Three Losses
- Part 1I thought I was 89th. I was 265th.
I lost Samsung SCPC 2026. My score climbed from 0.11 to 0.85 in ten days; the cutoff climbed faster. Two hours before the deadline I found out the rank I had been reading was wrong.
- Part 2I built a bot that collects draws
I placed 447th at NYPC 2026. I had a perfect 8/8 on the sample problems. Over eight days I shipped twenty-six bots and my rating gained nothing. That is what it costs to start without reading the reward function.
- Part 314 days of tuning gained 0.05. Two days of research gained 0.08.
Seventeen days took me from 0.138 to 0.90052, and 41st of 130. Three weeks of tuning gained less than the last two days of research. That is the price of never asking one layer up from the variable I was tuning.
- Part 4Two days after we called it over, I was 30th
A competition we had closed at 41st came back at 30th after top teams were disqualified. Twenty-three hours to the deadline, and the best model's weights did not exist anymore. How the model was retrained in twelve and a half hours and the report, code, and external validation went in.
- Part 4After three losses I didn't build a better model. I built a checklist.
305th, 447th, 41st. Putting the three post-mortems side by side turned up one thing in common: all three losses were decided on day one.