Ex-Meta ML leader: tail models like Grok 4.5 handle 90%+ of everyday queries
jquinonero · x · 2026-08-16
Veteran ML leader Joaquin Quiñonero (formerly head of applied ML at Meta) shared his experience learning to prune an apricot tree with Grok 4.5 — succinct and to the point. His takeaway: the hardest use cases still need the strongest frontier models, but 90%+ of everyday questions are fine with tail models. We're in an era of constant Pareto optimization across cost, performance and latency, with many great options to choose from.
More from Models
- NVIDIA Confirms Continued Nemotron Iterations, RTX 5090 Support — thursdai_pod · 2026-08-16
- ChatGPT Plus user gets bizarre error: 'Thank You Very Much?' — Virtual-Complaint521 · 2026-08-16
- Former OpenAI Researcher Says Kimi Outperforms Grok — Miles_Brundage · 2026-08-16
- Opinion: Google should release a 120B Gemma model to disrupt the market — EducationalCicada · 2026-08-16
- Teutonic-I claims to be the most performant decentralized LLM — const_reborn · 2026-08-16
- Claude Opus 5 dominates KernelBench, hinting at even stronger internal model — scaling01 · 2026-08-16