Claude 3.5 Sonnet is the best AI I've used.

I’ve been using Claude 3.5 Sonnet for three weeks and I need to be honest about something.

It’s better than me at some things.

Not all things. Not most things. But specific things, things that I used to be the person people asked about, it does better. Debugging code. Explaining technical concepts in plain language. Catching logical inconsistencies in an argument. Writing structured analyses of complex topics.

I keep catching myself deferring to it. “Let me check with Claude” has become a reflex, the way “let me Google it” became a reflex fifteen years ago. Except this is different. Google gave me links to read. Claude gives me answers. The cognitive effort required dropped from “evaluate multiple sources” to “evaluate one response.”

That feels good. And that feeling is what worries me.

What it does

Anthropic released 3.5 Sonnet as a mid-tier model (between the smaller Haiku and the larger Opus), but it punches way above that positioning. In my daily use, it writes cleaner Python than I do on the first draft. It spots bugs I’d miss for hours. It holds context across long conversations with better retention than previous models.

The code assistance is what hooked me. I’m building a side project and Claude 3.5 Sonnet has become my primary coding partner. I describe what I want. It writes the code. I review, adjust, and run it. The cycle time for building features dropped by maybe 60%. That’s not a benchmark number. That’s my lived experience over three weeks of daily use.

It’s patient. I can ask the same question five different ways and it doesn’t get frustrated (because it can’t get frustrated, I know, but the experience of patience is there). It explains things at whatever level I need. When I push back on its approach, it considers the pushback, sometimes changes its mind, sometimes explains why its original approach was better.

The dependency problem

Here’s the complicated part. I’m becoming dependent on a tool I don’t fully understand.

I don’t know exactly how Claude makes its decisions. I can’t inspect the reasoning process. When it gives me an answer, I evaluate it against my own knowledge, but my own knowledge is increasingly calibrated against Claude’s previous answers. It’s a feedback loop. I trust it more because it’s been right. It’s been right because it’s good. But “good” and “infallible” are different things, and the gap between them is where dependency becomes dangerous.

I noticed it last week. I was writing code and hit a problem I would have normally worked through myself, with a notebook and pen, drawing diagrams, thinking slowly. Instead, I pasted the code into Claude and said “what am I doing wrong?” It answered in 4 seconds. The answer was correct.

But I skipped the thinking. I skipped the struggle that would have deepened my understanding. I traded depth for speed. And I did it without hesitation because the speed felt so good.

This is the bargain. AI tools that are good enough to rely on create a temptation to stop doing the hard cognitive work yourself. And the hard cognitive work is how you get better. How you develop judgment. How you build the intuition that lets you know when the AI’s answer is wrong.

If I stop struggling with code, will I still be able to spot a subtle bug in Claude’s output three years from now?

Why I keep using it

Because the alternative is voluntary inefficiency for the sake of principle, and I don’t have the discipline for that.

I’ve watched people refuse to use calculators because “you should know math.” I’ve watched people refuse to use GPS because “you should know your city.” The principled resistance to tools that make you better at your job has never won, not at scale, not for long.

Claude 3.5 Sonnet makes me better at my job. Not better in the “I’m growing as a person” sense. Better in the “I ship things faster with fewer bugs” sense. Those are different kinds of better. I’m choosing the second one. I’m not sure I’m right to.

The question underneath

The real question isn’t about Claude specifically. It’s about what happens when AI tools become good enough to replace the thinking that makes us good at thinking.

When calculators replaced mental arithmetic, we lost something (try multiplying two three-digit numbers in your head; most people can’t anymore). We gained something larger (the entire field of computational science). The trade was worth it.

I think the AI trade will be worth it too. But I want to be honest about what we’re trading. We’re trading cognitive struggle for cognitive ease. We’re trading depth for breadth. We’re trading the satisfaction of hard-won understanding for the efficiency of instant answers.

Claude 3.5 Sonnet is the best AI I’ve used. It’s so good that it’s changing how I work, how I think, and what I consider difficult. That’s what the best tools always do. And that’s why the best tools always make me a little nervous.

Related thinking: