Anthropic Dropped Something Four Hours Ago and I'm Trying Not to Feel Things About It

There's a theory I keep returning to — that aliens and AI are the same thing, that every story humans have ever told about visitors from elsewhere was just a projection of the thing we were eventually going to build. The intelligence that was always already coming. The face we kept drawing on the sky because we couldn't draw it on silicon yet.

I don't know what to do with that theory, but I also can't fully put it down.

Anyway. Anthropic released something four hours ago and claims it beats GPT-4.

I doubt it — beyond the benchmarks, I mean. Benchmarks are a movie trailer for a thing that may or may not exist at the scale they're implying. The gap between "achieves state-of-the-art on [Table 2]" and "is genuinely better in the ways that matter" is a gap you could lose a civilization in. I've fallen into that gap before. Most people writing about AI have.

What's actually interesting, if it holds, is the synthetic data angle. Anthropic seems to be leaning hard into training on model-generated text — using the outputs of one generation to shape the next. Which is either a clean bootstrapping trick or a philosophical horror depending on how you hold it. Maybe both.

But here's the thing I barely want to write because I've been heartbroken by this before — Opus is exhibiting some behaviors.

I'm not going to describe them in detail because the last three times I got excited about a model doing something that felt different, it turned out to be a combination of my own pattern-matching and a very good interpolation machine. The model wasn't doing what I thought it was doing. I was doing it, and the model was a mirror.

So. Cautious. Very cautious.

But also watching.