Forget it Jake, its AI2027 town

Yes please

Take that 17 square-packers

Everything is downstream of Mondo 2000

Hot take: Anthropic’s safety obsession accidentally made the better enterprise product. Teaching chatbots to fastidiously follow rules, respect constraints, and defer to human judgment? Turns out corporations love that.

Yes because of the Markov Chain stationary distribution, you'll get to the end with probability 1. Lift it to an "expander graph" (adding something auxiliary like momentum) to get fast mixing. Then take away the actual randomness with a suitable pseudorandom number generator

leaving us with the eternal question in the line just before this, "Why does Rice play Texas?"

Claude thought I was pulling an April fools joke when I told it about the Artemis launch!

What's going to be the Barron's 'Burning Up' moment for AI?

No amount of testing will ever prove your software is secure. People are waking up to this. Formal verification isn’t academic anymore, it’s inevitable.

This is close to where people landed at a Bay Area LEAN meetup the other week. The room was mostly CS/verification people, not AI-for-math, and even they had no real counter-thesis. Just a kind of resigned recognition.

What the world needs is a Matt Levine but for AI. Someone who could make kv cache optimization and chain of thought reasoning genuinely funny while being precisely correct but also to make it read like a gossip column

Raises the question: Are there tax filings which are Turing uncomputable? Are there tax systems which are Godel incomplete?

The entire Dwarkesh/Jensen interview should be viewed through the lens that Jensen's primary goal is to commoditize his complements.

Dwarkesh didn't connect with Jensen's argument because he couldn't break from the nuclear arms race framing where every minute matters. Jensen's framing is AI is just regular technology. Then says medium-term there's no chip chokepoint, and restrictions just push China