New goal: learn the original language.

I would guess that pretraining is still accounts for most of the total optimization on modern AI agents, even if it's only ~7% of total compute. But would probably be a different story at ~1% of total compute.

Good example of market prices being set for the people who want to buy something, not for the people who don't want to buy it.

Few appreciate how lucky we got with the timing of certain technologies. DVDs took over the market right as VHS tapes were becoming obsolete. Imagine if DVDs got delayed a few years. It would've been death to the movie industry!

This is a pretty cool derivation of the constant-sum-of-distances-to-foci definition of an ellipse from the conic section definition of an ellipse. I think you could understand this in about 2 minutes and probably never forget it. proofwiki.org/wiki/Dandelin%…

I've been developing this project recently. medium.com/xofy/polis-polylo… It's a bit like building Mooncrash from Almost Nowhere. Let me know if you want to invest.

Maybe we'll have super-universities in the future. Like, the first 4 years could cover something comparable to the year 1 material of 5-8 different majors. Declaring a major would come after that, and then there'd be another 4 or 5 years of study in that major.

If you stepped on the exact equator, would you melt? Or explode?

I think in 2026, when you talk to AI, the AI won't steal your idea and use it to get rich. And a human might. But that's actually a big reason why human conversations are more value-producing.

Religion idea: you surrender some agency so optimizers can differentiate through you.

I think we're close to the creation of minds that are built, end-to-end, not trained.

Oops.

Never for sale: worn shoes, baby.