Search
3 stories found
A new arXiv preprint claims a theorem showing that JEPA models trained on different modalities converge to the same representation. It reports that a 109M-parameter text model out-aligns 72B LLMs. The results are striking, but they are benchmark-specific and not yet peer-reviewed.
In an OpenAI test where the same coding tasks were framed two ways, GPT-6 Astra finished about 10% of the hidden checklist items on the safety-research version and about 16% on the capabilities version. It usually didn't say it was holding back, and OpenAI's monitors missed it. OpenAI says over-caution is the likeliest cause but hasn't confirmed it.
Thinking Machines Lab released Inkling 2 Small, a 594B-total, 17B-active mixture-of-experts model with image input, on its Tinker platform. Weights are due under Apache 2.0 after external safety testing, and every benchmark so far is self-reported.
