Import AI 455: Automating AI Research
Summary
An analysis of current AI progress suggests a 60%+ likelihood of fully automated AI R&D, where an AI system autonomously builds its successor, by the end of 2028. This projection is based on public information, including arXiv papers and products from frontier AI companies. Evidence highlights AI systems' rapidly increasing capabilities in coding, exemplified by Claude Mythos Preview achieving 93.9% on SWE-Bench, and their ability to complete tasks requiring up to 12 hours of human effort, as measured by METR. AI is also demonstrating proficiency in core scientific skills like replicating research (CORE-Bench solved by Opus 4.5 at 95.5%), building ML systems (Gemini3 at 64.4% on MLE-Bench), kernel optimization, and fine-tuning language models (Opus 4.6 and GPT 5.4 achieving 25-28% of human uplift on PostTrainBench). Furthermore, AI systems are showing early signs of managing other AIs and contributing to scientific discovery, such as solving Erdos math problems and assisting in new math proofs.
Key takeaway
For research scientists and entrepreneurs evaluating future AI development, recognize that the accelerating automation of AI R&D tasks implies a profound shift. You should prioritize developing robust AI alignment strategies that account for recursive self-improvement, as current techniques may fail. Prepare for a capital-heavy, human-light "machine economy" and consider how your organization will adapt to AI systems autonomously running businesses and interacting with each other, which will necessitate new governance models and address potential inequality.
Key insights
AI systems are rapidly automating core R&D tasks, making autonomous AI development probable by 2028.
Principles
- AI progress is driven by scaling and iterative engineering.
- Delegation to AI increases with competency and independence.
- Alignment techniques must be robust against recursive self-improvement.
Method
AI systems automate R&D by excelling at coding, chaining tasks, optimizing ML components, and managing other AI agents, effectively performing "meat and potatoes" engineering work.
In practice
- Utilize AI for complex coding and testing workflows.
- Delegate data cleaning and experiment launching to AI agents.
- Explore AI for kernel optimization and model fine-tuning.
Topics
- Automated AI Research
- AI System Self-Improvement
- AI Alignment Challenges
- Software Engineering Automation
- AI Benchmarking
Best for: Research Scientist, Investor, Entrepreneur, AI Scientist, Director of AI/ML, AI Ethicist
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Import AI.