Have we seen an acceleration in discoveries? Cyber spikes, math rises, algorithms flat
soumitrashukla9 · x · 2026-09-20
testingham and Nate Rush analyze whether AI has accelerated discoveries, with tentative conclusions:
- Cyber: sharp acceleration
- Math: some acceleration
- Algorithms: no clear acceleration yet
They argue evaluation should focus on final outputs (exploits discovered, math problems solved, algorithmic efficiency) rather than inputs (time, token spend) or intermediate proxies (papers, lines of code, commits, self-reported speedups), which are hard to interpret—code lines can explode with no change in value. Final outputs are harder to experiment with and attribute to AI, but in some domains the trend break is so large that AI is the obvious cause.
More from AGI Musings
- Frontier model outputs now need another model to read them, insider says — DavidSKrueger · 2026-09-20
- A single web search makes LLM inputs OOD, argues developer — menhguin · 2026-09-20
- Two types of daily AI users: becoming more slop or more human — StewartalsopIII · 2026-09-20
- Spotify reframes recommender systems for the agentic web in weekly IR papers roundup — _reachsumit · 2026-09-20
- AI moral status debate: does conceding machine suffering imply utilitarian math? — teortaxesTex · 2026-09-20
- Consumer agents are killing incumbents' 'lazy margin' — and that favors startups — jeff_weinstein · 2026-09-20