The Ledger, every show

“The existing paper people are writing about multi-turn RL are not actually incorporating this, and it kind of, like, breaks all the math.”

“You can have a model trained on code, and a model trained on math, and a model trained on Spanish, and you can literally average the weights, and it works.”

“The updates made to model weights are orthogonal enough for specialized tasks that this is actually like totally fine. Things are nice and linear in most cases, things are nice and orthogonal, and you can get away with a lot of async updates to models that are…”

“GPT, 4.5 is like ginormous model, trillions of parameters, most likely. And that like, there's more room in the model to have these like little sparse connections materialize as you go through layers of the transformer. And I just haven't seen anything like th…”

“The quality bump over things that are much smaller is just like the, we're hitting diminishing returns on capital investment is a lot of it. Like they're taking out of the API because like the, they can sell other things with the same GPUs and make more money …”

“When Meta releases this behemoth model, I don't think anyone's really gonna run behemoth. For like their day to day stuff. It's just probably not gonna be worth it.”

“That's kind of like my bet is like, people are going to really want to train models to be agents. And I think you can get that to work well with a pretty small model.”

“I'm happy to call O three, 10 minute AGI. And I think like framing AGI in terms of like length of time, it takes a human to do a task is like more reasonable than like a global framing. Like, sure. There's a bar of like drop and replace for a human that we are…”

“Most of these pilots are very much intended to be churned. Like, they're not, they're very much in a, they're not being rolled out broadly. They are coming through in a kind of walled off environment for people who are, like, gonna be the beta testers.”

“Like Deep Seek is serving all of China on 2000 GPUs.”

“And it seemed like Anthropik's kind of attitude has been that extended thinking is an instance of tool use and that it's the kind of thing you want to equip the model with the ability to do. But it's not like, oh, it's a thinking model. It's just a sync for th…”

“A lot of the kind of headline anthropic like safety results, especially related to reward hacking and kind of deviation and alignment faking, Are all things to me that seem like a rock and a hard play situation where the model has two objectives it's given tha…”

“I think being an eval company puts you in a really hard spot. Some people are talking about this on Twitter, like just that to be an ed-all company, you kind of have to sell to the labs, but selling to the labs doesn't really, like kind of wrecks the revals.”

“I mean, I do think that like the best source of evals going forward is probably going to be academia.”

“GRPO is, like, great for, like, leaning heavy on highly parallel inference compute. It's more memory efficient for the actual training process. It's much easier to do in a distributed fashion because you have less gradient syncing and less model weight copies.”

“So it, like, feels like people are moving in the direction of model-based rewards, where you, either LLM is a judge where the judge sees the correct answer, or it has questions it's supposed to verify as properties of the response, just because that's much mor…”

“Whether or not you want to call this, like, end minute AGI, I kind of like the phrase, 10 minute AGI, for just, like, how to think about O-three is that anything that you can do as a human in 10 minutes, O-three is usually going to be able to do reasonably wel…”

“Where I think for a while some people were, like, speculating, like, oh, what if you have the model, like, generate images in its chain of thought reasoning where everything is, like end-to-end multimodal input and output, and it seems like you don't really ne…”

“I think people were, like, excited about multi-agent systems for a while, like, the crew AI sort of thing of, like, oh, I'm gonna put my coder agent, my finance agent in group chat, and, like, a lot of these are just kind of silly. They don't actually work ver…”

“Multi-agent RL's hard. I did five years of it in grad school. It's like not easy. And to the algorithms don't really even exist for the things you would really want to do.”

“Scaling to, like, million token contexts is, like, really, really hard. There, I don't think there are real, like, open source replications, open token context scaling, Beyond, like, tiny, like, academic model sizes.”

“And the most, one of the most promising ways, I think, towards doing this is having the reward models also be able to answer harder questions by themselves being reasoning models.”

“The reason O-three is good is because it's trained to use tools. The way you train a model to use the right tool for the job is reinforcement learning. And they've said as much, like, deep research, reinforcement learning.”

“A reason that Windsurf has been successful as a Cursor competitor is they lean way harder on enterprise than Cursor has. Like, they have really designed for enterprise integration, whereas Cursor really has not.”

“I mean, to me, the surprising thing was not any individual one thing. It's that each of these is maybe like a 30%, 50% gain, but they have like 10, 15 of them that all stack. And so getting all of these to stack nicely is what's hard.”

“Like they make, I think, still the best model suites for like doing research.”

“The thing that's going to make the next wave of stuff be powerful is just, like, everyone wants better agents. Everyone wants models that can, like, go off and do stuff. And, like, reasoning was kind of, like, a precursor to that a little bit.”

“I mean, I think these models should be the same model, and Anthropic knows what they're doing. Like, it's not that hard to, like, Quen did it in a very kind of, like, simple way, and they kind of talked about how they did it a little bit. But it's not, like, t…”

“So it seems like artificially truncating the thought is actually like fine. Like the model can, even if like it got cut off mid-sentence with an injected like think token, these are smart enough models that they can kind of finish with the best that they got f…”

“I think in chat interfaces, it probably won't stick around. Like, I don't think we're always going to have the dropdown of like Oath for many and Oath for many high. That feels silly.”

“The base model in general of LLM is not artificially constrained in any way. Like, with the right prompt, it'll do whatever up to its intelligence limit.”

“I never really got to the point where I found it was helpful for a thing that was like a large existing code base. But if it's like, hey, I want to like cook something up in a few hours for fun. Pretty good at that. But these become messy and they become hard …”

“It feels like people in the AI world, like, love Claude, or have grown type of Claude, but still had a phase where they were using it a ton. But it hasn't really broken out to general people in the way. And it feels like a lot of their marketing that I've seen…”

“If you set these models up to use tools, They just won't. Like if you say, hey, here's a question. You have access to these tools. Do as many rounds of tool calling as you want, and then submit your answer. They'll just submit their answer because they like ar…”

“If you want models to use thinking tokens, you kind of have to, like, incentivize that. You have to either do a little bit of, like, SFT warmup, or you have to, like, Reward them for doing it. Otherwise, they will not follow it a hundred percent of the time on…”

“If you start rewarding them for, like, tool use, They will use the tool, but they don't really want to, like, have to, they want to, like, be very safe with it... They would like do silly versions of tool use where they aren't actually using the tool to assist…”

“There really isn't currently an open source model that behaves in the way that these models do. We have things like R-one, which are great at kind of the single-term math and code reasoning problems. But they are not in the general purpose agents world yet.”

“And that, if you can crack that scaling direction of, like, pushing the boundary of how long these models can go out and do these things for, that is a plausible path towards things that become marvelous.”

“You probably want some kind of intermediate verification where you're not just waiting for the final answer at the end, But you want something like turn-level reward, potentially, where you want to be able to ensure that the model is getting credit for the mov…”

“You also ideally want to, whether you're doing this centralized or decentralized, move in the direction of, like, everything being async and overlaps because that is, like, otherwise you have these inefficiency bubbles that, like, pop up all over your compute …”

“Using more tool calls, searching the web more, gives you a nice scaling curve where you get better answers by putting in more effort, by, like, spending more time browsing the internet, essentially.”

“Coherent to this very extensively in their latest flagship model, command A”

“Like when people say program synthesis, like we're already there, like O three is program synthesis, but the programs are like JSON and Python.”

“The PowerPoint, like, Microsoft PowerPoint copilot is not great. It's not a thing that I have heard many people say saves them up time.”

“There's a surprising lot of alpha still on X just from places where you don't like, so there's places I have not found much alpha on LinkedIn. The group chats, the anons the open source like GitHub discussions, lots of really good stuff is buried in like a Git…”

“Like, the day GPT-IV launched, Morgan Stanley had integrations, because we had been working on it, and these were, like, we had press releases for these, like, we are ready to go”