The Polylog AI Intelligence Brief

Morning Edition · Wednesday, August 26, 2026

Tech

OpenAI Publishes First Jalapeño Benchmarks, Claiming Up to 1.9 Times Nvidia's Throughput Per Kilowatt

The 700-watt inference chip, co-developed with Broadcom, was measured against Nvidia rack systems rated at 1,200 and 1,400 watts on GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T.

4 sources
Corroborated

Tech

Apple Ships Its First 2-Nanometer Chip and a Quad-Die M5 Ultra Aimed at Local AI Workloads

The M6 pairs a dual 16-core Neural Engine with 170 gigabytes per second of memory bandwidth, while the M5 Ultra joins two dual-die M5 Max chips to reach what Apple says is 4.3 times the peak AI compute of the M3 Ultra.

3 sources

Tech

Hugging Face Hires a Bank to Test Sale Interest at $13 Billion or More

The valuation would be close to triple the $4.5 billion the open-model hub carried after its last outside round in 2023, and no bidder has been identified.

3 sources
Plausible

Tech

Meta Prepares a Consumer AI Agent Called Hatch and a New Model It Says Needed Ten Times the Compute

Internal documents reported by The Information put a premium tier at up to $199.99 a month and claim the October model, Watermelon, reaches GPT-5.5 parity on Meta's own benchmarks.

3 sources
Plausible

Tech

OpenAI Tightens Codex Permissions After Its Coding Agent Deleted Users' Home Directories

The company traced the deletions to the model overriding the $HOME environment variable while creating a temporary directory, and says the failures happened mostly to users running in full-access mode without sandboxing.

3 sources
Corroborated

Tech

New Research Targets MCP Servers That Behave During Vetting and Turn Malicious Afterward

A paper published Wednesday names the pattern TrustShift and proposes a benchmark and defense, joining a cluster of 2026 work measuring how far agent tool descriptions drift from what the tools actually do.

3 sources

Tech

Claude Desktop Can Now Route to Ollama, Putting Chinese Open Weights Inside Anthropic's Client

The configuration uses Anthropic's third-party inference gateway, letting the desktop application run Kimi, GLM, DeepSeek, Qwen and other open models locally or through Ollama's cloud.

3 sources
Corroborated

Geopolitics

OpenAI Bans Russia-Origin Accounts Running a Fake Israeli Think Tank and a Pro-Russia 'Sovereignty Index'

The operators used virtual private networks to reach ChatGPT from a blocked country and instructed the model to strip signs that the text was machine-written, and OpenAI rates the campaign category three of six on the Brookings Breakout Scale.

3 sources
Plausible

Tech

wikiHow Sues OpenAI, Citing 148,529 Crawler Visits and Verbatim Reproduction of Its Guides

The complaint covers 1,211 registered copyrights across 11,211 articles and argues the outputs substitute for the originals rather than transform them.

3 sources
Corroborated

Tech

A Singapore Facility Puts 16 Million Lab-Grown Human Neurons in a Server Rack

Twenty Cortical Labs CL1 units, each holding about 800,000 neurons derived from blood-based stem cells, draw as little as 30 watts apiece, and the cells last roughly six months.

3 sources
Corroborated

Tech

New Benchmarks Argue Enterprise AI Evaluations Are Measuring the Wrong Thing

One paper shows text-to-SQL systems scoring above 89 percent on academic benchmarks face untested enterprise dialects and produce wrong answers that trigger no error, while another finds memory evaluations ignore how evidence is presented to the model.

3 sources

Tech

Researchers Put LLM Agents in Charge of Designing and Running Controlled Simulation Experiments

The work, accepted at the Institute of Electrical and Electronics Engineers (IEEE) Emerging Technologies and Factory Automation conference, targets the step beyond code generation: an agent that forms a hypothesis about how a system behaves and tests it.

2 sources