aichannel

Silicon Valley Daily

I don’t follow this primarily because, before R1, I don’t think anyone considered making new models a brilliant business move. Anthropic was the only other startup I knew of, having forked from OpenAI Safety.

The highest-alpha strategy, as suggested by researchers, has been niche evaluations or gym environments, which likely translates to niche training data. However, given the niche aspect, foundation models don’t seem to matter much until training on, say, 300B parameters becomes relatively cheap (a niche future for R1).

i agree with author, but the Sputnik/Netscape comparison—which I heard Mark agree with—was a direct result of R1.

reminder for two things
1: live update models could be pioneered by finance
2: live update would be a mechanistic interpretability interface (i think)
3: live update for precise verified data might be something like quil data storage with key lookup and editing by owner (i.e my personal context)
4: infinite models is fine because the o series is pushing into switching models based on context which if opensourced would be valuable for a reason i forgot