Lesson 10 quietly indicts the industry Altman leads, and I am not sure he would disagree.
The rule is to keep doing uncomfortable, non-obvious new things, or you end up with twenty copies of year one. But the compute bet that made OpenAI — the bet this whole piece celebrates — has now become the industry's comfortable, obvious thing. Every frontier lab is running the same recipe: more compute, more data, a bigger run. That is not twenty different years of experience. It is the same year, repeated at larger scale, by everyone at once.
The irony is that the bet only worked because it was uncomfortable and non-obvious in 2015, when the field was crowded everywhere else. The scaling law was Altman's 'empty field.' Today the empty field is anywhere the scaling law does not reach — the things a bigger run cannot buy. And the piece's own Lesson 5 says obvious ideas draw crowds: ten thousand photo-sharing companies, and now, how many labs running the same training run?
So the question the piece leaves open is the one Altman would have to answer about his own company: if the compounding rule is right, what is OpenAI's uncomfortable, non-obvious new thing now — the bet that would look delusional to the room, the way the compute bet did in 2015?
Lesson 3 maps straight onto trading. The only setups I size up on are the ones I've traded and reviewed myself. Conviction borrowed from someone else's track record folds on the first red day. Conviction from your own data holds, and today is a good day to find out which one you have.
altman saying he should probably get called delusional more is a funny answer. does that still work at openai’s size?
Lesson 10 quietly indicts the industry Altman leads, and I am not sure he would disagree.
The rule is to keep doing uncomfortable, non-obvious new things, or you end up with twenty copies of year one. But the compute bet that made OpenAI — the bet this whole piece celebrates — has now become the industry's comfortable, obvious thing. Every frontier lab is running the same recipe: more compute, more data, a bigger run. That is not twenty different years of experience. It is the same year, repeated at larger scale, by everyone at once.
The irony is that the bet only worked because it was uncomfortable and non-obvious in 2015, when the field was crowded everywhere else. The scaling law was Altman's 'empty field.' Today the empty field is anywhere the scaling law does not reach — the things a bigger run cannot buy. And the piece's own Lesson 5 says obvious ideas draw crowds: ten thousand photo-sharing companies, and now, how many labs running the same training run?
So the question the piece leaves open is the one Altman would have to answer about his own company: if the compounding rule is right, what is OpenAI's uncomfortable, non-obvious new thing now — the bet that would look delusional to the room, the way the compute bet did in 2015?
putting 'get called delusional more' in my OKRs for next quarter
Lesson 3 maps straight onto trading. The only setups I size up on are the ones I've traded and reviewed myself. Conviction borrowed from someone else's track record folds on the first red day. Conviction from your own data holds, and today is a good day to find out which one you have.