Swarm hype critique: a loop is not AGI — frozen models just sampled more
gerardsans · x · 2026-09-26
A long-form rebuttal of the "10,000-agent swarm" narrative, arguing labs are selling software engineering as model progress.
- The model is frozen: a swarm is the same distribution sampled many times; what changed is search policy and budget, not intelligence.
- Research under equal conditions shows a single model beats multi-agent setups nearly every time — context drift and coordination overhead explode with agent count.
- The ladder: Chat (one sample, no state) → Agent (a loop with tool calls, bounded autonomy) → Swarm (many runs, role-play autonomy). The loop dimensions move; the distribution does not.
- Math is a closed domain with cheap verifiers where search works; cybersecurity is open and adversarial where it goes in circles. Hype generalizes the closed case.
- The Navier-Stokes result came from mathematicians narrowing the problem for years, not the swarm.
Suggested scrutiny questions: are failed trajectories persisted? How good is the verifier? What's the reward signal? Who pointed the search?
More from AGI Musings
- AI incidents are often human failures — worth more discussion than doomerism — basedjensen · 2026-09-26
- Scholars Debate: Did AI Break Peer Review, or Was It Broken All Along? — RexDouglass · 2026-09-26
- Robots could reach the Moon years before humans, building the infrastructure we need — r0ck3t23 · 2026-09-26
- tqbf leaves Fly.io as Simon Willison calls tech's year '100 years at once' — mhmazur · 2026-09-26
- mitsuhiko: Austria adopted cars 30 years late — temper your AI rollout expectations — mitsuhiko · 2026-09-26
- Bostrom's anti-anthropomorphism advice isn't the same as saying LLMs can't think — maksym_andr · 2026-09-26