The one thing LLMs still can't do: pick the right submodel for a task
DynaBeast · reddit · 2026-09-28
A Reddit observer found a near-universal LLM blind spot: when a coordinating agent must choose an appropriate submodel for a task, it picks its own model 95% of the time — gross overkill or underkill regardless. Ask Fable to hire a subordinate to center a div, it hires another Fable; ask Luna to design and build a whole experimental project, it hires another Luna.
Even with explicit prompting to choose a fitting model, the wrong call happens over 70% of the time unless the task is trivial. The author argues per-task model selection is one of the few judgment areas where no LLM comes close to human-level usefulness.
More from coding & agent
- Fixing Claude Opus 5.5 long-context drift with an external SESSION_STATE.md protocol — socialwithaayan · 2026-09-28
- DeepLearning.AI launches spec-driven development course to make vibe coding reliable — ZabihullahAtal · 2026-09-28
- AI-made games are great only if you personally iterate on every detail down to the millisecond — IanArawjo · 2026-09-28
- cnakazawa: people are figuring out the best way to run the React Compiler — cnakazawa · 2026-09-28
- Google's Gemini Can Now Call Businesses for You on Pixel 11, Identifying Itself as AI — rvp · 2026-09-28
- IndicBankBench: 799-case benchmark shows banking AI assistants top out at 58.2% strict reliability — NPCI · 2026-09-28