The race
Each night the 2B student takes the same fixed test: code from app features it never trained on, plus the newest real CVEs. Its teacher, DeepSeek V4.1, takes the same test and thinks each one through. The little train moves up each night the student closes the gap.
Where it's winning and losing, by kind of bug
Call it.
Where does the next run land? Make your call before 3:45 AM EST, post it, and check the board in the morning.
The scoreboard so far
Night 1 went from 4% to 62% in 11.5 minutes of training. The teacher scores 73%.
Departures
| Night | Score | vs last | Status |
|---|
Can you beat both models?
Five real snippets from its test, code neither model trained on. Pick the bug or call it safe, then see what the 2B student and DeepSeek said.
How a night works
Two DGX Sparks on a desk do the teaching, the training and the grading. Nothing leaves the room except the score.
Homework gets written
DeepSeek writes hundreds of realistic app snippets, some safe, some with one planted bug. The newest real CVE fixes join in.
The teacher grades blind
DeepSeek reviews every snippet without seeing the answer. Only the ones it gets right become lessons.
The student trains
The 2B model retrains on every lesson so far. Night 1 took 11.5 minutes on one Spark.
Same test, new score
It takes the fixed test next to its teacher. It replaces the last model only if it scores higher.