Writing software while sleeping to predict NFL games.
An AI agent, a Docker sandbox, and 53 minutes later: a working ML microservice.
Conference Championship predictions:
• Broncos over Patriots (100%)
• Seahawks over Rams (94%)
https://www.larsgregori.de/2026/01/20/nfl-ame-predictor-with-ralph-wiggum/
#VibeCoding #MachineLearning #NFL
AI writes code. AI writes tests. What could go wrong?
Mutation testing result: 85% of bugs caught.
After hiding the implementation from the AI: 92%.
Property tests found 11 real bugs all others missed.
https://www.larsgregori.de/2026/01/08/mutation-testing-the-vibes/
"Just use a different LLM for tests."
Doesn't work.
The moment any LLM sees your code, it reasons about what it does, not what it should do.
Independence needs different information, not different systems.
https://www.larsgregori.de/2026/01/06/independent-verification/
The T in vibe coding stands for testing.
At 39C3, Johann Rehberger () showed how AI coding assistants can be hijacked through prompt injection. He calls compromised agents "ZombAIs."
The attack surface exists because we stopped looking. When the AI writes code AND tests, that's not verification. That's confirmation.
@codepo8 don't worry, this calculation feels good and the 'F' in Fahrenheit stands also for feeling. We just name them Altman numbers and the -0.31 is correct.
After trying to explain some LLMs to implement the Altman Subtraction I found a solution: https://gist.github.com/choas/e3f6cad89bdb399a39c15ef75893202f
