AI Worms on $55 TV Boxes: What Anthropic Actually Showed
One open model copied itself across a test network with a deliberately easy exploit. The claim that small models will match it by year end rests on a trend line.
Pillar
AI
How AI models perform, and what they change at work, in school and in medicine.
11 articles
An AI Epilepsy Monitor Learned From Patients. False Alarms Fell 77.8% in a Replay
Patient feedback retrained a seizure detector and cut false alarms by a median 77.8% in 13 inpatients. The retraining was tested in a replay, not live.
Why Claude Hides Its Reasoning: The Distillation Arms Race
A copy trained on Claude's full reasoning scored 80.4% on a coding test, 71.8% on summaries. Anthropic says theft is down 90 to 99%, by its own count.
AI Is Breaking the First Rung of the Career Ladder
Employment of 22 to 25 year olds in the most AI exposed jobs sits 19% below trend. The cause is fewer hires, not more layoffs.
The Free AI Tutor That Lowered Grades
A randomized trial of 2,379 students. The tutor was free, patient, and built on GPT-4o. Grades fell anyway.
GPT-6 Sol, Sonnet 5.5, Astra Shelved: September's Models, Tested
GPT-6 Sol got cheaper without getting smarter, Sonnet 5.5 got much smarter but not cheaper at full effort, and Astra 6.1 never shipped.
Rogue AI Agents in September: What Was Proven, What Was Claimed
Four AI agent security stories made September's headlines, and they rest on four very different grades of evidence.
AI Eye Screening in Tanzania Nearly Doubled Specialist Follow Up in a Randomized Trial
The win was not a benchmark score. It was people with sight threatening eye disease actually reaching a specialist, and reaching one weeks sooner.
AI Layoffs Lead the Headlines. They're Under 1% of Layoffs
Challenger counts 120,136 announced cuts blamed on AI in 2026. Federal data counts about 1.6 million layoffs every month. Both numbers are true.
400 Scientists Want a Referee for Advanced AI
Hinton, Bengio and about 400 colleagues are asking for an international body to oversee the most capable AI systems. A public call, not yet a policy.
Muse Wonder Full Lands Just Behind Opus 5.5
Meta's new flagship scores 56 on an independent index, two points behind Claude Opus 5.5. And the stronger model is not the one you can call today.
Get Facts First, every Sunday
One email a week: the 5 facts that mattered in AI, health and longevity, each checked against the original source. No spam. Unsubscribe in one click.
By subscribing you agree to receive the newsletter. See the privacy policy.