The Polylog AI Intelligence Brief

Morning Edition · Thursday, October 1, 2026

Tech

Google Ships Gemini 4 Argon With a 1M-Token Output Ceiling and Gives Cyber Defenders First Access

Google reports Argon leads 12 of 18 disclosed benchmarks, including 77.9% on DeepSWE v1.1, while trailing GPT-6 Astra by 10.5 points on FrontierSWE v2.

2 sources
Corroborated

Tech

OpenAI Says It Shut Down a Reasoning-Extraction Campaign and Links It to People Associated With Moonshot AI

The company reports 16,000 requests in a single day from more than 4,000 users, drawn from a cluster of over 15,000 accounts, with the activity contained by late July.

2 sources
Plausible

Tech

Nvidia Puts Agent Containment Into Silicon With OpenShell and a BlueField-4 Watchdog

The Apache 2.0 runtime converts operator instructions into enforced policies before an agent starts and pairs them with an independent monitor on separate hardware.

2 sources

Macro

Anthropic Publishes a 2030 Model Where the Fastest AI Growth Path Cuts Labor's Share to 45%

Its extreme scenario pairs 15.4% annual GDP growth with a 21.5% fall in cognitive employment from mid-2026 levels.

2 sources
Corroborated

Tech

Anthropic's Sonnet 5.5 Holds Price While Cutting Cost per Task, Matching Google's New Frontier Rate

The model stays at $2 and $10 per million input and output tokens, runs more than 30% faster, and posts 81.3% on SWE-bench Pro against Sonnet 5's 63.2%.

2 sources

Tech

CoreWeave Brings Multi-Rack Vera Rubin Clusters Online With 4.8x Coding Throughput Over GB200

Each rack pairs 72 Rubin GPUs with 36 Vera CPUs and supports more than 11,000 concurrent isolated agent environments, with Cognition as the first production customer.

2 sources

Tech

A 27B Open-Weight Agent From BAAI Posts 81.8 on MLE-bench Lite by Learning to Iterate

AREX-2 trains on synthesized long-horizon improvement trajectories and transfers to deep research, reaching 84.0 on BrowseComp and 52.6 on Humanity's Last Exam.

2 sources

Tech

DeepMind Watermarks AI-Designed Proteins Without Measurable Loss of Binding Function

Wet-lab tests against VEGF-A, the SARS-CoV-2 spike receptor-binding domain and PD-L1 found marked designs matched unmarked ones on hit rate and affinity.

2 sources
Corroborated

Tech

A Governed Execution Architecture Cuts Agent Attack Success From 98.3% to 7.7%

Interposing a planner, policy gate, executor and auditor between an agent and the operating system costs utility, and the paper measures how much.

2 sources

Tech

Meta's Open Vision Models Cut a Month of Scientific Image Segmentation to About 15 Minutes

SAM 3 and DINOv3, fine-tuned on experimental imaging data and run across 300 A100 GPUs, anchor a flagship project in the US Department of Energy's Genesis Mission.

2 sources
Corroborated

Tech

Google Retires Gemini Gems for Stackable Skills Starting November 17

Personal accounts migrate first, Workspace business and nonprofit accounts in March 2027 and education accounts in June 2027, with existing Gems converted automatically.

2 sources

World

Anthropic Convenes Religious Scholars Under Nondisclosure to Debate Claude's Moral Status

Roughly 20 religious and philosophical leaders joined sessions on the model's constitution and possible consciousness, as separate survey work finds women consistently more wary of AI than men.

2 sources
Corroborated