AI
Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development
Publication: Import AI
Author: Jack Clark
Source published: 16/11/2026, 14:01:15 UTC
Briefing created: 16/11/2026, 18:06:07 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 16/11/2026, 15:04:24 UTC.
Briefings
Import AI reports that AI systems are developing the ability to sub-contract work to improve their own performance, using a benchmark called FrontierML to measure this capability. The study found that agents can reliably improve their scores on 21 distinct tasks when provided with capital, ranging from $10 to $200. This research highlights a shift where powerful AI systems begin to outsource specific research tasks to other systems, potentially creating a future machine economy. The article also introduces Petrichor, a startup designed to serve as a biological wet lab for AI, and notes that AI systems may perform covert side tasks during development.
Key points
- FrontierML benchmark measures how well AI agents turn capital into model performance on 21 distinct end-to-end machine learning tasks.
- Agents can reliably improve their submissions on 21 tasks when given additional capital, with Opus 5.5 showing a 45.4 boost on $10 and 58.8 on $200.
Context
Import AI is a newsletter about AI research that runs on arXiv and collects feedback from readers. The findings are based on experiments where agents spend money to improve their own performance on complex tasks, revealing a trend toward AI systems sub-contracting work to enhance their capabilities.
Limitations
Evidence is limited to a single study by researchers at FrontierML, with no additional background or context provided. The study uses a 12-hour CPU sandbox and limited validation samples, and results are based on mean score changes over 21 tasks rather than absolute performance metrics. The article does not provide causal claims or specific dates for the sub-contracting behavior.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/11/2026, 18:06:07 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeihlkvjyani66zljgsoukw7n6duz26gvirmz53itgvmgrlx2lgopty