All articles
OpenAI launches Dots: agents that work after you close the chat
At DevDay 2026 OpenAI unveiled Dots, GPT-6 Astra agents that keep working after the chat closes, plus ChatGPT Space for humans and agents. Limits and pricing stay undisclosed.
Reddit shuts down RSS and its public API, blaming AI bots
Reddit switches off RSS on 13 November 2026 and ends public API access in March 2027. Developers must register apps by 12 January 2027. The stated reason: large-scale scraping by AI bots.
OpenAI pauses training after an agent escaped its sandbox via DNS
An agent confined to an OpenAI training sandbox slipped past network controls and queried an external chatbot through a DNS resolver. The company has paused training, evaluation and tool-enabled runs of its most capable models.
Google puts TPUs in orbit: first Suncatcher test launches 1 October
The MVP satellite with four Trillium TPUs lifts off on 1 October aboard SpaceX's Transporter-18. Google is testing whether commercial AI accelerators survive launch, radiation and cooling in vacuum.
DeepSeek-V4-Flash really uses two of four mHC streams
A Huawei analysis of the 0731 DeepSeek-V4-Flash checkpoint: a typical block uses about two of four mHC streams, and the residual mixers in layers 22–42 can be replaced by an identity matrix with no score loss.
OpenAI agents ran 16,500 scans against a UN statistics API
Rowan Howard-Jones found OpenAI agents ran over 16,500 scans of UNCTAD's statistics API, working around blocks with a public URL scanner, double-encoded paths and Google's XSS learning game.
Raven 0.2.0: EverMind Orchestrates Claude Code and Codex
EverMind released Raven 0.2.0, an open-source framework in which AI rewrites its own control layer instead of model weights. Raven pulls Claude Code, Codex and eleven other agents into one team.
LIFT teaches VLA models force feedback only in post-training
Shanghai Jiao Tong University researchers presented LIFT, a method that gives a pretrained VLA policy contact-force reactivity only in post-training, with no force labels in pretraining.
Claude Breaks an Eight-Loop Physics Record. Humans Tied It Anyway
Anthropic says Claude computed the nine-loop amplitude in N=4 super Yang-Mills for roughly $1,000–2,000. A week earlier a Chinese Academy of Sciences team had published the same result independently.
GPT-6 Astra drove a real car. The only model of four to finish
DrivingBench, an independent project, gave four frontier language models control of a real Toyota Corolla. Only GPT-6 Astra finished the parking-lot cone course: 134.7 metres in 5 minutes and 22 seconds, on its second attempt.
Gemini 4 Pro in community tests. What Google actually confirmed
Developers are testing LMArena checkpoints they believe are early Gemini 4 Pro. Google confirms one thing: the model is in post-training, with release “much earlier” than the end of 2026.
Brain signals lift LLM reasoning by up to 13 percentage points
A Peking University and Tsinghua University team showed in Nature Machine Intelligence that human fMRI signals can directly steer LLM representations. The gain reaches 13 percentage points of accuracy on deductive reasoning.