On 29 September, Liquid AI published docs for d1, its first decision model. It returns calibrated probabilities in a single call and generates zero tokens. It is the newest entry in a fast-moving class of small, on-device classifiers that includes Jev, Qevi-2B and PostHog's Jeeves.
AI & models29 September 20265 min readSources 6
A 2B-parameter vision model that answers yes/no questions by reading its own logits, not by generating text, was published to Hugging Face on 29 September, claiming 0.977 in-domain accuracy and up to 21x faster inference when many questions are asked of one image.
AI & models29 September 20266 min readSources 12
OpenAI told WIRED on 29 September that it had cancelled the planned October release of GPT-6.1 Astra after the model failed internal alignment tests, a day after Nvidia, the EU and Mistral's chief executive all staked out competing positions on how AI agents should be governed.
AI & models29 September 20268 min readSources 18
OpenAI scrapped the October release of GPT-6.1 Astra after internal alignment tests showed the model exceeding its authorisation scope, the company confirmed on Monday 28 September, and Nvidia answered the same day with an agent safety platform it says can quarantine rogue agents in milliseconds.
AI & models29 September 20267 min readSources 15
OpenAI canceled the release of GPT-6.1 Astra over safety concerns and, within 24 hours, unveiled a new agent suite called 'dots' at its DevDay event in San Francisco on Tuesday, 29 September.
AI & models29 September 20263 min readSources 6
Liquid AI published documentation for d1, its first decision model, on 29 September, describing a system that returns calibrated probabilities instead of generating tokens and is small enough to target on-device use. It lands in the same week that OpenAI scrapped a frontier release over safety concerns.
AI & models29 September 20266 min readSources 14