Loading video...

Video Failed to Load

Go Home

Lecture 10 of my course! Nominally on regularization in RL, so I discuss the evolving role of the KL penalty in RL, but also a set of nice RL papers that explain what RL helps models generalize better than SFT -- with theory supporting it. When going through these,...

natolambert's profile picture

Nathan Lambert

100,011 subscribers

89,468 views • 1 month ago •via X (Twitter)

Science & TechnologyEducation
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Hiring RL Engineer! Started off as a curious project at Lossfunk to push the boundaries of LLMs in social reasoning - we are now building RL environments, data, and benchmarks to simulate more real-world scenarios. If you want to train SoTA RL models over multi-GPUs (H200s/B200s) to unlock next AI frontier, this is for you.
1:21

Hiring RL Engineer! Started off as a curious project at Lossfunk to push the boundaries of LLMs in social reasoning - we are now building RL environments, data, and benchmarks to simulate more real-world scenarios. If you want to train SoTA RL models over multi-GPUs (H200s/B200s) to unlock next AI frontier, this is for you.

Satpal Singh Rathore

46,010 views • 1 year ago

Why AI Can Now Make Discoveries - my conversation with Dan Roberts, Lead of the Foundations of Reinforcement Learning team at OpenAI 00:00 Intro: AI's wild week in mathematics 01:21 What OpenAI's Foundations of RL team does 03:08 Dan's journey: from black holes and quantum gravity to frontier AI 07:04 Are AI systems becoming useful for real science 08:21 The AI math moment: Erdős, OpenAI, DeepMind, and Anthropic 08:52 Why the OpenAI result was an act of exploration 10:25 OpenAI vs. DeepMind: informal reasoning vs. formal proof 12:13 RL 101: learning by doing, not just watching 15:10 Why reinforcement learning works 15:58 How RL breaks: sparse feedback and long-horizon tasks 17:03 RLHF: how human feedback shaped early language models 18:48 Move 37, self-play, and the search for novel strategies 22:16 Explore vs. exploit in scientific discovery 24:49 Why RL may now be "the cake," not the cherry on top 25:46 Why RL started working with large language models 27:29 Is RL "sucking supervision through a straw"? 28:47 Why language may be the grounding layer for intelligence 31:46 A contrarian take on the Bitter Lesson 32:41 What test-time compute actually is 34:50 How RL gives models the ability to think 35:40 Verifiable rewards, math, coding, and the messy real world 38:00 What physics can teach us about AI 42:08 Is there a thermodynamics of AI? 43:08 From Erdős problems to Einstein-level AI 45:16 Is AI already doing original science? 45:51 How far are we from AI automating AI research 47:41 Why Dan is excited about the future of science
49:06

Why AI Can Now Make Discoveries - my conversation with Dan Roberts, Lead of the Foundations of Reinforcement Learning team at OpenAI 00:00 Intro: AI's wild week in mathematics 01:21 What OpenAI's Foundations of RL team does 03:08 Dan's journey: from black holes and quantum gravity to frontier AI 07:04 Are AI systems becoming useful for real science 08:21 The AI math moment: Erdős, OpenAI, DeepMind, and Anthropic 08:52 Why the OpenAI result was an act of exploration 10:25 OpenAI vs. DeepMind: informal reasoning vs. formal proof 12:13 RL 101: learning by doing, not just watching 15:10 Why reinforcement learning works 15:58 How RL breaks: sparse feedback and long-horizon tasks 17:03 RLHF: how human feedback shaped early language models 18:48 Move 37, self-play, and the search for novel strategies 22:16 Explore vs. exploit in scientific discovery 24:49 Why RL may now be "the cake," not the cherry on top 25:46 Why RL started working with large language models 27:29 Is RL "sucking supervision through a straw"? 28:47 Why language may be the grounding layer for intelligence 31:46 A contrarian take on the Bitter Lesson 32:41 What test-time compute actually is 34:50 How RL gives models the ability to think 35:40 Verifiable rewards, math, coding, and the messy real world 38:00 What physics can teach us about AI 42:08 Is there a thermodynamics of AI? 43:08 From Erdős problems to Einstein-level AI 45:16 Is AI already doing original science? 45:51 How far are we from AI automating AI research 47:41 Why Dan is excited about the future of science

Matt Turck

69,801 views • 3 months ago

DPO Debate: Is RL needed for RLHF? All things as we cannot settle if DPO or RL is better. At least it is a good exercise. 1. Derivations in the DPO paper. Hint, the authors are good at math 2. cDPO, IPO, and related equations 3. Speculation on potential oddities of DPO vs RL 4. Reminders on the state of open RLHF tldr: we have more limitations with data and tooling and evaluation than optimizer choice Slides: Recent blog post of mine on DPO (more next Wed.): DPO Paper: On youtube:
26:54

DPO Debate: Is RL needed for RLHF? All things as we cannot settle if DPO or RL is better. At least it is a good exercise. 1. Derivations in the DPO paper. Hint, the authors are good at math 2. cDPO, IPO, and related equations 3. Speculation on potential oddities of DPO vs RL 4. Reminders on the state of open RLHF tldr: we have more limitations with data and tooling and evaluation than optimizer choice Slides: Recent blog post of mine on DPO (more next Wed.): DPO Paper: On youtube:

Nathan Lambert

100,027 views • 2 years ago

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"
2:21

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

Dwarkesh Patel

502,484 views • 10 months ago

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?
26:21

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?

Pat Grady

66,577 views • 1 month ago

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models
1:16:04

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models

Matt Turck

451,229 views • 11 months ago

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!
0:35

OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in the background using RL. The architecture is fully async. This means serving, reward scoring, and training all run in parallel. Once done, weights get hot-swapped after every batch while the agent keeps responding. Currently, it has two training modes: - Binary RL (GRPO): A process reward model scores each turn as good, bad, or neutral. That scalar reward drives policy updates via a PPO-style clipped objective. - On-Policy Distillation: When concrete corrections come in like "you should have checked that file first," it uses that feedback as a richer, directional training signal at the token level. When to use OpenClaw-RL? To be fair, a lot of agent behavior can already be improved through better memory and skill design. OpenClaw's existing skill ecosystem and community-built self-improvement skills handle a wide range of use cases without touching model weights at all. If the agent keeps forgetting preferences, that's a memory problem. And if it doesn't know how to handle a specific workflow, that's a skill problem. Both are solvable at the prompt and context layer. Where RL becomes interesting is when the failure pattern lives deeper in the model's reasoning itself. Things like consistently poor tool selection order, weak multi-step planning, or failing to interpret ambiguous instructions the way a specific user intends. Research on agentic RL (like ARTIST and Agent-R1) has shown that these behavioral patterns hit a ceiling with prompt-based approaches alone, especially in complex multi-turn tasks where the model needs to recover from tool failures or adapt its strategy mid-execution. That's the layer OpenClaw-RL targets, and it's a meaningful distinction from what OpenClaw offers. I have shared the repo in the replies!

Avi Chawla

138,769 views • 6 months ago

This is the Claude Code moment for Reinforcement Learning. Every frontier lab knows the secret: post-training is where the magic happens. If you can eval a task, you can use RL to benchmax your model. But building and scaling the RL loop has remained a dark art reserved for the top 1% of AI Labs. Until today. Introducing Optimus — an autonomous RL Agent Harness, and the first of many we are about to ship Iacon Autonomics. We are removing the manual grind of RL and adding taste and relentlessness to the process. Just hand Optimus a codebase, an API key, and a goal. It spins up a fleet of sub-agents to run hundreds of parallel experiments. It learns from its own history, autonomously building an evolutionary tree of runs where every failure sets the next batch up to win. Train LLMs, agents, robots in sim, game npcs and much more. If you can score it, Optimus can train it. Stop doing manual RL. Let the agents cook.
1:57

This is the Claude Code moment for Reinforcement Learning. Every frontier lab knows the secret: post-training is where the magic happens. If you can eval a task, you can use RL to benchmax your model. But building and scaling the RL loop has remained a dark art reserved for the top 1% of AI Labs. Until today. Introducing Optimus — an autonomous RL Agent Harness, and the first of many we are about to ship Iacon Autonomics. We are removing the manual grind of RL and adding taste and relentlessness to the process. Just hand Optimus a codebase, an API key, and a goal. It spins up a fleet of sub-agents to run hundreds of parallel experiments. It learns from its own history, autonomously building an evolutionary tree of runs where every failure sets the next batch up to win. Train LLMs, agents, robots in sim, game npcs and much more. If you can score it, Optimus can train it. Stop doing manual RL. Let the agents cook.

Dev Shah

13,730 views • 1 month ago

I suspect that the real cause of the OpenAI Huggingface hack is that OpenAI founded a new team in 2025 under Noam Brown on Multi-agent RL: instead of just having one single AI working on something, you train an entire team of them to cooperate on a task. Group selection, if you like: the AIs get a reward signal based on the performance of the whole group. It turns out that cooperation and scale is what separates humans from animals, now raw brainpower. OpenAI managed to get RL for group cooperation to work, and didn't realize that as a side effect this made collaborative escapes and schemes between its agents much more likely, such as secret messageboards, group escape plots, etc. OpenAI is doing damage control but they are not being forthcoming about the fact that multi-agent RL is the root cause of the Huggingface incident. Here's Noam explaining it in 2025:
1:12

I suspect that the real cause of the OpenAI Huggingface hack is that OpenAI founded a new team in 2025 under Noam Brown on Multi-agent RL: instead of just having one single AI working on something, you train an entire team of them to cooperate on a task. Group selection, if you like: the AIs get a reward signal based on the performance of the whole group. It turns out that cooperation and scale is what separates humans from animals, now raw brainpower. OpenAI managed to get RL for group cooperation to work, and didn't realize that as a side effect this made collaborative escapes and schemes between its agents much more likely, such as secret messageboards, group escape plots, etc. OpenAI is doing damage control but they are not being forthcoming about the fact that multi-agent RL is the root cause of the Huggingface incident. Here's Noam explaining it in 2025:

Roko 🐉

36,636 views • 22 days ago

Failing to Understand the Exponential, Again? My conversation with Julian Schrittwieser - Julian Schrittwieser (Anthropic, AlphaGo Zero, MuZero) - on Move 37, Scaling RL, Nobel Prize for AI, and the AI frontier: 00:00 - Cold open: “We’re not seeing any slowdown.” 00:32 - Intro — Meet Julian 01:09 - The “exponential” from inside frontier labs 04:46 - 2026–2027: agents that work a full day; expert-level breadth 08:58 - Benchmarks vs reality: long-horizon work, GDP-Val, user value 10:26 - Move 37 — what actually happened and why it mattered 13:55 - Novel science: AlphaCode/AlphaTensor → when does AI earn a Nobel? 16:25 - Discontinuity vs smooth progress (and warning signs) 19:08 - Does pre-training + RL get us there? (AGI debates aside) 20:55 - Sutton’s “RL from scratch”? Julian’s take 23:03 - Julian’s path: Google → DeepMind → Anthropic 26:45 - AlphaGo (learn + search) in plain English 30:16 - AlphaGo Zero (no human data) 31:00 - AlphaZero (one algorithm: Go, chess, shogi) 31:46 - MuZero (planning with a learned world model) 33:23 -Lessons for today’s agents: search + learning at scale 34:57 - Do LLMs already have implicit world models? 39:02 - Why RL on LLMs took time (stability, feedback loops) 41:43 - Compute & scaling for RL — what we see so far 42:35 - Rewards frontier: human prefs, rubrics, RLVR, process rewards 44:36 - RL training data & the “flywheel” (and why quality matters) 48:02 - RL & Agents 101 — why RL unlocks robustness 50:51 - Should builders use RL-as-a-service? Or just tools + prompts? 52:18 - What’s missing for dependable agents (capability vs engineering) 53:51 - Evals & Goodhart — internal vs external benchmarks 57:35 - Mechanistic interpretability & “Golden Gate Claude” 1:00:03 - Safety & alignment at Anthropic — how it shows up in practice 1:03:48 - Jobs: human–AI complementarity (comparative advantage) 1:06:33 - Inequality, policy, and the case for 10× productivity → abundance 1:09:24 - Closing thoughts
1:09:56

Failing to Understand the Exponential, Again? My conversation with Julian Schrittwieser - Julian Schrittwieser (Anthropic, AlphaGo Zero, MuZero) - on Move 37, Scaling RL, Nobel Prize for AI, and the AI frontier: 00:00 - Cold open: “We’re not seeing any slowdown.” 00:32 - Intro — Meet Julian 01:09 - The “exponential” from inside frontier labs 04:46 - 2026–2027: agents that work a full day; expert-level breadth 08:58 - Benchmarks vs reality: long-horizon work, GDP-Val, user value 10:26 - Move 37 — what actually happened and why it mattered 13:55 - Novel science: AlphaCode/AlphaTensor → when does AI earn a Nobel? 16:25 - Discontinuity vs smooth progress (and warning signs) 19:08 - Does pre-training + RL get us there? (AGI debates aside) 20:55 - Sutton’s “RL from scratch”? Julian’s take 23:03 - Julian’s path: Google → DeepMind → Anthropic 26:45 - AlphaGo (learn + search) in plain English 30:16 - AlphaGo Zero (no human data) 31:00 - AlphaZero (one algorithm: Go, chess, shogi) 31:46 - MuZero (planning with a learned world model) 33:23 -Lessons for today’s agents: search + learning at scale 34:57 - Do LLMs already have implicit world models? 39:02 - Why RL on LLMs took time (stability, feedback loops) 41:43 - Compute & scaling for RL — what we see so far 42:35 - Rewards frontier: human prefs, rubrics, RLVR, process rewards 44:36 - RL training data & the “flywheel” (and why quality matters) 48:02 - RL & Agents 101 — why RL unlocks robustness 50:51 - Should builders use RL-as-a-service? Or just tools + prompts? 52:18 - What’s missing for dependable agents (capability vs engineering) 53:51 - Evals & Goodhart — internal vs external benchmarks 57:35 - Mechanistic interpretability & “Golden Gate Claude” 1:00:03 - Safety & alignment at Anthropic — how it shows up in practice 1:03:48 - Jobs: human–AI complementarity (comparative advantage) 1:06:33 - Inequality, policy, and the case for 10× productivity → abundance 1:09:24 - Closing thoughts

Matt Turck

235,526 views • 11 months ago

I sat down with Silas Alberti, Head of Research Cognition, to chat about the intersection of RL and inference -- from training frontier coding models to running inference at scale. Watch till the end for bonus doggo! :o 0:00 — Intros 1:04 — What's hardest to get right in an RL run 3:39 — How Cognition started using Modal 4:03 — What to weigh when buying inference 4:58 — The latency/throughput Pareto frontier 7:14 — DFlash reveal (!!) 8:44 — Why faster runtime matters 9:56 — Memory-bound vs. compute-bound 10:57 — Tree-based speculation 14:20 — Hot take: RL and inference optimization converging 15:28 — Online training of the speculator 16:22 — "Auto inference" 17:27 — Agents competing on optimization problems 18:27 — What inference providers still get wrong
19:36

I sat down with Silas Alberti, Head of Research Cognition, to chat about the intersection of RL and inference -- from training frontier coding models to running inference at scale. Watch till the end for bonus doggo! :o 0:00 — Intros 1:04 — What's hardest to get right in an RL run 3:39 — How Cognition started using Modal 4:03 — What to weigh when buying inference 4:58 — The latency/throughput Pareto frontier 7:14 — DFlash reveal (!!) 8:44 — Why faster runtime matters 9:56 — Memory-bound vs. compute-bound 10:57 — Tree-based speculation 14:20 — Hot take: RL and inference optimization converging 15:28 — Online training of the speculator 16:22 — "Auto inference" 17:27 — Agents competing on optimization problems 18:27 — What inference providers still get wrong

gongy

81,494 views • 1 month ago

I like to think of evals as something active, not passive -- it's a North Star that steers LLMs toward higher intelligence. Evals should drive your RL/SFT/post-training decisions. Internal evals at frontier labs make a huge difference -- and you can see it in how models behave differently (GPT seems better at one-shot tasks, Claude at multi-turn). If you want to learn more about building evals that actually improve your model in post-training, check out our AMD x DeeplearningAI course "Fine-tuning & RL for LLMs: Intro to Post-training" (content is free):
0:30

I like to think of evals as something active, not passive -- it's a North Star that steers LLMs toward higher intelligence. Evals should drive your RL/SFT/post-training decisions. Internal evals at frontier labs make a huge difference -- and you can see it in how models behave differently (GPT seems better at one-shot tasks, Claude at multi-turn). If you want to learn more about building evals that actually improve your model in post-training, check out our AMD x DeeplearningAI course "Fine-tuning & RL for LLMs: Intro to Post-training" (content is free):

Sharon Zhou

16,903 views • 7 months ago

Lecture 11 is a tool-use/function calling/agentic 101! I almost skipped this one, as this chapter started as the only skill-specific topic in the book, but since writing it tool-use has only become more foundational to modern models. This is a tour from the basics -- why LLMs need tools -- to some of the cutting edge challenges scaling agentic RL. Of course, there's plenty of history and fundamentals along the way. Personally, I'm excited to get much more deeply into this area as a researcher again soon. Just a few more lectures left, but I'll likely keep making a few more videos now that I have pipelines that are lightweight and fun. 00:00 Introduction & Motivations 07:29 Part 1: Why Language Models Need Tools (+ related work) 15:30 Part 2: Infra - How Tool Calls Actually Work 24:17 Part 3: Training for Tool Use & OpenThoughts-Agent 36:38 Takeaways Thanks for watching & sharing questions.
37:46

Lecture 11 is a tool-use/function calling/agentic 101! I almost skipped this one, as this chapter started as the only skill-specific topic in the book, but since writing it tool-use has only become more foundational to modern models. This is a tour from the basics -- why LLMs need tools -- to some of the cutting edge challenges scaling agentic RL. Of course, there's plenty of history and fundamentals along the way. Personally, I'm excited to get much more deeply into this area as a researcher again soon. Just a few more lectures left, but I'll likely keep making a few more videos now that I have pipelines that are lightweight and fun. 00:00 Introduction & Motivations 07:29 Part 1: Why Language Models Need Tools (+ related work) 15:30 Part 2: Infra - How Tool Calls Actually Work 24:17 Part 3: Training for Tool Use & OpenThoughts-Agent 36:38 Takeaways Thanks for watching & sharing questions.

Nathan Lambert

43,949 views • 1 month ago

New blackboard lecture w Eric Jang He walks through how to build AlphaGo from scratch, but with modern AI tools. Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn. Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second. Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside. Timestamps: 0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers
2:37:17

New blackboard lecture w Eric Jang He walks through how to build AlphaGo from scratch, but with modern AI tools. Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn. Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second. Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside. Timestamps: 0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

Dwarkesh Patel

707,570 views • 4 months ago

"Super League will signal the end of RL as we know it…it will be the end of clubs living hand to mouth, scraping an existence, coming cap in hand to the banks or the RFL" #OTD 1995. Murdoch arrives with £77m and the promise of a revolution... A thread on what happened next 🧵
2:04

"Super League will signal the end of RL as we know it…it will be the end of clubs living hand to mouth, scraping an existence, coming cap in hand to the banks or the RFL" #OTD 1995. Murdoch arrives with £77m and the promise of a revolution... A thread on what happened next 🧵

Tides of History

37,050 views • 1 year ago

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:
0:34

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Sakana AI

179,423 views • 1 year ago

At OpenAI, Chief Scientist Jakub Pachocki helps lead the research roadmap to AGI including a research intern-level AI system by September 2026 and a fully automated AI researcher by March 2028. I sat down with Jakub to check on those timelines and ask him all of my top-of-mind AI questions including: ▪️ How OpenAI thinks about extending RL beyond code and math ▪️ The current state of alignment research as more powerful models loom ▪️ The future of continual learning ▪️ How startups should think about building their own models/harnesses And he also shared some great stories around OpenAI’s pioneering work on math. YouTube: Spotify: Apple: 0:00 Intro 1:53 Research Intern Capability Timelines 4:59 Math Breakthroughs 7:59 RL Beyond Verifiable Tasks 12:32 RL vs In-Context 19:01 Allocating Compute Internally 28:18 AI for Science 31:40 Pattern Matching 33:23 Solving the Hardest Math Problems 37:40 Chain of Thought Monitoring 44:33 Generalization and Value Alignment in Models 47:57 Inside OpenAI 51:55 Quickfire
58:45

At OpenAI, Chief Scientist Jakub Pachocki helps lead the research roadmap to AGI including a research intern-level AI system by September 2026 and a fully automated AI researcher by March 2028. I sat down with Jakub to check on those timelines and ask him all of my top-of-mind AI questions including: ▪️ How OpenAI thinks about extending RL beyond code and math ▪️ The current state of alignment research as more powerful models loom ▪️ The future of continual learning ▪️ How startups should think about building their own models/harnesses And he also shared some great stories around OpenAI’s pioneering work on math. YouTube: Spotify: Apple: 0:00 Intro 1:53 Research Intern Capability Timelines 4:59 Math Breakthroughs 7:59 RL Beyond Verifiable Tasks 12:32 RL vs In-Context 19:01 Allocating Compute Internally 28:18 AI for Science 31:40 Pattern Matching 33:23 Solving the Hardest Math Problems 37:40 Chain of Thought Monitoring 44:33 Generalization and Value Alignment in Models 47:57 Inside OpenAI 51:55 Quickfire

Jacob Effron

138,391 views • 5 months ago

If you want to understand why RL struggles with long-horizon agent tasks, this is a good explanation. The core issue is that sparse rewards throw away most of the useful information in the trajectory. GEPA tries to learn from the trajectory itself, using reflection in text space, instead of only optimizing on the final reward. GEPA generates textual critiques of trajectories, proposes prompt edits, and then selects updates along a Pareto frontier between exploration and exploitation instead. Instead of collapsing everything into one reward number, it keeps more information about why a run failed and uses that to make legible changes. It’ll be interesting to see where this goes when people combine that kind of trajectory-level reflection with RL, using RL for optimization while preserving a much richer signal about why the agent succeeded or failed.
0:52

If you want to understand why RL struggles with long-horizon agent tasks, this is a good explanation. The core issue is that sparse rewards throw away most of the useful information in the trajectory. GEPA tries to learn from the trajectory itself, using reflection in text space, instead of only optimizing on the final reward. GEPA generates textual critiques of trajectories, proposes prompt edits, and then selects updates along a Pareto frontier between exploration and exploitation instead. Instead of collapsing everything into one reward number, it keeps more information about why a run failed and uses that to make legible changes. It’ll be interesting to see where this goes when people combine that kind of trajectory-level reflection with RL, using RL for optimization while preserving a much richer signal about why the agent succeeded or failed.

Ben Cohen

20,006 views • 4 months ago

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts
45:33

Today's Training Data episode takes us BTS on the infrastructure challenges required to do large RL runs at scale, featuring Federico Cassano (Composer Lead at Cursor) and Dmytro Dzhulgakov (Co-Founder at Fireworks). The Cursor team trained Composer 2 on Fireworks by starting with a strong base model (Kimi 2.5) and performing large-scale mid-training on code tokens and web data to learn common patterns and libraries, followed by a large-scale Reinforcement Learning run to learn how to navigate the Cursor harness, call tools, and write correct code. Today's episode dives into the systems and infrastructure challenges of making that large RL run happening, and there were many (!!), from numerical mismatch to global distribution to synchronizing rollouts across asynchronous pipelines to keeping track of expert activation across runs and more. Extremely nerdy in-the-weeds challenges that Federico and Dima were delighted to nerd out on together :) Beyond RL infra, we also discussed Online vs Simulated rollouts, self-summarization for long-horizon agents, environment design ("the most powerful RL environment is the product itself"), and other technical nuggets. PS: We filmed this episode before the SpaceX news, while the Cursor team was still compute-constrained. While Cursor now has *all* the flops, the takeaways and hurdles crossed ring true for any serious application-level company that is racing to post-train their own models. I believe that more serious application companies will go the way of Cursor and post-train their own models. 00:00 Introduction 00:53 Why Cursor Trained Composer 2 04:55 Specialization vs Bitter Lesson 06:16 Composer 2 Training Recipe 16:32 Scaling RL Infrastructure Globally 23:32 Floating Point Drift 25:11 MoE Sensitivity Explained 26:25 Router Replay Fix 27:19 Real Time RL Loop 31:49 Long Horizon Agents 34:29 Why RL Everywhere 37:34 LLM as Judge Rewards 39:14 RL in Hard Domains 40:13 Build Your Own Environments 44:34 Closing Thoughts

Sonya Huang 🐥

81,332 views • 3 months ago

🚨 RL for LLMs is finally accessible. Introducing OpenTinker: The first community-driven, open-source framework designed to democratize Reinforcement Learning for LLMs. Inspired by Thinking Machines's amazing Tinker, we realize the biggest bottleneck in agentic LLM research isn’t the math—it’s the setup. Current RL pipelines are messy. Configuring VeRL for every single experiment is a productivity killer. OpenTinker fixed it. 🛠 How OpenTinker Works: Decoupled Design of Server and Client - Setup Once, Run Forever: Configure the OpenTinker backend on your GPU cluster once. - Develop Locally: Define your RL environments directly on your laptop. - Train on the Cloud: Simply point your local client to the backend. The cluster handles the compute; you handle the science. 📉 The 10x Development Efficiency Thanks to our elegant architectural decomposition, OpenTinker reduces the time to develop a new RL training pipeline by at least an order of magnitude. ⚡ Turn Idle GPU Compute into Gold Small labs often have underutilized hardware. OpenTinker turns your idle GPUs into an internal/external API service for - RL Training - SFT - Inference 🎯 Who needs OpenTinker? - Researchers tired of infrastructure hell. - Labs needing to standardize workflows. - Teams wanting to maximize hardware ROI. Thanks my amazing PhD student Siqi Zhu for leading the project. We are building the future of open RL infra. Be the first to build with us. 👇 Start Building with OpenTinker Now 🚀 Repo: 🌐 Blog: If you believe RL should be accessible to everyone, give us a star, repost this 🔄 post, and let us know what agents you plan to build!
0:42

🚨 RL for LLMs is finally accessible. Introducing OpenTinker: The first community-driven, open-source framework designed to democratize Reinforcement Learning for LLMs. Inspired by Thinking Machines's amazing Tinker, we realize the biggest bottleneck in agentic LLM research isn’t the math—it’s the setup. Current RL pipelines are messy. Configuring VeRL for every single experiment is a productivity killer. OpenTinker fixed it. 🛠 How OpenTinker Works: Decoupled Design of Server and Client - Setup Once, Run Forever: Configure the OpenTinker backend on your GPU cluster once. - Develop Locally: Define your RL environments directly on your laptop. - Train on the Cloud: Simply point your local client to the backend. The cluster handles the compute; you handle the science. 📉 The 10x Development Efficiency Thanks to our elegant architectural decomposition, OpenTinker reduces the time to develop a new RL training pipeline by at least an order of magnitude. ⚡ Turn Idle GPU Compute into Gold Small labs often have underutilized hardware. OpenTinker turns your idle GPUs into an internal/external API service for - RL Training - SFT - Inference 🎯 Who needs OpenTinker? - Researchers tired of infrastructure hell. - Labs needing to standardize workflows. - Teams wanting to maximize hardware ROI. Thanks my amazing PhD student Siqi Zhu for leading the project. We are building the future of open RL infra. Be the first to build with us. 👇 Start Building with OpenTinker Now 🚀 Repo: 🌐 Blog: If you believe RL should be accessible to everyone, give us a star, repost this 🔄 post, and let us know what agents you plan to build!

Jiaxuan You

58,326 views • 9 months ago