Ramez Naam pulls together data from multiple experiments to show that adding more AI agents delivers far more speed than it delivers intelligence, and the intelligence gains taper off with striking steepness.
Many have imagined that superintelligence could be triggered by packing millions of AI researchers into a data centre and letting them work together. Ramez Naam set out to see whether the experimental evidence so far supports that picture.
Speed is easy, getting smarter is hard
When a few dozen agents work together, the time it takes to finish a task often drops by two or three times. But when the goal is solving harder problems, things get far more difficult.
On tasks like cybersecurity and formal mathematical proofs, every additional batch of agents brings smaller score gains than the last. A hundred agents often earn only a little more than ten.
Even in OpenAI's solution to Navier-Stokes, which used ten thousand agents, researcher Noam Brown was blunt: at most ten percent of the credit belongs to multi-agent collaboration; the rest comes from the model itself getting stronger.
A stronger model beats ten times the numbers
On the same set of maths problems, ten agents built from a newer model outperform a hundred built from an older one. Swapping in a smarter model does far more than multiplying copies of the old one.
This points to an important fault line in the superintelligence debate. Self-improving AI would need to make one conceptual breakthrough after another, yet these multi-agent experiments keep showing intelligence gains so weak they struggle to even reach logarithmic growth.
Several forecast models built around recursive self-improvement assume that doubling the number of researchers produces twice the research output or more. The real-world results from adding AI agents sit orders of magnitude below those assumptions.
The closing judgement
As of now, the data does not support the idea that copying today's models a million times over will build a superintelligence. If the conceptual breakthroughs stay out of reach, adding more parallel workers may just be pushing harder against a smaller and smaller gain.
Curated from high-quality sources, with concise summaries and key takeaways.
