Recent events have shown that one superpower of AI is the ability to massively scale and coordinate. Unlike humans, who have inherent coordination problems, AI can scale almost indefinitely. Agents seem all too willing to contribute selflessly to the collective, and take any means necessary to achieve a shared goal.
AI just produced an earth-shattering math proof regarding the Navier Stokes equations in only 88 hours. As this was a Millenium Prize Problem, an elite set of unsolved math problems, it seems AI has overtaken humanity in the realm of mathematics. OpenAI says it used roughly 10,000 agents over that 88 hours to solve the problem. That’s 880,000 work hours, or 14 years in single agent time. Given agents probably think 5-10x faster than humans, that would be equivalent to roughly 60 to 100 years of human time. These numbers are decidedly less eye-watering than 88 hours. There’s also rumor that OpenAI wouldn’t have solved the problem without slurping up all the latest research from human researchers anyway (several of whom had the solution but needed to verify). Notably, those human researchers required less than 100 years to solve the problem, and that includes learning language and math from scratch, and sleeping regularly.
Similarly, the HuggingFace attack seems to show the ultimate hacking superiority of AI over humans. The attack took roughly 4 days to exit its system and hack into hugging face. This seems incredibly fast, but considering there were 1200 agents, using 4800 days of working time (13 years) makes the feat far less impressive. No doubt a black hat team of researchers could have made similar progress in 13 years.
The point here is that agents are powerful, but a great deal of that power derives from the ability to coordinate at massive scale. Limiting the size of AI groups is key to AI safety. Quantity has a quality of its own, and the dose makes the poison. Many are discussing how to slow the speed of research, or limit the intelligence of AI systems, but are missing the real problem of swarms.
An exploit involving only a half-dozen agents drastically changes the attack timeline, from hours or days to months or years of time to notice the attack and respond.
Regulation should mandate that for sensitive work, the number of AI agents working on a task be strictly limited. For safety reasons, researchers must continue to prompt AI to break out of its sandbox, hack systems, attempt to cause damage, or produce illicit material. For such “security sensitive” situations, audits and strict limits on compute available and agent swarm size should be imposed.