[State of RL/Reasoning] IMO/IOI Gold, OpenAI o3/GPT-5, and Cursor Composer — Ashvin Nair, Cursor

Summary

Latent Space: The AI Engineer Podcast

Episode: [State of RL/Reasoning] IMO/IOI Gold, OpenAI o3/GPT-5, and Cursor Composer — Ashvin Nair, Cursor
Guest: Ashvin Nair (Cursor, formerly OpenAI, Berkeley RL PhD)
Host: Latent.Space
Recorded: At NeurIPS, December 30, 2025

---

Episode Overview

This special NeurIPS episode features Ashvin Nair, a recent addition to Cursor and a former OpenAI researcher, discussing the evolving frontier of reinforcement learning (RL), reasoning in language models, code generation, and the shift from major labs like OpenAI to high-velocity startups. The conversation traces Ashvin’s trajectory through robotics and RL research, his time on OpenAI’s code and reasoning teams, the significance (and limitations) of recent language model achievements, and finally his move to Cursor, where product/model co-design and continual learning are at the heart of rapid model improvement for coding agents.

---

Major Discussion Topics

1. From Robotics to Language Models: The Transfer of Grit and Skills

• Robotics researchers frequently move into language models, carrying valuable real-world problem-solving skills.
  • "Robotics is a pretty good fit for LLMs because the Switch ends up being— you kind of do similar things... it's kind of hard to get stuff working in robotics world. I think it builds very gritty people." (01:31)
• The contrast between robotics’ “gritty” real-world focus and the sometimes “unhinged” simulation world is noted.
  • "Lex [Fridman] said: the most rounded people are the robotics people because they don't have a choice. They work with the real world. The most unhinged... are the simulation people." (02:01)

2. The State and Hype Cycle of Robotics vs. Software AI Agents

• Robotics funding is increasing, but practical breakthroughs and market size still lag behind software agents.
  • "It feels like LLM agents are going to be like a trillion dollar market before robotics is maybe even like a $10 billion market." (03:46)
• Robotics is compared to where language models were at the GPT-1/GPT-2 stage: impressive demos, but not at a generalization or economic inflection point. (04:54)

3. OpenAI, GPT Codex, and IOI/IMO Gold Medals

• Ashvin joined OpenAI’s CodeGen team before ChatGPT became public.
  • "I was like, okay, I'm going to go to this chill research lab... and then ChatGPT happens and everything blew up." (06:00)
• The significance and “moving goalposts” of AI benchmarks:
  • "If you told me we could have gotten IOI gold, then I would have just assumed... AI is solved, no point in working anymore." (07:07)
  • Yet, even winning at these benchmarks hasn’t translated to total automation or societal upheaval, which leads to meta-reflection on progress.
  • "I think that's actually what I spend a lot of time thinking about, is why is that the case?" (07:30)

4. Shifting the Goalposts and Overfitting to Benchmarks

• AI has a history of achieving “impossible” things (chess, Go, IOI gold), yet society adapts and demands more.
  • "I think shifting the goalpost to some extent is correct. We keep Goodhart-ing whatever goalposts we have." (08:13)
• Reinforcement Learning research itself often overfit to its own benchmarks during the “RL hype” in academia.
  • "We gave ourselves a lot of new knobs to tune and then... tuned those to fit the benchmarks. Everyone knew... but it's hard to appreciate that's happening at the community level." (09:09, 10:15)
• The “RL winter,” and how academia incentivized theoretical over simple, scalable solutions.
  • "One of the pitfalls of academia is it doesn't really reward simple ideas that work... the things that actually work tend to be kind of simple." (10:38)

5. The Scaling Era: Is It Over?

• Scaling works, but not as an infinite engine for automating all jobs.
  • "I think scaling is still happening, but it's worth seriously interrogating... why we’re not just automating all jobs right now." (11:45)

6. Limitations of RL in LLMs and Pathways to Practical Use

• RL in LLMs generalizes only narrow slices of tasks—what’s missing is bringing practically valuable context into scope.
  • "It really feels like, if RL is a tool, then a big thing that needs to happen is... bring the world of economically useful tasks in distribution for RL." (12:21)
  • Example: GDP-eval is a push to benchmark LLMs on real-world economically valuable tasks. (13:43)

7. General-Purpose Models vs. Specialist Models

• OpenAI has moved away from “one model fits all.”
  • "Fiji [Simo] writing a blog post: we are no longer doing one model fits all." (17:02)
• It’s more of an organizational constraint than a technical limitation; orgs invest in the data that reflects their needs.

8. The “Blip” at OpenAI and the Unsettled State of AGI Governance

• Ashvin recalls the chaos around Sam Altman’s temporary firing and the uncertainty regarding board/governance structure.
  • "It does feel like, no matter if we hit AGI in two years or ten or whatever, it’s not clear we have a good structure for governance." (19:36)

9. Origins of the Reasoning Paradigm (“o” Series at OpenAI)

• OpenAI’s “o” line models derive from a conviction that RL is the path to AGI, but RL truly started to click after pretraining improved.
  • "People have been convinced for a long time that RL would be the way to get there— it started to work once pretraining got good enough." (23:15)
• The “big leap” moments are often perceived externally, but from inside the labs, progress feels incremental and continuous.
  • "Internally at OpenAI it feels very smooth... you keep having different runs that get a little better each time." (26:07)

10. Cursor: Why Join a Startup Now?

• At Cursor, the small, tight org allows product-model co-design—something hard at scale in OpenAI.
  • "At Cursor... the product people sit right next to the ML people. There's a lot of potential." (33:07)
  • Ultra-fast RL policy updates (every 2 hours) are achievable because of this tight loop. (34:34)
• Continual learning in production is possible, but data curation and learning from mistakes (like humans do) remains an open problem.
  • "Humans are quite good about bad data... you see someone touch a hot stove and you don't have to do it again. Models will happily keep doing it." (35:25)

11. What Makes Composer at Cursor Different?

• Composer’s key edge is combining sufficient intelligence with speed, keeping the developer in the loop—contrasted with friction caused by long LLM response times.
  • "Smart enough you actually want to use it and it's also fast, so you kind of stay in the loop... all the other smart models are slow." (37:15)

12. Frontiers in ML/RL Productization

• Product/ML co-design can finally automate much of the process of software engineering rather than just answer prompts.
  • "Really automating software engineering as a process... you write code, you go look at datadog, come back, have hypotheses... rerun stuff." (38:08)
• A smaller, focused ML group and deep internal tooling lets Cursor move faster and understand its data/user needs more directly.

Notable Quotes & Memorable Moments

• On robotics transferable skills:
  "It kind of builds very gritty people who look at data a lot, that kind of thing." — Ashvin (01:31)

• On the IOI/IMO milestone and meaning:
  "If you told me we could have gotten IOI gold, then I would've just assumed that we could all go on vacation. It's all over. AI is solved." — Ashvin (07:07)
  "Feels like nothing that much has changed, right? Life is still the same." — Host (07:25)

• On the continuous “shifting goalposts” for AGI:
  "We keep Goodharting whatever goalposts we have." — Ashvin (08:13)

• On overfitting to RL benchmarks:
  "We gave ourselves a lot of new knobs to tune and then implicitly tuned those to fit the benchmarks." — Ashvin (09:09)

• On product-model co-design at Cursor:
  "There's unique opportunities to co-design the product with the model in ways I think we couldn't do unless we actually built the model ourselves." — Ashvin (33:07)
  "We have policy updates every two hours." — Ashvin (34:34)

• On the limits and opportunities in continual learning:
  "We're a few orders of magnitude in data efficiency away from... you do something once or make a mistake, models will happily just keep doing it." — Ashvin (35:38)

• On practical edge of Composer:
  "It's smart enough you actually want to use it and it's also fast, so you kind of stay in the loop... all the other smart models are slow." — Ashvin (37:15)

Timestamps for Key Segments

• Robotics & Transfer to LLMs: 00:25–02:55
• Robotics vs. LLM Agents Market: 03:27–04:54
• Joining OpenAI, CodeGen, and ChatGPT Launch: 06:00–06:32
• IOI Gold Medals and the Real-World Impact: 07:03–08:13
• Shifting the Goalposts & Benchmarks: 08:13–09:09
• RL Academic Hype & Overfitting: 09:09–11:23
• Is Scaling Over? Moving to Economic Value: 11:41–13:34
• Generalization & Organization Limits in Foundation Models: 17:02–18:36
• The ‘Blip’ at OpenAI (Altman’s Firing): 18:46–21:03
• Origins of OpenAI’s Reasoning Paradigm: 22:11–24:48
• Continuous Progress vs. Perceived Leaps: 26:02–27:22
• Cursor’s Edge: Co-design, Fast RL Updates: 33:07–35:13
• Challenge of Continual Learning: 35:13–36:20
• What Makes Composer Stand Out: 37:15–38:08
• On RL Insights & Internal Tooling: 39:09–40:21
• Scientific Curiosity About Weight Capacity: 42:16–43:51
• RL Interview Question: 44:00–44:38

---

Final Reflections & Hiring Call

• Cursor is hiring ML engineers, especially those interested in data and reward systems for code.

  • "If you're interested in working on... data and rewards for code, that's a huge need. Please get in touch." (44:43)

• Ashvin closes with a classic RL interview question:

  • "Why is off-policy RL unstable?" (44:04)

---

Takeaways

• Progress inside labs is incremental even when outside perceptions are of sudden leaps.
• RL has reached unprecedented impact in code and reasoning LLMs, but overfitting to benchmarks and lack of continual learning remain challenges.
• Product/model co-design and tight team structures can drive much faster improvement in applied ML, as evidenced by Cursor’s rapid iteration.
• The conversation about AGI governance is far from finished—the “blip” served as a reminder that foundational choices are still unresolved.

---

Listen or Read More

• Full show notes and transcript at: https://latent.space

---

Transcript

[A]

Okay, we're here at Neurips. We're recording a special lane space coverage of the folks at Neurips. And we're here with Ashwin from Cursa.

[B]

Welcome. Hi. Yeah, thanks for having me.

[A]

So I guess astronaut from Cursa is like a new identity. I didn't even know if I should say that because you only joined Cursa for three months. Before that, you're opening AI work in 0103. Before that, Berkeley. PhD in RL, but focus on robotics.

[B]

Robotics, yeah.

[A]

Is it weird searching for robotics to language models?

[B]

Okay, this is kind of interesting because a lot of people have been kind of doing this. I mean, OpenAI. Yeah, exactly. So actually, no, I actually was at OpenAI in 2017. Also working on robotics. Yeah, I was interning, like right before my PhD where I worked on robotics there.

[A]

2017. Is that Javan? And that's Javan over there. He was famously OpenAI's first intern.

[B]

Oh, really? Okay, then he might have been before, but yeah, there was 15 interns. It was a very different company. It was just like Robotics Dota and like 15 interns that summer, all having, like, pretty exciting individual projects. Like, yeah, that set of interns, if you look at where they are now, it's kind of cool.

[A]

Yeah. But yeah, anyone from that class that.

[B]

Like, you would shout out, like, there's just like a lot of cool papers that came out. Like Lero Pinto. Now is it nyu? The person who leads Reasoning at X? I forgot his name. Well, he left. No, Eric. Yeah, I forgot his name. But he worked on like, KFAC and stuff, I think.

[A]

Yeah, Vision dude. Greg.

[B]

Not Greg, but yeah. But yeah, it was like an exciting time to be there. But yeah, I think robotics is a pretty good fit for LLMs because the Switch ends up being pretty. You kind of do similar things. Like you want to look at a lot of data. It's kind of hard to get stuff. Get hard to get stuff. Working in robotics world, I think it kind of builds very gritty people who look at data a lot, that kind of thing. So, yeah, for whatever reason, I think that transfer is happening a lot and I think it makes a lot of sense.

[A]

One of my Neuros highlights so far, I had dinner with. It's like a small group dinner with Lex Freeman yesterday. And Lex used to be in robotics. And he was like, my assessment of robotics. People are the best to talk to at in Europe because they're most rounded, he says, because they don't have a choice. They work with the real world. And then the Most unhinged, the most detached from reality are the simulation people.

[B]

I see. Yeah, yeah, I think I agree. Yeah. And I kind of. I actually did a little bit of Both during my PhD. Like, I work in kind of like prototype ideas in sim and then get them working on real world robotics. And. Yeah, I mean, probably like, robotics is where you kind of feel AGI the least, right. Because it's just so far away from working now. I think, like, I think over the last year maybe there's been demos that have been super interesting from like, physical intelligence and like Sunday and stuff that. Yeah, I'm starting to be like, okay, like this.

[A]

This kind of feels like Sunday.

[B]

Robots themselves, I haven't seen them.

[A]

Apparently they've been doing demos. I mean, I'm pretty keen to seeing.

[B]

Yeah, I've seen the physical intelligence ones live and yeah, it's pretty impressive. Like, just on, like in like someone's living room, folding laundry and stuff. Like, you know, you can just like toss it in there.

[A]

Everybody must he materialized.

[B]

Yeah, yeah, yeah, yeah.

[A]

Okay. And Nathan, last thing on robotics and we can kind of pivot to 0103. Just OpenAI has like restarting a robotics team. Is that serious? Is that.

[B]

I actually know very little about it because, yeah, I was in like, a pretty difficult part of the org. So, yeah, I mean, I think it's serious. I think there's a ton of excitement around robotics right now. I'm actually kind of curious what drives it, because I don't think I fully understand. There's been crazy raises and stuff recently for robotics companies. I guess my own view on it. When I Left Robotics in 2022, I thought I would actually come back to robotics. But I think my view on it now is that it feels like LLM agents are going to be like a trillion dollar market before robotics is maybe even like a $10 billion market. And this is just because. So, I mean, you know, LM agents already create value out in the world. Yeah, robotics, it's like, kind of hard to make the case that like, you know, kind of AI Robotics, like, does anything that useful yet, and then once it does something useful, then you have to make the unit economics work out. And I think that's also quite hard, like reliability, you know, I mean, these robots have to be, like, fixed and these kind of things. So I think it's kind of hard.

[A]

I. I would say the market is kind of efficient in that the software LLM companies are raising tens of billions, and then the robotics companies are raising hundreds of millions.

[B]

So I Think this very recently it's been like single digit billions. Oh, really? Okay. Yeah. So I think that's like the maybe surprising thing to me is that like.

[A]

It feels like it's ahead of where it's actually at.

[B]

Yeah. Like, I would say that robotics isn't kind of like the GPT1 to GPT2 area right now. Okay. And I haven't worked on robotics closely.

[A]

What task would qualify as like, oh, that's, that's the inflection.

[B]

It's a little bit like, you know, when you see it. I thought the like Sunday demos were kind of cool. Like maybe it's like starting to get there where and the details matter a lot where it. It's kind of like it can't be. It has to be in like a new scenario. Like in one that you haven't seen before and maybe on like Object General. Yeah, exactly. And I think that was kind of what GPT2.2 was too.

[A]

Right.

[B]

Is kind of like you start to see hints of cool generalization. But. And I think that's fine. It doesn't have to work out of the box. But yeah, I think at this point especially, it still feels like in robotics you're not exactly investing in a technology. Probably you're just investing in a team. Yeah, I'm not in the space whatsoever, but that's kind of my impression of it.

[A]

It's actually nice when you're not in there because, you know, as much as most basically everyone else. So we just kind of speculate.

[B]

Exactly.

[A]

In my people, there's a robotics team at OpenAI. Yeah. So coming back to language models, did you join 401 or were you like.

[B]

So I joined right before ChatGpt in like I think September of 22. Yeah. So yeah, actually, yeah, I was like pretty burnt out for my PhD and I was like, okay, I'm going to go to this like chill research lab. And then like. Yeah, yeah, like ChatGPT happens and like, you know, everything kind of blew up and like a lot of stuff got kind of like refocused.

[A]

But what. I guess what? So OpenAI, obviously chatgpt surprised. OpenAI. What did they tell you they were looking for you to do? And then obviously it changed.

[B]

But yeah, I mean, so I joined on the code gen team. So.

[A]

Codex.

[B]

Yeah. Like you Codex. Exactly. Yeah, it was like the team that shipped Codex. But by the time we were working on, like by the time I joined, we were kind of more so working on the model, doing tool use and these kinds of things. And so like very related to the ChatGPT. Like we're kind of like a sister team to the team that made ChatGPT. Yeah, exactly. So we were just kind of working on making the models smarter, like kind of programming competitions, how to do sat for that, that kind of stuff.

[A]

The word and IOI gold have felt reachable in that title.

[B]

Oh yeah. Crazy. And this is something I've repeat to people again and again these days. If you told me that we could have gotten IOI gold, then I would've just assumed that we could all just go on vacation. It's all over. AI is sol. Like no point in working anymore.

[A]

We got feels like nothing's nothing that much has changed. Right.

[B]

Like life is still the same.

[A]

Yeah.

[B]

So I think that's super interesting. Yeah. I don't have a great way to explain it, but I think that's actually what I spend a lot of time thinking about is why is that the case? Because you kind of see this again and again in AI, right. With solving chess and then it doesn't really matter and solving go and. Yeah. So you keep seeing it, but yeah, I think it surprised you every single time.

[A]

Yeah. I think one is we keep moving to the goalposts. We're very good at that. And then two is I think actually our definitions of what constitutes AGI is bad. And we don't actually mean what we say. When we say, oh, when we have achieved this, then we have AGI. So clearly when we have achieved our goal with a language model, we have AGI. It's wrong.

[B]

Yeah. And I think shifting the goalpost to some extent is correct. We keep goodhearting whatever goalposts we have. And I think it's kind of hard to like to be good.

[A]

Hard is like too negative. It's like I will cheat to do what you asked me to do. But I don't think it was cheating. It was just scaling test time compute.

[B]

At a meta level. I think the community not cheating, but makes a lot of implicit decisions to go after the evals and benchmarks that matter the most.

[A]

Sub is very fine for sure.

[B]

Yeah, exactly.

[A]

But yeah, ioi hopefully not that good hearted. Well.

[B]

But it kind of clearly is to some extent. Right. Because most programmers in the world cannot do IOI at any decent level. But we're still struggling to automate most programming jobs or there's a lot left to do.

[A]

Language models are here like junior, senior, dev and then suddenly for IOI you're spiking.

[B]

Exactly. And there's something suspicious about that. Yeah. Okay. I kind of saw this at a meta level also with RL research. So yeah, I my PhD with Sergey Levin at Berkeley from 2017-22. And that era of RL research was super interesting because RL was super hyped. Right. Starting from about DQN in 2015. And a lot of the methods that people were really excited about is off policy learning, value functions, these kind of things. And somehow that stuff hasn't really panned out, I would say, and it's not exactly clear why, but in the academic literature we thought we were making a ton of progress. And I think in retrospect I had to say that we probably kind of overfit to the benchmarks pretty heavily. And how I see this in retrospect is that we gave ourselves a lot of new knobs to tune and then implicitly kind of tuned those to fit the benchmarks. Everyone knew that we were doing that at some level. But I think it's hard to appreciate that it's not just happening for a single paper at kind of a meta level for the whole community. That's happening too.

[A]

Yeah.

[B]

And I think the result is that like, I don't know, like a lot of the RL research that came out of that era I don't think is like that used, you know, and I think it's kind of for similar reason that basically were kind of like benchmark maxing.

[A]

I will full out say there was R.L. winter.

[B]

Right.

[A]

Like entire startups that were founded based on premise at the time basically gave up. Some of them died, some of them pivoted, whatever.

[B]

Yeah, yeah, yeah. I think in. So because I was in academia, there was still quite a lot of excitement over it. But yeah, it still felt quite academic. And yeah, I in that area was a little bit frustrated because I felt like one of the pitfalls of academia is that it doesn't really reward simple ideas that work and instead kind of tends to reward kind of math theory ideas. Those math theory ideas also give you these kind of implicit knobs to tune in that allow you to overfit, while the things that actually work tend to be kind of simple ones that have less knobs and just generalize to many.

[A]

Things with there's just less secret sauce to it apart from just throw a lot of compute in.

[B]

Exactly, exactly. But those are things that tend to.

[A]

It's not intellectually interesting.

[B]

Yeah, exactly. And from an academic point of view, it's like, oh, why am I sitting in school? I think for a lot of people who do PhDs, they're kind of Wired in a way. They want to think about interesting new stuff. And yeah, the scaling era kind of probably stuck to that.

[A]

Scaling era. Is the scaling era over since we're.

[B]

I think I've just been paged into that from Oyaski's interview. I don't think it's over, but there's definitely something interesting happening. The thing I was saying about IOI and imo, I think we'll still continue more or less on the same track. Like clearly, you know, like these labs are like releasing their new pre trained models and they're like still doing like, like much better than before. So I think, I think scaling is still happening, but I think it's happening in a different way or you know, it's like worth like seriously interrogating. Why is it that we're not just like automating all jobs right now? I think my view, it is something like RL, the way it's applied to LLMs right now is kind of a weird, funny tool where it doesn't really generalize beyond the training distribution that much. It generalizes to some extent and generalizes in interesting ways, but it's very picky. It can kill the training distribution completely. It can be best in the world at it with not that much effort really. But yeah, it doesn't really generalize. So I think what we had to do is bring the world of economically useful tasks in distribution for rl, if we commit to using RL as a tool. And it might be the case that maybe there's some cool continual learning thing or something that shifts a paradigm next year or something like that. But it really feels like if RL is a tool, then a big thing that needs to happen is it doesn't feel like intelligence of the models is the bottleneck. It's more like you just have products that bring the entire context of what someone wants to do into the product so that the alarm can like see it. And then you use RL on top of that.

[A]

Yeah. Have you seen GDP val?

[B]

Yeah, I've seen it, yeah. Yeah.

[A]

Is that basically what you're envisioning?

[B]

Yeah, I haven't, I haven't looked at GDP VAL closely actually. Haven't seen exactly like, like so roughly. Yeah.

[A]

Recap. Yeah, it's 128 tasks across. Like any white collar job that takes like more than 5% of GDP. Right. And they basically created all the context, the eval on it and evaluated every model. Yeah, famously. OpenAI's evals to you, whoever runs that one always finds the anthropics are the best one. Yeah.

[B]

Props to them for publishing doing that. Yeah.

[A]

It's actual science.

[B]

I think it's good.

[A]

But I think in the sense of generalizing beyond coding competitions to economically useful tanks, that is it. I think that is what is more important for GG6.

[B]

Yeah. What I'd like to do is kind of like, I haven't. I just haven't read the like GDP eval traces closely. It's not clear to me that, you know, like, what, like what does the job of an accountant entail? And like, what kind of context needs to be in the product?

[A]

Yeah.

[B]

So you can actually do it.

[A]

PDFs.

[B]

I see, I see.

[A]

Like, so they try to go as.

[B]

Close to source documents as I see. I see. Yeah. Yeah. So, yeah, I think like roughly operating in this kind of thing is what I envision.

[A]

Yeah. But because it can't be like an artificial, like, oh, let me clean up this data for you.

[B]

Exactly.

[A]

To make it easy for the LLM to process this. No. PDF in an agent go.

[B]

Yeah, I think that's roughly the right shape of the thing. And I guess how I imagine this being operationalized is that you'd want to co design the product and the model so that the product for whatever it is, coding is kind of maybe the easiest first step because most of the context that you care about is just your code base and being able to run stuff in the terminal and that kind of stuff. And still we're not that close really to automating it necessarily. But for all the other jobs, the context is like insane. Right. It's like all the conversations you've had with your co workers, like your Slack messages, you know, like, for my. So at OpenAI, I was working on kind of like hyperparameter scaling research and I actually wrote not that much code.

[A]

Grid search or neuro architecture search.

[B]

No, more like understanding how different. Like science of deep learning in like 2020 where it's like, oh, you had to like initialize the layers in a particular way to get good scaling. Love kind of the analog for that for rl. The thing is, I didn't write a ton of code. So the LLM writing code is not the bottleneck, but it's more like over the course of a year I run sweeps, look at the interaction between different hyperparameters and kind of build up that knowledge for a year of just different graphs. And to do my job, the model would also need all those things in context to successfully like, you know, kind of automate my job. And you'd Kind of want a product that allows you to like, bring all that context in.

[A]

Did you build it for yourself or is there an existing one?

[B]

Oh, no, I mean, like, no, I mean, I, like, I. You know, those graphs are just sitting in my head. Yeah, Right. So I think it would be pretty hard to like, go automate that job. But I think what you need to do is build a product that kind of. Yeah. Has. Brings that context in. And then you want to rl on top of that to understand, to teach the models. Use that context.

[A]

Yeah. Another conversation that I think has really come to face this year is kind of the depth of one model fits all. I feel like the point of the G in the AGI is like one model fits all. I think OpenAI has clearly abandoned that this year.

[B]

Oh, what do you say?

[A]

Fiji? Simo writing a blog post to the title that we are no longer doing one model fits all.

[B]

Okay, interesting. Okay.

[A]

And I think Mark Chen or one of the other senior people that are not Sam also saying this in the podcast. So basically the idea was you started with Codex, someone else was doing instrument GPT. Then we launched GPT 4 for O. I guess 01 and 01 was kind of supposed to be like a reasoning one model fits all. And there we merged the 40 and 0103 line into 5, and now we're splitting it out into 5 and 5 codecs.

[B]

Again.

[A]

It's like just a weird.

[B]

Well, OP is very guilty. I mean, you know, I don't think you should interpret those as like, scientific facts about the universe. It's just more like OpenAI has a tendency to ship the org chart, basically. Yeah, right.

[A]

The world has the tendency.

[B]

Yeah, exactly. So I think a lot of it is related to that. But. Yeah, I see what you mean by like. Yeah, actually I do wonder if. Yeah, like the current reasoning part, the current reasoning paradigm is just kind of fitting itself to this kind of peaky in certain areas thing. Right. I don't think it's so much a matter of model capacity, though. It's just more of another kind of organizational thing that if you care really a lot about coding, you probably don't have the data to do all the other stuff. I don't think it's so much a matter of like, if you had all the data, probably you would benefit from just like training on all of it and you'll get some generalization between these, but it's hard to find like one organization that cares about all these at once. Yeah, yeah.

[A]

So before I double click on just like the o series in OpenAI. I do like to ask OpenAI people who are there, do you have a favorite blip story?

[B]

Yeah, the blip was crazy for me. Like, yeah, I was. It was like Thanksgiving.

[A]

Like, everyone remembers where they were, what they were wearing.

[B]

Yeah, yeah, exactly. I was at Thanksgiving with two opening eye friends, actually. And then one of them on, like, Friday afternoon is like, oh, like, Sam Altman just got fired. We were just like, co. Working together. It's like, I'm like, what? Good joke. And then, yeah, it was crazy. And then, yeah, it was. It was just like a crazy weekend of just like, ups and downs. Like, you know, we thought. Yeah, like, we signed a letter.

[A]

Yeah, I did say 95% of people signed.

[B]

Yeah, yeah, yeah, yeah. I thought, you know, like, I bound to Microsoft or. Well, I think maybe I had a slightly more complicated. Like, I actually do think that governance feels really important to me.

[A]

Yeah.

[B]

Because it does feel like no matter if we hit AGI in, like, two years or 10 or whatever, it's not clear that we have a good structure for the governance of it.

[A]

Okay.

[B]

And so it is a question that I think we, like, probably should spend more time on. And I was like, during that period, just pretty willing to be like, you know what? Let's forget about the equity and stuff. I think it's good and healthy to have a conversation about how exactly the government should work.

[A]

Okay. You care about this. Yeah. Right. So now the OpenAI nonprofit has this secret shadow board of members that determine when we've reached AGI.

[B]

Yeah.

[A]

Is that better?

[B]

Yeah. I don't have a. Like, maybe I would say I don't have an answer. Like, you know, like, it's just.

[A]

It's not there above my pay grade. But, like.

[B]

Yeah. And even back then, I was kind.

[A]

Like, well, I don't care.

[B]

I. I do care quite a lot. When the blip happened, one of my reactions was like, well, you know, this nonprofit board stuff. Like, actually, if it takes such somewhat, like, surprising, maybe erratic actions, like, maybe you'd rather just have like, you know, a thing like the Microsoft board, which is kind of like, you know, like, probably like all the pensions in the world, but sure, like serious people, but also, like, you know, the stakeholders are kind of like the whole world because everyone's kind of, you know, through their pensions or something, invested in it. Like, maybe that is a bit more of a democratic way to run things and having, like, seven people run it. But, yeah, I don't. I don't really know. It feels like we haven't solved governance like at all though, right? Like, if we get AI, even stuff like unhealthy food or like social media, it kind of feels like, like whatever the kind of like capitalistic incentive is, like, doesn't actually capture kind of good outcomes for society. Maybe. Yeah.

[A]

So about the transition into reasoning, right. You shocked me by mentioning that the reasoning team is 300 people.

[B]

It's kind of like now that when O3 was kind of short as a product, I think it just kind of gets larger and larger. How many people work on it? I think I've lost track of the numbers, but yeah, a lot of people contribute to the different aspects of safety and whatever.

[A]

Eval original 01 I saw the video. It was like a dozen people.

[B]

Yeah. Even then, if you look at all the contributors, it was probably more like 50 to 100 people.

[A]

Okay, yeah, let's tell that story from your point of view, figuring out what does RL mean there. And I guess was this a branch of any other prior work that you want to credit?

[B]

Yeah. So I think setting the scene, I guess in 2023, people are kind of talking about, oh, are the scaling laws dead?

[A]

This kind of stuff every year? Every year, yeah.

[B]

Yeah. But especially, I think especially that year, it felt pretty like, serious, you know? Yeah. I think in general, OpenAI is really good about, like, having conviction in something and just like, really like from first principles, like, going after it. And I think, like, the people who are kind of most responsible for that is probably like Ilya Suskever and Jakob Pachocki. I think even like, like, Dota was kind of more or less the same template in some ways. Right. And that was 2017. And so a lot of the people there have kind of this like, AGI in their bones kind of point of view. And they've basically been convinced that, like, RL would be the way to get there. So I think for a long, long time people have been convinced that something like that should work. And it's just that it started to work once the. SO pre training got good enough.

[A]

Okay.

[B]

Yeah. I think human feedback is kind of like a bit of a side branch because you can't really pour that much compute into it. Right. It's like you take the model and you elicit it to be a little bit better in terms of personality. But the people that are really convinced that at some point it's not about copying the Internet, you can go do rl, and that's the path to getting much better intelligence. So I think there was kind of like A long line of kind of like returning to RL in different ways. And then it's just that around 2023 is when it started really clicking. And it's kind of interesting because even it's not like those initial models performed way better than the existing models because they're like smaller scale. But people were very good at being like, oh, this is kind of interesting. The reasoning trace that you see here is kind of not something that you've really seen be so accurate in other models like this one. Kind of similar to how I think a lot of people didn't really think of GPT or GPT2 as something that was super compelling, probably. I know that I personally didn't like GPT2 that much of GPT2. I was like, okay, whatever. And then I think. And then GPT3 happened. I'm like, oh, whoa. I feel a lot of FOMO sitting in my PhD. It's kind of that where I think it takes a bit of first principles conviction to decide that, oh, this thing, there's something here and we should really go scale it up. And OpenAI is really good about. Once you decide that something is good, then you just scale it up all the way.

[A]

Yeah. Was there an internal prototype pre 01 that was like, okay, this is the thing, we'll fund it to scale it up. Right? There usually is.

[B]

Yeah, yeah, exactly.

[A]

What was the thing? What was the demo that really sort of sold?

[B]

Just this running RL on even a pretty small model, producing very interesting reasoning traces and getting surprisingly good scores on math in a way that we couldn't have done without a bunch more pre training. And then once that looks good, then more and more resources went to just scaling up that new law and things like adding tool use and this kind of stuff. Yeah, yeah.

[A]

I think a lot of people make a lot of headlines on the large models, but I think it's very underappreciated the minis, how well this solution works. Any comments or just like discoveries on.

[B]

Yeah, nothing much to say there. I was also, like, not super involved in the mini stuff. I think maybe one thing not exactly related to that, but it seems like externally people are kind of very like, oh, research seems to come in these big leaps. But I think internally at OpenAI, it feels very smooth.

[A]

You have a bunch of experiments, some of them have inconclusive results, but maybe you stack them.

[B]

Yeah, exactly. You stack them and just you keep scaling. You keep having different runs that get a little better each time. Okay. So I think that's maybe one other Aspect that's a little underappreciated is that, I don't know, in the media there's just these wild swings between, like, oh.

[A]

It'S so like Googlers. We're in it.

[B]

Yeah, exactly. And I think internally at Big Labs, it's just kind of like, oh, we're just like chugging along. Maybe this month is a little better than last month or something, but it's not as crazy up and down.

[A]

I think the question is there used to be more of this and now I know there's less, which is. Well, the stuff we've released, we're like, you know, internally, we're like six months ahead. Part of the reason why people opening I wasn't that excited about ChatGPT's launch was because he already had GPT4. They're like, oh, we'll just put this out. We're already way ahead. I think now people are just releasing things as they have them. Like, I think.

[B]

Yeah, especially because there's some, like, competitive pressure. Right? Yeah. I think people are probably pretty worried that, like, if you, if you let a lead linger for too long, that will like, grab a lot of market share. Like, I don't know, like Nano Banana Pro right now is probably like, you know, it's like a month. Yeah.

[A]

So I would say like, now the lead internal to external lead time is about one to two months.

[B]

Yeah. Which. Exactly.

[A]

Tiny.

[B]

Pretty sure. Yeah.

[A]

Tiny. Anything else on reasoning side, I guess you can talk about specifically the work on coding. Anything surprised you or, like, is an External misconception on 103side before I go to cursor.

[B]

Well, not really. Like, yeah, I mean, it's. Yeah. Pretty cool. I think it felt already by like maybe early 2024, like, oh, wow, like this recipe really works and we can see how far we take it. And so I think it was like very steady progress. And by that point it was probably pretty predictable that we could really smash things like IMO or ioi. Yeah. One funny thing that kind of happened is while this was happening, I went to this conference called the Curve, which is about kind of AI progress and Joseph Gordon Levitt. Yes. I went last year. This is before the Owen stuff was released. Yeah. And I went to this thing where people were kind of making bets on where we would be on epoch AIs. Like the epoch a math exam and humanities last exam and stuff like that. And their estimates were like, oh, we'll be at like 10, 20% in 2027. And I think at the time there was models internally that were already better than their estimates. So it's off by two years or something. And the interesting thing is those are also people who are kind of predicting that there'd be dyson spheres by 2035 or something.

[A]

Okay.

[B]

You see what I'm saying? So the current estimate is way under.

[A]

Yeah, they're too pessimistic in the short term, too optimistic in the long term. Yeah.

[B]

Well, I don't know if there might be dyson spheres by 2035. I don't really. But I think that is one interesting aspect is that. Yeah, I think people still seem pretty miscalibrated in different ways. I do really appreciate how that community makes predictions though, because I think most of the rest of the world just kind of cynically says, oh, I saw this the whole time. I do appreciate that.

[A]

Is this EA adjacent?

[B]

Yeah, I think, yeah, exactly. It's like that group.

[A]

Yeah, yeah, I like that. They like to sort of register their opinions ahead of time.

[B]

And I think like broadly, the people who've been the capabilities predictions in that group have been broadly correct. If you look from 2015 to 2020 or something, where I think a lot of people kind of thought that AI was a sham or not really going to be that useful for a long time and actually it's somewhere in the 2030ish thing that it will probably reach human level intelligence.

[A]

Yeah. It's weird. I feel like a skeptic when I keep saying everyone always predicts that EJ happens in their lifetime. That's very convenient for whoever. And we have a consistent view of history where you see people in the 1800s and 1900s making predictions. It somehow always lands in their lifetime, whatever the thing is. But this time it might happen almost surely, right? I'm pretty sure, yeah. So it's an interesting observation. How different are we from our predecessors in terms of developing our technology? Did the Deep Seek moment this year also this year? Crazy. Change anything internally?

[B]

Not really. Yeah. I think more so. Just surprised that it created such a moment. It was kind of confusing, right? It was like Deep Seek shows that Nvidia chips are actually more useful than previously thought. And Nvidia's stock goes down a bunch of. It was kind of like.

[A]

I think it's more like, okay, well I'll do the Steelman that side, which is. Well, you don't need the top of the line Nvidias. You can just use the sort of previous generation or the shackled ones they sell to China to do an equivalent amount of work for a model.

[B]

I See? Yeah. But then it was also, I guess the feeling in OpenAI is that, like, well, I think we had a better model already at the time. Right. And it was quite valuable. Smarter models were clearly quite valuable. So you kind of wanted to be at the Frontier.

[A]

Okay. So I wasn't quite framing this as like a race dynamics thing between labs. It was just also more like, well, were they right? Would their approaches. Right. They had R10, which is kind of like a really cool branch. So more like commentary on what we learned about RL this year.

[B]

Yeah. Yeah. Well, it does seem like basically a lot of the labs have kind of converged onto some similar ish way of doing rl, and they're all kind of back at the same level of frontier again. Like, even the anthropic models, like the Opus 4.5, it has this kind of like, there's this, like, RKGI2 plot that looks exactly like the OpenAI ones.

[A]

Right.

[B]

Like, so I think everyone seems to be converging on a pretty similar form of rl. Yeah, it's kind of interesting. I think people basically figured out in one way or another to achieve more or less the same thing.

[A]

Let's talk about the move to Cursor. Why is Cursor accumulating enjoying so many cool RL people?

[B]

Yeah, I'd actually kind of already talked about this so far, I guess. Yeah. I think from the perspective of Cursor, it's nice not to be so dependent on external labs for everything. And I think there's also unique opportunities to co design the product with the model in ways that I think we couldn't do unless we actually built the model ourselves and had access to making it good. So, yeah, that's kind of broadly why Cursor's so excited.

[A]

I'll push back a little bit. Right. OpenAI has Infinity resources, Infinity Data has codecs. You could have just stayed.

[B]

Yeah, well, actually, right around when I was leaving is when I think people started actually using codecs a lot. So that was kind of like, it happened right after I left. So that was kind of funny.

[A]

So mostly people are using Cursor internally. Maybe a bit of Windsurf because it's left over from the previous thing.

[B]

Sure. Yeah, exactly. So it wasn't that obvious. But actually, I think more to the point, this thing I was saying about RL is kind of a tool that doesn't really generalize that well. So what you want to do is bring the entire kind of test distribution inside your training distribution. I saw the opportunity to do that at Cursor kind of directly. I think the Cursor folks are also just really excited about that kind of vision. And it's just like a small place where the product people sit right next to the ML people. And I think there's a lot of potential there. You can kind of see that recently Jacob Jackson had this blog post about online tab where we're doing policy updates every two hours. Exactly like a policy update every two hours or something. And I think that's the type of thing that I think it's a little hard to do. It's very hard to imagine that at OpenAI, for example, just because the product is this kind of complicated thing. And also the product people and RL people are pretty on different sides of the the org.

[A]

I think if you put your mind to it, you would. It's like, you know, Tab is an autocomplete. It's a smaller model. It's, you know, it's not as complex, I guess, below them.

[B]

Yeah, but I don't think that's really this like, you know, I think we. I don't think that's why Cursor was able to do it. It's actually more about like just the org itself being kind of like smaller and a bit more like focused.

[A]

Yeah. Okay. Well, I mean, since you're indulging this, I think the question about continual learning, which obviously is a big theme. It's always been a big theme. It's bigger this year is, well, don't you need to curate your data? You can't just like Chuck, whatever your users are doing in straight in because that tends to get you towards the middle of the distribution. Actually you want to spike it.

[B]

I guess it depends how you're thinking about continuing. I mean, I don't know. Humans are quite good about dealing with bad data too. Right. You can see someone doing something dumb and decide like you're not going to do it.

[A]

Filter it out.

[B]

Yeah, yeah, but it's not even actually filtered out. You have to. Presumably some kind of value function that says that if you see someone touch a hot stove, you're not going to go. You don't need to. It's not just filtering it out. You're actually not going to do it. Right.

[A]

You could rediscover hot stoves on first business.

[B]

Yeah, but you don't need to. So I think there's something pretty deep there. Yeah, it seems like we're kind of like a few orders of magnitude of kind of data efficiency basically away from that kind of. You do Something once or you make a mistake, you introduce a bug in your code, you're not going to do it again. But the models will happily just keep doing it even within the same context. But definitely of course across context. So I think there's something interesting and deep. There is maybe, yeah, I suspect that it'll be kind of like paradigm shifting in the next year or something, but I had no idea what it might be.

[A]

So primarily you worked on Composer Tab and maybe Search.

[B]

So I've actually just worked on Composer and that's kind of like the main focus of the company basically. Or like the ML group is shipping a better shipping embedded composer.

[A]

Can you describe I guess the impressive brag a bit about the ML group?

[B]

Yeah, yeah. I mean I think the ML group is great. It's like, you know, it's just like 20, 25 people and you know, I was like, honestly like pleasantly like very, very surprised at like how good Composer is like given the size of the group and you know, it's not like a big research lab yet. And yeah, it's, I think it's like a really good model. You can kind of see that in the reception and I think it's kind of the start of hints of like co design with the product in some ways because I think one of the reasons that people really like it is it's smart enough that you actually want to use it and it's also fast so you kind of stay in the loop with the model while you use it. Because I think all the other smart models have this kind of they're slow that you want to go kind of context switch away and come back. And that sucks as a programmer it just sucks to kind of context switch. It kind of gives you adhd. It's really terrible. I agree and I think yeah, it's like one step in the direction of being able to be more sync and I think that's basically the whole company is just really full of people who want to code. Even the co founders. Actually the co founders are often some of the best high taste testers which also kind of gives you a lot of reassurance that you're going to ship good stuff.

[A]

Any example test that maybe Composer doesn't solve yet but you're really motivated to solve?

[B]

Yeah, Ironically I feel like I'm actually a low taste tester in some ways because I don't know, I just write stuff. Slow machine learning code and just think about algorithms and stuff all day. I think more broadly I'm super excited about co designing the product. So that you can actually not just right now, we're getting better and better at answering user prompts, and I think that's why Composer one is quite good. But what we're really aiming for is more automate software engineering as a process where you write code, you go look at datadog, look at what's happening, then come back and maybe have some hypotheses about what's better. Rerun stuff. I think that's the type of thing that we actually want to make the model do. And I do think that Cursor is kind of uniquely positioned to do that in the sense of if we can kind of. If a lot of what a software engineer does kind of ends up in the product, I think we can use that to get better and better at, you know, not just writing code, but kind of like the whole job.

[A]

Yeah, I think that's very inspiring. Just to double click on just any sort of RL insights. Sasha and Lee have talked a lot about like the internal tooling that you've had for all the, like the cluster visualizations. Is that helpful? Is that what every lab has?

[B]

Yeah, I think the tooling at Cursor is actually really good, I think, because, you know, it's just kind of like people are just down to like code stuff. They like, do test their own stuff. Like, so we just have like a lot of good tooling where you can, you know, like have like a SSH session into like our own like user environment or something and like, you know, see if like code runs the way that like users got it to run, like this kind of thing. I think that's actually, yeah, quite nice. I think basically one of the big lessons in ML in general is that you want to be like really close to your data and understand your data well. And yeah, I think there's like kind of. Yeah, again, kind of like uniquely positioned to do that.

[A]

Well, especially because all internal tooling, you're not buying anything.

[B]

Yeah, it's just like internal. And part of it is just that we're also working on a product where you can understand it really well because of the code product. Well, like, you know, if. I don't know in OpenAI, if I was like to look at like a biology question, I have no idea, like, you know what. What this is about.

[A]

Yeah, yeah, yeah. Interesting. Okay, so I think that's a good overview of everything, I guess, other than the we covered OpenAI and cursor, just interesting RL work that other people are doing that you're like, still mulling over Its influencer to your thinking. Good papers. Anything like that?

[B]

Yeah. You know, unfortunately, I kind of got in the habit, especially at OpenAI of like, not reading that much external work and just like reading people's like, slack posts internally as like the main, like, way to like, you know, like, learn new stuff. No. Super inspiring. Recent things have popped up to me. I do think that this like, kind of vibe of like, yeah, continue learning, like, does feel like. I think there's something super interesting there. And like, it feels like maybe even in academia people could make like a.

[A]

Big crack at it and continual learning, specifically meaning kind of what tab is doing.

[B]

Yeah, maybe what tab is doing, but also just like kind of like in context learning, but with like infinite memory or something, so that you don't. Once you experience something in context, it should just like, be in your weights and you shouldn't have to like, make that same mistake again, that kind of thing.

[A]

Why do you think there's. Okay, so it should be in your weights, but there's a finite capacity for the weights to remember things. You will forget things if you do that too much. Right.

[B]

I mean, you start out by memorizing or learning from trillions of tokens. Now you're going to experience thousands or maybe millions of tokens and somehow we can. And those million tokens are kind of.

[A]

You only need one epoch.

[B]

Yeah, exactly.

[A]

Crazy.

[B]

Yeah. So it feels like if you could learn enough about those million tokens that you're actually in deployment on. I don't think you should need, like, I don't think there's a risk of overloading the capacity of your model. Right. Because you can trade on a trillion tokens and it's like, fine.

[A]

Right, right. So proportionately it's a drop in the water.

[B]

Yeah, exactly. Water bucket. Yeah.

[A]

Unless you run it for years and you know, at some point it's maybe. Yeah, yeah. So basically, I find it very curious. I've only had one podcast on information theory of language models. Like, what is the theoretical capacity? How much are we using? And you should probably track that.

[B]

Yeah, that's a good idea.

[A]

Yeah. Treat the weights. If you want to store things in weights. Okay. Treat it as a hard drive. What's the capacity of the hard drive? How much can be stored in there? We know the capacity. It is the number of bits that's occupied by the language, by the parameters.

[B]

Yeah.

[A]

Physically cannot store Internet.

[B]

Yeah, yeah, yeah. And it's. Yeah. I've heard that there's this kind of like someone recently at cursor, Jacob, kind of Brought up this view. I don't know if it's like a more public view. It's like, oh, there's kind of like a hard drive view of, you know, neural networks and kind of like a CPU view of neural networks where, you know, is, is what's happening the weights like. Yeah. Memorizing stuff or is it like you're like having some like few circuits that do a lot of work? Yeah, this kind of thing and. Yeah, I don't know. Yeah, you know, I would love to. Yeah, there's like actually so many of these kind of more science y questions that I would like love to explore sometime, but then it really kind of conflicts with like empirical stuff. You know, like unfortunately at any given moment in time it doesn't seem like the most fruit for like, you know, improving something in the, especially in the short run, but even in the next like couple years is like understanding some of these questions. So yeah, I mean, I guess this is technically supposed to be the role of academia, but it's also hard to explore those ideas there without enough compute. But yeah, actually I would love to go at some point return to exploring these kind of fundamental science ideas.

[A]

Okay, I'm kind of springing this on you so you can take some time. What is a good RL interview question that if somebody can answer they should join cursor immediately?

[B]

Ooh, it's a hard question.

[A]

Assuming you do interviews.

[B]

Yeah, well, actually at Cursor we do like work trials and it's like two day worth trials that I actually think that that's like more representative because you.

[A]

Plug in and see how they behave.

[B]

Exactly. So I actually think it's like more valuable. This is honestly less of a thing about how you understand RL and bit more like were you around in the like 2017-22 era? But it's like why is off policy? RL unstable is kind of, I think like a good question to like dive into.

[A]

I don't actually know. So dig into it.

[B]

Yeah, cool.

[A]

Thank you. That was great conversation. Do you have any sort of call section?

[B]

Yeah, I mean we are definitely hiring a cursor. So yeah, if you're interested in working on especially like kind of data and rewards for code, I think that's like a huge need. Yeah. Please get in touch. Yeah, that's it. Yeah.

[A]

Thank you. Sweet.

[B]

It.