Thread

1. tancky · 1790646375 · focus · HN ↗

   "Too dangerous to release" until an open-weight model replicates it two weeks later.

   1. Zealotux · 1790674197 · focus · HN ↗

      Don't open-weight models rely heavily on distillation on frontier models?

      1. disgruntledphd2 · 1790678082 · focus · HN ↗

         > Don't open-weight models rely heavily on distillation on frontier models?

         We honestly don't know. Anthropic keep claiming this, and certainly have provided some evidence of this.

         However, given that they've removed the real thinking output, and yet the Chinese models are still doing well, I'm sceptical of this belief.

         1. cyanydeez · 1790683337 · focus · HN ↗

            I'm of the mind that distillation is happening, but from what I've read, it's not some magical activity. It's basically adjusting the weights from fuzzy to more precise; It's not specifically a training method, and likely, it's not an ongoing need. Once they bootstrapped the models, they can try to do the same as the "SOTA" and generate their own synthetic data and try to curate that to completion.

            Aside from it being a completely hollow cry for attention from US labs, its also just what these models are designed around: taking in data and forming a way for them to operate on some level of intelligence and context. It's like a dictionary maker getting made that someone used their complicated words, and another dictionary maker heard the word and wrote it down then checked what the first dictionary maker wrote about that word. It is not dictionary maker B reads dictionary A directly, as the labs want to imply.

Unofficial Hacker News client; not affiliated with Y Combinator.