Can an LM, starting from random init (!!), learn to generate all of its pretraining data? Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities. A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.

Gave Opus 5.5 donald's prompt, Midjourney, and a moodboard 12 hours later, woke up to this:

"Hey Opus, make a 100th anniversary video for @SpaceX. Here's some API keys, knock yourself out."

How the fuck did they train Opus 5.5

Claude Opus 5.5 writes a fugue in the style of Bach. I called Astra’s fugue the most impressive musical feat I’d ever seen from an LLM, and this is just as good. The circle-of-fifths sequence beginning at 1:13 is beautiful. The interplay of voices is natural and satisfying. LLMs are getting to the point where their music is good enough to listen to for pure enjoyment.

'An average of 72 percent of respondents said they felt curious, happy or excited about A.I., compared with 41 percent who felt worried, sad or angry.' @nytimes's Damien Cave uses @Gallup data to remind us the rest of the world digs AI. (Annoyingly, NYT doesn't link to Gallup directly, link below.) nytimes.com/2026/09/23/world… news.gallup.com/poll/714593/…

An opposing camp of mathematicians should put out an open letter asking OpenAI to just release all of their math solutions People are going to work each day to tackle problems already solved—a great waste, for those who care about math progress over status hierarchy and prestige

Claude Opus 5.5 will be the first Opus meant to fall back to a less capable model for "a small set of capabilities related to the development of frontier LLMs, such as kernel development.."

We’re working with an independent advisory group of mathematicians to help OpenAI responsibly share advances in AI and mathematics. The group will advise on how we assess and communicate new mathematical results, uphold academic and professional standards, and build tools that support mathematical research and learning. Through this work, we want mathematicians to be at the center of shaping how AI supports mathematical understanding and how its benefits reach the wider community. openai.com/index/advisory-gr…

Things are heating up on Terry Tao’s blog. In “If Math Is More Than Proof, We Need to Better Celebrate the Rest of It,” Grant Sanderson of @3blue1brown proposes “open exposition problems” -- rewarding the work of making math genuinely understandable. terrytao.wordpress.com/2026/…

3/4 papers I reviewed and 2/3 reviews I got this ARR cycle were AI slop. Fraudsters are already outnumbering honest researchers. Unless drastic measures are urgently adopted, conference publications will have zero value or even become negative signal, and academia will be toast.

If it's isn't obvious already, it is the end of papers as a measure of productivity, expertise or accomplishment. The odds of publishing in "glam" conferences & journals also very quickly approaches random & the time for gate keeping review increasing exponentially. 1/