7 hours ago
- Proof of work relies on guaranteed success with enough computational resources, unlike bug detection in code.
- Bug detection in LLMs is limited by model intelligence, not just computational effort, due to saturations of code states and sampling paths.
- The OpenBSD SACK bug demonstrates that weak models cannot find it; they only hallucinate or pattern-match without true understanding.
- Future cybersecurity will depend on better models and faster access, not merely more GPUs.