AI Systems Poised to Start Building Themselves, Says Jack Clark
Key Takeaways
- ▸Claude's coding performance improved from 2% to 93.9% on SWE-Bench in less than 18 months, effectively saturating the benchmark
- ▸Jack Clark estimates 60%+ probability of fully automated AI R&D by end of 2028, representing a fundamental shift in how AI systems are developed
- ▸AI systems now demonstrate ability to chain multiple coding tasks together iteratively, a key prerequisite for autonomous AI development
Summary
In a significant analysis published in Import AI, researcher Jack Clark argues that we are approaching a critical inflection point where AI systems will become capable of autonomous research and development, potentially building their own successors. Clark estimates a 60%+ probability that fully automated AI R&D could occur by the end of 2028, marking what he describes as crossing a "Rubicon into a nearly-impossible-to-forecast future."
The evidence for this projection centers on dramatic improvements in AI coding capabilities, exemplified by the SWE-Bench benchmark where Claude has progressed from a 2% success rate (Claude 2, late 2023) to 93.9% (Claude Mythos Preview). Clark argues that all the engineering components necessary for automating AI development are already in place, and that continued scaling improvements should enable AI systems to not only replicate human AI development workflows but potentially contribute novel research ideas, fundamentally changing the trajectory of AI progress.
Clark emphasizes that while he expects to see proof-of-concept examples of models training their successors within 1-2 years, the full transition to automated AI R&D represents a civilizational inflection point with massive implications that society may not be adequately preparing for.
- This represents a critical inflection point that could lead to recursive self-improvement and unpredictable AI capabilities trajectories
Editorial Opinion
Clark's analysis, rooted in public benchmarks and deployed AI capabilities, presents a compelling case that automated AI R&D may be considerably closer than widely assumed. The exponential improvement in coding-related benchmarks suggests the engineering prerequisites for self-improvement are rapidly being met. However, the profound implications—including questions around safety, alignment, and whether such systems will emerge with adequate safeguards—demand urgent attention from researchers, policymakers, and the broader public.
