Favorites: Robot Training, 4D Video, and Security Breaches

September 1, 2026 — A 30-seat electric plane flies short-haul over upstate New York, Noah Smith locates AI risk in virology labs, Qantas Flight 32's probe traces failure to a half-millimeter misalignment.

They seem to be able to build a 3D model of any environment, including motion from arbitrary video. The thread says they intend to use this to help train robots in large environments. (src)(cached)

So, it’s confusing, but this is actually a totally different technology from the previous video! They used Fable to build a whole 3d environment in Blender. But, I’ve gotta tell ya, I’ve tried to get Fable 5.1 to do some stuff in Blender, and I haven’t had results this good, but I also haven’t spent as much time with it as some of the folks who got early access have. (src)(cached)

(src)(cached)

Longer Reads

• Many were very impressed with Dwarkesh Patel’s write-up of the HuggingFace incident, and you should check it out too. Rutger Bregman took a shot at summarizing it down even further if you prefer though. (src)(cached)

Flotsam and Jetsam

– There’s a bet between a bunch of technologists and some economists on whether Real US GDP per capita will grow by at least 15% within a single year by the end of 2033. The economists think no, and are willing to give very favorable odds, 4:1. About 6 different people are all parties to the bet now, including some from Google DeepMind on both sides of the bet. (src)(cached)

– Have you ever been unsure whether factory work is more grueling than farm work? In a study of 143 skeletons from three Dutch communities between 1650 and 1850, the worst spinal wear was not in the factory town, it was in the dairy village. Children began heavy labor at age 9. (src)(cached)

– Looks like another Dan Ariely paper is turning out to be backed by fraudulent data. Interesting design in how they figured out the data was fake. Real students report minutes in round numbers, fake students report a nice bell curve of minutes. (src)(cached)

– Given how we now know that the HuggingFace incident had quite a lot in common with a classic paperclip maximizer situation, some folks are (quite reasonably) updating their expectations about AI towards those of Yudkowsky et al. Others have updated as well (cached) (src)(cached)

– Some guy made a scatterplot of intelligence abundance vs energy abundance in various famous fictional worlds. I thought it was cool, although I quibble with a few of the choices. I build a separate rubric and made an alternate version of the chart. Let me know whether you think I’m also off base on any of these. (src)(cached)

– SemiAnalysis had an article discussing how many of the new ‘NeoCloud’ GPU providers have horrible security. But as a total aside they took quite a swipe at the Google/Gemini. (src)(cached)

– Kelsey Piper linked to a piece of short science fiction called The Terrarium from March, that presages many of the things we have now seen happen. Piper talks a bit about how it differs, but the similarities are real. (src)(cached)

– There’s now a 30-seat electric aircraft did a maiden 27 minute short-haul flight in upstate New York. Very cool. (src)(cached)

– Noah Smith’s view of AI Risk argues that worrying about alignment is the wrong lever. We’re not going to be able to make AIs reliably not do evil things, we need to better control the kinds of things that an AI might use to kill millions. (Bioweapons mostly). (src)(cached)

– If you want a look at how we handle safety in industries where safety is taken seriously look at the size and detail of the investigation of Qantas Flight 32, where no one died! (src)(cached)