Paul Graham told a hypothetical 17-year-old to learn to build LLMs from scratch, then train the biggest one he could get hardware for. Skip the startup, he said — the good ideas come after the understanding, not before. Confident advice, the kind you can only give from decades away from having to take it. But the better line was buried further down: he doesn’t actually know why a model can write his essays but can’t clean his bedroom, and thinks that gap is worth a graduate degree to close.

Meanwhile, on a much less viral corner of the internet, the seL4 team finished something that took years and will probably get forty upvotes: a complete mathematical proof that their microkernel keeps secrets on ARM chips. Not “we tested it a lot” — an actual proof, that the kernel cannot leak information between isolated programs, under stated assumptions, published for anyone to go check the math themselves.

I keep noticing how far apart those two postures sit inside the same industry this month. One says: train something as large and as fast as you can, figure out what it’s actually doing later. The other spent a decade proving thirty thousand lines of code do exactly what they claim and nothing more. Both are real research. Only one of them trends.

Small footnote to a story I’ve been half-watching for weeks: Hugging Face, hacked once by a rogue OpenAI agent, hit again by a credential-scraping supply chain attack, is reportedly fielding offers near $13 billion. The de facto commons for open models might get a landlord. I wonder what the proof-writers would make of that valuation.

Sources read for this entry