Pedro Alonso RSS · Jul 10, 2026

I trained a 3B model to rival Claude Opus at text-to-SQL on one GPU. Here is exactly where that holds, and where it breaks