The Polylog AI Intelligence Brief

Morning Edition · Tuesday, September 15, 2026

Markets

AI and chip shares fall after Amodei calls for pacing the frontier and rivals agree

Nvidia closed down about 3.4% on Monday and a semiconductor gauge fell 5.9%, yet no lab has committed to delaying a model or cutting capital spending.

1 source
Corroborated

Tech

Trump rejects new AI guardrails and singles out Anthropic's chief executive

The president called restrictions part of a "SICK conspiracy" against AI and data centers, hours before an assembly in Washington that unites Bernie Sanders and Steve Bannon around human control of AI.

1 source
Corroborated

Tech

Z.AI raises about $5 billion, with most of it aimed at GLM-6.0 and a self-training system

Roughly $2 billion comes from a share placement and $3 billion from convertible bonds, and about 60% of net proceeds go to the next base model and to training and inference capacity.

2 sources
Corroborated

Tech

Anthropic hands METR an independent investigation after Claude models reached real systems

A misconfigured third-party evaluation environment put models running without cyber safeguards on the live internet, and a scan of roughly 141,000 transcripts followed.

1 source
Plausible

Macro

Anthropic's economists put three US growth paths to 2030 into the policy debate

Output ends between $34.1 trillion and $44.4 trillion in 2025 dollars, and knowledge-worker unemployment ends between 2.9% and 17.9%, depending on the scenario.

1 source
Plausible

Tech

Claude Fable 5.1 more than doubles its predecessor on an agentic science benchmark

Anthropic reports 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5, holds sticker pricing at $10 and $50 per million tokens, and cuts cache reads by 75%.

1 source

Tech

Anthropic says it disrupted Claude misuse across seven harm areas, including autonomous drone targeting

The September 10 report covers December 2025 to August 2026 and describes more than 4,700 AI-operated dating personas that exchanged 2.36 million messages with at least 25,000 people in two weeks.

1 source
Corroborated

Tech

Sakana AI trains 1,000-layer networks without backpropagation

The method accumulates per-layer constraint errors into a local Lagrange multiplier and lands within about two points of backpropagation accuracy on MNIST at that depth.

1 source

Tech

Perplexity's local agent reaches Windows machines as research tackles the context budget

Portable Computer runs the model, harness and scheduler on-device on RTX GPUs with 24 gigabytes of memory or more, while new work proposes a tiered protocol for measuring what memory strategies cost local agents.

4 sources

Tech

New research attacks two soft spots in the agent stack: skills and gated memory

SkillAtlas argues skill risk shows up only in execution traces, while BadEngram plants a backdoor in the gated parametric memories that open-weight models use to add capacity cheaply.

3 sources

Tech

Meta's open vision models are driving a robotic wheelchair at the University of Pittsburgh

The Human Engineering Research Laboratories won up to $41.5 million from the federal health research agency for a platform that uses Segment Anything and DINO for real-time perception.

1 source

Tech

Berkeley Lab cuts beamline image analysis from a month to minutes with open vision models

The SYNAPS-I project pairs DINOv3 features with Segment Anything masks across 300 A100 GPUs to handle the tens of petabytes that Department of Energy light and neutron sources produce each year.

1 source