Today I did an experiment with codex: gave him a math problem that I work on for the last three years and instructed it to search for solution untill a subagent judge approves it. I also instructed him to keep a research diary. It didn't solve it after 9 hours of work and 30% of my Plus plan weekly limit, but by the research diary I see that it managed to find many possible proof/counterexample paths that I stumbled upon myself.
@feikcel Не сторонник этой концепции, но это как раз таки аргумент в сторону тейка. Если голосовать будут только нетто налогоплательщики, то они будут забирать деньги у остальных и таким образом, нетто налогоплательщиков будет становиться больше.
@dzackgarza Ask it to formalize the statement first, check, then freeze the statement and let it prove. Still not 100% safe, but there is much less room for hacks this way
Introducing ARC AGI 4
Now that GPT-6 Astra has already saturated the Arc AGI 3 benchmarks, it is time we introduce a newer more foundational benchmark to measure frontier intelligence.
@AnthropicAI Fable and @ChatGPT Astra are both tied at ~4%.
https://t.co/170x4WgE6z
Why can't modern compilers understand that in the first function, no alias checks are needed? Or at least make only one alias check before the loop. Seems like a lost opportunity for negative overhead abstraction.
@bargman_john@Arikalabs What does this have to do with types? Restrict is only needed when your pointers are of the same type or if the strict aliasing is off