Justin Lebar
· SemiAnalysis
· May 28, 2026 at 00:22
· ⏱ 9 min read
| Read on Substack ↗
Summary
The article describes how the author used AI agents (Opus, ChatGPT) to fuzz compilers (LLVM, NVIDIA's ptxas, AMD's GPU backend) and found hundreds of bugs, including a critical atomic-store miscompile. It argues that AI-powered bug hunting is now feasible at high cost ($10,000+ per session) and that this capability gap will widen, but it offers no financial trade ideas or market implications.
•The author spent over $10,000 in a single afternoon running AI subagents (Claude Opus 4.7) to read LLVM source code and find bugs, yielding one bug every four minutes.
•A vibe-coded fuzzer using ChatGPT 5.5 found 80+ miscompile reproducers in NVIDIA's ptxas and a similar rate in AMD's GPU backend within days.
•One critical bug found by agents: LLVM would turn an atomic store into two non-atomic stores, causing silent data corruption in production.
•The author found bugs in LLVM's x86 backend at a rate of nearly two per minute via code-reading agents, with no signs of slowing down.
•The fuzzer for ptxas was cheaper (~$200/month ChatGPT Pro plus ~$1,000 for Opus tokens) than the agent-based approach, which cost $10,000+ in hours.
•SemiAnalysis spent an order of magnitude more on AI tokens than on the author's salary the day the agents ran, marking the first time the author felt the AI delivered more value than their own labor.