The way I usually spend my ChatGPT usage is I give tasks to a Sol instance, and when it struggles I upgrade it to an Astra. My feedback since 6.2 is, I don't know how Astra-6.2 compares to Astra-6 because Sol-6.2 has not been struggling on anything. 🤷♂️
Claude and ChatGPT are very bad at estimating how much time it will take to achieve a goal. They are much better at estimating how many pull requests it will take to achieve a goal. So just ask them about PRs when you are really curious about time.
Kind of depressing that the project to extend Caltrain from 4th and King an extra mile to 1st and Mission, which started in 2014, is currently projected to cost $8 billion and be ready in 2036. sfist.com/2026/11/13/caltrai…
I don’t think this will work perfectly today, because agents do miss obscure bugs while reimplementing. But once formal verification is cheap, rewrites will be cheap as well. Just verify that the new implementation is equivalent to the old one.
I'm convinced that at this point, you could have an Opus 5.5 agent re-implement all of CPython probably within two weeks without any human input, against just the tests, and probably with full compatibility with including ABI, no GIL, a jit and probably beat it on all measurable metrics. It's completely crazy.
"If I had to put probabilities on it, 50/50 oblivion. And then the other 50 percent — the wildest thing you can imagine."
Hans Moravec, on p(doom)
(via @slatestarcodex )
"Site Reliability Engineer" and "Forward Deployed Engineer" are similar. Both of them were invented to take an existing role, "sysadmin" or "sales engineer", and make a higher-paying, higher-status, more-softwarey version. I wonder what else could be transformed like that?
Imagine a programming language designed for maximal performance tuning, over simplicity. Explicit memory layout and threading, down to individual CPU instructions.
If the AI is writing it, why not?
Progress on P vs NP ! The dangerous sort: a faster algorithm for solving NP problems.
Hopefully someone out there has a plan for what to do if encryption doesn't work any more!
An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4
It's underrated how many bugs you can fix by enforcing that a program adhere to *any* spec, even a somewhat wrong one. It eliminates the whole category of "undefined behavior".