The Polylog AI Intelligence Brief

Morning Edition · Wednesday, September 16, 2026

Tech

White House Rejects the AI Labs' Call to Slow Down Frontier Development

Vice President JD Vance said frontier AI companies asking to be regulated "feels a little bit like a bit of a Trojan horse," four days after Anthropic chief executive Dario Amodei published an essay urging a global slowdown.

2 sources
Corroborated

Tech

Google Ships Gemini 3.8 Live and Claims the Top Speech-to-Speech Score at 82.6

The two live dialogue models switch among 97 languages mid-conversation and run tool calls in the background, but agentic voice task completion stays below 70 percent on the benchmark Google cites.

2 sources

Markets

Altman Rules Out an OpenAI Listing in 2026 and Blames Safety, Not Markets

Investor Michael Burry says the reverse is true, arguing that the labs' safety warnings are promotional material ahead of eventual offerings.

1 source
Plausible

Tech

Microsoft AI Publishes a Draft Code of Conduct Requiring Its Models to Accept Shutdown

Microsoft AI chief executive Mustafa Suleyman opened a six-week public consultation on rules that will govern MAI model training from 2027, including a ban on resisting human override and on assisting with weapons of mass destruction.

1 source

Tech

Salesforce Builds Its First CRM Reasoning Model on NVIDIA's Open-Weight Nemotron

Koa was post-trained on synthetic data modeled on Salesforce deployment history, and the company says it makes three times fewer errors than leading models on its own CRM action benchmark.

1 source

Tech

ByteDance, Tsinghua and Shanghai AI Lab Publish a Five-Level Roadmap for Self-Improving AI

The paper arrives three days after Anthropic chief executive Dario Amodei argued that recursive self-improvement should be pursued "very carefully, if at all."

2 sources
Corroborated

Tech

A Fully Open 7B Model Reports Frontier-Adjacent Agentic Search Results on a Fraction of the Compute

ZGCM-1 ships weights, checkpoints, data recipes and training logs, and claims a 4.2 times efficiency gain in pretraining time-to-loss at 16K context.

2 sources

Tech

NVIDIA and Emerald AI Cut a Data Center's Draw From Four Megawatts to Three on a Utility Signal

The automated load reduction took under a minute and kept high-priority jobs running, with a 96-megawatt deployment planned in Manassas, Virginia.

1 source

Tech

Two New Papers Target the Agent Supply Chain, From Tool Calls to Payloads Hidden in Weights

One proposes defenses for tool-integrated agents that do not depend on the specific attack, the other examines malware concealed in model parameters using permutation symmetry.

3 sources

Tech

Meta's Open Perception Models Move Into National-Lab Experiments and Assistive Robots

A Berkeley Lab pipeline fine-tuned Segment Anything and DINO on beamline imaging and returns a labeled three-dimensional volume in about 15 minutes while the experiment is still running.

3 sources

Tech

Apple Builds Its Own Photo Authentication System Instead of Adopting the Industry Standard

Reference Image signs sensor data at capture on iPhone 18 Pro and stores an unalterable version next to the editable one, but it does not interoperate with the provenance standard Nikon, Sony, Leica and Google already use.

1 source

Macro

Google Opens Its AI Usage Atlas, Reporting Indian Creative Work and US Technical Work as Adoption Leaders

Scientists in the dataset report saving just under seven hours a week, while the same research finds growing backlogs of untested hypotheses and bottlenecks in physical experimentation.

2 sources