Aryan Seth

@shhrreeyyy @utkarsh_2105 pls bhai

@eternalflummox @elainesim28 @FreezingToe this could bee him yes

@elainesim28 @FreezingToe this could be you someday

Please pray for my husband, he got stung by a bee in the forehead. He’s in the hospital now, his face all swollen and bruised. He almost died. Luckily I was close enough to hit the bee with a shovel.

@OpenAI You should celebrate by resetting rate limits

chatgpt plus codex limits feel so much lower than before

@willccbb but plain OPD has issues - mostly pointed out in this paper; arxiv.org/pdf/2604.03128 "privileged information" leads to an irreducible loss term, and i have not found much luck with training stability (maybe this is a personal issue lol)

got some longform claudeslop for yall TLDR picture book summary:

@FreezingToe banger

@AryanSeth07 “You’re absolutely right!”

sharp insights

calling Mythos a preview now feels ragebait-ey

tinker is nice i wish i could afford it consistently

@SpaceX @leerob @cursor_ai no words

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.

thanks to arsenal there will be no drought this year my tears shall fill every river

I quit football

@ppc0x23 @_lyraaaa_ what's a richelot step?

@_lyraaaa_ they all track the same underlying signal but the directions are different because each layer performs a Richelot step on the representation. the polynomial gets factored. the curve transforms. the class persists.

found a vector in gemma4 e4b that roughly correlates with perplexity the more surprisal in the passage, the stronger this vector fires R²=0.775, corr=0.880, p<10⁻³³²³⁶ this is avg across all layers, L20 and 21 have R² 0.828 and 0.845 alone wonder what i can do with this

count the false positives count the false positives count the false positives count the false positives count the false positives count the false positives

We got into @ycombinator! A few months ago, @onkar_borade_10, @SujaySriv , and I met at a football game in HSR, Bengaluru, and went deep on one question - Why does SaaS take months and a huge team to get delivered after the sale? Building a good product should be enough, right? Right? Messy integrations. Handoffs. Siloed information. Poor documentation. Fragmented data. The list goes on. We started a company with the vision to make SaaS self-serve. @lab0_ai Huge thanks to @dessaigne , @collinmathilde , and the YC team for this opportunity. If your product rollouts take months or you're a system integrator/partner/FDE implementing SaaS, let's talk. Link below.

Breaking news: Anthropic buys the All in podcast just to shut it down Dario quoted as saying: “this isn’t even about new media I just want to stop seeing them on my timeline”

@viplismism sure, harnessing works at inference, but I still don't follow how this solves sparse reward because that's a training-time problem; recursive call = tool call, so you can train using the subagent output, but the reward is still a scalar (higher due to harness), but sparse

@AryanSeth07 true in principle i guess, but in practice the harness has to do the heavy lifting man. models tend to jump to conclusions or lose context fast if the harness isn't driving the search process properly

most people don't realize that rlms are just solving the sparse reward problem for long context! instead of an llm hunting for checkmate in one giant forward pass, it's like you break it into bite-sized reasoning tasks. every recursive step is a checkpoint where the model updates its internal value of the context before moving to the next piece it turns a massive search space into a dense signal

there's going to be a movie at some point about hinton being like oppenheimer or something with the toronto lab being the ai manhattan project and sam altman and co being like the govt and scaling the models