HFT Bread

Be a kid again and dream bigger.

Our world models are action conditioned, and hence causal. The concept of world model for planning goes back to the 1950s in optimal control (before I was born). I didn't just discover it. But training action-conditioned world models from sensory inputs (like video) requires new techniques.

This guy is literally at the state of the art of this field, he lives in bumfuck Texas and no one knows him. It’s fucking hilarious.

Simple. 1. US students are not interested in solving hard problems, so they don’t go to grad school. 2. The US military needs hard problems solved, so they pour money into universities to solve them. 3. Chinese students take those positions, and are paid to learn from the best in the world and work on the most important problems. 4. Upon graduation, we KICK ALL OF THEM OUT back to their home country, after paying between $250k-$500k to train them. 5. Those students go on to innovate for their home country. In engineering, 60-90% of graduate students working on a thesis are non-citizens, paid for by US grants. China is able to out-compete domestic innovators because we are training them to be superior and then forcing them to compete with us. It’s not a mystery, and yes it is insane. Source: am a professor in a “top” research program in the US.

OK I’m actually mad about this now. Diffusion models are a perfect glimmering jewels of applied stochastic calculus and these clowns are ruining them by turning them into trillion parameter overfit transformer slopdels. Fuck off.

When I say AMD is dogshit at software: this is what I mean. Tensor cores, used for AI, are simpler than GPU and CPU and AMD is perfectly capable of making competitive ones. But they are useless without proper software support. And not only AMD is incapable of delivering on this front, but they are even unable to take advantage of people doing the legwork for them. Perfect exemple of wasting billions to save millions. That being said, there is so much upside. If they are willing, they can 10x their market cap in a few years just by hiring a competent leader for their software division, and, core importantly, getting out of the way so he can get things done.

Huge if true