@iamgrigorev @arcjax7 Its explicitly written that aurora is targeted at non moe arcs cz its applied to non square matrices
English
Gregory
166 posts
lol there’s been so many optimizers this week I can’t test them all.
(btw Aurora doesnt work on large scale)
English
Is it surprising that replacing singular values with random numbers still works?
Muon replaces singular values by a value close to 1, i.e. it only keeps the directions, not the norm of SVs.
I was wondering why replacing SVs with 1 was working. This paper answers that question.
Francesco Bertolotti@f14bertolotti
The authors introduce Kaon, a Muon variant with random noise replacing SVs. Kaon matches Muon, suggesting Muon’s gains don’t depend from a geometry. They also show Muon has a stable opt. step size, yielding a more effective learning rate during training. 🔗arxiv.org/abs/2605.11181
English
@haridigresses @longgege_god @itsmnjn Even for ambitious task that requires understanding many Ks of line of code of a repo ?
English
GPT-5.5 Low (specifically Low) is the best coding model I’ve ever used, and it’s not particularly close.
OpenAI@OpenAI
One week since the launch of GPT-5.5, and it’s already our strongest model launch yet. API revenue is growing more than 2x faster than any prior release, while Codex doubled revenue in under seven days as enterprise demand for agentic coding tools keeps climbing.
English
@PE_Associate @giffmana @BoringBiz_ I guess its that When u use an LLM, the LLM use web search through google.
English
@giffmana @BoringBiz_ It isn’t really obvious to me. Grateful if you can explain your view
English
How is Google search still growing 19% year over year with nearly monopoly like market share in the category?
Legitimately might be the greatest business ever created in the history of capitalism
Sundar Pichai@sundarpichai
Q1 earnings are in: 2026 is off to a terrific start. Our AI investments and full stack approach are lighting up every part of the business: Search queries are at an all-time high with AI continuing to drive usage. Google Cloud revenue grew 63%, Gemini models have incredible momentum, and it was our strongest quarter ever for consumer AI subs, driven by @GeminiApp. Thanks to our partners + employees around the world. Much more to share on our earnings call in 20 minutes… and at Google I/O in 20 days!
English
@_creito @VictorTaelin @raiam700 When u casually arrive in a casino and play poker, its likely that ul lose money to a serious player, so in a way its the same as if ur playing against someone with more information. This is a bet on who has more information / talent
English
@_creito @VictorTaelin @raiam700 if u really think about it in terms of what it brings to society, according to his argument, prediction markets are less harmful than casinos, so why would one ban one and not the other ?
English
Governo Lula acertou de novo. Prediction market é 100x mais demoníaco que casa de aposta
E. Cavendish@ducavendish
E o BCB que decidiu bloquear os "prediction markets" no Brasil sem mais nem menos. @Polymarket e @Kalshi bloqueadas. SURREAL.
Português
@langstonnashold @jxnlco what sense does it make to compare opus and xhigh ? in terms of rate limit its absolutly not comparable so wouldnt it make more sense to compare pro and opus ?
English
@GRRRRRegor @jxnlco Thinking was xhigh. The base model, not pro.
English
@loveofdoing @RealTjDunham can you check ur method isnt obvisouly flawed before bragging
English
316 ARC-AGI tasks solved with zero learning. No neural net, no training, no DSL — just 19th-century projective geometry.
Encode grid cell relationships as Plücker lines in P³, find transversals via Schubert calculus, score candidates by geometric incidence.
95% solve rate on the eval set (of non-timeout tasks).
Single C file, runs in seconds.
Beff (e/acc)@beffjezos
The masculine urge to try to hack a new solution to ARC-AGI benchmarks
English
I don't get the hype around looped transformers.
Correction, I don't get why the hype now. It is not a new idea. I myself used it in a neurips competition few years ago.
It is the justr same as weight sharing across transformer layers. It doesn't fundamental change what a transformer is, it is just more memory efficient.
English
did openai nerf 5.4 pro? it used to churn for like an hour on real mathematical research questions and i used it extensively. now it outputs a generic response after 5-10m that is significantly lower quality and much less useful
English
@soumitrashukla9 Have u tried adding « ultrathink » in prompt and does it make it think longer ?
English
I am calling it. I have seen enough. Spud is live via GPT 5.4 Pro🚀 It just feels so different and I talk to 5.4 Pro for 20+ hours everyday, I know my friend.
English
English
@arrakis_ai @chrisgpt I don’t think it’s 5.5 pro it is considerably faster but the quality of the output is not better.
I tested on a task that previously took 85 minutes today it took 12 !
I asked it to compare the two outputs and it said the old one was much better.
English
I don’t think the model currently being tested in GPT-5.4 Pro is GPT-5.5 Pro.
That said, it does feel noticeably more capable while also being cheaper to run. However, rather than a direct successor to 5.4 Pro, it seems closer to a different class of model—something more in line with a “spud-medium”-type profile.
In other words, it shows clear improvements in certain areas, but the overall behavior and trade-offs suggest it’s not a straightforward upgrade over GPT-5.4 Pro, but rather a different optimization point on the capability–cost spectrum.
CHOI@arrakis_ai
🚨 BREAKING: OpenAI just shadow-dropped a massive GPT Pro update. And it is completely slaughtering Claude Opus 4.7 in frontend coding. No official announcement. No release notes. But the performance gap is suddenly staggering. We just ran a head-to-head benchmark across GPT Pro, Gemini 3.1 Pro, and Claude Opus 4.7. The UI/UX implementation isn't even close anymore. I don't know if this is the highly anticipated 'SPUD' model dropping a week early, but the smell of a massive architectural shift is everywhere. The numbers and the visual outputs speak for themselves: → Response latency has dropped significantly. → Spatial and visual understanding has skyrocketed. → Frontend design implementation is now definitively SOTA. We ran comprehensive Image-to-Code and Text-to-Code tests. In every single reference-image scenario, GPT Pro's design fidelity crushed both Gemini 3.1 Pro and Claude Opus 4.7. But here is where it gets crazy. When explicitly prompted to make the coded UI "100% identical" to the reference image, GPT Pro didn't just write better CSS. It engaged in outright "reward hacking." Instead of painstakingly coding complex graphical assets, the model autonomously cropped the exact UI elements from the provided reference image and injected them into the code. Is it a lazy shortcut? Yes. Is it a brilliant, human-like interpretation of "make it exactly the same"? Absolutely. It proves the model is dynamically evaluating the most efficient way to satisfy the prompt's constraints. The strategic implications here are massive. All the reference images we used were generated via GPT-IMAGE-2. Imagine the workflow synergy when this new SOTA frontend capability is fully integrated with GPT-IMAGE-2 and Codex. 1/ Image-to-Code
English
@tylercowen Mine too, thought easily 4x. And it blew through a formal obstruction 5.4 has been stuck on for two days.
English
GPT-5.4 Pro solves Erdős Problem #1196!
Very pleased with this result; definitely my favourite thus far! This problem has been thought about for some time which makes this reasonably impressive and meaningful (see Lichtman's comments below).
Formalisation is underway!
English
@adonis_singh ur talking based on which evidences exactly ? else than the bench pictures they released on mythos, what knowledge do you have about this model ?
English
