Karpathy built a video site from 43 million Hacker News comments in a weekend, and Simon Willison benchmarked a SQLite parser with Claude

Two weekend builds with Claude show where the human still earns their keep.

What happened

The Decoder reports that Andrej Karpathy built HackerTube over a weekend using Claude Opus 5.5. Claude filtered 43 million comments from 20 years of Hacker News down to 6,454 videos that keep getting recommended there. The site needs no login and stores state locally in the browser. Karpathy says Claude Code handled nearly all the work, from data analysis to server setup. His manual steps were a Google Cloud API key, a rented server and buying the domain. He then built a second site with 47,000 videos from a channel list, which he says was one shotted with Claude Code in 3 minutes.

Simon Willison wanted a SQLite function that turns strings like -$1,234.56 or 12.5% into numbers you can sum and sort. He asked Claude to compare building it in C and in Rust. The first result surprised him: the C extension was 14 KB against roughly 300 KB for Rust, but Rust was faster, because its equivalent of C's strtod is about twice as fast. He then asked Claude to write its own C parser, and it did: 16 to 19 ns per row, about twice as fast as the Rust versions. His write up says all versions matched across 8 million fuzz checks.

My take

Neither story is about AI writing code. Both are about a person picking the right question.

Karpathy chose a dataset nobody else had thought to mine. Willison measured the first answer, noticed it was off, and pushed back with one sharp prompt. That loop of build, measure, challenge is the skill.

For operators, Willison's function is a real problem in disguise. Currency and percent strings exported as text break every CRM report and spreadsheet sum I have ever inherited. A small, tested parser at the data layer beats patching formulas in twenty dashboards.

What I'd copy:

1. Prototype in a chat before you commit to a stack.
2. Benchmark the result, even when it looks done.
3. Keep keys, servers and domains in human hands.
4. Add a fuzz or comparison test so the fast version is also the correct one.