Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy
Who chooses what AI gets to do?
Welcome to Import AI, a newsletter about AI research. In this edition, we explore three key stories: swarm scaling in AI, Google DeepMind's efforts to watermark AI-generated biology, and the capabilities of AI systems in partially automated scientific labs.
Swarm scaling refers to the use of multiple agents working in parallel to achieve faster results. Toby Ord, a researcher, explains that swarms can be seen as a new form of inference-scaling. While swarms require more tokens than single agents, their parallelization allows for completing tasks in less wall clock time. Ord notes that scaling up the number of agents in a swarm by 10x does not result in a proportional increase in performance, due to diminishing returns.
This concept is similar to managing large teams of human workers, where coordination incurs a "stepping on toes" tax. Ord's analysis suggests that swarms could potentially increase the likelihood of an RSI-driven intelligence explosion, rather than reduce it.
In other AI news, Google DeepMind has developed SynthID Bio, a watermarking system specifically designed to enhance biosecurity and scientific integrity in synthetic biology. The watermarking method adjusts its approach based on the type of data, such as tweaking amino acids in sequences or adjusting atomic coordinates for 3D structures.
DeepMind's testing across three target proteins demonstrated that watermarked designs maintain the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions. The goal of SynthID Bio is to prevent bioterrorism by creating a reliable signal for detection. This development is one of many efforts to ensure the responsible use of AI in biological research and applications.
Written by urgent.news from Import AI's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
