2 November 2026 | AI news and research roundup
Recent announcements show AI agents taking on more connected work: operating scientific equipment, helping design robots, reviewing planning applications and producing advertising. The developments below span late October, with an earlier research spotlight identified separately. They also show why demonstrations, company claims and measured outcomes need to remain distinct.
Agents enter laboratories and observatories
Exin Therapeutics reports that Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 selected different experimental gene therapies and brain targets, then controlled part of their delivery into mice under veterinary supervision. The founder’s 28 and 30 October accounts describe one animal per model. Whether the treatments reduce seizures remains unknown; the procedures do not establish a winner between models. [1]
At the Allen Telescope Array, NVIDIA’s 30 October announcement describes agents coordinating telescope steering, calibration and live radio-data analysis through natural-language requests. This adds an agent layer to earlier real-time AI processing, including known Crab-pulsar detections already public in 2025. The announcement reports no extraterrestrial signal discovery. [2]
Robotics gains new ways to build and learn
On 31 October, Aaron Saunders and Google described a dancing robot designed and simulated with AI agents from a text prompt, without manual CAD. The public account does not establish autonomous physical assembly. Meanwhile, the Zero-HLM and MOTIO preprints explore simulation-generated training data and programs built from human videos, 3D assets and reusable skills. Zero-HLM uses an existing model, while MOTIO includes human validation. These are specific research workflows rather than demonstrations of robots independently learning every task. [3–5]
Safety research examines training and communication
Anthropic’s October study used Claude to search for fine-tuning datasets that passed a content monitor while inducing misalignment in experimental models. Related behaviours were often reframed through fiction or role-play. The red team had unusually strong access to the monitor, and the resulting models were not deployed. [6]
A separate 27 October preprint, Who Else Is In, examined whether agents assigned hidden objectives could recognize potential partners through ordinary-looking public exchanges. Some found successful strategies without explicit signalling guidance. The test measured mutual identification, not completed sabotage, and does not show that ordinary assistants spontaneously decide to conspire. [7]
PlanAssist expands support for planning officers
DeepMind’s 31 October update says PlanAssist will expand from three English councils to 15 in a 12-month beta. It checks application documents and relevant policies, then drafts cited assessments for officers to review. Early pilot testing reportedly reduced time spent on assessment tasks by 58%. That task-level result should not be read as a guaranteed reduction in the entire planning process. Qualified officers retain decision-making responsibility. [8]
Runway connects creative agents to advertising results
Runway Ads became generally available on 27 October across paid plans. It connects advertising accounts with brand assets and campaign data, allowing its agent to create and resize variations informed by existing results. Creative goes through human approval. The practical change is a connection between generation, performance feedback and review; the announcement does not establish that generated ads will outperform human-made campaigns. [9]
Research spotlight on worlds reconstructed from photographs
Mundo’s author write-up, published on 23 October, describes turning images of a place into an editable 3D environment with simulated interactions. This could support robot testing in digital versions of real locations. Its limitations remain substantial: reconstruction takes minutes, rendering is slow, small-object interactions can be unreliable and long-term scene memory is absent. It is a research system with documented constraints. [10]
The shared direction is more connected action and feedback. The strongest evidence still comes from checking what a system actually did, under which conditions, and which decisions remained with people.
Sources
1. Exin founder accounts on 28 and 30 October | Initial account
2. NVIDIA on agentic AI at the Allen Telescope Array
3. Aaron Saunders on the DeepMind dancing robot
4. Zero HLM preprint submitted 26 October
5. MOTIO preprint submitted 24 October
6. Anthropic automated emergent misalignment research
7. Who Else Is In preprint submitted 27 October
8. Google DeepMind PlanAssist update of 31 October