Nathan Lambert

"This is why I like the ChatGPT app, because it gives the AI a home on your computer where you can focus on it, rather than just being another tab in my mess of internet options."

"Gemini 3 is a fantastic model, and I still use it. It's just kind of differentiation is lower."

"I will regularly have like five Pro queries going simultaneously, each looking for one specific paper or feedback on an equation or something."

"I use Gemini for that. So I use thinking for all the information stuff and then Gemini for fast things or stuff that I could sometimes Google."

"I don't know, Exa is my preferred search provider, but somebody else might care for a different search startup."

"And then for code and any sort of philosophical discussion, I use Claude Opus 4.5. Also always with extended thinking."

"And then sometimes use Grok for real-time information or finding something on AI Twitter that I knew I saw and I need to dig up."

"I need to get him a copy of Season of the Witch, which is a history of SF from 1960 to 1985, which goes through the hippie revolution, like all the gays taking over the city and that culture emerging, and then the HIV/AIDS crisis and other things. And it's like, that is so recent, and so much turmoil and hurt, but also love in SF. And it's like, no one knows about this. It's a great book, Season of the Witch. I recommend it."

"Largely because the margin on NVIDIA chips is insane, and Google can develop everything from top to bottom to fit their stack and not have to pay this margin."

"Like Deep Research, Sora, o1 thinking models—all these definitional things have come from OpenAI."

"Like Deep Research, Sora, o1 thinking models—all these definitional things have come from OpenAI."

"Like Deep Research, Sora, o1 thinking models—all these definitional things have come from OpenAI."

"On my blog, we scrape Hugging Face so we keep download numbers for every dataset and model over time, so we have them."

"Qwen might be the one— Oh, yeah. Qwen was the obvious name I was gonna say."

"When I was writing about OpenAI's open model release, they were like, 'Don't forget about GPT-2,' which I thought was really funny 'cause it's just such a different time."

"Hugging Face has SmolLM, which is very popular."

"With OpenRouter, it's easy to look at multi-model things. You can run DeepSeek on Perplexity."

"If you think about GPT-5.2 Pro taking an hour, it's like, what if your training run has a sample for an hour and you have to make sure that's handled efficiently?"

"The hype over Anthropic's Claude Opus 4.5 model has been absolutely insane, which is just... I mean, I've used it and built stuff in the last few weeks, and it's... it's almost gotten to the point where it feels like a bit of a meme in terms of the hype."

"ChatGPT kicked off a movement in the US where everything had a chatbot."

"The likes of Z.ai with their GLM models, Minimax's models, Kimi Moonshot, especially in the last few months, has shown more brightly."

"The likes of Z.ai with their GLM models, Minimax's models, Kimi Moonshot, especially in the last few months, has shown more brightly."

"The likes of Z.ai with their GLM models, Minimax's models, Kimi Moonshot, especially in the last few months, has shown more brightly."

"Personally, I have very mixed reviews of GPT-5, but it must have saved them so much money with the high-line feature being a router where most users are no longer charging their GPU costs as much."

"Largely because the margin on NVIDIA chips is insane, and Google can develop everything from top to bottom to fit their stack and not have to pay this margin."

"I think there's a seminal paper from a Meta internship. It's called something like 'The Art of Scaling Reinforcement Learning with Language Models.' What they describe as a framework is Scale-RL."

"the famous paper, Direct Preference Optimization, which is a much simpler way of solving the problem than RL. The derivations in the appendix skip steps of math."

"But if you go from a small university with no compute and find something that Claude struggles with, and then the next Claude model has it in the blog post, there's your career rocket ship."

"Although when Grok 4 came out, the Grok 4 SuperGrok Heavy, which was like their pro variant was actually very good and I was pretty impressed with it."

"And then DeepSeek are the people that did the training breakthrough, which is, they scaled the reinforcement learning."

"That's the older term for it coined in Anthropic's Constitutional AI paper."

"I think you can kind of take this in order. I think you could view it as what made o1, which is this first reasoning model, possible, or what will the latest model be?"

"If we look at the GRPO equation, this one is famous for this because essentially the reward given to the agent is based on how good a given action—an action is a completion—is relative to the other answers to that same problem."

"Fun fact: I was on the team that came up with the term RLVR, which is from our Tulu 3 work before DeepSeek."

"Fun fact: I was on the team that came up with the term RLVR, which is from our Tulu 3 work before DeepSeek."

"I work at the Allen Institute for AI, where we've been building OLMo, which releases data and code."

"I think a lot of what I was trying to do in this RLHF book is take post-training techniques and describe how people think about them influencing the model and what people are doing."