OpenAI Internal AI Models Automate Training and Research Processes
OpenAI's internal AI models can now automate the training of new experimental models, including writing GPU kernels and optimizing training code. Multiple AI agents collaborate without human intervention, reducing experiment time from years to about a week. Separately, Claude discovered a new CRISPR-like enzyme system, demonstrating AI's potential in biological research.
Key insights
AI Automates Model Training at OpenAI
According to The Information, OpenAI's AI models can now automate the training pipeline for new experimental models, including writing GPU kernels and optimizing training code. Researchers only need to provide an optimization example, and the AI spends weeks implementing and testing similar improvements. Multiple AI agents collaborate to solve problems without human involvement. This capability has improved significantly in recent months, reducing experiment time from years to about a week.
Claude Discovers Novel CRISPR-like Enzyme System
Anthropic's Claude independently discovered ART, a CRISPR-like enzyme system in bacteriophages, using 950 agents running for 21 hours. The system includes a reverse transcriptase, an unknown partner gene, and repeated DNA arrays. This demonstrates AI's ability to make significant biological discoveries, with potential implications for gene editing. The discovery was verified by human scientists in Anthropic's wet lab.