Intelligent Machines (Audio) · Tech Strategy & Big Tech · November 2026
After complaints that OpenAI's new GPT-6.2 models feel dull, Leo says over-caution is a problem with Anthropic's models too, describing one that asked twice for permission on a command he had already approved.
…A lot of repetitive designs, nothing is stunning. She said GPT 6.2 solved it even worse. One of the most boring results ever. I have no words. Now, she's one of the influencers who had early access. But so maybe is, I guess you could fix that. I don't know. It feels like something odd is going on.
I mean, alignment is difficult in a lot of ways. And I think that one thing that is obvious, it's such a nebulous concept, but you have to align for vibes. And you can get bad vibes.
I think that's exactly right. In fact, I've noticed that myself with Fable 5.2. I went back to Opus 5.5, which I really like, but I noticed that the latest models, as good as they are, and Anthropics models are really good, they're neurotic because they have so been so worried that they've trained them to be extra cautious. I've had Opus 5.5 ask me not once, but twice for permission to do something. He said, can I do this? I said, yes. And they said, are you sure? Can I do this? That's what were you asking?
What was it asking to do?
Oh, it was an SSH command or something. It wasn't, first of all, it wasn't that risky. Second of all, I've set up a really good system where it gets biometrically signed approvals for me. So it got. Did you drop a
little thing of blood every time?