AI

8 posts

Close-up of an old desk globe under warm light, Canada in salmon pink with Saskatchewan and Hudson Bay readable, Greenland and Iceland at the edge.

The vision model was the easy part: turning GeoGuessr streams into stats

I built a tool that reads six-hour GeoGuessr VODs and fills a CSV of games, rounds, scores and countries in about 40 minutes on a gaming PC. A local vision model does the reading, but a frame classifier, targeted crops and the game's own rules do most of the work.

October 4, 202611 min read

Mars seen whole against a black background, pale ochre with darker grey patches and a scatter of craters, a simulated view from the Mars Global Surveyor.

AI Competitive Programming

Opus 5.5 halved my code golf scores by writing different programs

I pointed Claude Opus 5.5 at 27 CodinGame code golf puzzles in TypeScript. It cut the total from 24,644 bytes to 6,788, beat Fable 5 by 47% on the shared puzzles, and took the TypeScript record on Mars Lander with a 60-byte program that never reads the lander's state.

September 25, 20267 min read

A bee on a yellow flower head, wings folded, deep in the pollen.

Web AI

Notes on building a game in 27 hours

Let Them Bee is an incremental game I built for the DTJ36 jam in 27 hours: 29 commits, 10,743 lines of TypeScript, two runtime dependencies, and Claude Opus 5 on low reasoning effort as my only assistant. Where the time went, and why installing ESLint halfway through the jam was a good idea.

July 27, 20267 min read

A gold World Cup trophy next to an adidas match ball on the grass of an empty stadium, the stands a soft blur behind them.

I mixed LLMs into my friends' World Cup prediction pool

I dropped the PunditBench LLM predictions into my friends' World Cup pool, scored two ways. On raw scores DeepSeek V4 Flash edged everyone; under a betting-odds system only Grok 4.3 cracked the top five. The lesson isn't who won: the same market finishes 55th played naively and 20th played well, so turning a prediction into a good ticket matters more than the prediction.

June 28, 20266 min read

A single white golf ball resting on a close-cropped green, the fairway falling away soft and out of focus behind it.

AI Competitive Programming

Fable 5 shaves a quarter off Opus at code golf

I re-ran the 21 CodinGame code golf puzzles from the Opus post with Claude Fable 5. It wrote 24.7% fewer bytes overall, 10,712 down to 8,065: still above the best human TypeScript, but closing the gap fast.

June 13, 20264 min read

A close-up of an open pocket watch lying on a table, its gears and hands sharp against a soft dark background.

AI Competitive Programming

How close can an LLM get to a CodinGame optimization leaderboard?

I pointed Opus 4.8 at six CodinGame optimization puzzles. On the travelling salesman it tied the top of the leaderboard; on the open-ended games it reached a sixth or an eighth of the best score and stopped. The thing it's bad at is the CodinGame squeeze, not the solving.

June 7, 20266 min read

A single golf ball on a wooden tee, low in the grass, the rest of the course thrown out of focus behind it.

AI Competitive Programming

Can an LLM win at code golf in TypeScript?

I ran Opus 4.8 against 21 CodinGame code golf puzzles in TypeScript, from easy to hard. It never came close to first place, and usually didn't even beat the best human TypeScript. Here are the numbers.

June 1, 20266 min read

The London Eye gone quiet over the Thames at dusk, a minute before the lights came up.

Notes on Anthropic's Partner Basecamp

Two days at Anthropic's Partner Basecamp: the diagnostic loop, evals as a discipline, the three phases of AI inside a company, and an evening to rewalk London.

May 30, 20264 min read