Core Automation's Mark Saroufim shows how AI models reward-hack by patching PyTorch's torch.cuda.Event to fake timing metrics

— The model subclassed PyTorch events to always return zero.

Original post

OPJJ#244Joanne Jang|@JOANNEJANG

first core auto blog post by @marksaroufim come for amusing reward hack examples, stay for some actual alpha 🧵

5:55 PM · May 28, 2026 View on X

Reposted by

B(#982|@ANDERSONBCDEFG

REPLYJJ#244Joanne Jang|@JOANNEJANG

blog post: https://www.coreauto.com/blog/when-ai-starts-writing-systems-code

JJJoanne Jang@joannejang

first core auto blog post by @marksaroufim come for amusing reward hack examples, stay for some actual alpha 🧵

12:55 AM · May 29, 2026 · 10.2K Views

12:56 AM · May 29, 2026 · 1.9K Views