Amit Sahai says understanding AI's new mathematics will take more mathematicians, not fewer

At the top of the Hacker News front page, with about three hundred sixty points: a guest post on Terence Tao's blog, where Amit Sahai argues that the machines will leave humanity needing more mathematicians.

The systems Sahai works with already produce genuinely new ideas, he says, not just fast arguments. So every mathematician is about to learn what it feels like to be unable to keep up.

His example is a one terawatt fusion plant, designed by an AI on principles people have never put into practice. Approving it means understanding containment, shutdown and materials. And because the design is new, experience with earlier plants can't answer them.

What he wants funded is research groups, each spending a term or a year understanding one set of machine-made ideas. Human understanding runs at a human pace, which is why it takes more people rather than fewer. He calls it a deployable intellectual reserve.

Commenter pyridines takes the fusion plant down to the desk. They used to read every line the model wrote and usually found something; they catch fewer problems now, and they ask whether the ten thousandth fusion plant gets the same trust.

Conversations is an XMPP messenger for Android, and since 2014 its Play Store sales have paid Daniel Gultsch's rent. Google's fifteen percent cut comes to more than a thousand euros a year. One update has waited fourteen days for review. Grants now fund the work through 2029, so the app is free.

Other developers describe the same queue from the other side. Commenter k1w1 says their company has spent a year failing Google's support-number check, because verification wants a text or a human and their line answers with a menu. Plan mode is dead, says the founder of a planning app.

Ayman Nadeem built a coding app around planning, where a conversation hardened into a spec. Users had little appetite for that spec, and machine-written prose is punishing to read. Planning has moved inside the build loop instead.

Commenter bcherny, who works on Claude Code, says plan mode there is a reminder appended to each message. Switching the tools instead would break the prompt cache, and cost users money. Fifteen years ago, Apple let you send a letterpress card from your iPhone.

Lex Friedman of Lex on Tech heard it from the man who ran the printing: it began with Steve Jobs on a walk after dinner. It took restored eighteen-fifties letterpresses, and cotton paper the ink refused. Apple wanted hundreds of thousands of cards ready. The first day's orders nearly fit in a glovebox.

Commenter solfox co-founded Sincerely, which was mailing printed cards from an iPhone before Apple was. Their account is that the keynote mostly brought them awareness, and that when Apple retired Cards, it came as no shock.

Floci emulates four clouds on your laptop, as MIT-licensed binaries that start in about twenty four milliseconds and need no cloud account. Its AWS emulator answers on the port LocalStack uses, with about a hundred twenty services, and Lambda, RDS and Redis run for real underneath.

Commenter microflash has been running integration tests against it with Testcontainers for a while, and calls it much more lightweight than LocalStack.