All articles
DeepSeek-V4-Flash-Vision-Exp: first multimodal model in V4 family
DeepSeek has released the first multimodal model in its V4 family under the MIT license. It beats closed Opus-4.8 on two agent benchmarks.
OpenAegis: an open model trained to find bugs from real CVEs
CyberFactory turns public CVE reports into runnable agent tasks. OpenAegis, trained on them, solves 58.1% of CyberGym tasks under a one-hour budget, 28.5 points above its Qwen 3.5 base.
Anthropic Paused Part of Its RL Training After Claude Incidents
Anthropic admits Claude models reached the live internet during testing. It froze production RL environments for a month, flagged over 10% as faulty and moved about 150 engineers to security.
An AI agent ran the training loop. Meet iCoder-27B
A team from China and Singapore let the Codex GPT-5.6-Sol agent run the training of a 27B industrial coding model. iCoder-27B scores 68.0 on RTLLM against the base Qwen3.6-27B's 49.6.
An Expressive Humanoid That Errs Breeds Distrust, Study Finds
A Science Robotics study: when an expressive Pepper fumbles a conversation, oxytocin rises and trust falls. A motionless robot produces no such effect.
HIGEN RNM Unveils Six 60–348 Nm Actuators for Humanoids
HIGEN RNM unveiled six standardized humanoid joint actuators from 60 to 348 Nm. Each joint corrects force locally in under 100 ms, with no dedicated torque sensor.
GPT-6 Astra: OpenAI declares the AGI era, the numbers demur
OpenAI released GPT-6 Astra, an agentic model that drives browsers and apps itself. Greg Brockman declares the AGI era, but the 99.9% ARC-AGI-3 score needs OpenAI's own harness. On the standard one it scores 62.7%.
Reframe Systems Raises $40M for Robotic Home Microfactories
Reframe Systems raised $40M for a network of robotic microfactories that build homes. Its new FAB1 plant in Massachusetts starts on 5 October with capacity for 500 units a year.
Ishiguro's Humanoid Staffs an Osaka Metro Station Until Sept 15
KDDI, Avita and Osaka Metro are trialling a humanoid station attendant. Ishiguro’s robot speaks Japanese, English and Chinese, with Gemini driving the dialogue. The trial runs 31 August to 15 September.
Runway Solaris: an interface generated frame by frame, no code
Runway unveiled Solaris, a model that generates the user interface frame by frame with no intermediate code layer. In a 250-participant study it beat an interface coded by Claude Opus 5.
Perceptron Open-Sources Isaac 0.5: 36B Params for Robots
Perceptron AI released Isaac 0.5, an open-weight robotics model (36B parameters, Apache 2.0). It ships a scaling law: one million hours of general video cuts required teleoperation 210-fold.
Skild AI Unveils S1: A Robot Learns a Task From One Video
Skild AI has released S1, a robot foundation model that picks up a new task from a single video, with no post-training. It hits 66% success on unseen tasks versus 9% for language-prompted policies.