The state encodes the task information. On the sphere a token's only free quantity is its direction, so we track each cell's angle to its own input. On Sudoku, clues rotate toward their input, blanks end near orthogonal, and the final angle flags wrong answers. (5/9)

i have many thoughts on this - you can have a jev feedback loop to figure out when you're done - you can recurse on bboxes and get smaller and smaller - you could mask objects as you detect them and then ask for the next instance of object if you need to keep it sequential tldr: Box is a very interesting primitive for sys1 models

Mirror is a Rust agent core with a python TUI using an RPC. The same agent core can run headless behind an HTTP API just as easily as in your terminal. Every session is a single append only log, and everything else (the transcript, what the model actually sees, token usage, extension state) is just a fold over that log. Compaction appends a new context window with a summary and the recent turns, and a rewind is a fork that points back at the parent’s log up to where you rewound, so nothing ever gets copied or thrown away.

RLCF: reinforcement learning for canine feedback

Of all the buckets of manuscripts, I’m currently most interested in bucket 376 of 377, “Universal computation in forced Navier-Stokes flows”, building on the work of @evamirandag “Can a force that eventually stops changing continue to perform an arbitrarily long computation?”

somebody should make a language that's good for systems programming, fast compile times, good concurrency, pretty simple mostly just gets out of your way, easy to distribute cross arch, very unfortunate that we all _MUST_ use rust all the time, maybe some day

Can LMs discover & fix bugs if you don't tell them what went wrong? To proactively maintain a repo, agents need to find problems before users do. SWE-sweep benchmarks this on 100 repos, 22 languages, 4k real bugs (numpy, php interpreter, lean kernel). Top models get <5%.

did a single person predict that coded music videos would take over diffusion

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.

I made this with one prompt using Opus 5.5 I spoke to my computer for 5mins, claude worked for 12 hours, and I woke up to this full prompt:

I asked Opus 5.5 to explain camera focus by building an interactive lens lab Here's what it came up with after 1 hour 26 minutes in one shot, $25.66 API cost lens.lab.sael.net Move the focus ring and you can see the glass elements shift the sharp plane through the scene