Import AI 455: AI systems are about to start building themselves.
Credible expert forecasting of automated AI R&D by 2028 sharpens the urgency of AI governance frameworks that assume meaningful human oversight in the loop.
Key points
- Jack Clark estimates 60%+ probability of fully automated AI R&D—no human involvement—by end of 2028.
- Benchmark evidence shows AI task horizons expanding from 30 seconds in 2022 to 12 hours in 2026, with 100-hour tasks projected by end of 2026.
- The analysis is expert opinion and probabilistic forecasting, not confirmed policy or technical fact—treat as informed signal, not settled consensus.
Summary
Import AI founder Jack Clark argues, based on public research benchmarks and observed capability trends, that there is a greater than 60% chance that AI systems capable of autonomously training their own successors will exist by end of 2028. He marshals evidence from SWE-Bench coding saturation, METR task-horizon data showing rapid expansion of autonomous AI work duration, and progress on scientific reproducibility benchmarks. Clark frames this as a potential 'Rubicon' event with consequences difficult to anticipate, and notes that nearly all engineering components of AI development can already be partially automated. The piece is analytical and speculative rather than reporting a completed development.
Implications for Australian agencies
- Monitor AI governance and strategy teams may want to monitor the 'automated AI R&D' capability trajectory, as it directly challenges human-oversight assumptions embedded in current APS responsible AI policy.
- Consider Agencies developing or reviewing AI risk frameworks could consider whether existing oversight and accountability models remain adequate if AI systems begin meaningfully substituting for human researchers in iterative model development.
Implications are AI-generated. Starting points, not advice.
"Import AI 455: AI systems are about to start building themselves." Source: Import AI – Substack (Jack Clark) Published: 4 May 2026 URL: https://importai.substack.com/p/import-ai-455-automating-ai-research Import AI founder Jack Clark argues, based on public research benchmarks and observed capability trends, that there is a greater than 60% chance that AI systems capable of autonomously training their own successors will exist by end of 2028. He marshals evidence from SWE-Bench coding saturation, METR task-horizon data showing rapid expansion of autonomous AI work duration, and progress on scientific reproducibility benchmarks. Clark frames this as a potential 'Rubicon' event with consequences difficult to anticipate, and notes that nearly all engineering components of AI development can already be partially automated. The piece is analytical and speculative rather than reporting a completed development. Implications for Australian agencies: - [Monitor] AI governance and strategy teams may want to monitor the 'automated AI R&D' capability trajectory, as it directly challenges human-oversight assumptions embedded in current APS responsible AI policy. - [Consider] Agencies developing or reviewing AI risk frameworks could consider whether existing oversight and accountability models remain adequate if AI systems begin meaningfully substituting for human researchers in iterative model development. Retrieved from SIMS, 18 May 2026.