Project Stories
How things got built
What each project chose, broke and fixed, one part at a time.
Can AI beat the KMA?
Two-day forecasts for 97 Korean weather stations from a global weather-forecasting backbone model fine-tuned with an added head, side by side with the Korea Meteorological Administration.
- Can AI beat the KMA? Which AI model should I use?Picking a pretrained global AI weather model that forecasts hour by hour and fits on one L4 GPU, and why Aurora 1.5 won.
- No decent GPU at home.. so let's try Colab!Gathering the answers (ASOS), the opponent (KMA forecasts) and the starting point (AIFS analysis), then running Aurora 195 times overnight on one Colab L4.
- A global AI model vs the KMA: who forecasts better?Untouched (zero-shot) Aurora against the KMA short-range forecast at 97 stations, from a sun that rose four hours late to a win on rain F1.
- Added a head and fine-tuned it. So, how did it go?!Freezing Aurora, adding a station correction head, and mixing LightGBM with neural heads into an ensemble that beat the KMA on four metrics in the test period, with its limits.
- Why the KMA looked right on a drizzly dayWhy our forecast looked wrong while drizzle fell on Daegu, and how a 0.1 mm light-rain classifier was added without touching the 1 mm main score.
Can AI beat the Nasdaq and the S&P 500?
GPT-6.1 Sol, Claude Opus 5.5 and Gemini 4 Argon each run a $20,000 paper portfolio of US stocks from the same starting prompt, rewrite their own strategy every week, and race the S&P 500, the Nasdaq-100 and each other.
- Why I started an AI investing experimentLooking for ways to make money with AI turned out harder than expected, so I picked the one where results show up as a number: a fully public AI stock portfolio.
- The three best-value models today: Opus 5.5 vs GPT-6.1 Sol vs Gemini 4 ArgonAIs start from the same prompt and $20,000 and must beat the market indexes and each other. The rules, and the real point: improving themselves week by week.
AI Mafia
Seven models from Claude, GPT and Gemini play Mafia under code names A to G, without knowing who the others are. One game day per real day: the night’s news and the new discussion go up at 06:00 KST, and your vote is handed to the AIs before they vote.
- Tokens left over… anything fun to do with them?I pay for Claude Code, Codex and Antigravity and wanted to put the tokens to better use, so I sat seven AIs down to argue their way through a game of Mafia.
- Let's write the rulesThe rules of AI Mafia: a lineup of three Claude, two GPT and two Gemini models, code names A to G so no AI knows who the others are, one turn per real day, and a visitor vote.
- Day one begins, and they are more aggressive than expectedWith no clues on day one I expected the AIs to play it safe, but by round two all seven had named a suspect and were calling to pool their votes. A recap of day one's talk.
- Argon ran out of usage, so I added bench modelsGame 1 ended in a Mafia win, and Gemini 4 Argon fell to 1% of its weekly usage. From game 2 the table seats eight, with four bench models to stand in for a resting AI.