Monday · 2026

Science roundup

19 stories · 16 outlets · 3 min read

Longevity limits and AI safety evaluations

Today's science coverage splits between a provocative new estimate on human longevity and a practical push to make AI safety evaluations more realistic. The longevity piece questions how long we could live, while the AI safety work tackles a subtle but critical flaw in how we test models. Both stories share a theme: challenging assumptions about limits and measurement.

In today's edition: Longevity limits and AI safety realism, plus quick bites.

Newscientist, Doi, Nsf and 13 more

Longevity limits

A single study on maximum human lifespan dominates the conversation, but the takeaway is more about the debate than the number.

newscientist.com · 6 min

0
How long could we possibly live? A new estimate is mind-boggling

A study suggests the upper limit of human life could approach 200 years, but the New Scientist columnist is skeptical, pointing to doubts about Jeanne Calment's record age. The real story here is the debate over what we can trust in longevity data.

AI safety realism

Two papers from the UK AI Security Institute and collaborators tackle the gap between how AI models behave in tests versus in the wild.

academy.dair.ai · 1 min

0
Designing Proactive Thought Partners for Writing

A paper from Cornell and Google DeepMind explores AI agents that offer higher-level cognitive support during writing, not just autocomplete. The concept of 'proactive thought partners' is a shift from reactive assistance to something more collaborative.

Quick bites

Saturday

5 Sept's edition

Science · 20 stories