korrents

Nathan Lambert

@nathan-lambert · 22 positions · 0 changes of mind

Research scientist at the Allen Institute for AI, where he works on post-training open language models, and the author of the AI blog Interconnects.

Nathan Lambert did not write this page.

We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.

  1. Our reading

    Data processing, filtering and quality, not architecture, is the number one determinant of how good a language model turns out.

    Their wordsAnd we'll get into the details of the models and again and again as we try to get deeper into how the models were trained, we will say things like the data processing, data filtering data quality is the number one determinant of the model quality.
    spoken · official transcript hear it at 0:06:17 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 1st of 44 in this recording

    Interesting

  3. Our reading

    The Bitter Lesson is misread as being about scale; it is about refusing to add human priors, because a clever specific fix always loses to a simple scalable one.

    Their wordsThe scale word gets a lot of attention in this. The interpretation that I use is effectively to avoid adding the human priors to your learning process. And if you read the original essay, this is what it talks about is how researchers will try to come up with clever solutions to their specific problem that might get them small gains in the short term while simply enabling these deep learning systems to work efficiently, and for these bigger problems in the long term might be more likely to scale and continue to drive success.
    spoken · official transcript hear it at 0:40:20 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 2nd of 44 in this recording

    Interesting

  5. Our reading

    Post-training is the best place to work in AI right now, because cheap runs mean a far higher share of your experiments can be all-or-nothing bets.

    Their wordsThis is why you want to work in post-training because the GPU cost for training is lower. So you can make a higher percentage of your training runs YOLO runs.
    spoken · official transcript hear it at 0:49:00 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 3rd of 44 in this recording

    scaling laws

    Interesting

  7. Our reading

    A published training-run cost always understates the truth, because any notable model takes two to four times that run's compute in experiments alone.

    Their wordsAccepted practice is that for any given model that is a notable advancement, you're going to do two to 4x compute of the full training run in experiments alone.
    spoken · official transcript hear it at 0:56:33 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 5th of 44 in this recording

    Interesting

  9. Our reading

    Export controls cannot stop China from training frontier models; what they can do is cap how much AI China gets to run.

    Their wordsThere's not many worlds where China cannot train AI models. I think export controls are decapping the amount of compute or the density of compute that China can have.
    spoken · official transcript hear it at 1:02:02 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 6th of 44 in this recording

    China

    Interesting

  11. Our reading

    Language models are already a form of AGI; what the labs are chasing is a further step, not the arrival of general intelligence.

    Their wordsI think my personal definition of AGI is much simpler. I think language models are a form of AGI and all of this super powerful stuff is a next step that's great if we get these tools. But a language model has so much value in so many domains that it's a general intelligence to me.
    spoken · official transcript hear it at 1:08:05 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 8th of 44 in this recording

    AGI

    Interesting

  13. Our reading

    Language models have not made misinformation meaningfully worse, because distribution, not the cost of writing convincing text, was always the limiting factor.

    Their wordsThere's some research that shows that the distribution is actually the limiting factor. So language models haven't yet made misinformation particularly change the equation there.
    spoken · official transcript hear it at 1:12:49 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 9th of 44 in this recording

    Interesting

  15. Our reading

    Making safety your number one goal costs real release speed, and on a steep progress curve that shows up directly as a worse-looking model.

    Their wordsWe know that a lot of the American companies are very invested in safety, and that is the central culture of a place like Anthropic. And I think Anthropic sounds like a wonderful place to work, but if safety is your number one goal, it takes way longer to get artifacts out.
    spoken · official transcript hear it at 2:17:46 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 18th of 44 in this recording

    AnthropicAmerica

    Interesting

  17. Our reading

    Open frontier models will keep being released whether or not anyone wants to stop them, and stopping them would only leave the world less prepared.

    Their wordsthese open models are probably going to keep coming for the time being, whether or not we want to stop them, and stopping them might make it even worse and harder to prepare.
    spoken · official transcript hear it at 2:21:25 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 19th of 44 in this recording

    Interesting

  19. Our reading

    Censoring a fact out of a language model is practically impossible, because you would first have to remove that fact from the internet.

    Their wordsI almost think it's practically impossible because you effectively have to remove them from the internet.
    spoken · official transcript hear it at 2:34:18 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 21st of 44 in this recording

    censorship

    Interesting

  21. Our reading

    RLHF is not a tax on capability: preference tuning also raises maths and code scores, which is why the labs keep reaching for it.

    Their wordsAnd the important thing to say is that no matter how you want the model to behave, these RLHF and preference-tuning techniques also improve performance. So, on things like math evals and code evals, there is something innate to these, what is called contrastive loss functions.
    spoken · official transcript hear it at 2:39:47 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 22nd of 44 in this recording

    Interesting

  23. Our reading

    Reasoning is not taught to a model by people: it emerges from reinforcement learning on questions with checkable answers, with no human preference data at all.

    Their wordsAnd these reasoning behaviors emerge naturally. So these things like, "Wait, let me see. Wait, let me check this. Oh, that might be a mistake." And they emerge from only having questions and answers.
    spoken · official transcript hear it at 2:43:31 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 23rd of 44 in this recording

    Interesting

  25. Our reading

    The faster AI progresses, the better NVIDIA does, so a shock like DeepSeek should expand its market rather than shrink it.

    Their wordsThe more progress that AI makes or the higher the derivative of AI progress is, especially because NVIDIA's in the best place, the higher the derivative is, the sooner the market's going to be bigger and expanding and NVIDIA's the only one that does everything reliably right now.
    spoken · official transcript hear it at 3:18:53 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 27th of 44 in this recording

    semiconductors

    Interesting

  27. Our reading

    Distilling from another lab's model is standard industry practice, not cheating; who you distill from just depends on whether you have products to protect.

    Their wordsDistillation is standard practice in industry. Whether or not, if you're at a closed lab where you care about terms of service and IP closely, you distill from your own models.
    spoken · official transcript hear it at 3:26:04 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 29th of 44 in this recording

    Interesting

  29. Our reading

    OpenAI's claim that DeepSeek trained on its outputs is narrative management: other startups bootstrapped exactly that way and were never banned for it.

    Their wordsI think that they're trying to shift the narrative. They're trying to protect themselves. We saw this years ago when ByteDance was actually banned from some OpenAI APIs for training on outputs. There's other AI startups that most people, if you're in the AI culture, were like they just told us they trained on OpenAI outputs and they never got banned.
    spoken · official transcript hear it at 3:30:29 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 30th of 44 in this recording

    OpenAIstartups

    Interesting

  31. Our reading

    Stealing AI code and data is hard and stealing ideas is easy, because Silicon Valley already moves ideas legally, inside the heads of the people it hires.

    Their wordsCode and data is hard, but ideas is easy. Silicon Valley operates on the way that top employees get bought out by other companies for a pay raise, and a large reason why these companies do this is to bring ideas with them.
    spoken · official transcript hear it at 3:33:55 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 32nd of 44 in this recording

    Interesting

  33. Our reading

    The biggest untapped business in AI is advertising inside model outputs, and nobody has yet worked out technically how to place an ad there.

    Their wordsThe short-term that company that could make the most money is the one that figures out what advertising targeting method works for language model generations.
    spoken · official transcript hear it at 4:19:37 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 36th of 44 in this recording

    Interesting

  35. Our reading

    Reasoning will generalize the way instruction tuning did: add enough verifiable domains and, at some point nobody can yet locate, the rest start working on their own.

    Their wordsAnd the history of NLP and language processing instruction, tuning and tasks per language model used to be like one language model did one task, and then in the instruction tuning literature, there's this point where you start adding more and more tasks together where it just starts to generalize to every task. And we don't know where on this curve we are.
    spoken · official transcript hear it at 4:29:46 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 37th of 44 in this recording

    Interesting

  37. Our reading

    AI will not push software engineers off a cliff; it will flatten their growth curve, the way Meta's stories flattened Snapchat.

    Their wordsThe big picture is that I don't think it's going to be a cliff. I think a really good example of how growth changes is when Meta added stories. So Snapchat was on an exponential, they added stories, it flatlined.
    spoken · official transcript hear it at 4:33:47 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 40th of 44 in this recording

    Interesting

  39. Our reading

    Humans keep their place in AI as judges rather than authors, because telling which of two answers is better is far easier than writing a good one.

    Their wordsAnd humans are actually very good at reading or judging between two things versus... This goes back to the core of what RLHF and preference tuning is that it's hard to generate a good answer for a lot of problems, but it's easy to see which one is better.
    spoken · official transcript hear it at 4:35:32 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 41st of 44 in this recording

    Interesting

  41. Our reading

    Open-source AI is still an ideological mission rather than an ecosystem, because it has none of the feedback loops that make open-source software compound.

    Their wordsuntil there are feedback loops of open source AI, it seems like mostly an ideological mission. People like Mark Zuckerberg, which is like America needs this and I agree with him, but in the time where the motivation ideologically is high, we need to capitalize and build this ecosystem around, what benefits do you get from seeing the language model data?
    spoken · official transcript hear it at 4:46:12 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 42nd of 44 in this recording

    open sourceAmerica

    Interesting

  43. Our reading

    AI takeover is not the thing to worry about: physical constraints bound recursive self-improvement, and humans reliably act once a risk becomes immediate.

    Their wordsAnd for that reason, there's physical constraints to things like AGI, like recursive improvement to kill us all type stuff. For the physical reasons and for how humans have figured things out before, I'm not too worried about AI takeover.
    spoken · official transcript hear it at 5:02:49 · all korrents from this recording

    ↗DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459youtube.com 43rd of 44 in this recording

    AGIrecursive self-improvement

    Interesting