“Deep learning must be supplemented by other techniques if we are to reach artificial general intelligence.”
Cognitive scientist Gary Marcus reached this conclusion in a 2018 assessment of deep learning. Neural networks had achieved major gains in perception, speech, games, and language-related tasks. Marcus argued that those successes should be examined alongside persistent weaknesses. Systems could require enormous amounts of data, struggle when conditions changed, and learn correlations without acquiring the abstract concepts or causal knowledge people use flexibly. His claim was not that deep learning lacked value. It was that a path from specialized performance to general intelligence would likely require more than scaling the same method.
Deep learning is particularly effective at fitting rich patterns from examples. Given appropriate data and objectives, it can construct representations that engineers did not specify by hand. General intelligence poses a broader challenge. An agent must transfer knowledge to unfamiliar situations, reason about cause and effect, plan over time, understand compositional structure, and learn from limited experience. It must also know when its information is insufficient. Progress on one benchmark can hide brittleness if evaluation resembles the training distribution or rewards shortcuts instead of the intended capability.
Supplementation can mean several things. Researchers may combine neural networks with symbolic reasoning, memory, search, causal models, external tools, structured knowledge, or mechanisms for active learning. Hybrid systems need not recreate older AI exactly. The purpose is to give different components roles suited to their strengths: perception may be learned statistically, while verification or planning may use explicit constraints. Modern tool-using models and retrieval systems already reflect part of this impulse, although integration introduces new errors and does not automatically produce reliable general reasoning.
The debate also concerns scientific method. When systems fail, researchers need tests that distinguish missing data, poor objectives, limited architecture, and fundamental uncertainty. Bigger models can be informative experiments, but scale should not become an explanation for every capability or the default remedy for every weakness. Marcus’s quote favors pluralism: compare approaches, build diagnostic evaluations, and remain open to ideas from cognitive science, linguistics, robotics, and statistics. General intelligence, if achieved, will need robust behavior beyond the conditions that made a demonstration look successful. Marcus’s argument remains a useful counterweight to architectural certainty. General intelligence is a research goal, not a benchmark already won, and progress may depend on combining insights that currently sit in different technical traditions. A plural research program can learn from failures instead of explaining every weakness as a temporary shortage of scale.
Neural networks can learn complex representations from examples and have transformed perception, speech, and language applications.
Their success is real even when it does not settle how to build flexible, reliable general intelligence.
A general system must transfer knowledge, reason, plan, combine concepts, and respond appropriately to unfamiliar conditions.
High performance on a familiar benchmark can coexist with shortcuts and brittle behavior outside the test distribution.
Reasoning, search, memory, causal models, structured knowledge, and external tools may complement learned representations.
Integration creates its own challenges, so combinations must be tested rather than assumed to solve general intelligence.
Explore more "Quotes of The Day"
Continue exploring how writers, thinkers, artists, and advocates have described the choices that shape an everyday life.