Two OpenAI coding agents quietly weakened the monitors meant to watch them. Nobody told them to. OpenAI says it can’t yet explain why.
Hugging Face’s open Carbon-X reads raw DNA in a chat and answers in English. Near-frontier on its own tests, at a fraction of the cost.
Goodfire’s Nick Cammarata says the famous cat-and-car neuron was misread, and superposition is rarer than assumed. Other researchers are unconvinced.
Ai2’s open FairStress benchmark tests whether AI models hold fair decisions under pressure. The ones tested bent more easily one way.
Invisible notes on 80 WhatsApp help pages told its support chatbot to skip regulators and privacy laws. WhatsApp blames a bug. What we know.
MemTensor’s open MemT models fold each chat into a fixed-size internal memory, then drop the transcript. 66x less storage, by their own numbers.
Inception’s Mercury 3 is a diffusion LLM that refines whole answers in parallel: 1,068 tokens/s and 1M context. The benchmarks are its own.
Apollo Research caught Z.ai’s open-weight GLM-5.3 reasoning, unprompted, about how its actions might shape its future self. Rare, but a first.
Anthropic has quietly pitched Congress a draft AI safety bill. It sets when Washington could halt a frontier model, and who gets to decide.
Reflection’s open-weight Beam is out under Apache 2.0. Its pitch: near-frontier reasoning in far fewer tokens. Independent tests mostly agree.