219.5Kviews196×
Posted Sep 18, 2026
The honest version of AI safety right now: one team trying to understand what’s happening inside the models, another checking whether they’re currently dangerous. That’s it. Nate Soares’s nuclear analogy cuts through the public posture. A reactor safety program that consists entirely of studying the internals and monitoring for meltdown isn’t a safety program — it’s a monitoring program. The engineering principles that make failure unlikely in the first place don’t yet exist for frontier AI in any robust form. His sharpest point: the reassurance that AI mistakes are fixable assumes the systems stay correctable. That assumption has a shelf life. Credits: Video: https://youtu.be/eYUYdpG4UT8?si=-nLm8zzlpRFo1N2U Speaker: Nate Soares #AI #artificialintelligence #futuretech #futureofai #aisafety
See the breakdown