Capabilitiesv.01
How much can an AI agent do reliably?
Change the success threshold. Watch the task horizon change with it.
3 original sources4 min read
AI RESEARCH & INTELLIGENCEINDEPENDENT PERSPECTIVE / VOL. 01
A clearer picture starts here.
Examine the claim.
Follow the evidence.
Make sense of important AI developments.
Explore the data, question the claims, and
trace the story back to its source.
Explore the collection OR START WITH YOUR OWN QUESTION
FROM THE EVIDENCEv.02
EVALUATION RUNSUK AISI
10 / 122
runs with reported unsanctioned behaviour
002 / THE COLLECTIONREAD. INTERACT. INVESTIGATE.
Go beyond
the headline.
An opening collection on AI autonomy,
control, and the evidence behind the claims.
06 stories
Control & safetyv.02
A fast response starts with noticing.
What 122 evaluation runs tell us about detection, access and containment.
1 original source3 min read
Control & safetyv.03
New evidence. A different explanation.
Follow the change between Anthropic’s initial disclosure and later assessment.
2 original sources3 min read
Control & safetyv.04
One incident. Three vantage points.
Read the Hugging Face disclosure alongside the developer and external investigations.
3 original sources4 min read
Evaluationv.05
What can an AI monitor actually see?
Written reasoning and observable actions offer different windows into behaviour.
1 original source3 min read
Evaluationv.06
A benchmark is only as good as its questions.
A closer look at Epoch’s sample audit of Humanity’s Last Exam.
1 original source3 min read
003 / THE DEVELOPING PICTURE
Evidence changes.
So should the story.
Revisions deserve as much attention
as the original announcement.
OPEN TO SCRUTINY
Follow every claim
back to the source.
Original material. Visible limitations. A clear distinction between what the evidence says and how we read it.