François Chollet, the creator of ARC-AGI, one of AI's most respected benchmark tests, has a message for the industry: the next big fight isn't just over larger models. Today's high-performance agents are already complete systems built from a 'model plus program' — and the program side is about to matter a whole lot more.
According to Chollet, the model handles understanding and judgment, while the outer program manages the workflow, calls tools, and runs code. He describes this neuro-symbolic architecture as a 'symbolic sandwich': programs drive neural networks, and neural networks call programs in return.
Neither side is going anywhere. Large models will keep getting smarter, Chollet argues, but the surrounding program scaffolding will only grow in importance. Complex tasks, he says, will ultimately be completed by both parts working together.