VizPy cracks analog placement at 97%

This feels less like prompt hacking and more like a compiler pass for LLM behavior: it extracts rules from failures, validates them, and keeps only the ones that survive. The caveat is scale: the benchmark is small enough that the result reads as a strong proof of concept, not a universal ceiling.

• –Analog placement is a brutal test because symmetry, routing, parasitics, and area all pull against each other.
• –The learned rules are refreshingly legible, with heuristics like column-based stacking and drain-pair alignment that engineers can inspect.
• –The dataset is only 9 circuits with a 6/3 split, and the expert gap is 0.018, so the 97% figure is exciting but still fragile.
• –The RL baseline overfit the training circuit while falling to 0.502 on test, which makes the prompt-optimization win look efficient rather than brute-force.
• –Product Hunt discussion suggests ContraPrompt will shine when you can express relative preferences, while PromptGrad needs numeric scores and a cleaner eval signal.