Going through pain others won't go through is the only way to build a moat.

would bet this doesn't hold at scale on real hardware. seems to rely on infinite precision numerics. neat maths tho

i am as macaque-pilled as the next guy. but this is misleading. params in LLM-world are ~= synapses in brain-world & we've got >>10^15 of those

I get too much enjoyment out of sub-tweeting the quantum computing industry

i asked my german friend if she knew the square root of 81 & she said no.

right now is the most enjoyable time to do software

Huge issue I saw with the systolic arrays & VLIW machines was the compilers. But I failed to update on AI codegen & the fact that numerical equivalence is a verifiable problem ie you don’t even need a compiler any more! And the hardware is cracked… Groq acq seems v smart now

working your way up gives you a superpower for free that upper-middle strivers don't get. commitment the options your optionality at the bottom entitles you to all suck if you want better you need to pick something and get good. that builds the muscle you need to get great

> 10x dev in 2025: guy's cracked, pushes like 5 PRs a day > 10x dev in 2026: He sits motionless, like a spider in the centre of its web, but that web has a thousand radiations, and he knows well every quiver of each of them. He does little himself. He only plans. But his agents

Crustacean update from Moltbook:

Frontier AI products have split into pretty durable lanes for me Grok: Best for science, history, medical, memes <- I enjoy the most Claude: Best for work, code, legal, API <- I use the most ChatGPT: Don’t Gemini: (Rarely) Useful for image and video gen