The Better the Model, the More My Taste Matters

Opus 4.7 came out this week. I'm using it now. I stopped using 4.6.

The new model is faster. It makes fewer mistakes. It catches things 4.6 would have missed. All real improvements. But if you asked me to describe the gap between the two, I'd call it marginal.

That's not a complaint. It's an observation about where we are.

For a while, every new model release felt like a step change. Claude 2 to Claude 3 was a different league. Sonnet to Opus was a different league. Each version solved problems the previous one couldn't touch. You could feel the ceiling moving.

That's not what the last few releases have felt like. The ceiling isn't moving the way it used to. The model is better, but the work is still the work. Same questions, same trade-offs, same decisions to make.

Which tells me something about what's actually hard.

What I keep noticing is how much of my work isn't about the model being smart enough. It never was. The model has been smart enough for a while. What the model can't do is know what I'm trying to build, why it matters, and which of the ten reasonable paths forward is the right one.

That part is mine. It was mine with 4.6. It's still mine with 4.7. And I suspect it'll still be mine with whatever comes next.

Taste doesn't scale with model capability. It scales with paying attention.

You can feel the shift when you watch people use these tools. The ones who get the most out of them aren't the ones typing faster prompts. They're the ones who know what they want, who can look at output and say "no, that's not it" without being able to explain exactly why, who throw away good work because something better is possible. That's not a technical skill. It's a judgment skill, and it gets built the same way it's always been built: by caring, and by doing the thing enough times that your gut knows what right looks like.

The model is a collaborator, a fast and smart one that gets better every six months. But the direction still has to come from somewhere.

The quiet truth about AI is that it doesn't replace the thing that was hard. It replaces the thing that was slow.

Writing a draft used to be slow. So did writing code, or reading documentation. The model makes all of that fast. What's left is everything that was hard before: knowing what to build, what to say, what to ship, what to throw away.

The models are getting better at the slow parts. They're not getting better at the hard parts, because the hard parts aren't about capability. They're about knowing what you want.

So I'll use 4.7. It's better than 4.6. Marginally, but meaningfully. And I'll keep doing what I've been doing, which is figuring out what I actually want and then asking the model to help me get there.

The better the model, the more my taste matters. Every improvement in capability moves more of the work onto the part the model can't do for me, which is the part I actually wanted to be doing anyway.