François Chollet is an AI researcher associated with the ARC benchmark family and a critic of treating raw model performance as equivalent to general intelligence. In an August 2026 X thread, he argued that a large inference-time harness orchestrating thousands of neural-model calls is, by definition, a neurosymbolic architecture.source: francois-chollet-neurosymbolic-agent-architecture-2026.md
His key architectural claim is that present-day high-performing agents use a symbolic outer layer: ordinary software drives the neural model, invokes tools, and runs model-generated scripts. Chollet calls this a “symbolic sandwich” and contrasts it with the older end-to-end neural ideal in which almost all task logic lives inside a differentiable model.source: francois-chollet-neurosymbolic-agent-architecture-2026.md
The thread connects directly to harness-engineering and agent-loops: capability should be attributed to the complete runtime system—model, orchestration code, tools, state, search, and verification—rather than to model weights alone. It also exposes a contested terminology question: critics argue that Python glue code and API routing are ordinary programming, not symbolic reasoning in the stronger classical sense of logic, rules, theorem provers, or knowledge graphs.source: francois-chollet-neurosymbolic-agent-architecture-2026.md
In a reply about ARC-AGI-3, Chollet distinguished the public demonstration set from the benchmark itself and said that the discussed ARC-3 harness was not automatically an admissible private-set submission; eligibility depended on the leaderboard rules in the ARC-3 paper.source: francois-chollet-neurosymbolic-agent-architecture-2026.md
Related pages: neurosymbolic-agent-systems, harness-engineering, agent-loops, test-time-compute-evaluations, production-llm-reliability.