Sonya Huang 🐥
@sonyatweetybird • 21,119 subscribers
funding big computer @sequoia
Videos
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 AI). 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 🐥76,501 просмотров • 19 дней назад
"Member of the technical staff" is the hottest job title in SF right now. What's behind the name? OpenAI chose this title deliberately to blow up the previous industry dichotomy between researchers and engineers. The best researchers in AI right now aren't academics in a pure lab environment; they get their hands dirty technically writing code and digging into the data and implementations e.g. Alec Radford Bob McGrew, former Chief Research Officer at OpenAI and one of U.S. Army's newest recruits to Detachment 201, joined us on Training Data to share more about the secret sauce behind leading OpenAI's research org, the three legs of the stool to AGI, and why he thinks we're already there.
Sonya Huang 🐥358,508 просмотров • 1 год назад
Some of the most iconic consumer products -- Tide Pods, M&Ms, etc -- were born out of user research studies. LLMs democratize deep user research to every decision. Synthetic audiences will go even further. Alfred Wahlforss of Listen Labs shares more on Training Data cc Konstantine Buhler
Sonya Huang 🐥16,173 просмотров • 12 дней назад
Claude Code and Suno have more in common than you might think: "It's fun to build things, and it's fun to use what you build." AI lets people be creative in almost any domain, from coding to making music. Today on Training Data, Mikey shares his thesis for why generative AI is the newest form of active entertainment (the next 'gaming'), music as a cultural phenomenon vs creative expression platform, and more. My favorite part was Mikey's explanation of why Suno learns music theory implicitly vs explicitly: "In Western music, there are 12 tones. If you tell the model there are 12 tones, it will only ever produce those 12 tones. You will be forever limited. And if you tell the model there's 200 instruments, those are the only sounds that you'll ever be able to make." The more you constrain a model with what humans already know, the less capable it becomes. By treating everything as pure sound, Suno built what Mikey calls a totally generalized "music-making machine." Such is the power of neural nets.
Sonya Huang 🐥20,605 просмотров • 1 месяц назад
Don't sleep on Google DeepMind in AI... This week on Training Data, Google Labs VP Josh Woodward gave us the BTS on Google's imagination playground for AI, from Notebook to Mariner (computer use agent) to Veo (video models). Thanks Josh for the spicy convo and hot takes :)
Sonya Huang 🐥183,326 просмотров • 1 год назад
Baumol cost disease is real, AI is the solution Take law -- imagine a world where consumers have plentiful access to high-quality legal services Crosby is building an AI-native law firm towards that vision. @Ryanjdaniels John Sarihan share more on Training Data!
Sonya Huang 🐥69,577 просмотров • 9 месяцев назад
Best technology M&A of all time has to be NVIDIA's $6.9 Billion acquisition of Mellanox in 2020. It was a special treat to interview Michael Kagan, CTO of NVIDIA and co-founder/CTO of Mellanox, for today's Training Data episode. Michael has been driving forward the Compute Frontier for more than 40 years now, first as Chief Architect at Intel in the 90s, then Co-Founder and CTO at Mellanox, and for the last 5 years CTO at NVIDIA. There's nobody better positioned than Michael to share the complete history of the compute frontier and what's ahead, from decades pushing forward Moore's law (squeezing more transistors on a chip) to the last decade of work scaling beyond single chip physics limitations (scaling out to 100K+ GPU clusters). Interconnect is the secret sauce enabling compute to scale beyond chip-level Moore's law. Connecting a fabric of 100K+ GPUs to function as a single unit of compute is the enabling technology for today's intelligence explosion. But other things break at the 100K+ GPU cluster scale: individual chips inevitably fail, power and networking become more complex, etc etc. Net effect of scale out: we've inflected the silicon frontier from Moore's Law (2x every 2 years) to Huang's law (~10x a year). Very excited about today's episode! Learned so much from Michael with Pat Grady
Sonya Huang 🐥56,175 просмотров • 7 месяцев назад
Today on Training Data: Sanjit Biswas, founder & CEO of Samsara (NYSE:IOT) and former Sequoia Capital backed founder of Cisco Meraki Sanjit shares the ups & downs of running neural nets on constrained compute and power footprints in the real world, ~2-10 watts Physical AI is hard 🫡
Sonya Huang 🐥35,714 просмотров • 6 месяцев назад
Today on Training Data, the OpenAI team behind ChatGPT agent explain how Agent Mode works, combining: 1) Deep Research (text based research agent) 2) Operator (GUI/action based computer agent) 3) Other new tools (terminal, computer apps) 4) Tied together with shared state to create an agent that's highly capable at most tasks that humans do on a computer: data science analysis, analyzing spreadsheets, making slides, etc. Thanks for joining us Isa Fulford Casey Chu Zhiqing Sun Lauren Reeder!
Sonya Huang 🐥44,149 просмотров • 10 месяцев назад
Our most exciting episode of Training Data yet 🍓🍰 OpenAI’s o1 represents a major leap forward by giving models time to "think." Inference-time compute is the next big research frontier. Thrilled to have Noam Brown, ilge, and hunter on the show Pat Grady Sequoia Capital
Sonya Huang 🐥51,628 просмотров • 1 год назад
Today's Training Data episode features our newest investment 🤗 the fal guys Gorkem Yurtseven, Burkay Gur, and batuhan the fal guy on where generative media is headed~ When computer animation first arrived, the film industry was skeptical at best and downright hostile at worst. But technology doesn’t stop, and now some of the highest grossing films and most celebrated artistic achievements in film are CGI. In this episode we discuss how Hollywood’s stance toward AI has shifted fast, the rise of AI-native studios, why IP holders aren’t sitting still, and so much more. Excited to dig into the ai video compute supercycle with the fal guys.
Sonya Huang 🐥17,181 просмотров • 6 месяцев назад
This week on Training Data: robots! I really admire @Physical_int for their open publishing spirit. Karol Hausman and Tobias Springenberg joined me and Alfred Lin to chat about pi*0.6, learning from experience, long-horizon robot performance, and more.
Sonya Huang 🐥12,093 просмотров • 5 месяцев назад
Is English the hottest new programming language or assembly language? Filip Kozera shares how Wordware is building tools for "word artisans" - people who can communicate their creative vision to AI systems. What Excel did for numbers, Wordware aims to do for AI.
Sonya Huang 🐥23,738 просмотров • 1 год назад
Can we map the mind of an LLM? Our first mechanistic interpretability episode on Training Data featuring Goodfire founder Eric Ho (and our first cameo from Roelof Botha!) Goodfire is building an independent mech interp lab, led by some heavyweight researchers from the field (e.g. Lee Sharkey who has led a lot of important work in sparse autoencoders to "unscramble" LLMs and resolve superposition, Nick who has been a key pioneer behind auto interpretability) On this episode, Eric gives us a flyover of the technical results so far from this nascent field (universality, superposition), what's ahead in the research (going from circuits to weights, going from understanding to increasingly surgical editing), a preview of the real-world work they're doing already with Arc Institute, and the impact he expects Goodfire and the broader field to have on steering, safety, editing and more.
Sonya Huang 🐥19,371 просмотров • 11 месяцев назад
What if 8 hours of research could be done in 5 minutes? OpenAI's Deep Research is an agent trained end-to-end w/ RL fine-tuning on my favorite task: internet sleuthing👩💻 Isa Fulford & Josh Tobin joined us to share what's under the hood + the future of OpenAI's agent roadmap.
Sonya Huang 🐥20,210 просмотров • 1 год назад
Hearing @raiza_abubakar describe NotebookLM's 💩💨💩💨💩💨 viral moment was hilarious. Raiza is whip smart and refreshingly real.... it's no wonder she's behind Google's biggest Gen AI success so far. Listen to Raiza & Jason Spielman on this week's episode of Training Data 🎥
Sonya Huang 🐥20,305 просмотров • 1 год назад