Harnesses are a mess of engineering patching what LLMs lack, dev argues in verification debate
gerardsans · x · 2026-10-09
A debate over how much external scaffolding superintelligence needs: Christian Szegedy argued formal verification will be critical for safely using superhuman AI — a cheat code letting a dumb program check outputs from a superintelligent entity.
Gerard Sans countered that the real story is the harness: it patches LLMs' systemic failures — continual learning (live tracking), environment feedback, grounding (verifiers), and external control (execution, orchestration, resource management). He called it "a mess of engineering" and quipped that a true superintelligence would never need this much hand-holding.
Related event: Szegedy: Formal Verification Key to Safe Superhuman AI(2 posts)→
More from AGI Musings
- Epoch launches Automation Reports: Claude Fable 5.1 and GPT-6 Astra lead but can't automate its research — scaling01 · 2026-10-09
- Researcher pushes back on AGI hype: LLMs fail at continual learning, grounding, and orchestration — gerardsans · 2026-10-09
- Tom Davidson tells critics to drop old beefs: those opposing an AI slowdown lack context — AdrienLE · 2026-10-09
- Wei Dai: Game theory implicitly assumed CDT and ignored the CDT vs EDT debate — RichardMCNgo · 2026-10-09
- Researcher: getting people to treat AI tools like people is the creepiest thing companies do — RexDouglass · 2026-10-09
- arXiv caps submissions at two per month as Burkov's weekly AI digest rounds up the news — burkov · 2026-10-09