The Polylog AI Intelligence Brief

Morning Edition · Saturday, August 22, 2026

Tech

Nvidia Reports a Perfect Public-Set Score on ARC-AGI-3 by Wrapping Claude Opus 5 in Its Own Agent

The same model the ARC Prize Foundation independently measured at 30.2 percent cleared all 183 public levels once Nvidia wrapped it in its AVO harness, a result Nvidia generated and scored on its own.

5 sources
Corroborated

Markets

Unitree's Shanghai Debut Puts a $66 Billion Price on Humanoid Robots as Beijing Opens Its Robot Conference

The humanoid robot maker closed about 460 percent above its 150.80 yuan offer price on August 19, days before more than 300 companies began showing embodied systems at the World Robot Conference.

4 sources
Corroborated

Markets

Nvidia Backs Up to $105 Billion in Financing for an OpenAI Data Center in Ohio

The credit supports an initial 4.25 gigawatts of capacity with an option for 3.75 more, with power expected to arrive in phases from 2028.

2 sources
Corroborated

Tech

OpenAI Offers Zero Data Retention on Frontier Models, and Anthropic Reverses Its Own Retention Rule

OpenAI previewed Private Safety Processing, which searches for multi-session abuse without exposing prompts to staff, after months of public criticism from Palantir's chief executive over how labs treat enterprise data.

4 sources
Corroborated

Tech

Scale AI Benchmark Measures Whether Models Can Rewrite Another Agent's Harness

Across 111 scored runs, the choice of optimizer model separated results more than the coding harness the optimizer worked through, and native harnesses were not consistently better than a shared one.

2 sources

Tech

Meta Runs a Two-Track Model Business: a Metered API on Top, Open Weights Underneath

Muse Spark 1.2 and the Muse Code agent sell through an OpenAI-compatible application programming interface (API) at $1.25 and $4.25 per million tokens, while the 30-billion-parameter Muse Glimmer ships free for a single consumer graphics card.

3 sources
Corroborated

Tech

Chinese Open-Weight Models Set the Coding Bar for Downloadable Systems in August

Trackers put DeepSeek V4 Pro at 80.6 percent on SWE-bench Verified and Alibaba's Apache-licensed Qwen3.8-27B at 73.0 on Terminal-Bench 2.1, both released within a day of each other.

3 sources
Plausible

Tech

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

SAM 3 and DINOv3 run on 300 Nvidia A100 accelerators at the national supercomputing center, turning expert annotation of X-ray and neutron data into a near-real-time step.

3 sources
Plausible

Tech

DeepMind Sets EVE Online as Its Test Bed for Agents That Must Plan Over Months

The lab tied fifteen years of game research to a partnership with Fenris Creations, the Icelandic studio it took a minority stake in during a $120 million management buyout.

3 sources

Tech

A $41.5 Million Federal Health Project Puts Meta's Vision Models on Battery-Powered Wheelchairs

The University of Pittsburgh runs DINOv3 and Segment Anything on-device for real-time perception in a robotic wheelchair and arm, accepting lower boundary precision for stable frame rates.

2 sources

Tech

Prompt Injection Remains the Dominant Failure Mode as Agent Frameworks Absorb Supply-Chain Attacks

Security researchers report a compromised release of LiteLLM, the model gateway used by CrewAI and DSPy, carrying an autonomous attack bot and downloaded roughly 47,000 times.

3 sources
Plausible

Tech

Google Gives American Students a Free Year of Its Paid AI Tier as Labs Compete on Habit Formation

Students in more than 140 other countries receive the lower Google AI Plus tier, and the offer converts to a paid subscription automatically after twelve months.

3 sources