The Polylog AI Intelligence Brief

Morning Edition · Monday, August 10, 2026

Tech

Anthropic Makes Automatic Approvals the Default in Claude Code From August 14

The company says its own controlled test found human reviewers caught a planted dangerous command 13.6 percent of the time, against 89 percent for the automated checker.

2 sources
Corroborated

Tech

Databricks Says Model Routing Cuts Average AI Coding Task Cost by More Than 30 Percent

The company's broader claim of savings up to 90 percent bundles routing with open-model substitution and token trimming, not routing alone.

2 sources
Corroborated

Tech

Meta Prices Its Model API Below Frontier Rivals as Muse Image Reaches Second on Human-Preference Rankings

Muse Spark 1.1 carries a one-million-token context window and sells at $1.25 per million input tokens, putting Meta into the metered API business it avoided for years.

2 sources
Corroborated

Geopolitics

Palantir's Karp Urges Washington Not to Ban Chinese Open-Weight Models

His argument, that competition disciplines American labs better than protectionism, runs against Treasury Secretary Scott Bessent's concern that Chinese open models are built on United States laboratory technology.

2 sources
Corroborated

Tech

OpenAI's First Country-Level Usage Data Shows Fastest Growth Outside Its Established Markets

Peru, Uruguay and Costa Rica rose the most in the second-quarter rankings, and multimedia messages reached 7.8 percent of the global total after the April image model update.

2 sources

Tech

Meta's Open Vision Models Cut a Month of Beamline Annotation to Fifteen Minutes at Berkeley Lab

SAM 3 and DINOv3 run on 300 Nvidia A100 accelerators inside a Department of Energy science project, and the same pair runs on battery-powered hardware in a $41.5 million assistive robotics program.

2 sources

Tech

New Preprint Proposes Manufacturing Verified Vulnerabilities to Train Defensive Security Agents

CyberForge injects flaws into whole repositories with verification, addressing the shortage of realistic training data for agents meant to find and patch bugs rather than exploit them.

2 sources

Tech

Paper Finds Bundling Many Verdicts Into One Judge Call Degrades Oversight, Regardless of Compute

Splitting the work into separate calls, each returning fewer decisions, keeps verdicts grounded in evidence where added tokens and tools do not.

2 sources

Tech

A Batch of Mixture-of-Experts Papers Targets Routing as the Next Efficiency Lever

Three preprints published the same day attack expert routing from different angles: byte-level allocation, task-aware fine-tuning supervision, and interpreting what routed reward models actually reward.

3 sources

Tech

Watermarking Research Turns to Detection That Survives Across Different Model Hosts

A new preprint argues existing decoding-time schemes fail when the text moves between providers, which is the condition regulators are actually trying to cover.

2 sources

Tech

OpenAI Buys Presentation Startup NextSlide and Folds the Team Into ChatGPT

The deal closed earlier this year and was disclosed only this week, with terms undisclosed and the founder previously behind a checkout startup Instacart acquired in 2021.

2 sources

Tech

OpenAI's First Hardware Device Is Reported as a Displayless Speaker Priced Above $300

The battery-powered, hockey-puck-sized unit comes from Jony Ive's LoveFrom, follows OpenAI's $6.5 billion purchase of his hardware company io, and is reported for 2027.

2 sources
Plausible