I loaded the CSVs from the release_2025_03_27 folder into a SQLite database so I could explore them in Datasette Lite:<p><pre><code> cd /tmp
git clone https://huggingface.co/datasets/Anthropic/EconomicIndex
cd EconomicIndex
git lfs pull
cd release_2025_03_27
for file in *.csv; do sqlite-utils insert data.db "${file%.*}" "$file" --csv -d; done
</code></pre>
<a href="https://lite.datasette.io/?metadata=https://gist.github.com/simonw/37c9ef4ad2404b9c3d7ddb639ccd3266&url=https://static.simonwillison.net/static/cors-allow/2025/release_2025_03_27.db#/release_2025_03_27" rel="nofollow">https://lite.datasette.io/?metadata=https://gist.github.com/...</a>
Claude is amazing for UNIX janitoring. Having it go through installing torch and various other things, automatically making venvs, and so on, made me realize how much of my life I've wasted dealing with stuff like cpan and pip and npm.
I suspect "Education, Instruction and Library" bar on the graph mostly amounts to cheating and more broadly defeating the point of unsupervised study. But it's nice we have a metric for that now.
Exciting to see these numbers.<p>Anthropic if you’re listening - how about a breakdown by number & taxonomy of tools being made available to Claude? Or number of tool invocations in a given session?<p>That’ll show how much use it’s getting via agentic systems vs. as a chatbot. By proxy it should show how quickly agentic use is growing.