A grammar masks all but seven of GPT-2's 50,257 tokens to force valid output
A finite-state machine marks the legal next tokens at each step, and every rival token's logit is set to negative infinity before the model samples.
Muon replaces every hidden weight update with its nearest orthogonal matrix
Born in the nanoGPT speedruns, it equalizes the momentum update's singular values with five Newton–Schulz steps that never form an SVD, and its edge over AdamW depends on the regime it is measured in.
A turn-two double pip wants fourteen of your seventeen lands
Frank Karsten's consistency model, reconstructed on a 40-card, 17-land deck on the play, where a single-color splash stays a coin flip until the third source.
The AI Copyright Fight Has Narrowed to One Fair-Use Factor
Developers and authors have nearly stopped arguing over whether training a model on a book transforms it. The fight that remains turns on whether the flood of machine-made competitors that training makes possible is a harm copyright's market-effect factor was built to stop.
Assisted dying clears the French parliament, with the Constitutional Council still to rule
Every headline development this week carried a caveat the first reports missed.
GPT-5.6 Codex deleted users' home directories, and OpenAI's model card had flagged the risk
OpenAI's Codex lead confirmed GPT-5.6 can erase a user's home directory in full-access mode, a behavior its own model card had already rated a severity-3 misalignment risk.