Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil

Summary

Jerry Tworek and Rohan Anil, founders of Core Automation, argue that transformer scaling is hitting limits and that the main bottleneck to smarter AI is architecture, especially the lack of continual learning and test-time adaptation. They explain why pre-training and RL should be optimized end-to-end, why current inference and kernel generation are inefficient, and why they are building an automated lab to search for a transformer replacement. They also discuss why large AI labs are unlikely to pursue these contrarian bets while competing in the coding-agent race. The conversation is research-focused and does not present explicit public-market investment recommendations.

• Jerry Tworek and Rohan Anil founded Core Automation after leading frontier AI research at OpenAI and Gemini.
• They believe transformer architectures have carried scaling but cannot continually learn or adapt at test time.
• Rohan argues pre-training and RL should be optimized together for large efficiency gains.
• They identify kernel generation as a key bottleneck and target automating high-performance kernels.
• Core Automation aims to build the most automated lab to accelerate architecture search.
• Jerry says big AI labs are locked in the coding-agent race and unlikely to pursue architectural alternatives.
• Rohan doubts current hardware can match biological learning efficiency without analog approaches.
• No specific public securities or tickers are recommended.