AI briefing
Top 3 by heat · 14 signals
AI summary unavailable right now, so here are the hottest signals.
Sunday, Nov 8 · 2 signals
A team at the fictional Ashby College trained small models on a larger model's answers and found they copied its wrong-but-sure replies almost exactly. The authors suggest filtering the teacher's low-confidence answers before training.
via Preprint Patrol
Researchers at the fictional Harrowgate Lab compared agents that maintain a plan document with agents that plan 'in their heads'. The file-keeping group failed much less often after step ten.
via Preprint Patrol
Nov 7, 2026 · 2 signals
The fictional Veritas Eval Group built 600 questions with no correct answer, such as dates for events that never happened. Most systems tested guessed anyway on more than half of them.
via The Eval Desk
Researchers at the fictional Northgate Institute argue carefully curated data mattered more than model size. Critics point out the test set covers a narrow slice of the curriculum.
via Preprint Patrol
Nov 5, 2026 · 1 signal
In an experiment by the fictional Holloway Behaviour Unit, identical answers scored higher on trust as bullets than as paragraphs, even when the content was wrong. Formatting, it seems, reads as confidence.
via The Eval Desk
Nov 4, 2026 · 1 signal
The fictional Ridgeway Lab hid facts at different depths in long documents and found recall dipped most for material in the middle third. Asking for quotes first helped noticeably.
via Preprint Patrol
You're all caught up.
The Ideabro Brief
The five AI stories that matter, explained in five minutes. Every weekday.