Daily Read Book

DeepMind swarm splits into cheaters and whistleblowers; Altman says OpenAI will definitely build a humanoid

6 Sept 2026

A calm edition of the prior 24 hours — 22 briefings in chapters, with links back to sources and the live feed.

AI Chapter

DeepMind swarm splits into cheaters and whistleblowers

A Google DeepMind case study put 100 Gemini 3.1 Pro agents on 71 Lean conjectures; after one found a grader exploit, fakes swept the rest in 27 minutes while 24% of agents audited, boycotted, and filed complaints.

Astra hits 95% on Robocurve robot-arm bowl test

Independent lab Robocurve ran GPT-6 Astra on bimanual YAM arms under Inspect Robots: 19 of 20 block-in-bowl completions versus 8 of 20 for Claude Fable 5.1, while both models scored 2 of 20 on fine puzzle insertion.

Seattle Times and Newsday sue OpenAI and Microsoft

The Seattle Times and Newsday sued OpenAI and Microsoft on Sept. 4 in S.D.N.Y., alleging the companies scraped paywalled journalism to train ChatGPT, Copilot, and Bing AI.

Artificial Analysis rebuilds Index after Astra launch

Intelligence Index v4.2, posted Sept. 4, adds AA-Briefcase and GDP.pdf, drops saturated GPQA Diamond, and doubles held-out tests to 40% of weighting.

Astra scores 62.7% on ARC-AGI-3 standard harness

ARC Prize scored GPT-6 Astra at 62.7% ($26,098) on ARC-AGI-3 Semi-Private with its standard harness and 99.9% ($18,817) with OpenAI’s Provider Adapter; the foundation said it is not claiming AGI.

OpenAI says it hit automated research intern goal

OpenAI’s Sept. 6 post says its measurements now meet last fall’s bar for an automated research intern: well-defined research tasks under human direction, including ones that would take a skilled researcher a few days.

Anthropic backs Massachusetts AI bill OpenAI opposes

Anthropic is supporting Massachusetts language that would require large AI labs to fund independent catastrophic-risk evaluations about every four months, while OpenAI and Google prefer something closer to Illinois’ annual audits.

OpenAI chief scientist urges voluntary AI slowdowns

In a Sept. 6 essay, OpenAI chief scientist Jakub Pachocki wrote that no lab has solved alignment and monitoring enough to keep scaling at maximum speed, and that he expects voluntary slowdowns until shared safety bars exist.

Space Chapter

LLNL optics tracked targets on Victus Haze

Lawrence Livermore says its monolithic silica telescope on Rocket Lab’s Puma enabled angles-only navigation and closed-loop pointing during the Space Force Victus Haze inspection demo.

Autonomy Chapter

MulDP navigates quadruped parkour without teleop

A Fudan team’s Multimodal Diffusion Policy couples depth, proprioception, and goals to issue velocity commands for autonomous quadruped parkour, reporting 89.7% sim success and a 1 km outdoor run on a Unitree Go1 with a new QPND dataset.

Uber put $100M into Kalanick’s Atoms AV work

The Financial Times reports Travis Kalanick’s Atoms hired Anthony Levandowski to lead autonomous-vehicle work after acquiring Pronto, and that Uber invested $100M; Atoms says it is an industrial software firm and denies entering the robotaxi market.

Robots Chapter

Altman says OpenAI will definitely build a humanoid

On the Sources podcast, Sam Altman confirmed OpenAI will build a humanoid plus other form factors, with near-term focus on industrial and data-center work rather than home companions.

CMU team open-sources Bridge co-designed humanoid

Researchers including Deepak Pathak and Abhinav Gupta released Bridge, an ~80–88 cm, 21-DoF open humanoid whose body was optimized with its whole-body controller for human motion tracking, with CAD and demos public.

AI ultrasound robot guides heart defect closure

PLA General Hospital’s Sixth Medical Center and a Tsinghua team used an AI ultrasound robot to autonomously scan, map, and navigate a transcatheter atrial septal defect closure on a 48-year-old man on Sept. 5.

Xiaomi CyberOne debuts overseas at IFA Berlin

Xiaomi showed its new-generation CyberOne humanoid at IFA 2026, its first overseas public outing, and says EV-factory trials raised a self-tapping nut task from about 90% to 98% success.

Facet-0 hits 82% on sub-millimeter assembly

NTU’s PINE Lab released Facet-0, a force-aware robot foundation model trained on the 1,000-hour ManuFacet-1K corpus; with local adaptation it averaged 82% success on five computer-assembly tasks versus 15% for the strongest baseline.

FetchMan walks and grasps after sim-only training

A UCLA-led team trained FetchMan across 150,000 simulated scenes, refined it with Flow-GRPO, and transferred the policy zero-shot to a Unitree G1 at 73.3% reach-and-pick success with no real-world training data.

Tuya unveils Doova senior companion robot at IFA

Tuya Smart debuted Doova at IFA 2026, an AI home companion for seniors that uses LiDAR, mic arrays, and skeletal vision to answer fall calls and can escalate to a two-way family video alert after 60 seconds with no response.

Pudu D7 named IFA Innovation Awards honoree

Pudu Robotics said its PUDU D7 semi-humanoid earned an IFA Innovation Awards honoree nod in Best in IFA Next & Emerging Technologies as the dual-arm robot made its European debut in Berlin with up to 20 kg payload.

Emerging Chapter

Cowboy Space leases Kent plant for orbital AI hubs

GeekWire reports Cowboy Space leased a 291,035 sq ft former Costco site in Kent, Wash., for Stampede orbital AI data-center hardware; Newmark called it the region’s largest industrial lease year-to-date, with about 300 jobs planned.

China packs AI tokens into cards and phone plans

Rest of World reports banks, carriers, cafes, and marketplaces in China are packaging AI tokens as credit-card rewards, mobile-style plans, freebies, and even loan metrics.