Agent Swarms Speed, Not Smarts

Early experimental evidence suggests that running many copies of present‑day AI agents in parallel chiefly accelerates task throughput (speed) rather than producing sustained, compounding jumps in problem‑solving ability (intelligence). Multi‑agent gains often show rapid diminishing returns on the hardest, conceptual problems that would be needed for recursive self‑improvement to take off. — If true, this undermines popular claims that simply scaling agent counts (millions/billions) is a viable route to near‑term superintelligence and should shift policy and investment from panic about takeoff to governance of scaled deployment and safety testing.