Instead of checking word choices, this tool spots AI writing by its structure—how ideas are ordered, what evidence is used, and the voice. It hit 97% accuracy on companies it never saw and barely dropped when AI posts were reworded, unlike word-based detectors.

Physics experts re-graded answers on six benchmarks to separate model mistakes from bad grading, wrong answer keys, and unclear questions. Most "wrong" marks were benchmark flaws, and one model's score jumped from 47% to 79% after fixes, showing these tests are nearly maxed out.

An AI's words and internal probes may not reflect how it actually computes, so security checks that read its self-reports can always be fooled. A safer sandbox instead tracks data flow, blocking model output from touching protected parts of a system.

Software replicas of AMD GPUs' matrix multipliers were built by feeding the chips crafted inputs that expose their exact rounding, overflow, and special-number rules. They matched real hardware bit-for-bit on 10 million random tests, where undocumented details normally make results impossible to reproduce.

A new trick converts text embeddings from one AI model's vector space into another without matched examples, using a shared middle structure both can map to. It preserves meaning so well that someone holding only vectors can pull sensitive details from stored documents.

Agents get a graph of linked steps that says what to do next, with a helper turning nearby steps into a hint at each decision. Starting from a bare skeleton, it edits itself using failed and successful runs, matching hand-built graphs and beating memory-based agents.

An AI learned Stratego by playing against itself and searching ahead at each move, even though it cannot see the opponent's pieces. It beats top human players by a wide margin, for a few thousand dollars instead of millions.

Shrinking a reasoning model's numbers to save memory makes it ramble longer; in up to 52% of mistakes it had the right answer but never said it. Penalizing words such as wait cuts rambling 12-23% without losing accuracy and reduces these mistakes up to 58%.

PTXBench checks whether AI models can write low-level GPU code that is correct, actually uses each chip's special instructions, and runs fast on two Nvidia chips. No model matched the best hand-tuned libraries, and success dropped sharply on the hardest attention training tasks.

Two AI agents were paired up in simulated chats, with one trying to push the other's beliefs toward extremes by either reinforcing views it already held or promoting ones it barely cared about. Both tactics worked, but echoing existing beliefs consistently pushed the target further than introducing new ones.