Jerry Tworek, Core Automation CEO on AI Reasoning, Agents and Research

Core Automation bets its $532m seed against the transformer

Core Automation is betting that the transformer is not the final architecture for AI. It wants models built differently so they keep learning from each customer's data after deployment, the way a new hire does, instead of staying frozen until the next release. Tworek thinks AI tools now make building a lab fast, and that this may be the last window for a new foundation model company.

Takeaways

• The transformer is not the end of AI. Tworek says researchers grew complacent and came to treat the transformer as the end of the road. Core Automation is testing new ways to arrange neurons and memory on GPUs, betting that architecture research is severely undervalued and due a comeback.
• Models should keep learning on the job. Core Automation wants models that learn from each customer's own data on the fly, like onboarding an employee, instead of labs buying data for each profession. Today's agents cannot meaningfully relearn skills and only improve when a new model ships.
• Fine-tuning open-weight models is a poor substitute. Tworek calls fine-tuning open-weight models hard, labour intensive and not guaranteed to work. Core Automation uses different algorithms and architecture to deliver what fine-tuned transformers struggle with, and hopes to need less data, though data remains necessary.
• The seed money goes straight to NVIDIA. Gong cites a $532m seed round. Tworek says nearly all of it flows to NVIDIA for GPUs, with little spent on the office. NVIDIA is also an investor, and he prefers its chips to building custom hardware because they are the best for training.
• This may be the last foundation-model cohort. Gong reports the industry view that this is the final batch of foundation model companies to get funded. Tworek agrees Core Automation may be among them, since big incumbents will move faster as AI removes their internal political obstacles.
• A seven-person team drafted for specific roles. Tworek hires for defined roles, like drafting a football team, rather than for niceness and intelligence. He argues no lab monopolises talent and big companies have grown bureaucratic, while a clearly articulated, ambitious research plan pulls in excellent researchers.
• Desk jobs will be automated roughly together. Tworek expects a post-labour society, with the main split between desk work and physical work, where robots will lag. He sees the future company as largely AI operated, with humans setting objectives while AI processes information and proposes decisions.

How Core Automation plans to operate and sell

• Research comes first. The company aims to deliver breakthrough technology before focusing on commercialization, and stays out of the existing market of training and deploying transformer agents.