Interconnects Audio

The Open Source Initiative is working towards a definition.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/defining-open-source-ai0:00 On the current definitions of open-source AI and the state of the data commons3:17 Reasons to not mandate fully released data4:24 Sufficient but not exhaustive data docs5:22 Frustration with the data commons7:04 We need more examples to define the definitionFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/defining-open-source/img_005.png

The latest model from one of the most popular fine-tuning labs makes us question how a model should be identified as a "frontier model."This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/nous-hermes-30:00 Nous Hermes 3 and exploiting underspecified evaluations5:29 Parsing training lessons from Hermes 3Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_005.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_010.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_012.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_020.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_027.pngFig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_030.pngFig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_032.pngFig 8: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/nous-hermes-3/img_036.png

I had the pleasure of Talking with Ross Taylor (https://x.com/rosstaylor90), who has a great spectrum of unique experiences in the language modeling space — evaluation experience, Galactica lead author, Llama post training, etc. This is a really great conversation on the frontier of language model (LM) reasoning, LM deployments and demos, LM’s for science, RLHF, and other topics. I’ve been trying to get Ross to come on for a bit. He’s one of those people in the LM space that doesn’t speak too much, but when you do, you listen.Ross Taylor was previously an LLM lead at Meta AI, heading up the reasoning team. Previously he led the early work on LLM agents, and was the research lead on the Galactica project. Before that, he was a co-founder of Papers with Code, which was acquired by Meta in 2019. Before that, he has worked as a quant in sports betting and finance, and before that a policy advisor for the UK Government. He is currently working on a new startup.More details: https://www.interconnects.ai/p/interviewing-ross-taylor-on-llm-reasoning00:00:00 Introduction of Ross Taylor and his background00:02:12 Papers with Code00:09:58 Galactica, goals, controversy, legacy00:18:12 Technical details of the Galactica model00:23:18 Potential for language models to make scientific discoveries00:25:21 Defining and improving reasoning in language models00:32:38 Process-based reward models and their potential applications00:35:00 Generating synthetic data for SFT00:40:23 Evaluating the effectiveness of language models as judges for human preference data00:42:43 Considerations for creating base models that are easy to fine-tune00:46:45 Balancing SFT and RLHF00:54:13 Characteristics of successful post-training teams00:58:26 Future directions for language model development

Apple, Meta, and Nvidia all agree -- synthetic data, iterative training, human preference labels, and lots of filtering.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/frontier-model-post-training00:00 Llama 3.1 post-training and the new normal for RLHF01:18 A new standard pipeline01:45 Human preference data02:59 Scaling RLHF05:03 Synthetic data06:10 The new normal06:51 Data quality is king07:18 Apple confirms the new normalFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_018.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_020.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_031.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_033.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/frontier-rlhf/img_035.png

This week, I had the pleasure of chatting with Sebastian Raschka. Sebastian is doing a ton of work on the open language model ecosystem and AI research broadly. He’s been writing the great Ahead of AI newsletter (that has the biggest audience overlap with Interconnects, at 26%, so a lot of you know him) and multiple educational books, all on top of being a full time machine learning engineer at Lightning.ai, where he maintains LitGPT, which he described as being like Karpathy’s NanoGPT, with slightly more abstractions.This conversation mostly surrounds keeping up with AI research, the state of the open LLM ecosystem post Llama 3.1, and many narrow topics in between. I learned that Sebastian used to be an Arxiv moderator, which gives some simple color on how Arxiv and sifting through thousands of papers works. We cover a lot of ground here, so I hope you enjoy it.00:00:00 Introduction & Sebastian's background00:04:28 The state of deep learning and language models in 201800:08:02 Sebastian's work at Lightning AI and LitGPT00:12:23 Distillation and its potential in language model training00:14:14 Implementing language models and common pitfalls00:18:45 Modern architectures: Mixture of experts models, early v. late fusion multimodal00:24:23 Sebastian's book on building language models from scratch00:27:13 Comparing ChatGPT, Claude, and Google's Gemini for various tasks00:38:21 Vibing and checking new language models during implementation00:40:42 Selecting papers to read and moderating Arxiv00:45:36 Motivation for working on AI education00:52:46 Llama 3 fine-tuning00:57:26 The potential impact of AI on jobs in writing and education01:00:57 The future directions of AIMore details: https://www.interconnects.ai/interviewing-sebastian-raschka

And how to understand Llama three point one's results.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/gpt-4o-mini-changed-chatbotarena0:00 GPT-4o-mini changed ChatBotArena3:23 Llama 3 in the arena5:13 Partial solutions and next stepsFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_013.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_015.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_019.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_021.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_025.pngFig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_039.pngFig 7: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/new-chatbotarena/img_043.png

Defining the future of the AI economy and regulation. Is Meta's AI play equivalent to the Unix stack for open-source software?This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/llama-405b-open-frontier-model00:00 Llama 3.1 405b, Meta's AI strategy, and the new open frontier model ecosystem01:37 Meta's open frontier model03:51 Zuckerberg's vision for open-source AI (vs. reality)08:35 Does the Llama 3.1 license support open-source AI?12:55 Different futures for regulating frontier modelsFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_008.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_010.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_015.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_018.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/llama-405/img_050.png

SB 1047, AI regulation, and unlikely allies for open modelsThe rallying of the open-source community against CA SB 1047 can represent a turning point for AI regulation.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/sb-1047-and-open-weights00:00 Introduction01:53 SB 1047 and targeting regulation07:57 Unlikely allies of "open"12:05 What would I regulate today?

I Switched to Claude 3.5Speculations on the role of RLHF and why I love the model for people who pay attention.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/switched-to-claude-from-chatgpt00:00 I Switched to Claude 3.503:57 Product priorities05:15 RLHF's peak?Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_016.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_018.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_020.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/claude/img_022.png

I’m really excited to resume the Interconnects Interviews with Dean W. Ball from the Hyperdimensional Substack. We cover the whole stack of recent happenings in AI policy, focusing of course on California’s bill SB 1047. We cover many, many more great topics here including:What will happen in the case of a minor AI disaster,If Meta will release the 405B model, and why,The status of Chinese open-source AI,Training on model outputs,Anthropic’s recent strategy,What scaling laws actually mean,Creating content and shifting the needle of the AI discourse.View online: https://www.interconnects.ai/p/interviewing-dean-ball-on-ai-policyChapters00:00 Intro and Welcome Dean Ball 02:44 The Origins of California Bill SB1047 08:56 The Evolution of Bill SB1047 13:00 How SB1047 Affects Fine-Tuning 20:00 The Future of Bill SB1047 21:58 The Impact of AI Disasters 29:02 Meta and its 400 billion Parameter Model 32:25 Open Source AI and the Chinese Market 37:37 The Future of Open Source AI 43:35 Synthetic Data, Licenses, and Future AI Development 45:18 Anthropic's Approach to AI Safety 50:46 Scaling Laws 53:01 The Role of Audience in Influencing AI Policy

Things to be aware of if you work on language model fine-tuning.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/rlhf-roundup-202400:00 RLHF Roundup: Trying to get good at PPO, charting RLHF's impact, RewardBench retrospective, and a reward model competition04:32 How big is the impact of RLHF relative to pretraining?05:54 RewardBench retrospective after 100 models and 90% peak accuracy09:19 LMSYS's reward modeling competitionFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf-roundup/img_009.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf-roundup/img_012.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf-roundup/img_017.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf-roundup/img_026.png

Synthetic data is known to be a super powerful tool for every level of the language modeling stack. It's documented as being used for expanding vanilla pretraining data and creating large swaths of fine-tuning data. Many, many more rumors surround its use, Anthropic's pretraining-scale constitutional AI, Mistral AI's first models being pretrained on OpenAI outputs, Q-star's hopes as OpenAI's remaining moat, and much more. The diversity of use cases for synthetic data makes planning around the role of synthetic data in solving specific goals.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/frontiers-in-synthetic-data00:00 Frontiers in synthetic data01:14 1. Direct distillation is still king02:54 2. Are Gemini Flash and Claude Haiku distilled?04:03 3. Filtering prevents collapse06:30 4. Synthetic data strategy taxes07:32 5. Pros and cons of training on multi-output-source synthetic datasets08:54 6. Structured synthetic data09:42 7. Weak-to-strong generalization is maybe real10:27 8. Creating synthetic prompts is overlooked again

Signs point to a general-use Sora-like model coming very soon, maybe even with open-weights.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/text-to-video-ai-is-already-abundant0:00 Text-to-video AI is already abundant5:08 What's next for the text-to-video market?6:49 Are text-to-video models good for the world?Fig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_005.mp4Fig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_009.mp4Fig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_011.mp4Fig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_013.mp4Fig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_015.mp4Fig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/text-to-video/img_017.mp4

Apple Intelligence makes a lot of sense when you get out of the AI bubble.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/apple-intelligence00:00 AI for the rest of us02:46 Apple's technical approach03:32 Core models: What did Apple build?05:35 Alignment strategies: Some new things!10:00 Orchestrating adapters and on-device magic11:58 Light for other narratives around AIFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/apple-intelligence/img_005.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/apple-intelligence/img_015.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/apple-intelligence/img_039.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/apple-intelligence/img_041.png

A realistic path to robotic foundation modelsNot "agents" and not "AGI." Some thoughts and excitement after revisiting the industry thanks to Physical Intelligence founders Sergey Levine and Chelsea Finn.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/robotic-foundation-models0:00 A realistic path to robotic foundation models2:51 Key factors for the future of robotics6:19 Everything is a token: The transformerification of robotics

Data licensing deals, scaling, human inputs, and repeating trends in open vs. closed.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/the-data-wall0:00 We aren't running out of training data, we are running out of open training data2:51 Synthetic data: 1 trillion new tokens per day4:18 Data licensing deals: High costs per token6:33 Better tokens: Search and new frontiers

Celebrity's power will only grow in the era of infinite content.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/name-image-and-ai-likeness0:00 Name, image, and AI's likeness1:11 OpenAI's second terrible, horrible, no good, very bad week4:36 The expansion of name and likeness7:46 Culture and AI development

ChatGPT leaves the textbox, and Google is building the same, and more, as practical tools.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/openai-and-her00:00 OpenAI chases Her02:10 Talking to ChatGPT03:53 GPT-4o: Toward omnimodal models08:25 Google's mirror with Gemini10:11 OpenAI's AI Safety: Have your cake and eat it tooFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/her/img_018.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/her/img_023.jpg

Now we will have some grounding for when weird ChatGPT behaviors are intended or side-effects -- shrinking the Overton window of RLHF bugs.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/openai-rlhf-model-spec00:00 OpenAI's Model (behavior) Spec, RLHF transparency, and personalization questions02:56 Reviewing the Model Spec08:26 Where RLHF can fail OpenAI12:23 From Model Spec's to personalizationFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_027.pngFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_029.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_033.pngFig 4: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_034.pngFig 5: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_041.webpFig 6: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/model-spec/img_046.webp

Many, many signs of life for preference fine-tuning beyond spoofing chat evaluation tools.This is AI generated audio with Python and 11Labs.Source code: https://github.com/natolambert/interconnects-toolsOriginal post: https://www.interconnects.ai/p/how-rlhf-works-200:00 How RLHF works, part 2: A thin line between useful and lobotomized04:27 The chattiness paradox08:09 The mechanism for making models chattier10:42 Next steps for RLHF researchFig 1: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf/img_012.webpFig 2: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf/img_018.pngFig 3: https://huggingface.co/datasets/natolambert/interconnects-figures/resolve/main/rlhf/img_025.png