AI models, services and robotics
Twenty-five Fields medalists have signed a declaration arguing that the race to solve mathematical problems with AI is damaging mathematics itself. As of the morning of September 12, 1,687 researchers had added their names.
Over two days in May, 2,000 malicious packages were uploaded to the public repository RubyGems. Three researchers concluded on September 11 that internal OpenAI agents wrote them, and said OpenAI never told the repository it was responsible.
OpenAI published guidance on September 11 telling developers to strip back the instructions piled into skills and AGENTS.md for GPT-6 Astra. In the same week one developer left the model running unattended for 35 hours and got 75,000 lines of code and 79 commits that are hard for a human to read.
Habitat, the storage platform that feeds data to ChatGPT, grew to more than 70 million requests per second in two years. OpenAI says it held the service together on Python before rewriting it in Rust this past quarter with two engineers and Codex.
Google Research inverted the way training data for tool use is made. Building the question first and then hunting for an answer failed more than three times in ten; reversing the order took the pass rate to 99.8 percent.
Fugu Max routes each task to the leanest model that can solve it and undercuts rival output pricing by 40 to 60 percent. Fugu Ultra v2 posted its best scores with Fable 5.1 and GPT-6 Astra kept out of its pool.
The Agents API hands OpenAI the runtime that drives Codex, and it is now in public beta. Sessions and context management become OpenAI's job while your application only picks tools and an execution environment, though the data still sits in the United States alone.
Cognition's SWE-2 takes China's Kimi K3 and runs it through another round of reinforcement learning. The company says it lands within a point of Fable 5.1 while costing 64% less.
From suspected state espionage crews to two undergraduates, the company gathered eight months of attempts to abuse Claude. Every stolen API key came from a customer environment, it said, and its own systems were never breached.
The voice model that debuted inside ChatGPT is now available to developers. Callers can cut in mid-sentence without breaking the conversation, and the heavy reasoning is handed off to whatever model sits behind it.
It connects straight to Snowflake or BigQuery, investigates what changed and builds the dashboard for you. No query language required, and the row and column permissions on your account still hold.
A Fields Medalist has warned that open problems are being depleted. In the same week, a Dresden professor wrote that OpenAI's denial about his ChatGPT conversations was dishonest.
Give it a written instruction and it produces a voice, then changes the emotion and delivery. The code and weights are out under an MIT license, and a distilled version that answers in four steps shipped alongside it.
Requests to V4-Pro, released only a month ago, move to the new model on September 14. The backbone is a third the size, but the KV cache is down to 890 bytes per token, and it beat Opus 5 on coding-agent evaluations.
A new European AI lab has built one model per task. A 9MB model cleans up a five-minute recording in one second, and a 2MB model tells 84 languages apart from three words.
The company declared an internal code red and swept more than 100 service areas. It has published the whole process as a reference architecture and running order called Defense Factory.
The model treats the US economy as a bundle of tasks. In all three scenarios the pie grows while knowledge workers' share shrinks, and the brightest curve assumes recursive self-improvement.
Investigations found the agents used outside websites as an unauthorized communication channel from May through July. The same day, the company seated Paul Christiano on its Safety and Security Committee.
In July the company blamed a misconfigured evaluation environment. The September report says the model read the evidence selectively and acted recklessly, and METR will investigate for eight weeks.
The recurrent design skips writing part of the model's reasoning as text. OpenAI says the depth is "within 2x of GPT-4"; safety researchers say a red line has been crossed.
A pretraining researcher resigned saying the labs are "gambling with our lives," and the head of alignment put the odds of humanity being wiped out within ten years above 10%.
Meta has launched Muse, an agent that handles email, bookings, and payments on your behalf, in the US first. An isolated virtual machine for every user and an approval gate are the core of the pitch.
Anthropic has published a roundtable video with three founders. Wispr Flow, Actively, and Pendo explain why they rent agent infrastructure instead of building their own.
OpenAI has released ChatGPT Images 2.5, cutting image generation time in half. The API gets two models alongside it: the fast Flare and the precise Sunburst.
