AGI scenario: 20,000 agents per 2,000 researchers could kick off recursive self-improvement
haider1 · x · 2026-09-10
Blogger haider1 extrapolates from OpenAI's reported 3.1 agent workdays per human researcher workday: with enough compute, 2,000 human researchers paired with 20,000 Astra-level agents doing AI research nonstop (a conservative estimate that could scale to 200,000) would cross the threshold where recursive self-improvement begins. The core constraint is compute supply.
More from AGI Musings
- Investor: falling behind on agent product cycles may be impossible to recover from — matt_slotnick · 2026-09-10
- Observers spot a coordinated "AI will kill us" talking-points push on X amid rapid progress — Dr_Singularity · 2026-09-10
- Forethought's Tom Davidson: data bottlenecks will slow, not stop, an intelligence explosion — connoraxiotes · 2026-09-10
- Experts Warn AI Could Kill All Humans Within a Decade — How Seriously Should We Take Them? — tw1st3d_m3nt4t · 2026-09-10
- OpenAI Researcher: I'd Prefer Shutting It All Down to Letting AI Rip — Just-Grocery-2229 · 2026-09-10
- Data eval practitioner pushes back on AI2027 panic: the capability frontier is jagged, not imminent superintelligence — geoffwolfe · 2026-09-10