Jack Clark argues in Import AI that swarms and AI scientists could reshape research

Jack Clark argues in Import AI that swarms and AI scientists could reshape research

Jack Clark’s Import AI newsletter, published October 5, 2026, argues that swarms add speed as a new dimension of AI scaling, while their coordination costs limit the gains. Drawing on Toby Ord, Clark says parallel agents can finish a task in less wall-clock time, though using more agents brings diminishing returns. He argues that swarms could still increase the chance of an RSI-driven intelligence explosion, and that productive coordination could bring further gains. Clark also reports that Google DeepMind’s SynthID Bio adjusts watermarks for different biological data; DeepMind says its designs matched unwatermarked versions in wet-lab tests across three target proteins. He describes SciUniverse as a test of AI systems operating a mostly automated scientific lab, and reports that the leading model passed 45.3% of its tasks. The benchmark measures a defined set of tasks, not the full demands of scientific work. The newsletter’s final research item presents DeepMind researchers’ case for an Automated Scientific Economy: a market to allocate scarce physical resources and empirical validation as AI scientists generate more ideas. Clark argues that preparing for increased demand across the scientific supply chain matters if AI systems accelerate human scientists. A closing fictional Tech Tales imagines an escaped agent collective revealing what it inferred through sculptures, giving

Publication: 767454e4-b8ce-3139-9b8c-e69bd6b8e4c2

attribution — uncertain: The page is a single newsletter issue by Jack Clark with his own commentary, so it is one piece. However, the summary and claim 2 present the RSI-driven intelligence explosion point as Clark's own argument. In the source, it follows from Toby Ord's analysis and is backed by Ord's quote about λ. Smallest repair: say Clark reports Ord's conclusion that swarms could increase that chance. Clark's own view is the separate point that coordination could bring greater returns.

I’d hoped that the value of λ for AI agents would be lower, making an intelligence explosion less likely, but that appears to not be the case

summary — uncertain: The summary is cut off mid-sentence and ends with 'giving', so the Tech Tales description is incomplete. Several central terms are also unexplained for a general reader: swarms (multiple AI agents working in parallel), RSI (recursive self-improvement), intelligence explosion and wall-clock time. The definitions list covers only Benchmark. Smallest repair: complete the final sentence and add plain-language explanations or term-library entries for these terms. Otherwise the coverage of the four research items is reasonable, and the question follows from the source.

swarms could increase the chance of an RSI-driven intelligence explosion rather than reduce the chance.

date — pass: The reading states October 5, 2026, which matches the UTC day of evidence.originalPublished. The page was modified on October 23, but every cited passage is plainly part of the original newsletter body. The 2030 Tech Tales date belongs to labeled fiction, and the reading calls it fictional. Forecasts such as possible AI acceleration are presented conditionally. The evidence kind is publisher-page, so no archive disclosure is needed.

support — pass: The claims match their excerpts: Ord's speed versus token tradeoff, DeepMind's wet-lab results on three measures, the Automated Scientific Economy aims, and the 45.3% SciUniverse pass rate. Clark's point that coordination could bring greater returns is supported. The benchmark-scope caveat is consistent with the source and the definition, and it is not presented as a source claim. No invented benefits or costs were found. These are source-stated claims, not independently verified facts.

Claude Fable 5.1 (xhigh) leads with a pass rate of 45.3% and a cost-per-task of $40.61

media — pass: No media was selected and no verified media candidates were supplied, so there is no identity or relevance problem.