GPAI.WIKI
roadmaps/R-002··v1.0·contested

The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence

Gary Marcus
Author
Gary Marcus
Version
1.0
Status
contested
Horizon
2020–2030

Argues that scaling statistical learning alone will not produce robust intelligence, and proposes hybrid architectures with explicit symbolic machinery, large-scale knowledge, and cognitive models as prerequisites.

The four steps

  1. Hybrid architectures combining learned and symbolic representations.
  2. Large-scale structured knowledge, including a formal ontology.
  3. Reasoning machinery capable of operating over that knowledge.
  4. Cognitive models rich enough to support causal and counterfactual inference.

Why it is worth keeping on file

The predictions are unusually specific for a position paper, which makes it useful regardless of whether one agrees. The failure modes it names — brittleness under distribution shift, absence of a stable world model, poor compositional generalisation — remain the standard diagnostic vocabulary, and the disagreement is now about how much of each survives scaling, rather than whether the categories are real.

Read alongside R-001, which shares the diagnosis and rejects the prescription.

reactions
no reactions yet
react on GitHub