of them will be a part of business courses for decades to come. It truly is an interesting thing that happened where so much creativity and output happen in such a relatively small area. Three of the books that I have enjoyed are listed below.Bilton, N. (2014). Hatching Twitter: A true story of money, power, friendship, and betrayal. Penguin.Frier, S. (2021). No filter: The inside story of Instagram. Simon and Schuster.Wiener, A. (2020). Uncanny valley: A memoir. MCD.You can zoom out a bit and grab some classic silicon valley reading like:Isaacson, W. (2014). The innovators: How a group of inventors, hackers, geniuses and geeks created the digital revolution. Simon and Schuster.A lot of scholars over the years have focused their attention on Twitter for a variety of purposes. You can imagine that my interest and the interest of those scholars overlap around the ideas of AI and sentiment analysis. Digital agents abound within the Twitter space and some of them are doing some type of sentiment analysis with what scholars are identifying as artificial intelligence. That second part of the equation makes me a little bit skeptical about the totality of the claims being made. We will jump right into the deep end of Google Scholar on this one anyway [1].Papers from a search for “Sentiment analysis Twitter artificial intelligence” [2]Kouloumpis, E., Wilson, T., & Moore, J. (2011). Twitter sentiment analysis: The good the bad and the omg!. In Proceedings of the international AAAI conference on web and social media (Vol. 5, No. 1, pp. 538-541). https://ojs.aaai.org/index.php/ICWSM/article/download/14185/14034 Ghiassi, M., Skinner, J., & Zimbra, D. (2013). Twitter brand sentiment analysis: A hybrid system using n-gram analysis and dynamic artificial neural network. Expert Systems with applications, 40(16), 6266-6282. Giachanou, A., & Crestani, F. (2016). Like it or not: A survey of twitter sentiment analysis methods. ACM Computing Surveys (CSUR), 49(2), 1-41. https://arxiv.org/pdf/1601.06971.pdf Alsaeedi, A., & Khan, M. Z. (2019). A study on sentiment analysis techniques of Twitter data. International Journal of Advanced Computer Science and Applications, 10(2). https://www.researchgate.net/profile/Abdullah-Alsaeedi/publication/331411860_A_Study_on_Sentiment_Analysis_Techniques_of_Twitter_Data/links/5c78175ba6fdcc4715a3d664/A-Study-on-Sentiment-Analysis-Techniques-of-Twitter-Data.pdf I had considered some evaluation of searches for both “opinion mining Twitter artificial intelligence” and “artificial intelligence analysis of public attitudes” [3][4]. It’s possible some papers from both of those searches show up later. Generally, all of that argument and content could be broken down into two camps of intelligence gathering related to advertising and general opinion mining geared at understanding sentiment. One divergent thread of research from those two would be some of the efforts to identify fake or astroturf content. You can imagine that flooding either fake or astroturf content could change the dynamic for advertising or sentiment analysis. Advertising to a community of bots is a rather poor use of scarce resources.Links and thoughts:Top 5 Tweets of the week:Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=twitter+artificial+intelligence&oq=twitter+artif [2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=Sentiment+analysis+Twitter+artificial+intelligence&btnG= [3] https://scholar.google.com/scholar?q=Opinion+mining+Twitter+artificial+intelligence

This week it seemed like a good idea to take a look at some of the most highly cited AI papers. Back during week 81 of this journey into writing on Substack, I took a look at some of the most highly cited ML papers [1]. I was expecting a lot more overlap, but was pleasantly surprised at the differences. One of the papers really stood out based on the total number of citations and it’s up first. Intellectually I can accept that a paper has more than 100,000 citations, but in practice that is an awful lot of references for an academic paper to have and a representation of a degree of asynchronous interaction between researchers that helps bring the intellectual space called the academy to life.Papers with over 100,000 citations:* Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. https://arxiv.org/pdf/1412.6980.pdf Papers with over 50,000 citations:* Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28. https://proceedings.neurips.cc/paper/2015/file/14bfa6bb14875e45bba028a21ed38046-Paper.pdf * Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30. https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Papers with over 20,000 citations: * Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., ... & Hassabis, D. (2015). Human-level control through deep reinforcement learning. nature, 518(7540), 529-533. https://daiwk.github.io/assets/dqn.pdf * Ioffe, S., & Szegedy, C. (2015, June). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning (pp. 448-456). PMLR. http://proceedings.mlr.press/v37/ioffe15.pdf Papers with over 10,00 citations:* Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., ... & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. nature, 529(7587), 484-489. * Arjovsky, M., Chintala, S., & Bottou, L. (2017, July). Wasserstein generative adversarial networks. In International conference on machine learning (pp. 214-223). PMLR. http://proceedings.mlr.press/v70/arjovsky17a/arjovsky17a.pdf * Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. https://arxiv.org/pdf/1609.02907.pdf Without question these levels of citation are an indication that the works are being read and actively referenced within the scholarly community. I’m referencing the citation numbers here to give you a sense of scale when it comes to considering the AI community and how many people are researching and considering the things happening in this space. This is a very crowded and vibrant place in the academy where a lot of time and effort are going into moving things along toward building very real and deployable technology in this space. Given the sheer volume of people working in this space it’s only a matter of time before somebody will shout “Eureka!” and we see practical deployments in production which will influence our daily lives.What would ChatGPT create?If you were wondering what ChatGPT from OpenAI would have generated with the same prompt, then you are in luck. I had that output generated over at https://chat.openai.com/chat by issuing a prompt.Links and thoughts:Top 5 Tweets of the week:Footnotes:[1] Week 81 of The Lindahl Letter:What’s next for The Lindahl Letter? * Week 108: Twitter as a company probably would not happen today* Week 109: Robots in the house* Week 110: Understanding knowledge graphs* Week 111: Natural language processing * Week 112: Autonomous vehiclesIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead.Lindahl, N. (2023). The Lindahl letter: 104 Machine Learning Posts. Lulu Press, Inc. https://www.lulu.com/shop/nels-lindahl/the-lindahl-letter-104-machine-learning-posts/ebook/product-y244ep.html This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

You probably were wondering how long into the new year before a bunch of focus and attention were placed at the efforts of Hugging Face. You won’t have to wait any longer as this missive will dig into BigCode and some other efforts to use AI to build out code generating systems [1]. I was reading a TechCrunch article from Kyle Wiggers and wondering about how many different systems existed [2]. That article references 5 code generating systems that you could in practice elect to go evaluate. For completeness I’m listing Codex and Copilot separately in this last given that the interfaces are holistically different. * BigCode - Hugging Face & ServiceNow’s R&D division [3]* AlphaCode - DeepMind [4]* CodeWhisperer - Amazon [5]* Codex - OpenAI [6]* Copilot - GitHub (Codex based) [7]One of the things you might be interested in learning about at this point would be a dataset called “The Stack” which happens to be a collection of 6 terabytes of permissive code data that covers 300 programming languages [8]. The permissive code part of the dataset is interesting. The GitHub archive was roughly 69 terabytes of data that they filtered by licensing which they considered permissive and ended up with that 6 terabyte collection. Understanding how the dataset that feeds the code generating system was built is very important. All my contributions on GitHub are intended to be MIT license which I think should be permissive [9]. You have to deeply consider that a lot of propriety code writers and corporations employing said coders would not have given permission to use their code in a code generation system. Generative coding systems will abound shortly and are in an early and developing state at the moment. We are getting to the point where you can instruct Codex to build something code related in terms of creating an application and you might get a great result. It’s not a universal code generation engine at this point. However, we are getting closer and closer to conversational code generation or some flavor of that outcome which I would classify as a generative coding system. It will be a seismic shift in code generation based on democratizing the creation of applications. What would ChatGPT create?If you were wondering what ChatGPT from OpenAI would have generated with the same prompt, then you are in luck. I had that output generated over at https://chat.openai.com/chat by issuing a prompt.Links and thoughts:Top 5 Tweets of the week:Footnotes:[1] https://huggingface.co/bigcode [2] https://techcrunch.com/2022/09/26/hugging-face-and-servicenow-launch-bigcode-a-project-to-open-source-code-generating-ai-systems/ [3] https://www.bigcode-project.org/[4] https://alphacode.deepmind.com/[5] https://aws.amazon.com/codewhisperer/ [6] https://openai.com/blog/openai-codex/ [7] https://github.com/features/copilot [8] https://huggingface.co/datasets/bigcode/the-stack [9] https://github.com/nelslindahlx What’s next for The Lindahl Letter? * Week 107: Highly cited AI papers.* Week 108: Twitter as a company probably would not happen today* Week 109: Robots in the house* Week 110: Understanding knowledge graphs* Week 111: Natural language processingIf you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.Lindahl, N. (2023). The Lindahl letter: 104 Machine Learning Posts. Lulu Press, Inc. https://www.lulu.com/shop/nels-lindahl/the-lindahl-letter-104-machine-learning-posts/ebook/product-y244ep.html This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

We are now starting the third year journey of The Lindahl Letter on Substack. You can find the first two years worth of that effort combined into one manuscript right here or at any other online retailer:Lindahl, N. (2023). The Lindahl letter: 104 Machine Learning Posts. Lulu Press, Inc. https://www.lulu.com/shop/nels-lindahl/the-lindahl-letter-104-machine-learning-posts/ebook/product-y244ep.html During the course of this next year I’m probably going to write a book about how I would set up a modern polling company or to put it more bluntly, a book about how to monetize the sentiment measuring process in modern America. Part of that book writing effort could end up being a part of this series or it could end up being a stand alone set of content. Yes, that would mean 2023 would yield at least two books of exceptionally enjoyable reading content. Overall, this new writing direction is one of those things that I’m going to really dig into throughout 2023. This would be inline with my research trajectory and current five year writing plan. My thoughts have just become a little more focused on what exactly I would produce within that writing window. Naturally, that is a good thing to start forming up before the writing project begins. I’m going to need to sit down and write up the topics to include within my literature review on this one. That part of this analysis will be key and has to be pretty darn good at this point. Within that work effort I’m considering a few article titles to include in the writing plan:* “Sentiment pooling: Applied multimedia polling”* “Beyond the phone: A study of multimedia polling methods”* “When nobody answers: Evaluating the efficacy of phone based polling”* “Automated sentiment analysis compared to respondent based sentiment”* “Beyond the paywalls of academic journals: how information is being shared more broadly.”We will see which of those ends up getting produced as either book chapters or true stand alone articles. This last week of planning and review was actually pretty productive. I have sketched out a writing plan for about the next 25 weeks. This new Substack series now has a backlog to work against as posts get put into a five week planning and review cycle. As we move from 2022 to 2023, my primary weekly Substack based content creation is going to shift a bit from ML related coverage to a stronger AI focus mixed with more content on modern polling methodologies. You might be wondering why the combination of AI and sentiment analysis (polling methodologies) happens to be important to me as something to consider. As the intersection of technology and modernity occurs certain things will happen. Part of understanding civil society and figuring out what happens to civility during that intersection will require better active sentiment analysis. Critical thought and analysis into understanding the nature of measuring modern sentiment will be essential on a go forward basis based on the automation that AI will introduce throughout business.I was looking at the proposed content from weeks 105 to 132 and realized that ethics as a part of the AI journey and process were left out as an early topic. I’m going to try to work in some evaluation of and references to academic works that consider ethics and responsible AI each week as we move forward. As we close out here at the start of this series it would be only fitting to share what I do consider to be a foundational AI book to consider as key reference material. Probably the best book about AI you could grab is a textbook called, “Artificial Intelligence: A Modern Approach,” by Stuart Russell and Peter Norvig [1]. The current version appears to be the 4th edition [2]. You can get a copy on eBay or some other used book proprietary for a lot less than the academic asking price. Some people prefer the findings of a certain alternative deep learning historian. I don’t want to deprive those readers of a link to what they might be looking for as fresh reading material. I am aware that Jürgen Schmidhuber has started writing some histories. It does present some alternative findings and different points of view. You can now access and read 75 pages of that content that have been recently published and are pretty easy to download from arXiv.Schmidhuber, J. (2022). Annotated History of Modern AI and Deep Learning. arXiv preprint arXiv:2212.11279. https://arxiv.org/ftp/arxiv/papers/2212/2212.11279.pdf Outside of those two works a book that would be more conversational and easier overall to absorb about where AI is at current would be this one:Kissinger, H. A., Schmidt, E., & Huttenlocher, D. (2021). The age of AI: and our human future. Hachette UK.This is the start of a series that will last 52 weeks. It should be an interesting journey. What would ChatGPT create?If you were wondering what ChatGPT from OpenAI would have generated with the same prompt, then you are in luck. I had that output generated over at https://chat.openai.com/chat by i

My appreciation for you reading this sentence right now is very real. That holds true in the moment of writing and within my intentions moving forward. At the start of this project a bit of writing on Substack happened two years ago now. During the course of starting that effort, I had not considered the amount of effort sustaining it would require. Probably the single most sustained writing project I have ever completed was my doctoral dissertation [1]. That writing process took less time than producing the 104 posts that are currently a part of The Lindahl Letter. I know some future writing projects could involve more effort, but that seems unlikely given the amount of time that was devoted to this one. You can take a moment and reflect on the realization that we have reached the point in the program where 104 weeks of content creation has occurred. Over time the writing process ended up following a planful approach to getting things done. The structure ended up being a block of writing, links/thoughts, top five tweets, and footnotes. * Weekly topic coverage. This is the heart of what is happening. Five weeks in planning or review is what makes the magic happen. For the most part, I ended up with a backlog with a bunch of planned posts. You should always keep a writing backlog of topics that deserve attention but might not have been advanced before. This method of keeping a writing topic backlog let me work on a couple of different weeks of content at any one time. Believe it or not, this is really important to keep things moving along. Momentum in writing is a real and present element of the creative process. If you run into a bit of writer’s block, or the inspiration is not showing up for a given topic, then you can move ahead and shift to a topic that allows the creation to continue. I’m not one to sit back and brew some word tea, but switching topics and trying to mix things up with a bit of stream of consciousness is certainly an option. * Links and thoughts. During the course of the week, I end up selecting some links and from research. Depending on where I’m at during the planning and review process, these links are either very timely or could have been pulled a week or two before. * Top five tweets. This section is really just for my own amusement. Based on the data from the last two years, nobody ever clicks on any of these tweets. I’m getting so close to having published 10,000 tweets. It will probably happen toward the end in the next couple of months. * Footnotes. Most writers in the academic space are comfortable with footnotes. Sometimes it takes people a bit of reading to get used to the links, references, and footnotes that I drop into the things being published. It is something that I plan on continuing with to help provide the context that we stand on the shoulders of giants in terms of our intellectual library of thoughts and considerations. The academy lives based on the interaction of scholars. That is something that cannot be forgotten or set aside to write prose without acknowledgement of the contributions of others. * Podcast. Adding the podcast element changed my posture to working to stay a week or two ahead. A part of that change was that I had to now plan ahead for vacation windows. That five-week planning and review cycle was key to keeping the publishing streak alive for what is approaching two years. After Post 104, I’m going to switch things up and focus more on sharing the content I find interesting in the AI and technology space [2]. Part of that is planned and was noted out in my five-year writing plan. In case you were wondering, I have included my five-year writing plan as of March 3, 2022: Year 1: Keep a heavy machine learning focus for the rest of 2022. Finish writing a collected series of machine learning/AI essays on Substack and combine them into a manuscript, The Lindahl Letter: 104 Machine Learning Posts. This manuscript should include both Years 1 and 2 of the Substack series. * Keep writing weekly Substack posts. * Take time for the manuscript generation process at the end of the year. * That manuscript will need to be edited by a professional before the print edition goes live. * Rework last year’s speaking engagement talks into academic papers. This could be one combined paper or potentially five different papers depending on how the initial effort shapes up. * “What Is Machine Learning Scale? The Where and the When of Machine Learning Usage” * “The Machine Learning Scale Problem: Thinking About Where and When to Use Machine Learning, ROI Models, Synthetic Data, Repeatable Frameworks, and Teams” * “Applied Machine Learning ROI: Understanding Machine Learning ROI From Different Approaches at Scale” * “Demystifying Applied Machine Learning: Building Frameworks and Teams to Operationalize Machine Learning at Scale” * “Figuring Out Applied Machine Learning: Building Frameworks and Teams to Operationalize Machine Learning at Scale, V3” * Rerun the MLOps GitHub re

You probably have gotten a sense of the tremendous and overwhelming flood of publishing that is happening in the machine learning space. So much content is being created right now that nobody could possibly consume it all. The flood of content is real and overwhelming. That is one of the reasons that I work really hard to distill the complex topics I select into a readable format for people to consume. I really try to provide a pathway to people who want to be a part of this journey. Maybe those two observations are inherently in conflict, but I think helping people navigate the deluge of information to focus on key things helps break down the conflict. As I sit down to rethink the future of machine learning, my thoughts are circling back to some of the original things I was writing about use cases and how people select what they want to work on based on ROI. Right now, machine learning is being built into everything, and we are seeing a creative explosion of people using machine learning to generate things on the fly that otherwise would have never been possible. Some of the models within the image- and video-creation spaces are really changing how we interact with the world. My honest guess about the future of machine learning is that it will become seamless within the background of our daily lives. It will be ever present and constantly just on the edge of how we perceive and interact with the world around us. I grabbed five papers where the future of machine learning was discussed from 2001 to 2018. Zhou, Z. H. (2016). Learnware: On the future of machine learning. Frontiers in Computer Science, 10(4), 589–590. https://www.lamda.nju.edu.cn/publication/fcs16learnware.pdf Mjolsness, E., & DeCoste, D. (2001). Machine learning for science: State of the art and future prospects. Science, 293(5537), 2051–2055. https://www.researchgate.net/profile/Eric-Mjolsness/publication/11789794_Machine_Learning_for_Science_State_of_the_Art_and_Future_Prospects/links/09e415147695a42e12000000/Machine-Learning-for-Science-State-of-the-Art-and-Future-Prospects.pdf Choy, G., Khalilzadeh, O., Michalski, M., Do, S., Samir, A. E., Pianykh, O. S., Geis, J. R., Pandharipande, P., Brink, J. A., & Dreyer, K. J. (2018). Current applications and future impact of machine learning in radiology. Radiology, 288(2), 318–328. https://doi.org/10.1148/radiol.2018171820 Obermeyer, Z., & Emanuel, E. J. (2016). Predicting the future—Big data, machine learning, and clinical medicine. The New England Journal of Medicine, 375(13), 1216–1219. https://doi.org/10.1056/NEJMp1606181 Handelman, G. S., Kok, H. K., Chandra, R. V., Razavi, A. H., Lee, M. J., & Asadi, H. (2018). eDoctor: Machine learning and the future of medicine. Journal of Internal Medicine, 284(6), 603–619. https://doi.org/10.1111/joim.12822 What’s next for The Lindahl Letter?* Week 104: That 2nd year of posting recapI’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

You end up with people who are working in a field and people who study the field of inquiry academically. Sometimes and especially within the machine learning space you end up with people who are actively working as a practitioner. Those very same practitioners of the craft of machine learning are publishing academic articles at a rate never before seen within any field of study at large within the academy. The best way to describe that effort would be to call them the pracademics of the machine learning space. Intellectually it’s probably good to have people write about things who really understand how they are occurring. You certainly get solid firsthand accounts of what people are creating. Research is the process of digging into things, investigating, studying, or maybe just seeking to understand the things better. Original research is often brought out to describe novel inquiry. Some of these pracademics are pushing things into a new frontier for machine learning. The DALL-E-2 AI system was introduced and changed the way people think about how an AI system would create realistic images and art based on a prompt [1]. During the course of working on this Substack post, you can imagine I was surprised to find a webpage called “AI Brain Drain” [2]. That site opens up with a clear, “Welcome to the AI Brain Drain Index.” They pretty much are tracking the number of AI faculty that have left academic areas to go work within industry. You can see charts and other figures that sort of run from 2004 to apparently 2018. The researchers must have stopped making charts in the last few years, but the ones they made are pretty nice. All of that research seems to have yielded a paper you can read [3]. According to the SSRN website where I downloaded the paper, 928 downloads have occurred on this one. Gofman, M., & Zhao, J. (2022, July 31). Artificial intelligence, education, and entrepreneurship. Journal of Finance [Forthcoming]. https://doi.org/10.2139/ssrn.3449440 or at SSRN: https://ssrn.com/abstract=3449440 An article from Inside Higher Education called “AI Academy Under Siege” from author Oren Etzioni took a look at AI experts leaving institutions of higher education and what might be some potential solutions to that situation [4]. You could find more from the work of Professor Michael Gofman [5]. You could check out an editorial published in Springer by Lars Kunze called, “Can We Stop the Academic AI Brain Drain?” [6]. Outside of those sources, one of the articles that I really liked was from Ben Dickson titled, “What Is the AI Brain Drain?” [7]. You could pivot to an article that I enjoyed less called, “Brain Drain of AI Researchers: Academia vs Industry” [8]. Let’s close this one out with an interesting look at, “AI Brain Drain to Google and Pals Threatens Public Sector’s Ability to Moderate Machine-Learning Bias” [9]. That article links back out to a paper that was interesting. Jurowetzki, R., Hain, D. S., Mateos-Garcia, J., & Stathoulopoulos, K. (2021). The privatization of AI research(-ers): Causes and potential consequences. https://regmedia.co.uk/2021/02/04/drain.pdf Top 5 Tweets of the week:Footnotes:[1] https://openai.com/dall-e-2/ [2] http://www.aibraindrain.org/[3] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3449440[4] https://www.insidehighered.com/views/2019/11/20/how-stop-brain-drain-artificial-intelligence-experts-out-academia-opinion[5] https://simon.rochester.edu/blog/deans-corner/brain-drain[6] https://link.springer.com/article/10.1007/s13218-019-00577-2 [7] https://bdtechtalks.com/2019/09/26/artificial-intelligence-brain-drain/ [8] https://medium.com/codex/brain-drain-of-ai-researchers-academia-vs-industry-8e385e8fd517[9] https://www.theregister.com/2021/02/04/ai_brain_drain/What’s next for The Lindahl Letter?* Week 103: Rethinking the future of ML* Week 104: That 2nd year of posting recapI’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

One of the core topics within the machine learning space that I have studied happens to be ROI or, more to the point, an examination of just how well spent money would be in the space. You can certainly spend money on machine learning and AI related efforts as a part of a think tank, an independent lab, or a pure research institution. Within corporate spaces, pure research and development is one thing, but generally to expend the precious capital resources of an institution, organization, or group, you want to know that some type of return will be accumulated from those activities. It’s inherent to the nature of the venture into running a corporation compared to running some other type of organization. ROI for machine learning is a topic that deserves consideration. You can evaluate a variety of potential machine learning use cases to solve problems. Going from being a special product to an operationalized business process that utilizes technology to get things done will be dependent on the ROI associated with the use case. If it costs more to do it, then you could reasonably expect that use case is going to make most responsible actors pump the brakes on the project. However, it appears that a lot of use cases went forward anyway. I spent some time trying to find a good accounting of how much money has been spent on machine learning projects overall and how many of them have actually yielded solid ROI. You probably will not be all that surprised to learn that very few rigorous studies exist of successful ROI in the machine learning space. If some of those types of studies exist and I just missed them during my search, then by all means feel free to share them in the comments and let me know [1]. It won’t hurt my feelings or anything, and I’d actually be a little bit relieved that research on the subject exists. Before we conclude here, I do want to share one paper that does seem to be directly addressing this question. It has only been cited by three other papers since 2020. I had hoped it would lead me to a cluster of academic research. That was not the case. The search for solid academic research on machine learning ROI is still ongoing. Mizgajski, J., Szymczak, A., Morzy, M., Augustyniak, Ł., Szymański, P., & Żelasko, P. (2020). Return on investment in machine learning: Crossing the chasm between academia and business. Foundations of Computing and Decision Sciences, 45(4), 281–304. https://sciendo.com/article/10.2478/fcds-2020-0015 Links and thoughts:“We Talked To A VP At Microsoft - WAN Show December 23, 2022”Top 5 Tweets of the week:Footnotes:[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=%22return+on+investment%22+%22machine+learning%22&btnG=What’s next for The Lindahl Letter?* Week 102: ML pracademics* Week 103: Rethinking the future of ML* Week 104: That 2nd year of posting recapI’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

You might remember back to Week 75 when I tried to explain the nature of overcrowding within the machine learning space and what exactly that is doing to engineering colleges. Back in 2020, Igor Susmelj wrote a piece in Towards Data Science, “How to Keep up With the Latest Research and Trends in ML” [1]. Within that paper, Igor raised awareness of a chart from Jeffrey Dean that showed machine learning arXiv papers versus Moore’s law growth rate [2]. A quick Google Scholar search of that paper will show that it is cited by 72 papers at the moment [3]. I had hoped that a batch of papers would help me find more academic work about overcrowding and machine learning, but that did not really turn out to be the case. I have been trying to figure out the right words to describe intellectual overcrowding and machine learning, but for the most part I have not been able to figure out the secret decoder ring settings to locate a bunch of papers on the subject. I noticed a chart on page 10 of a paper called, “An Overview on Applications of Machine Learning in Petroleum Engineering,” which showed the number of publications with AI/machine learning within the SPE OnePetro digital library [4]. Just like the chart from Jeffrey Dean, it showed a clear takeoff point after 2010, where the terms just skyrocket like a hockey stick. That is probably a fairly consistent trend, and things will shake out over time to a key set of academic articles that get referenced a lot and a core set of topics that are covered within the AI/machine learning academic community. At the moment, however, we are at the peak of the inflection point, where intellectually everybody rushed to be first in the pool and kept on swimming. With a rush of academic focus in the area of machine learning, we will see both a great deal of progress and a potentially calamitous fallout from intellectual overcrowding. Only so much progress and ultimately only so many faculty positions are going to exist within engineering programs. I do think we will see a pretty significant oversupply in the number of people seeking those faculty positions at the more prestigious set of academic institutions. Right now it is a lot more lucrative for people who are able to get jobs within industry to do that and to enjoy some pretty solid compensation. That is probably the element that has made the overcrowding less problematic during the rise of machine learning implementations. People being able to work in the private section and people working within academic spaces have been able to find work. You can find a lot of articles about brain drain within academic institutions related to both AI and machine learning. Some of those pieces make some interesting arguments. A lot of private organizations have run labs that are doing things of a more academic nature than perhaps applied use case development. Links and thoughts:Top 4 Tweets of the week:Footnotes:[1] https://towardsdatascience.com/how-to-keep-up-with-the-latest-research-and-trends-in-ml-a45a356b1001 [2] https://arxiv.org/ftp/arxiv/papers/1911/1911.05289.pdf [3] https://scholar.google.com/scholar?cites=3848695121612936760&as_sdt=4005&sciodt=0,6&hl=en [4] https://www.researchgate.net/publication/339952951_An_Overview_on_Applications_of_Machine_learning_in_petroleum_Engineering What’s next for The Lindahl Letter?* Week 101: Back to the ROI for ML * Week 102: ML pracademics* Week 103: Rethinking the future of ML* Week 104: That 2nd year of posting recapI’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com