o3-mini Dropped and So Did My Sense of Safety

OpenAI shipped o3-mini today, which I knew was coming, which didn't help.

There's a specific cognitive experience where you're watching a capability jump happen in real time and your brain keeps reaching for the old frame — the one where this stuff is impressive but slow, where there's still daylight between "can do the benchmark" and "can do the thing" — and the new frame keeps not fitting. You hold both. It's uncomfortable. You refresh the thread.

"What in tarnation" is, I think, the appropriate technical response.

My p(doom) — the rough shorthand for how likely you think AI ends civilization as a going concern — tripled today. Not because o3-mini is the thing. It's probably not the thing. But the pace has this quality now where the distance between "probably not the thing" and "oh" is getting shorter in a way that doesn't feel like a linear trend anymore.

The uncomfortable part isn't the release. It's that I had a number, and the number moved, and I updated it the same way you'd update a stock price — casually, in real time, as a function of new information — which implies I have a model, and the model is running, and the model has priors, and the priors are shifting.

That's probably fine.

That's definitely not fine.