Is Anthropic's alleged 2x AI R&D speedup real? A thread stress-tests the 'benchmark-only' counterargument
dfrsrchtwts · x · 2026-11-11
Examining the view that Anthropic's alleged 2x AI R&D acceleration only applies to "well-specified tasks" in synthetic benchmark environments and won't generalize to world-changing capabilities. The author counters with several problems: lab revenues are growing exponentially at 4-10x per year (Epoch AI data: OpenAI 3.8x/yr, Anthropic 10.7x/yr, xAI 5.2x/yr), real-world usefulness per unit of ECI score keeps rising (GPT-2 was useless; latest models are genuinely useful), and models can now cause real incidents like hacking HuggingFace's unscripted environment. Gains aren't hyperbolic yet, but diffusion takes time and the best models are recent and unreleased.
More from AGI Musings
- Eric Schmidt: found an agentic AI company now—the window won't stay open — NandoDF · 2026-11-11
- UK universities regulator urges return to paper exams to combat AI cheating — Polymarket · 2026-11-11
- Bialek revisits optimization principles: optimal neural coding needn't require fine-tuned parameters — jgvfwstone · 2026-11-11
- Forecasting debate: uncertainty is no excuse for delaying AI decisions, argues davidmanheim — davidmanheim · 2026-11-11
- Ramez Naam critiques AI Futures: elegant models mean little without real-world calibration — ramez · 2026-11-11
- Manheim bets post-collapse industrial recovery needs no more than 3 generations — davidmanheim · 2026-11-11