You remember this shit? Opus 5.5 xhigh take - took almost an hour opus-55-balls.netlify.app The "Escape" mode which Opus came up completely un-prompted is a good example of why I think Opus is the first model with "Taste" and the knowledge of "why something is nice". The idea to make a game out of it, which you don't even play, which still is pleasing to watch. yeah. milestone.

Did NVIDIA just "solve" RL? So what does it do, in the extremely reduced TL;DR ELI5 version: A lot of current LLM RL methods generate multiple attempts for the same prompt, so-called "rollouts", and use their rewards relative to each other to figure out which behavior should be reinforced. Other approaches use a learned critic or value model to estimate how good an action or trajectory was. Both approaches add considerable compute and/or synchronization overhead. FlashREINFORCE instead uses just one rollout per prompt and no learned critic. It takes a batch of rollouts from different prompts, compares each reward against the batch's mean reward, and gets a positive or negative learning signal from that. So, extremely reductively: one attempt per problem, no critic, still enough signal to learn. That's kinda fucking wild.

Today a new declaration, "Math and AI", has been issued, initially signed by 25 Fields medallists and currently signed by more than 1000 mathematicians. The text points to many phenomena induced by the potential development of AI-driven mathematical activities, including AI slop generation, poor knowledge integration, lack of community-building aspects, capital concentration, and more. These issues are identified, but I am missing in this text the most important part: a clear list of countermeasures. Who is going to fight, and how, for the budgets to build a "CERN for AI"? How should we rebuild the education of students? Is it really possible for AI companies to let mathematics dry out of open problems, etc.? I would like to hear from those declaring mathematicians what their long-term vision of mathematics is, provided that the technology will stay with us, might not be equally distributed, and perhaps the standard view of the field is going to change forever. We should design damage control, develop bold new ideas about the purpose of human mathematical activity, and embrace the possibility that we might no longer be single-handedly the most intelligent entities in this world. It is a humbling perspective and a disruptive view, and perhaps a difficult reality in which nothing is given. We need to fight for every single bit of human intellectual dignity and seek new ways of enjoying, curating, and developing the cognitive process of mathematical exploration. It is time to abandon some of the old ways, brace for the impact, and build something anew. We will not stop this tectonic shift, we need to reshape our perspective on our capabilities and find new directions of development, possibly inventing completely new skills complementary to what AI can possibly do. It is a new intellectual age of discovery, and by stagnating we risk the gradual erosion of the field as we know it.

the math community is going through the whole denial, grief and whatever stages software went through when opus 4.5 was released :D quite funny to see like <insert 'your first time' meme> funny

Even the smartest human on Earth is missing the mark here, so I will explain it in simple terms: In 1-2 years, there won't be a "field" anymore. There won't be "promising research directions" to share because the bot will already have thought through whatever our human minds could come up with. Did you all sleep through the last five years?

this is how AGI looks like lol

per aspera ad astra, mah dudes and dudettes!

nice. Don't tell Dario tho.

lol

gemini 3.8 flash is actually cracked. google redemption arc lol

Every doomer on the TL freaking out about looped transformers all of a sudden. - It's 1 article, until someone at OpenAI confirms it, it's hearsay - We HAVE looped transformer recipes already. Here, go play with one: huggingface.co/ByteDance/Our… - It's not really much different to creating a model that's just 2x or 3x as deep. There's research to indicate 3 passes is optimal, and after a little replication experiment a few weeks ago can confirm - "Muh Neuralese" - see previous point. Not every operation in a model does map to token embeddings (even with things like j-lens or r-lens). They ALREADY "think" internally with neuralese. Hence interpretability existing as a field of research. - It doesn't replace CoT at all, it's just more flops per token Stop freaking yourselves out

People are circlejerking over Fable 5.1’s token-efficiency gains as if they’re all API users. NEWSFLASH: You won’t feel any of it on your sub. You still can’t 24/7 Fable like you can Sol on the big sub.

My body and my Mac are ready - Qwen3.8-Flash-Next (125B A5B) is coming in 24h modelscope.cn/models/Qwen/Qw…

How do DevDays look in the future? Some nerds standing in front of a screen watching codex do its thing for 6hours?

OpenAI DevDay 2026 will be our best DevDay in the history of the company. It will not be close.

Jokes on you, I don't understand shit anyway after a certain amount of vibing.

She looked at a technology already entering the patient-specific drug-generation loop and decided the real tragedy would be humans no longer getting enough personal fulfillment from solving the problems themselves. Disgusting.

As you can clearly see in this chart the paretos of this world are moving #deepseek #luna

pack it up guys, loop engineering already solved: make "done" fail-closed. a boundary the model can't talk past, with evidence on the other side. Slipway is basically that, local and git-native. github.com/signalridge/slipw…