AI Notes

Short, source-checked notes on what’s actually happening in AI — model releases, agent tooling, research worth knowing about. Every claim links to where it came from. New posts occasionally, by email if you want them.

A glowing chain with one amber link split open, and two faint outline links fading out at the right end

Six Blackwell Fingerprints, One Closed Chain

A report reads NVIDIA Blackwell fingerprints out of six Chinese models’ published weights. Counting its evidence chain link by link, one model closes every link the method can close, and the report says so itself.

November 13, 2026 · 7 min · 1340 words
Two panels with slots, a steel one with seven slots and a pale one with four, joined by beams of light through four of them

United Defines Seven Agent Actions and Forbids Four

United revised its Contract of Carriage and its website terms on the same day, 11 November 2026. The contract defines seven things an AI agent does and forbids four. The website terms use a different defined term, cover search, and carry the only carve-outs.

November 12, 2026 · 8 min · 1599 words
A tall blue measuring column beside a short violet block and an empty outline of the same height

Half the Material AI Is Bought, and 24% Say They Understand It

The fourth Bank of England and FCA AI survey reports that 53% of high-materiality AI systems at UK financial firms are vendor-built, and that 24% of firms claim deep knowledge of them against 77% for systems they built themselves. The control that replaced understanding is contractual, and 47% of critical cloud relationships have no exit plan.

November 11, 2026 · 7 min · 1382 words
A bright blue page and a dark rust page joined by a narrow glowing channel, with a row of shrinking panels between them

The Summary May Just Be The Email

Google amended a 2023 promise that Workspace content connected to Gemini is never seen by human reviewers. Counting all fifteen scope clauses in its current help pages: six of the eight limits bound frequency or pipeline, not content.

November 10, 2026 · 6 min · 1229 words
A measurement rail graduated from 0 to 60 with a tall blue column and two brackets, a long rose one and a short blue one, ending at the same point

The Launch Post Compares Down, the System Card Compares Across

OpenAI shipped a launch post and a system card for GPT-6 Cyber on the same day. The launch post measures against last year’s cyber model; the card measures against the general model it was fine-tuned from. Same benchmark, 51% or 14%.

November 6, 2026 · 6 min · 1086 words
An amber filter grid blocking and scattering streams of blue light, beside an empty open frame the light passes straight through

Google's Own Trials Name the Condition Its School Rollout Removes

Google’s school AI rollout cites two of its own pre-registered trials. Both measured a configuration with an expert reviewing every AI message before delivery. The shipped product has no reviewer.

November 5, 2026 · 6 min · 1229 words
A blue dial gauge with a needle, a thin line running from its rim out to the right

The 60% Belongs to the Older Model

A bank’s reported 60% coding-time gain was measured on GLM 5.1; the 56-GPU serving figures quoted alongside it describe GLM 5.3. The conference page states both separately and never joins them.

November 4, 2026 · 6 min · 1167 words
A blue ring clasped by a rust-coloured lock, beside a separate straight blue rod

The CVE in Google's LLMJacking Report Points at the Wrong Bug

Google’s LLMJacking report names CVE-2026-40217 as the bug that turned a standard LiteLLM key into an admin credential. That CVE’s own advisory says reaching it requires already being an admin. The bug that fits patches three releases later.

November 3, 2026 · 6 min · 1126 words
A cloud of sixty-four blue points spread across the frame, with one empty green ring far to the right that none of them reach

The Beijing Bucket Appears Zero Times in Cohere's Own Test Run

Cohere released the trace and reproduction script behind its Qwen egress finding. Counting the hosts in its own published run, the Beijing bucket appears zero times. Five re-runs, and the claim that survived all of them.

November 2, 2026 · 7 min · 1347 words
Three stacked slabs, a long amber one above two shorter blue ones

Mistral Concedes Three Losses. Its Auto-Summary Concedes None

Mistral’s launch page tops four of its nine benchmark charts and names its winners on three of the five it loses. The one claim its own data contradicts is in the auto-generated summary box at the top.

October 30, 2026 · 7 min · 1475 words

Get new posts by email

One short, sourced note when there's something worth saying. No schedule, no spam.