Norman Casagrande (personal)Feb 23
dragosr

I gave LLM agents an Ultra9 285K and no token budgets, then asked them to optimize prime number sieves. When they started, finding and counting all the primes representable in 64 bit integers (216,289,611,853,439,384) was going to take ~4,800,000 years.

Four days later, now down to 202 seconds. They went from optimizing seconds, to miliseconds, now nanoseconds, and still going. Follow along at:

https://github.com/secwest/fast-prime

GitHub - secwest/fast-prime

Contribute to secwest/fast-prime development by creating an account on GitHub.

GitHub
Feb 21 at 9:04pmWeb
Show threadAdam Shostack  Feb 21
@dragosr Have they invented anything or just gotten better at accessing algorithms in their training?
Show threaddragosrFeb 21
@adamshostack They are brute forcing, profiling, testing EVERY algorithmic and architectural optimization possibility they can find on the internet. The robots are relentless. And still going.
Show threaddragosrFeb 21
@adamshostack the neat part is that they are also re-evaluating previous optimizations at each step, sometimes some optimizations lead to new paths. They find lots of dead ends and failures. The detailed optimization logs for each algorithm in that repository document the whole journey.
Show threaddragosrFeb 21
@adamshostack As I typed this, they found another path that gets it down to 134 seconds. They'll check it in soon.
Show threaddragosrFeb 21
@adamshostack now 129 seconds....