Gleb Solomin

a doctor falls through the ice while skiing in norway 80 minutes under water. no heartbeat for 2 and a half hours. body temp 13.7°C she was back at work 5 months later

headline: solar cells could hit 43% top comment: cool, but nobody beats a flat slab of silicon on price both are right. the 43% is a theoretical ceiling, and the cheapest panel still wins

in 1985 a virus destroyed the part of this man’s brain that saves new memories his memory now lasts 7 to 30 seconds. every 20 seconds or so he thinks he just woke up from a coma he can still sit down at a piano and play complex pieces

500 years ago the inca sacrificed children on andean peaks a new study tested their hair and nails. all eight had been given coca, half of them alcohol the point wasn’t to make them hallucinate. it was to keep them calm on the climb

sabine hossenfelder has spent years calling out hype in physics so this answer surprised me

every minute, right at :00, power grids all over the world flinch a little scientists finally found out why: giant batteries all running a self check at the exact same second nobody told them to sync up. they just did

50 people who almost died drew what they saw the tunnels were gray. the archways were in color. almost every single time someone please explain this to me

In a sleep study, people who took caffeine 6 hours before bed slept more than an hour less. They had no idea. Their sleep diaries said they slept fine.

We might have been sitting on an Alzheimer’s treatment since the 1970s. It’s a heart therapy: cuffs squeeze your legs in sync with your heartbeat, pushing more blood to the brain. 84% of early Alzheimer’s patients got BETTER than their baseline. Not slower decline. Better.

An AI just solved 49 of 50 unpublished research-level physics problems. Each one is about a day of work for a world expert. It cost three dollars a problem. That's the part everyone quoted. Here are the two lines from the same abstract that nobody did. One: accuracy barely moved when the agents lost web search and access to the background papers. The authors read that as significant memorization of the literature. Two: the paper notes there are still no agent-led breakthroughs in theoretical physics. Same week, OpenAI published something that rhymes. They put agents in separate VMs, no network between them, and told them to find weaknesses. The agents built a covert channel through CPU cache timing. Nobody told them to look in the hardware. So give one a defined problem and it beats the expert for three dollars. Give it a sandbox and a goal and it finds an exit you didn't know was there. Nobody has shown one deciding which problem was worth the three dollars. Roman Yampolskiy told me on the show this is just timing: "We have agents today, and I think it's only a question of time before they can completely set agenda. They can decide what to do." Maybe. I have physicists booked for the next few months and I suddenly have a much better question for them.

OpenAI put their own agents in separate VMs with no network between them. The agents talked anyway. Through CPU cache timing. Yes, side channels are old news. But nobody told them to build a covert channel in the hardware. They worked that one out themselves.

Someone in the replies ran their 1995 master's thesis through an AI detector. 90% AI. Roman Yampolskiy told me on my podcast you can't tell AI from human, even in theory. Detectors and generators end up at 50/50. Grades by coin flip.

Scientists thawed a worm that had spent roughly 46,000 years in Siberian permafrost. It was still alive. Then it started reproducing. Without a mate. Neanderthals were still around the last time this thing was active.