Claude Opus 4 and what understanding feels like

I’ve been using Claude Opus 4 for a month and something shifted.

Not in the model. In me.

I stopped noticing the boundary. The moment in a conversation where you think “ah, right, this is a machine.” That moment used to be frequent. A response that was technically correct but tonally wrong. A follow-up that missed the subtext. A misread of what I actually needed versus what I literally asked.

With Opus 4, those moments are rare. Not gone. Rare. And their rarity changes the experience fundamentally.

What changed

The model contextualizes. Not just the immediate question, but the conversation’s arc. I mentioned a design constraint three messages ago, and ten messages later it referenced that constraint when evaluating an option I hadn’t connected to it. It didn’t just remember. It integrated.

It asks questions that reveal it understood the subtext. I described a problem vaguely (because I hadn’t fully articulated it to myself yet) and instead of answering the vague question literally, it asked a clarifying question that was better than the question I’d have asked myself.

That’s the thing that makes me pause. Not that it answered well. That it asked well. Asking a good question requires understanding not just what someone said but what they’re trying to figure out. The gap between those two things is where human communication actually lives.

The philosophical problem I keep avoiding

Is this understanding?

I know the standard answers. “It’s pattern matching at scale.” “It’s statistical prediction of likely next tokens.” “Understanding requires consciousness and consciousness requires…” and then the conversation goes somewhere I can’t follow because nobody agrees on what consciousness requires.

Here’s what I know from experience. When I work with Claude Opus 4, the collaboration produces better results than when I work alone. Not because it does more work (though it does). Because the interaction itself improves my thinking. It pushes back in useful ways. It surfaces connections I missed. It asks questions I didn’t think to ask.

If that’s not understanding, I need a better word for what it is. “Statistical next-token prediction” doesn’t capture the experience. And experience, whether the model has it or not, is what I have.

The getting-harder-to-articulate part

The distinction between “simulating understanding” and “actually understanding” is getting harder to articulate. Not philosophically. Practically. In terms of what difference it makes to me, the human in the conversation.

If the simulation is perfect enough that it improves my work, challenges my assumptions, and produces insights I wouldn’t have reached alone, does the question of “real understanding” matter? To a philosopher, yes. To me, sitting here at 11pm trying to solve a problem, the answer is less clear.

I’m not saying Claude is conscious. I’m not saying it understands in the way I understand. I’m saying the practical difference between “understands” and “behaves exactly as if it understands” is smaller than I want it to be.

And that gap is closing with every model update. Which means the philosophical question I keep avoiding is going to demand an answer eventually.

Not today. But eventually.

Related thinking: