Interconnects

Last minute title change from: The tech industry can't agree on what open-source AI means. That's the process.How to read what multiple people mean by the word openness and see through the PR speak.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/flavors-of-open-source-ai0:00 The tech industry can't agree on what open-source AI means. That's the process.2:45 1. Effective Accelerationists, Techno-Optimists, capitalists, etc.3:39 2. Scientists, promoting understanding and transparency5:16 3. Inclusion, public interest, and fighting concentration of power6:19 4. Freedom advocates7:25 Dissecting "openness"Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/openness/img_004.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Databricks' new model is surpassing the performance of Mixtral and Llama 2 while still being in a size category that's reasonably accessible.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolshttps://www.interconnects.ai/p/databricks-dbrx-open-llm00:00 DBRX: The new best open model and Databricks' ML strategy03:36 The DBRX narrative07:33 Databricks' open LLM (and AI) strategy09:42 Playing with DBRX Instruct14:54 Digging for detailsFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_007.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_012.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_023.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_045.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_047.pngFig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_059.pngFig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_066.jpegFig 8: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/dbrx/img_068.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Evaluation is not only getting harder with modern LLMs, it's getting harder because it means something different.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/evaluations-trust-performance-and-price00:00 Evaluations: Trust, performance, and price (bonus, announcing RewardBench)03:14 The rising price of evaluation05:40 Announcing RewardBench: The First reward model evaluation tool08:37 Updates to RLHF evaluation toolsYouTube code intro: https://youtu.be/CAaHAfCqrBAFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/evals/img_026.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/evals/img_030.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/evals/img_034.pngFigure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/evals/img_040.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Where moats are tested now that so many people have trained GPT4 class models. Claude 3, Gemini 1.5, Inflection 2.5, and Mistral Large are here to party.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/gpt4-commoditization-and-moats00:00 Building LLM moats despite the commoditization of GPT404:38 The Open's opportunities08:02 It's amazing people still think LLMs aren't going to be useful09:50 Things that are comingFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/moats/img_004.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/moats/img_028.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/moats/img_032.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

A proposal for a new definition of an "open source" LLM and why no definition will ever just work.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/an-open-source-llm00:00 The koan of an open-source LLM03:22 A new naming scheme for open LLMs07:09 Pivot points and politics08:16 Claude 3, arms race, commoditization, and national security10:01 Doomers debunking bio risks of LLMs themselves11:21 Mistral's perceived reversal and the EU13:22 Messy points: Transparency, safety, and copyright13:32 The muddling of transparency15:22 The muddling of "safety"16:30 The muddling of licenses and copyright20:12 Vibes points and next stepsFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/open-source/img_046.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/open-source/img_064.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

This interview is available on podcast players and YouTube.I’m excited to bring you another interview! This one is a deep dive right in my wheelhouse — all things RLHF. Louis Castricato is probably the hidden star of RLHF in the open. I’m not sure anyone who can speak freely knows as much as him. As I’ve said again and again on Interconnects:Giving a voice to researchers is the best way to cut through the noise and understand what is happening with core developments of LLM technologies.Louis recently has been founding a new startup focused on synthetic data for alignment, Synth Labs, and is a researcher at Eleuether AI. This interview should speak for itself, and it’ll need re-listens, even for myself. The list of topics we cover touches on pretty much every major and minor issue facing model fine-tuning. Please reach out or comment if there’s a paper we mention that I didn’t link before. Happy to dig it up for you. For more on Synth Labs, there was a profile in Bloomberg from Rachel Metz.This post is very technical, more than usual. If you’re having a hard time with it, I suggest you listen to my RLHF 201 post on Latent Space first.ChaptersThese are generated with smol-podcaster with moderate edits.High-level chapters* 00:00:00: Introduction* 00:01:24: Gemini News and RLHF’s Part in it* 00:09:05: Long Context, In-Context, and Multimodal RLHF* 00:21:20: What are people missing about RLHF these days?* 00:30:30: OpenAI's Influence and the Need for Alternatives* 00:39:20: Synth Labs and the Future of Alignment* 00:55:00: Evaluation Talk p2: Open-ended Evaluation and Data Diversity* 00:59:20: Algorithm Roundup: PPO, DPO, KTO, IPO* 01:18:38: CarperAI, Early Days of RLHF, Reflecting on ChatGPTDetailed chapters* 00:00:00: Introduction and Overview of RLHF* 00:02:02: Gemini News, Custom Demographics in Image Prompts, and Controllability Issues in AI Models* 00:05:21: Fixing Biases in AI Models Post-Training, Representation in AI Data* 00:09:00: Multimodal RLHF and Video RLHF* 00:16:09: Evaluating Long Context Behavior in AI Models* 00:17:05: The Potential of In-Context RLHF* 00:21:24: Shift from PPO to DPO in RLHF* 00:23:19: Generalization and Evaluation in RLHF, Human Evaluation* 00:27:03: The Discrepancy Between Research and Company Needs in Alignment* 00:29:20: Impact of ChatGPT and Language Model Outputs on Data Sets* 00:31:39: The Concept of Uncensoring Models* 00:34:05: Lack of Safety Data Sets in Instruction Tuning* 00:35:23: LMSYS ChatBotArena, AlpacaEval, MT Bench p1* 00:39:25: Introduction to Synth Labs and Alignment Platform* 00:43:05: Developing OpenCAI Constitutional AI Data Set* 00:49:41: The Need for Open-Ended Evaluation in RLHF, eval p2* 00:54:13: The Importance of Releasing Models for RLHF Research* 00:58:17: Self-Instruction and Self-Rewarding LMs* 01:01:03: Working on RLHF at Carper AI* 01:04:25: Scaling PPO in RLHF* 01:08:01: The Impact of ChatGPT on Carper AI* 01:10:56: The Potential of KTO (Kahneman-Tversky Optimization)* 01:17:39: The Importance of Implementation Details in RLHF* 01:20:14: The Initial Focus at Carper AI* 01:23:36: The Future of RLHF and Open Science CollaborationInterconnects is a reader-supported publication. Consider becoming a subscriber.Papers & artifacts we discuss* Recursively Summarizing Books with Human Feedback* Needle in a haystack recent example repository. * Urial paper: The unlocking spell on base llms: Rethinking alignment via in-context learning* Misha paper from Deepmind: In-context Reinforcement Learning with Algorithm Distillation* Museli Optimizer: Muesli: Combining Improvements in Policy Optimization* Unintended Impacts of LLM Alignment on Global Representation* Pink Elephants Problem: Suppressing Pink Elephants with Direct Principle Feedback* Cut the Carp: Cut the CARP: Fishing for zero-shot story evaluation* MT Bench data for correlating human to GPT4 preferences Full transcriptNote: this is generated by smol-podcaster and has minor bugs post human edits.Nathan [00:00:01]: The ticker's going up. Welcome, Louis. You're the second guest on the InterConnects podcast, I think. It's an interesting one for me because everyone kind of points to me now as the person that is in the face of RLHF and I get a lot of questions and to me Louis has represented that person. I think Louis provided a lot most of the information on the first RLHF blog post that I wrote for Hugging Face back in the day. If there's somebody that I want to ask questions about RLHF, it generally goes to him. So now you all are gonna know this in the open. We're gonna cover a lot of things. As always, I'm trying to talk with researchers on the ground and people actually doing things in these topics. I think we're gonna cover a lot of things today. We're in the Latent Space podcast. If you're watching on video, you may have noticed that we're in the Latent Space studio and they reminded us we've got to start off with covering the Gemini news and what that means for RLHF and then most of this is

Basic tips on how to assess inbound ML content and cultivate your news feed.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/making-a-ml-feed00:00 How I assess all these AI releases01:22 1. Model access and demos are king of credibility02:31 2. Focus your feed on depth or breadth03:09 3. Examples of using the model normally show its usable, shockingly04:10 4. Leaderboards as the single leading claim is often anti-signal05:00 5. Basic deep learning conceptual checks will often save you06:13 6. If it's not even remotely reproducible or verifiable, it's not science07:10 7. Don't over-index on Twitter08:32 8. Data sharing, licenses, communication clarity, and small things add up08:58 9. Research papers, technical reports, blog posts, and Tweets all serve different purposes09:49 10. Socialize your information and build relationships This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Google rejoins the open model party and gets some backlash for a frequent problem for generative AI.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/gemma-google-ships-it00:00 Google ships it: Gemma open LLMs and Gemini backlash03:12 Getting to know Gemma07:11 Alignment details08:55 Aside: What is REINFORCE? Some history of RL11:08 Implementation details and RLHF12:18 Terms of use: RAIL Licenses history repeated14:05 Is Google back on top? Gemini's woesFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/gemma/img_008.webpFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/gemma/img_014.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/gemma/img_035.pngFigure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/gemma/img_051.pngFigure 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/gemma/img_055.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

10 Sora and Gemini 1.5 follow-ups: code-base in context, deepfakes, pixel-peeping, inference costs, and moreThis is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/sora-gemini-follow-up00:00 10 Sora and Gemini 1.5 follow-ups: code-base in context, deepfakes, pixel-peeping, inference costs, and more00:46 1. Deepfake detection of Sora01:59 2. Playing with long-context, problem settings, and prompting03:39 3. Gemini paper snooping: contamination and citation games05:42 4. Training data and token estimates of YouTube07:42 5. Unlocking model-based RL and downstream research08:52 6. Midjourney style matching, V-JEPA, replicating Sora in the open10:09 7. Architectures and academic links10:57 8. Pixel peeping from the arts11:58 9. Inference costs13:24 10. Pressure on Llama and Mistral14:03 11. Sound effects, physics, and the complete pictureFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_003.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_007.mp4Figure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_009.mp4Figure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_011.mp4Figure 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_037.mp4Figure 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_044.pngFigure 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_047.pngFigure 8: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-2/img_049.mp4 This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Emergency blog! Three things you need to know from the ML world that arrived yesterday.This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/sora-gemini-and-mistral-next0:00 OpenAI's Sora for video, Gemini 1.5, and a secret Mistral model0:53 Sora: OpenAI's text-to-video model4:59 Gemini 1.5: Google's effectively infinite context length8:01 Mistral-next: Another funny release methodFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-gemini-mistral/img_015.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-gemini-mistral/img_023.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-gemini-mistral/img_026.pngFigure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/sora-gemini-mistral/img_036.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

In an era dominated by direct preference optimization and LLMasajudge, why do we still need a model to output only a scalar reward?This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: In an era dominated by direct preference optimization and LLM-as-a-judge, why do we still need a model to output only a scalar reward?Podcast figures:Figure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/reward-models/img_004.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/reward-models/img_009.png0:00 Why reward models are still key to understanding alignment This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Scale's making over $750 million per year selling data for RLHF, who's coming to take it?This is AI generated audio with Python and 11Labs. Music generated by Meta's MusicGen.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/alignment-as-a-service00:00 Alignment-as-a-Service upstarts taking on Scale AI04:25 The competition with humans-in-the-loop06:05 Scaling Alignment-as-a-Service via AI feedbackPodcast figures:Figure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/aaas/img_008.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

A small model at the beginning of big changes.This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/olmo0:00 Open Language Models (OLMos) and the LLM landscape6:24 Thought experiments7:51 The LLM landscape heading into 2024Figure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/olmo/img_010.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Note: some of the audio in the second half is a little wonky, but the general voice was upgraded so hopefully it's a little less "poppy" until then!I'm trying to fix little pronunciation problems on a weekly basis. Thanks to my early fans! It'll keep improving. E.g. some of the months were wonky.When what seems like pure LLM black magic is actually supported by the literature.This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/model-merging00:00 Model merging lessons in The Waifu Research Department02:21 How and why does model merging work?07:13 Aside: merging vs. ensembles vs. mixture of experts08:21 Why are people doing this?11:22 Tools & Links11:51 Brief (visual) literature review12:07 Full model merging and recent methods15:55 Weight averaging during pretraining17:18 LoRA merging17:53 More backgroundFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_005.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_016.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_042.pngFigure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_051.pngFigure 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_055.pngFigure 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_058.pngFigure 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_060.pngFigure 8: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_062.pngFigure 9: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_065.pngFigure 10: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_075.pngFigure 11: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_077.pngFigure 12: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-merging/img_084.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

Local LLMs: the latency solution, Meta's open AGI, personalization myth, and moats X factorThe deployment path that'll break through in 2024. Plus, checking in on strategies across Big Tech and AI leaders.This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/local-llms0:00 Local LLMs: the latency solution, Meta's open AGI, personalization myth, and moats X factor4:15 The personalization myth7:13 Meta's local AGI and moats X factorsFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/local-llms/img_026.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

A fun demo on how generative AI can transform content creation, and tools for my fellow writers on Substack!This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/multimodal-blogging-tools0:00 Multimodal blogging tools2:57 Stratechery, passport, and wonderful customer experiences5:51 Wrap-up, features, and next stepsFigure 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/multimodal-blogging/img_006.pngFigure 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/multimodal-blogging/img_008.pngFigure 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/multimodal-blogging/img_012.pngFigure 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/multimodal-blogging/img_020.png This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

A sampling of recent happenings in the multimodal space. Be sure to expect more this year.This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/multimodal-rlhf00:00 Multimodal LM roundup: Unified IO 2, inputs and outputs, Gemini, LLaVA-RLHF, and RLHF questions02:46 Unified IO 2: Scaling multi-input, multi-output model pretraining07:47 Collecting preference data for images09:31 LLaVA-RLHF: The first experiments in multimodal RLHF fine-tuning13:20 Multimodal RLHF questions, ideas, and resources This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

And why the comparisons don't really matter. Repeated patterns in the race for reproducing ChatGPT, another year of evaluation crises, and people who will take awesome news too far.This is AI generated audio with Python and 11LabsSource code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/open-gpt4-limitations00:00 Where 2024's "open GPT4" can't match OpenAI's03:19 Models vs. products04:51 RLHF progress: Revisiting Llama 2's release and potential in 202408:30 Smaller scale open RLHF10:33 Opportunities12:24 Commentary This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.interconnects.ai/subscribe

This interview is on YouTube and podcast players.Giving a voice to researchers is the best way to cut through the noise and understand what is happening with core developments of LLM technologies. I’m excited to get to talk with Michael Poli (Stanford PhD student + research at Together AI) and Tri Dao (incoming professor at Princeton + Chief Scientist at Together AI). This builds on the mega-post from yesterday on the same topics, though the interview is obviously less math heavy:Interconnects is a reader-supported publication. Consider becoming a subscriber.Topics: Introductions | Why Attention works and may not scale | Quadratic scaling in attention | What is Striped Hyena | What is Mamba | Mamba hardware optimization | Predictions for 2024 architectures | More predictions for AIIntroductions[00:00:00] Nathan Lambert: Okay. Hey, everyone. Welcome to the first interview that we're going to post on interconnects. I'm really trying to bring more scientific voices into the AI discourse as media is covering a lot these days. I'm happy to be here with Michael Poli and Tri Dao, experts in some of these non attention architectures that have been really blowing up in the last few weeks of December.So, Michael, do you want to introduce yourself first?[00:00:25] Michael Poli: Sure. Thank you. Thank you, Nathan. For inviting me, I, do research at Together AI. And I was also a PhD student at Stanford, working with Stefano Ermon and, and, Chris Re, that's, that's how I met Tri as well. if I had to choose maybe, I moved to a few different areas of research.if I had to choose one, I like to, do research at the intersection of signal processing, dynamical systems, and deep learning, and most recently, luckily, there's been more interest in, in kind of efficient architectures that use some of these principles. to improve scaling, along different axes and to, to get sort of new, new trade offs at inference time.[00:01:13] Nathan Lambert: Great. And Tri?[00:01:16] Tri Dao: Everyone, thanks Nathan for, for, hosting us. really excited to be here. I'm Tri. I, just finished my PhD at Stanford. and I'm being assistant professor at Princeton, and right now I'm chief scientist at Together AI. it's, it's a startup working on AI infrastructure. And, yeah, I've been working at the intersection of machine learning and systems, so designing algorithms that take advantage of the hardware that, that they run on.I'm interested in, long range, dependencies, how to encode that into a model, designing architectures that can, can, learn long range dependencies. yeah, really excited to be here.Why Attention works and may not scale[00:02:01] Nathan Lambert: Okay. I think I'm going to, I have some questions dive right into this. I think you two will kind of both answer them or someone can answer longer, whatever you want.I think to start with, we should talk about maybe why attention works and what the limitations of attention are. I think. Almost every person in tech broadly now knows that a transformer is a model built with attention and chat GPT does that but like, why is this so good, even like how much of a transformer is built with attention are there other things going on, and what might be challenges there.[00:02:35] Tri Dao: Right. so, transformer which is this. Contexture that powers most of the exciting applications that we're seeing nowadays, as you mentioned, and so on. attention is kind of the core layer there, and attention actually became, earlier, around 2014, 2015, and then transformer came out, incorporating that, focusing a lot on, on, attention, with these, MLPs, interleaving, MLP and, and attention.And I think the success largely has been, They are, they seem to be able to scale really well so that you can scale up the models, with more, more parameters, with more data. And that has been the recipe for, for success. It sounds obvious now, but I think, five years ago that wasn't, that wasn't clear.so it seems like, you know, Transformer Architecture is, is a hugely successful one. and, you know, a couple of reasons why it's successful. I think it's like General enough that it's able to learn a lot from data. And two is, is pretty friendly to hardware. You can, there's no kind of sequential dependency like previous RNNs.so as a result, you can run it well on GPUs, TPUs. you can scale it up. It's very hardware efficient. I've personally have worked on making it more hardware efficient as well. So it's just kind of the recipe for, for success, which is, general architecture that scales well. if you're an NLP person, maybe you, you, you said, you know, incorporate some kind of inductive bias for, for, to protect, personally, I think it's a very general architecture that, that scales well and it's hardware friendly.[00:04:16] Nathan Lambert: Yeah. Yeah. It's, it's remarkable how it seems so obvious now and it's like. I think one of the things that studying this work is the context length becomes a really interesting access to stud