The Polylog AI Intelligence Brief

Morning Edition · Tuesday, September 8, 2026

Tech

OpenAI's Chief Scientist Calls for a Coordinated Slowdown in Frontier Model Development

Jakub Pachocki argues that no lab's safeguards can support full speed scaling for much longer, and he wants voluntary commitments turned into mandatory thresholds enforced by outside auditors or governments.

2 sources
Corroborated

Tech

Claude Produced a Machine-Checked Proof of Fermat's Last Theorem in 13 Million Lines of Lean

Anthropic says a team of Claude agents worked largely autonomously for 11 days and proved 29,500 intermediate theorems, a formalization more than five times the size of Lean's main mathematics library.

2 sources
Corroborated

Tech

Anthropic's Claude Fable 5.1 Doubles Its Score on an Agentic Science Benchmark and Cuts Cache Reads 75%

The model reports 52.6% on Terminal-Bench-Science against 24.7% for its predecessor, while the sticker price stays at $10 and $50 per million tokens and cached input drops to $0.25.

2 sources

Tech

OpenAI Rates Astra "Critical" for Cyber Capability and Restricts Access

The model scored a perfect result on ExploitBench and found two zero-day vulnerabilities on its own, the first time OpenAI has placed a system in the top risk tier of its Preparedness Framework.

2 sources
Corroborated

Tech

Unitree Shows Humanoid Robots Fighting Under a World Model With No Human Pilot

UnifoLM-X2-1.0 replaces the teleoperator in the company's boxing demonstration, running real-time prediction and planning instead of scripted motion or remote control.

1 source
Plausible

Tech

OpenBMB Releases MiniCPM5-2B and Opens Its Training Data and Reinforcement Learning Stack

The 2.52-billion-parameter model averages 53.9 across 34 benchmarks and ships under Apache 2.0 alongside the pretraining corpora, supervised datasets and reinforcement learning samples used to build it.

1 source
Plausible

Tech

Anthropic Opens a Research Preview of a Standard for AI Agents to Operate Laboratory and Factory Equipment

Early testers report a laser stabilization improvement from 58% to 99.3% on QuEra's quantum computers and an imaging experiment compressed from weeks to a day at Janelia.

2 sources
Plausible

Tech

Token Prices Keep Falling and Consumption Keeps Rising Faster Than the Cuts

OpenRouter data shows agentic token volume climbing from 0.51 trillion to 7.3 trillion since February, with Chinese open-source models now accounting for more than half of the routed traffic.

1 source
Corroborated

Tech

Berkeley Lab Runs Meta's Open Vision Models on 300 GPUs to Cut Beamline Analysis From a Month to Minutes

Segment Anything 3 and DINOv3 now handle image segmentation for X-ray and neutron science under the Department of Energy's Genesis Mission, replacing work that took a month of expert annotation per time step.

2 sources

Tech

Anthropic Watermarks Claude's Text Output by Biasing the Sampler, Not by Inserting Characters

The method builds on Google DeepMind's SynthID-Text and applies to models launched on or after 2 August, driven by transparency commitments Anthropic signed under European Union rules in July.

2 sources

Tech

Meta Sells Frontier Model Access Through Its Own API at a Fraction of Anthropic's Rate

Muse Spark 1.1 carries a one-million-token context window and lists at $1.25 per million input tokens and $4.25 per million output tokens on the Meta Model API.

2 sources

World

OpenAI Funds an AI Adoption Program for Ukrainian Newsrooms Operating Under Wartime Conditions

Ten news organizations receive hands-on implementation support and API credits through a joint program with the World Association of News Publishers and a Ukrainian publishers' association.

2 sources
Corroborated