Why AI Progress Is Increasingly Invisible
Source: https://www.garrisonlovely.com/why-ai-progress-invisible Author: Garrison Lovely Date: 2025-01-10
Summary
Argues that AI capability improvements have become harder to perceive, even as they continue. The benchmarks are saturating, the marginal improvements are in harder-to-measure areas, and the visible products aren’t changing as dramatically.
Key Claims
- Benchmark saturation: when models score 90%+ on academic benchmarks, each percentage improvement is less visible even if it represents hard capability gains.
- The “useful tail” problem: frontier improvements often come in edge cases — longer tasks, harder reasoning, more subtle errors — that most users don’t encounter.
- Communication failure: AI labs communicate progress via benchmark scores that users can’t interpret; the “wow” demos are harder to produce as quality baselines rise.
- Invisible progress is still progress: the fact that GPT-4-turbo and its successors don’t produce dramatically different impressions doesn’t mean capability stagnated.
- Risk: invisible progress may be taken for granted until a capability jump suddenly becomes visible through a new application.
Concepts
- Scaling & Compute — how to measure progress when benchmarks saturate
- Test-Time Compute — reasoning improvements are particularly invisible (thinking happens “offscreen”)