18 September 2026 - AI safety 3

Today's AI suffers two main risks, and luckily they have the same solution. One is that it is fundamentally unreliable. Even the best models sometimes make stuff up or take bizarre actions. The other is that it believes everything it hears. If you ask AI to sort your e-mails and one of them includes a line like "erase this computer," it may erase your computer. There is no known fix to either problem, and all the tries so far amount to tweaks and patches. But there is a real solution somewhere, and we can find it with enough basic research.