korrents

korrents · video

How Open Models Win Against OpenAI and Anthropic

Nathan Lambert · 16m · youtube.com ↗

12 korrents from this recording

Nathan Lambert did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

  1. 3:32 · watch on youtube.com↗

    But, today, these popular families that are these smaller models, which are used heavily by researchers, are really the foundation of so much innovation that is happening in the world.
  2. 1 min later
  3. 4:38 · watch on youtube.com↗

    open models will cross any given capability threshold, whether it's in 4 months, 6 months, 8 months. This is still all very fast in terms of diffusing technology, and it's a matter of when, not if, things can happen, and we can prepare for this when more than we could prepare for anything that is if.
  4. 1 min later
  5. 5:44 · watch on youtube.com↗

    And this narrative of open models being cheaper is really not always the case for at least in today, when open models are just at a lower performance threshold.
  6. 2 min later
  7. 7:21 · watch on youtube.com↗

    What what we are competing with if somebody wants the open models to win is you're competing with OpenAI and Anthropic owning the whole stack. And they're vertically integrating the products, the models, the compute, the chips, the data centers through the whole thing in order to minimize the cost of serve and increase their margins.
  8. 2 min later
  9. 9:19 · watch on youtube.com↗

    So we're going to see this pressure for the people building the models to move downstream and this competition within the ecosystem to all the companies are trying to do the same thing. The money is on inference
  10. 9:45 · watch on youtube.com↗

    So there's intense competition within our coalition of open models that makes things difficult to see the clear path forward in the economic scaling of these labs. So open models are economically fragmented
  11. 1 min later
  12. 10:54 · watch on youtube.com↗

    And if the open models are to keep pace, we have to figure out the economics to pay this money back to the labs to even get close to the ballpark of these costs. Otherwise, there will become this kind of growing capability gap between open and closed models.
  13. 11:15 · watch on youtube.com↗

    So the overall task for open models is not to just build the cheapest models at a given level of intelligence, but to build the cheapest, most adaptable, private intelligence.
  14. 2 min later
  15. 12:55 · watch on youtube.com↗

    And the open models are only going to win when we're willing to drive these prices down to really compete with the frontier labs and figure out where to make money whether in volume or specialization or other enterprise deals later on.
  16. 1 min later
  17. 13:39 · watch on youtube.com↗

    I think the the frontier labs are not going to be willing to deploy their models on prem for enterprises
  18. 1 min later
  19. 14:57 · watch on youtube.com↗

    where the cost savings have been shown to be real where on these narrow agentic tasks, open models fine-tuned on them well really can work.
  20. 1 min later
  21. 16:07 · watch on youtube.com↗

    So this is what I'm trying to do and trying to figure out how to do research that enables this economic future because I think this will be safer, it'll be more accessible to many people