Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy

Toby Ord published a short post analyzing AI swarms as a form of inference-scaling. He writes that a 4-agent swarm needed about twice the total tokens to match single-agent performance, but only half as many tokens per agent, so it can theoretically complete the same task in half the time. He also describes a diminishing-returns effect he calls the "stepping on toes" parameter, comparing swarms to human teams: scaling agents 10x yields only 3x to 5x the performance, and the shortfall accumulates at larger scaleups. Ord states he had hoped the value would be lower, making an intelligence explosion less likely, but that appears not to be the case.

Polling from the Center for Shared AI Prosperity (CSAIP), with a sample size of 2498, found that 61% of Americans consider voluntary corporate self-policing on AI "not enough," including 53% of Trump voters. Additionally, 54% of voters (61% Harris, 48% Trump) said the government should set and enforce AI rules. The polling follows voluntary self-policing commitments announced by the Trump administration and companies including Anthropic and OpenAI.

Google DeepMind developed SynthID Bio, described as a family of watermarking methods for synthetic biology intended to strengthen biosecurity and scientific integrity. The approach selects different amino acids for sequences and adjusts atomic coordinates for predicted 3D structures. DeepMind writes that in wet-lab testing across VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1, watermarked designs matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions.

C5R Corp introduced SciUniverse, a benchmark with 92 tasks across 17 task families covering sample preparation, instrument control, protocol adaptation, learning across experiments, facility management, and interpretation of real measurements. Tasks span chemistry, biology, and materials science. Claude Fable 5.1 (xhigh) leads with a 45.3% pass rate at $40.61 per task, followed by GPT-5 Astra (xhigh) at 32.5% and $52.37, then Claude Opus 5 (xhigh) at 30.5% and $46.31.

DeepMind researchers published a paper proposing an Automated Scientific Economy, arguing AI scientist development will likely be bottlenecked by physical resources and empirical validation rather than idea generation. They propose four components: proof of ideation, ex-ante evaluation, brokerage and trade, and validation payout.