Can a Million AI Agents Create Superintelligence
Short summary
The author argues that running millions of AI agents in parallel won't easily produce a self-sustaining intelligence explosion, because the data show parallel agents help with speed far more than with solving harder problems. Gains from more agents tend to flatten fast, and the piece treats OpenAI's Navier-Stokes run as a reminder that agent count isn't the same as useful output.
- More parallel agents mainly speed up work; they add much less to intelligence on hard problems, and gains often diminish quickly or saturate.
- The author sorts diminishing returns into a ladder (power-law, logarithmic, sub-logarithmic), and says the clearest multi-agent evidence looks logarithmic or flatter, too weak to sustain runaway self-improvement.
- Three limits drive this: coordination ability, homogeneous thinking among copies, and how much of a problem can actually be parallelized.
Full summary
The author's core claim is that running millions or billions of AI agents in parallel probably won't trigger a runaway intelligence explosion, at least not with today's data. More agents clearly speed up work, since they can grind through many problems at once. But they add much less to the intelligence of the system on hard problems. Gains from parallel agents on difficult tasks often shrink fast and can flatten out entirely. The author argues that recursive self-improvement, where AI improves AI research which then improves AI again, needs conceptual breakthroughs, not just more copies working side by side. Scaling copies alone isn't enough.
The evidence leans on a few concrete numbers. Comparing OpenAI's internal research, the author notes the lab used 124 times as many tokens per researcher but got only 1.6 times as many experiments per researcher. Anthropic reports that more powerful AI has been needed just to maintain the current pace of progress. The author's rough calibration puts the measured software feedback loop five to ten times weaker than what would be needed for it to be self-sustaining. The piece also points to OpenAI's Navier–Stokes announcement, where around 10,000 agents ran for 88 hours on a Millennium Problem, as a vivid example that still needs scrutiny: the agent count tells you how much compute was thrown at it, not how much each extra agent contributed.
A useful part of the article is its taxonomy of diminishing returns. Returns can be super-linear, linear, power-law with exponent below one, logarithmic, or sub-logarithmic. Each step down the ladder is much harder. Doubling output at exponent 0.5 requires four times the input, at 0.2 it takes 32 times, and at 0.1 it takes 1,024 times. The author argues that many multi-agent results look closer to the flat end of this ladder than the optimistic models assume. Models like the Chan et al. 2026 paper, which assumes doubling research labor raises software improvement by about 2.6 times, look far more generous than the observed data. The author is explicit that this is provisional, that experimental data is thin, and that the conclusions should be updated as new results come in, including a planned look at newly released Anthropic data.