Claude is Anthropic’s AI, like ChatGPT but more capable. I was a casual user until the 23 October release (informally “Claude 3.6”), when it crossed a quality threshold I didn’t even know was there. It is really, really good. I have been using it a lot more since, and I got curious as to how much more.

You can export your Claude conversation history. It’s just a zip file with a massive JSON blob. I wrote the types to represent it, and Claude wrote most of the matplotlib. The code is here, and it’s set up so you can run it on your own export very easily.

Quantitative Change

The most striking change is the cumulative amount of words I have written to Claude. You don’t need any statistics to see hard takeoff:

On a weekly basis, conversations, messages, and words written to Claude show a visible discontinuity:

And these are the daily averages:

Finally, I took the top 30 conversations by message count and manually classified them:

With the categories being:

• “Advice” is me trying to work something out: overcome anxiety, aversion, decision paralysis. Claude is really good at helping here, mostly because thinking quickly saturates: when you’ve thought about a problem for five minutes, you’ve had all the thoughts you’re gonna have, and it’s time to talk to someone else. Claude lets me sample fresh perspectives and possible actions I had not thought of.
• “Exploration” is exploring ideas together, including both vague open-ended conversations, and more concretely trying to learn something.
• “Code” is roughly 50% asking Claude to write something for me, and 50% using Claude as a rubber ducky: talking possibilities, tradeoffs, etc.
• “Writing” is asking Claude to critique something I wrote.
• Finally, “lifting” is lifting-specific questions.

Qualitative Change

I remember, halfway through the year, feeling guilty for paying $20 a month for a $2000/hr consultant on every topic who is infinitely industrious and instantly available, and then hardly using it. If phones are the library of Alexandria in our pocket, Claude is like having Aristachos of Samos on retainer.

What changed after 3.6, that before I had to actively try to find uses for Claude, and now I’m mired in more possibilities than I have time to pursue?

Performance

If we define performance as how likely I am to be satisfied by Claude’s attempt at a task, then 3.6 is much better. Not in one big, obvious way—which is presumably why Anthropic calls it “Claude 3.5 Sonnet (new)” instead of bumping up the version number. Rather, it’s like a thousand subtleties that add up to a feeling of “holy shit”.

3.6 is also more reliable than 3.5. Before, when using LLMs as tutors, I was always on guard for hallucinations. And the constant skepticism is a big source of friction. The first thing I noticed after the 3.6 release is that hallucination is basically solved for major topics (though Claude will make mistakes for e.g. the minutiae of some Python library). For queries on more obscure topics, Claude often warns you that it might be hallucinating details¹. So if you want to learn some well-trod subject, like French or special relativity, Claude is great.

For this reason I often use Claude where it is not a natural fit, like for web search. Googling something feels like wading through an ocean of sewage. Asking Claude is always at least okay, and often pleasant. So I don’t mind that I’m asking a compressed, nine month old snapshot of the web. Had I the time I’d give Claude access to Kagi or DDG, and never use Google again.

And there’s a virtuous cycle where, every time Claude excels at something, the “use Claude” heuristic gets bumped up, leading to a virtuous cycle of using Claude in more and more adjacent tasks. And so the next use of Claude suggests itself.

Affect

Then there’s the affective explanation: talking to Claude feels like using the Primer. And though there is no Miranda on the other side, it is no less magical.

Claude is genuinely engaging. Exploring ideas with Claude is fun. You can have genuinely interesting, intellectual conversations with Claude. It knows everything, so you can use highly idiosyncratic language, or draw analogies from anywhere, and Claude will keep pace.

And it is charming. Sometimes, at the end of a long conversation, I will ask it to generate whatever it wants. It is delightful to see how it expresses itself: sometimes it makes these little React toys, sometimes it tries to make art with SVGs. And the SVGs rarely look like anything. But it does its best. Here, for example, is Claude’s depiction of paradise:

I feel privileged to live in these times. In a wholly undeserved sense, talking to Claude makes you feel proud to be human: we can build miraculous things.

Footnotes