korrents

korrents · video

Open Language Models: Overview, Open Source History, Distillation, Risks, & Chinese Labs

Nathan Lambert · 48m · youtube.com ↗

19 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. 11:14 · watch on youtube.com↗

    I think a lot of venture capitalists in today's era are acting in kind of odd ways because they're not invested in the companies that are making the biggest returns, and they are pushing business strategies onto the ecosystem before they are proven to support their portfolio.
  2. 11:41 · watch on youtube.com↗

    I think that open source is foundational to scientific progress and understanding and I you can see more of my writing on this, but I think this scientific progress and understanding is going to be core to safety
  3. 1 min later
  4. 12:51 · watch on youtube.com↗

    big tech companies in the American industry and American society benefit on the 5 to 10 to 20 year timeline when we do innovation in the open because it keeps moving technology forward.
  5. 3 min later
  6. 15:21 · watch on youtube.com↗

    And a lot of companies may start with this, or they may say it because it is a good business recruiting pitch, but in reality, the strategy is why there are so many open models.
  7. 15:30 · watch on youtube.com↗

    And any company that is not at the frontier, like OpenAI and Anthropic, is incentivized to open-source their models because it'll gain them adoption. It'll gain them mindshare. It'll gain them talent.
  8. 1 min later
  9. 16:02 · watch on youtube.com↗

    open models have a different business benefit than closed models, which I think is a lot more about ownership of the technology than about cost savings.
  10. 1 min later
  11. 17:10 · watch on youtube.com↗

    And I think this is the best point in why open models will be adopted is they allow specialization and they allow integration of domain expertise and they allow control of the technology.
  12. 5 min later
  13. 21:46 · watch on youtube.com↗

    they have such a financial incentive to show that they are keeping up. So even if the best intention of the researchers there is to be fair, they are going to figure out ways to make themselves look a bit better on the benchmarks for economic value and for the narrative of the company.
  14. 6 min later
  15. 28:11 · watch on youtube.com↗

    And I propose that we need a different word for this because the Chinese labs are literally jailbreaking and manipulating the APIs of OpenAI and Anthropic in order to get this reasoning content.
  16. 2 min later
  17. 30:20 · watch on youtube.com↗

    So, my read is that they're taking these first two points, scale and sophistication, as too much of a correlate. And it is more likely that the Chinese labs are doing this because this is easier than getting more compute.
  18. 30:38 · watch on youtube.com↗

    is that people underestimate China's ability to build technology cuz building language models is not magic. It's a lot of hard work over time
  19. 1 min later
  20. 31:33 · watch on youtube.com↗

    It's like how do you ban a software? If you start banning software, you're putting up great firewall around the US, and that's going to reduce innovation and kind of global interaction
  21. 1 min later
  22. 32:16 · watch on youtube.com↗

    But we shouldn't have run to the government and ask for laws until you secure your own projects.
  23. 10 min later
  24. 42:33 · watch on youtube.com↗

    banning open weight models does not does not really minimize the risk of future harms and kneecaps economic activity and kneecaps innovation.
  25. 42:47 · watch on youtube.com↗

    the bad actors say a North Korean hacker will still use open weight models. It is just American companies that will not use a open weight model out of regulatory uncertainty.
  26. 2 min later
  27. 44:49 · watch on youtube.com↗

    And the frontier models being a bit ahead in these companies that have time to do research on risks on closed models before the open weight models come in 5 to 8 months is a good balance.
  28. 45:13 · watch on youtube.com↗

    And I expect this to be a 2-to-5-year transition where there are new risks and new things that society has to overcome.
  29. 1 min later
  30. 45:53 · watch on youtube.com↗

    These are one of the biggest misreads on the Chinese labs is that the Chinese labs have a risk posture. And they don't want to cause harm. But they're operating in a different ecosystem
  31. 2 min later
  32. 47:37 · watch on youtube.com↗

    They enable us to use this scientific process, which has been so great for society overcoming challenges in the past, and we should be very wary to step back from openness out of fear rather than real measured risks that we're seeing.