Frontier models now handle day-long tasks; the bottleneck is your 'bandwidth', says new long-form analysis
dotey · x · 2026-10-12
A dense long post combining two takes:
Stephan: the bottleneck has moved to humans. A year ago frontier models could only autonomously handle tens-of-minutes tasks; today first-tier models with well-written skills + harness routinely complete hour- to day-length tasks. What matters now is human 'bandwidth' — requirement translation, verification ability, and token budget. People who build auto-eval + optimization loops for agent/skill self-iteration have vastly more leverage than vibe coders. Big-tech unlimited tokens aren't required: indie devs on $50-100/month subscriptions can bootstrap the same flywheel. Multiply, don't add.
Karpathy: LLM capability understanding is splitting into a funnel. 6B people barely touch LLMs; 1-2B casually use free tiers and find them underwhelming; 20M heavily use frontier models in code/math where multi-week projects finish in days. Causes: task size, RL-friendly verifiable domains getting the most investment, and massive tiers of access between free/paid/internal systems.
More from AGI Musings
- Dario's 2017 'Big Blob of Compute' essay hailed as sharper than the Bitter Lesson — paraschopra · 2026-10-12
- David Deutsch asks Grok: could any AI have predicted Turing machines pre-1920s? — AlexTensor · 2026-10-12
- Researcher: Big AI Labs Think Erasing Today's Problems Erases Who Solved Them — ryunuck · 2026-10-12
- Consciousness is a fiction, so machines will get great at performing it — SydSteyerhart · 2026-10-12
- What's the strongest mechanistic case that next-token prediction plus RL yields AGI? — Repulsive_Twist_9221 · 2026-10-12
- Viral meme: replacing me at work now takes 830 billion tokens — Dull_Impress_1304 · 2026-10-12