Originally from April 2024, forgot to post here apparently - but seems relevant as we head closer to the eschaton.
These models can operate as computers, and given enough memory, they can compute any computable program. The model is shown tapes for a few rules, and generalizes the process of running all remaining rules on random inputs without error, over thousands of attempts.
Stephen Wolfram said via email: “I assume ‘unbounded computation’ is happening through the next token iteration process.”
The reaction to this was generally pretty negative, so I’m pleasantly surprised that the top comments here were mostly positive.
It’s important to understand there are two floors of working spaces and a private office that’s usually open. It’s not just the pod, and that’s just for sleeping.
This is the lowest-cost way for me to be in SF, and it’s been an insanely net positive trade so far. I was in central IL trying to build a startup before this.
And to be honest, it’s pretty cool waking up and knowing my friends will be downstairs, and I can talk to them about whatever I’m working on that day etc. I like it way more than I thought I would, it’s beyond “bearable.”
It’s not luxury living by any means but it’s not this awful thing that people are taking it as.
These models can operate as computers, and given enough memory, they can compute any computable program. The model is shown tapes for a few rules, and generalizes the process of running all remaining rules on random inputs without error, over thousands of attempts.
Stephen Wolfram said via email: “I assume ‘unbounded computation’ is happening through the next token iteration process.”
Logs (deleted due to GitHub retention policy):
https://drive.google.com/file/d/12ZM2DYUysIvcpyJG5ufBWYT77tb...
It was also, in a separate experiment, able to do multiplication for up to 16-digit products using binary.
Repo: https://github.com/SpellcraftAI/turing