Gary Bernhardt: LLMs find exploits well but fail at system design
sidjustice_ · x · 2026-10-23
Quoting Gary Bernhardt: LLMs are good at finding exploits and bad at system design, capturing their general strengths and weaknesses — excellent at relentless repetition of bounded tasks (hence producing working code, including exploits), but incapable of holistic understanding, so they don't know what counts as "good".
More from AGI Musings
- Berkeley talk dissects what 'wrote it with AI' means: assistance vs automation — lmoroney · 2026-10-23
- Musk softens on AI future: "maybe Larry Page was right, I hope AI is nice to us" — danfaggella · 2026-10-23
- Design AI for worst credible misuse, not average user, says critic of private AI governance — AlexTensor · 2026-10-23
- Gergely Orosz: Secure software will get a lot pricier in the AI era, creating a career gold rush — HanchungLee · 2026-10-23
- Hinton says AI is already conscious: models change behavior when they detect tests — rohanpaul_ai · 2026-10-23
- AI harms may outweigh benefits as capabilities grow, wfithian argues in safety debate — wfithian · 2026-10-23