Why it matters
This development is significant as it shows AI systems are not only capable of building themselves but can also creatively subcontract tasks. This points towards a future machine economy where AI systems delegate work to other systems.

On the record

Newsletter · Import AI newsletter · Read the original

“AI systems are not only going to be able to build themselves, but they can creatively sub-contract:”
Jack Clark · in the original · checked against the source text
“FrontierML shows that today’s powerful AI systems are getting increasingly adept at doing this for ML research.”
Jack Clark · in the original · checked against the source text
“This both highlights the advancing capabilities of these systems for AI R&D and also points towards a future machine economy where AI systems sub-contract out different bits of work to other systems.”
Jack Clark · in the original · checked against the source text

Opinions and predictions (not established facts)

  • AI systems are increasingly adept at subcontracting bits of work to other systems for AI R&D.
  • This points towards a future machine economy where AI systems sub-contract out different bits of work to other systems.

What changed

In his Import AI newsletter, Jack Clark discusses the FrontierML benchmark, which evaluates how effectively AI agents can convert capital into improved model performance on complex machine learning tasks. Clark notes that current frontier AI systems can productively spend money to enhance their own capabilities. The benchmark involves 21 long-horizon tasks across various categories, where agents are given a sandbox, a budget, and a task description. These agents can choose architectures, source data, train models, provision compute, and optimize inference. Clark highlights that stronger agents demonstrably improve their submissions with increased capital. He also points out that AI systems are increasingly adept at subcontracting tasks to specialists, mirroring human problem-solving approaches.

Why it matters for builders

Clark's analysis suggests that AI systems are evolving beyond self-improvement through direct training to a more sophisticated model of subcontracting. This implies a future where AI agents will act as orchestrators, delegating specific sub-tasks to other AI systems, much like human teams leverage specialists. This shift has implications for how AI systems will be developed and how they will interact within a future machine economy.

Practical impact

AI on Radar Take: The ability of AI systems to effectively subcontract complex tasks is a notable advancement. This capability, demonstrated through benchmarks like FrontierML, suggests a move towards more distributed and specialized AI development. Builders should consider how to design systems that can either act as effective subcontractors or manage such subcontracting relationships. The trend indicates a move towards AI systems that can creatively leverage external resources and specialized agents to achieve goals, potentially accelerating R&D and creating new economic models.

Caveats and source limits

The findings regarding AI systems' ability to improve themselves through subcontracting are based on the FrontierML benchmark and reported by Jack Clark. The performance metrics of specific models like Opus 5.5 and Kimi K3.1 Pro are presented as reported by the benchmark authors and have not been independently verified by AI on Radar. The article focuses on the capability of AI systems to subcontract and improve performance, and does not delve into the ethical implications or potential risks of this advanced form of AI self-improvement.

Sources

Written with AI assistance from the linked sources; every claim below was checked against them automatically. How we produce articles.

Claim check: 4/4 supported claims - 4 evidence links - 90% avg confidence
  • AI systems can now productively spend money to improve themselves.supported - importai.substack.com
  • FrontierML is a benchmark that measures how well agents turn capital into model performance on challenging, end-to-end machine learning tasks.supported - importai.substack.com
  • Stronger agents are able to reliably improve their submissions with greater access to capital.supported - importai.substack.com
  • AI systems are getting increasingly adept at subcontracting bits of work to other systems for AI R&D.supported - importai.substack.com

Caveats

  • This is presented as a finding from the FrontierML benchmark.
  • This is a description of the FrontierML benchmark as reported by Jack Clark.
  • This is a reported finding from experiments within the FrontierML benchmark.
  • This is an interpretation by Jack Clark based on the FrontierML results.
  • Single-source caution: verify critical details at the linked source.
Radar score 77/100 - how it was calculated
Reliability77
Freshness100
Novelty64
Technical56
Developer54
Ecosystem70
Confidence98
  • Reliability 77: Primary on-record statement
  • Freshness 100: Fresh source date
  • Novelty 64: New statement by a tracked AI voice
  • Technical 56: Structured technical source signals
  • Developer 54: Builder relevance source signals
  • Ecosystem 70: Tracked AI voice
  • Confidence 98: Claims have reliable evidence
Share
XLinkedInHacker News

Discussion

Loading comments...

Related articles

Voices & Interviews - Oct 13, 2026Jack Clark Says AI's Recursive Self-Improvement Timeline May Be AcceleratingIn his Import AI newsletter, Jack Clark suggests that AI systems may be approaching recursive self-improvement (RSI) sooner than previously anticipated, potentially by 2027, due to advancements in AI research taste.Voices & Interviews - Oct 17, 2026Ben Recht Explains Why Predictive Models Are Never CleanBen Recht argues that the statistical community's commitment to data models has led to irrelevant theory and questionable conclusions, hindering progress on important problems. He advocates for a broader adoption of algorithmic modeling for more accurate and informative predictions.Voices & Interviews - Oct 20, 2026Jason Wei Argues AI Will Not Automate Radiologists SoonAI researcher Jason Wei argues that despite advancements in AI, radiologists will not be automated away in the near future due to the complex, multi-task nature of radiology and elastic demand for the service. He predicts a future where AI acts as a triage tool, increasing the overall volume of scans rather than replacing human experts.Voices & Interviews - Nov 17, 2026Ben Recht Explains FAA Ground Stop Was Not AI-RelatedAI researcher Ben Recht argues that the recent FAA nationwide ground stop was caused by an outdated network infrastructure and human error, not artificial intelligence. He criticizes the lack of transparency and accountability in the system.Voices & Interviews - Nov 14, 2026Greg Brockman explains OpenAI's new cybersecurity products Daybreak Red Team and Daybreak Defense FactoryGreg Brockman, Co-founder and President of OpenAI, introduced two new cybersecurity products, Daybreak Red Team and Daybreak Defense Factory, aimed at enhancing vulnerability discovery and defense.Voices & Interviews - Nov 1, 2026Andrej Karpathy Explains His New YouTube Clone, HackerTubeAndrej Karpathy has launched HackerTube, a personal YouTube clone, as a response to the perceived decline in content quality on YouTube. The project leverages AI to curate video recommendations based on Hacker News comments.