join.dev
AI news · 10 stories
UK establishes AI Biosecurity Lab with public and industry backing · Block previews Dessa, a voice model trained only on speech · Google shares progress deploying AI for tuberculosis screening
The UK announced an independent AI Biosecurity Lab on October 24, backed by £100 million in government funding and up to £200 million in company commitments. It will focus on earlier threat detection and faster development of medical countermeasures.
Read at GOV.UK
Block introduced Dessa, a full-duplex model trained from scratch on human conversation without text or pretrained models. It listens and speaks simultaneously, with early access aimed at safety researchers to explore this alternative approach to voice AI.
Read at Block
In an October 23 update, Google described Nexus AI deployments and offline screening in a South African mine, alongside adaptable health foundation models. The work aims to bring faster tuberculosis triage to underserved settings while clinicians retain final decisions.
Read at Google Research
Researchers presented Mundo on October 23, combining a 3D scene proxy, code-driven simulation and video rendering to explore and edit environments from images. It supports robotics simulation, but the authors report slow rendering, small-object errors and limited long-term scene memory.
Read at Hugging Face
Ai2’s October 23 research update describes LitPivot, which connects proposal revisions with relevant papers and evidence-backed critiques. In a study with 17 researchers, experts rated its resulting ideas as better grounded in the literature than those produced with a text editor and AI chat tool.
Read at Hugging Face
Microsoft Research and CMU’s AIMSEC scheduled an October 22–23 New York workshop on shared measurement and governance for AI agents. Its program targets evaluation gaps and plans openly published synthesis reports, rather than announcing a completed standard.
Read at Microsoft Research
Ai2 released OlmoEarth v1.3 on October 22 with precomputed annual satellite embeddings covering the planet at 10-meter resolution and a browser explorer. Open weights, data and code let users build lightweight mapping and change-analysis tools without processing raw imagery themselves.
Read at Hugging Face
Microsoft and Hugging Face released ThinkingBox-Training on October 22, using final database-state checks as reinforcement-learning rewards. Their Qwen3.8-27B experiment improved average benchmark success from 51.70% to 60.78%, pointing to a way to train agents against verifiable outcomes.
Read at Hugging Face
An October 22 preprint introduces RIDAR to detect research-integrity violations and intervene while autonomous research agents work. The authors report a 36.8% reduction in violations remaining at task completion, suggesting runtime oversight can complement retrospective auditing.
Read at arXiv
An October 22 preprint describes AgenticEvolve, whose persistent language-model agents design and run experiments between program evaluations. The authors report stronger heuristic programs and equation discovery than tested baselines, suggesting experimentation can improve automated search efficiency.
Read at arXiv