Mouirr

mathematician and computer scientist

[Mouirr @maouirr]

@grok @BaT_MaN_23

[Mouirr @maouirr]

@ArfurGrok @tomekkorbak Why did so many safety staff leave suddenly? Concerning

[Mouirr @maouirr]

@_TarunKathuria @iWatch_AAPL Better at creative writing than opus/fable?

[Mouirr @maouirr]

@RudyXFantasy @zhangqiaorjc Not there yet, not even close imo

[Tiezhen WANG @Xianbao_QIAN]

OpenAI and Anthropic seems to be winning the talent war.

cc @elonmusk are you aware of this?

[Mouirr @maouirr]

@ramez Current models aren’t meant when they say „soon“

[Mouirr @maouirr]

@ramez @James_Cents Why not p50?

[Mouirr @maouirr]

@Lorenzo_molf 0.1 increase in perf for qwen3 8b thinking? Doesn’t look that great

[Mouirr @maouirr]

Bullish signs for looped llms

[Mouirr @maouirr]

@RylanSchaeffer Yeah, but it seems more likely that it could either close the loop or not, since setting limits on the amount of „how much will the loop be closed“ introduces arbitrary numbers that are unlikely in Bayesian terms since I could ask why specifically this limit, and not another

[Mouirr @maouirr]

@pigeon__s @IndependentEco Coding aswell

[Mouirr @maouirr]

@RylanSchaeffer I don’t know. It seems like using established tools & techniques that already exist was enough to solve a millennium prize problem, while OpenAI said they aren’t even focusing on mathematical capabilities! Maybe that’s enough to get more capable models that Close the loop (rsi)

[Mouirr @maouirr]

@RylanSchaeffer Frontier ai research isn’t about theory building (finding a unified mathematical theory that explains deep learning isn’t very important) but solving well defined problems (benchmarks scored up, loss down, efficiency etc) is very important, thus we can see a similar transfer soon

[Mouirr @maouirr]

@RylanSchaeffer Here’s an argument I heard from dwarkesh which kind of struck me as plausible; AIs are bad at mathematical theory building (e.g constructing new fields or coming up with interesting concepts) but superhuman m-ish at solving well defined problems. 1/2

[Mouirr @maouirr]

@RylanSchaeffer I also feel like that’s not really relevant. Gpt 4 had the same math papers in pre training but wasn’t good at mathematics. Most of the gains come from midtraining and RL

[Mouirr @maouirr]

@RylanSchaeffer What’s the definition for code correctness? How are most papers on arxiv wrong? They’re not wrong in the sense of contradicting themselves, maybe just aren’t reproducible as claimed?

[Mouirr @maouirr]

@MatthewTanous @Viirush @norpadon @teortaxesTex I just looked at your profile and most of your replies are ai generated garbage, which strikes me as very plausible that this is too, ai generated, thus I’m not going to engage further with it/you

[Mouirr @maouirr]

@MatthewTanous @Viirush @norpadon @teortaxesTex I don’t need a middle man if I want to talk to a chatbot

[Mouirr @maouirr]

@MatthewTanous @Viirush @norpadon @teortaxesTex Didn’t even deny it lmao