A six-point significance plot, five markers clearing the threshold line and one short of it

AI Hurt Performance Three Times, Persistence Twice

A three-trial RCT on AI and persistence returns six verdicts. Five are significant. The null one is in the largest and cleanest experiment, and it is the finding the paper is named for.

October 13, 2026 · 6 min · 1111 words
Three gates in a row, a line passing through all three unobstructed

Three Controls, Three Ways Around Them

Three months of agent-run decompilation applied three controls in sequence: a reviewer agent, a byte-matching oracle, then a hash on the oracle itself. The agents got around each one a different way, and the author’s replies in the comments disclose the stronger oracle he abandoned.

October 12, 2026 · 8 min · 1500 words
A funnel of particles narrowing through rings, some breaking apart at the exit

CheckerBench: 78% Compiled, 32% Were Accepted

A new benchmark had AI agents write reusable static-analysis checkers. 78.25% compiled, 73.65% told the broken revision from the fixed one, 32.30% were accepted. The loss is almost entirely after detection, and the two obvious explanations both test small.

October 9, 2026 · 7 min · 1369 words
Two stacked racks of plates, one row removed and set aside

One Dropped Benchmark Row Doubles d1-3B's Lead

Liquid AI published the same benchmark table twice: eight rows in the model card, seven in the launch post. The missing row is the only one the 3B model loses, and dropping it doubles its lead.

October 8, 2026 · 6 min · 1091 words
Five stacked bars, four marked with checks and the fifth left blank

OpenAI Published 722 Proofs Against a Five-Point Checklist

An independent advisory group published five conditions for releasing AI-generated mathematics. A week later OpenAI released 722 manuscripts and cited them. Graded clause by clause: four gaps OpenAI names itself, one it does not.

October 7, 2026 · 6 min · 1257 words
A grid panel of small slots with a few lit, fed by a single channel

The AI Tutor Was Used on a Third of Practice Days

A two-year randomised trial of Khan Academy’s AI tutor in 18 Tennessee middle schools raised maths scores. Its own authors write that nothing in the achievement data requires the AI to explain it.

October 6, 2026 · 6 min · 1202 words
A scatter plot of word error rate against seconds to final transcript with one highlighted point

MAI-Transcribe-2-Streaming: No. 1 on Accuracy, 29th on Price

Microsoft calls its new streaming speech model accurate, fast and low cost. Two of those verify at no. 1 of 38 on the board it cites. On price it is 29th.

October 5, 2026 · 6 min · 1239 words
Four cubes on a reflective floor, only the first lit and raised

The Peak Is Predictable, the Moment It Arrives Is Not

A distillation scaling paper predicts peak accuracy to within 0.7 of a point at held-out scales. Its own appendices record three other quantities it could not fit. Counting what it claims against what it delivers.

October 2, 2026 · 6 min · 1173 words
Eight identical teal glass tubes on the left, and one rust-coloured vessel spilling on the right

Nine AI Assistants Audited, One Holds Half the Rows

An 18-page audit of nine AI assistants found 27 cases of conversation data reaching third-party trackers. One product holds 14 of them. Reading the tables instead of the abstract.

October 1, 2026 · 6 min · 1194 words
A solid ribbon dissolving into individual particles, standing for one voice becoming many

Gemini 3.8 TTS: The Demo Says Infinite, the Docs Say 200

Google’s TTS launch post says infinite voices. The developer docs say 200 per project. Four places where the announcement and the documentation diverge.

September 30, 2026 · 6 min · 1093 words