4 test series · 3 analyses · 1 proposal
Methods and progress differ. An analysis or proposal is not an experimental result.
Decision LLMs 4
How do models classify texts? Test series on short texts, research reports and checking evidence.
Small decision models
How reliably do small models classify short texts?
283 posts and 60 purpose-written test cases
Read study
Model answers in the research archive
How do models differ when classifying the same research reports?
35 questions across a growing report collection
Read study
Expanded test set and blind review
Which test cases are unambiguous, and where do reviewers disagree?
6,814 blind judgments across 33 questions
Read study
Checking answers against evidence
Can a decision model identify whether evidence answers an open question?
312 questions and 903 question–evidence pairs
Read study
AI safety 3
Analyses of safety questions, verifiable commitments and possible investigations.
An Alien Mind: model perspectives
What safety questions do models raise, and how well do they support their claims?
Eleven contributions in the historical comparison corpus
Read analysis
Pace the Frontier: safety questions and possible actions
What conditions are needed for verifiable safety commitments?
Six thematic roadmaps in a combined analysis
Read analysis
RSI: proposal for a controlled investigation
How could recursive self-improvement be measured without uncontrolled acceleration?
Study proposal with no experiment conducted
Read proposal
AI & design 1
How AI models turn references into shapes, characters and movement. Descriptive studies make processes and differences visible.
Models and run information are available in the model catalogue. Open model catalogue →
Language and review status
Main text: EN. Review notes apply only to the content they name.
Existing review notes appear with the respective content.
See language coverage and data snapshot →