Applicability is the Next Bottleneck for AGI: Knowing When to Use Capabilities
joecole · x · 2026-08-17
This post explores the distinction between "capability" and "applicability." The author argues that as base models saturate benchmarks but still make mistakes in practice, the real bottleneck is "applicability"—the skill of knowing when to deploy which capability. This boundary is highly contextual and likely a learned policy rather than a simple rule. The author draws an analogy to the router in Mixture-of-Experts models, suggesting that learning when to activate a capability is harder than acquiring the capability itself.
More from AGI Musings
- Trade skills may win in the AI era — ___Mufasaa · 2026-08-17
- Bridgewater Execs Warn Unreleased AI Models Could Cause Significant Damage, Urge Preemptive Action — austinc3301 · 2026-08-17
- Ex-Google exec: AI recursive self-development to disrupt math and intelligence apex — rohanpaul_ai · 2026-08-17
- Opinion: Anthropic Model 2 gains underwhelming, open source gap narrows — ricklamers · 2026-08-17
- From Manual to FSD: Solving Real-World AI vs Upgrading Drivetrains — EngineeringInsd · 2026-08-17
- Treat AI as the smartest person you know — jdjohnson · 2026-08-17