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

Import AI highlights recent advances in AI scaling and application. Toby Ord’s analysis of swarm scaling shows that while a 4‑agent swarm requires roughly twice the total tokens of a single agent, the parallel execution can halve wall‑clock time, making swarms attractive for time‑sensitive tasks. However, Ord notes a diminishing returns effect—termed the “stepping on toes” parameter—where scaling the number of agents by tenfold yields only a 3‑ to 5‑fold performance increase, suggesting coordination overhead limits large‑scale swarm efficiency. In parallel, DeepMind’s SynthID Bio introduces a family of watermarking methods tailored for synthetic biology. The system modifies amino acid selection and adjusts atomic coordinates in predicted 3D structures to embed a detectable signal. Wet‑lab tests on VEGF‑A, the SARS‑CoV‑2 spike protein RBD, and PD‑L1 showed that watermarked designs matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions, indicating no loss of biological function. C5R Corp’s SciUniverse benchmark evaluates AI models on 92 tasks spanning chemistry, biology, and materials science. Claude Fable 5.1 achieved the highest pass rate of 45.3%, followed by GPT‑5 Astra at 32.5% and Claude Opus 5 at 30.5%. These results demonstrate that frontier models can perform a range of laboratory operations, though performance varies by task family. Finally, DeepMind proposes the creation of an Automated Scientific Economy to address the bottleneck of physical resources and empirical validation for AI scientists. The paper argues that a market‑based system would align scientific pursuits with scarce lab resources, make trade‑offs explicit, and incorporate public interest considerations. Together, these pieces underscore the growing integration of AI into scientific workflows and the accompanying technical and governance challenges.