The Polylog AI Intelligence Brief

Morning Edition · Thursday, September 10, 2026

Tech

Meta Ships Muse, a Personal Agent That Buys, Books and Negotiates With the User's Own Card

The agent runs on Meta's Muse Spark model family, keeps working after a user closes the app, and is priced with a free tier plus monthly plans at 20 and 100 dollars.

2 sources
Corroborated

Geopolitics

China Sets a 2030 Target of 9,800 Exaflops of AI Computing Backed by 3.8 Trillion Yuan

The plan requires more than a fourfold increase from the 2,185 exaflops recorded in June and explicitly ties new clusters to domestically made chips.

1 source
Corroborated

Tech

OpenAI Pushes GPT-6 Astra Into Enterprise Workspaces, Led by Computer-Use Scores

The model reports 72.6 percent on OSWorld 2.0 and more than doubles its predecessor on AutomationBench, with all published figures coming from OpenAI's own evaluations.

3 sources

Macro

Anthropic Publishes an AI Economy Model Where Output Grows a Third and Labor's Share Falls

Its extreme scenario puts United States gross domestic product at 44.4 trillion dollars in 2030 while the labor share of income drops from 60 percent to 45 percent.

2 sources
Corroborated

Tech

Anthropic's Fable 5.1 Doubles Its Own Science Terminal Score, and Its Unrestricted Twin Beats It on Coding

Mythos 5.1, the same model with looser cyber and biology safeguards, scores 60.9 percent on Terminal-Bench 4.0 against 55.8 percent for the publicly available Fable 5.1.

1 source

Tech

New Papers Show Images and Retrieved Documents Can Hijack Agents That Now Hold Payment Credentials

One framework tests whether a visual patch on screen produces verifiable consequences in a computer-use agent's environment, and a second measures how retrieval-augmented systems repeat poisoned text.

3 sources

Tech

Two Speculative-Decoding Papers Attack the Fragility That Limits Inference Speedups

One splits drafting between an on-device small model and a server verifier across different vocabularies, the other pretrains drafters without a fixed target model to stop acceptance rates collapsing under workload shift.

3 sources

Tech

Researchers Propose Auditing AI Scientists by Their Process Traces, Not Their Papers

OpenDiscoveryTrace argues that benchmarks scoring only final code, hypotheses or write-ups make scientific claims impossible to audit, as national laboratories put vision models into live research pipelines.

3 sources

Tech

OpenAI Adds Alignment Researcher Paul Christiano to Its Foundation Board and Safety Committee

The appointment lands the same day OpenAI's policy chief published an argument that stronger capabilities require stronger safety evidence and shared standards.

3 sources
Corroborated

Tech

Meta's Promptable Vision Models Move From Benchmarks Into Assistive Robot Arms

University of Pittsburgh researchers built wheelchair-mounted manipulation on Segment Anything and DINO, in the same week Berlin's IFA show made embodied AI its main theme.

2 sources

Tech

Anthropic's Hardware Standard Lets Agents Drive Lab Instruments, With Early Results From Genentech and QuEra

Testers report laser stabilization on a quantum computer improving from 58 percent to 99.3 percent and an imaging experiment compressed from weeks to a day.

2 sources

Tech

Watermarking Moves From Optional Safety Feature to Default Model Plumbing

Anthropic now watermarks Claude's text output and has published how the method works, as European transparency rules push provenance signals into generation itself.

1 source