In 1957, Frank Rosenblatt, a psychologist at Cornell, unveiled a machine called the “perceptron.” Modeled on neurons in the human brain, the device learned patterns on its own when shown examples. Rosenblatt was optimistic that, with enough development, this approach could match human-level cognition. Around the same moment, John McCarthy and Marvin Minsky were walking a completely different road — encoding logical rules of the form “if A then B” directly into the computer. The first approach is called connectionism, the second symbolism. From the day artificial intelligence was born, it was already split, twin-like, into two lineages. The split was not merely technical. It was a philosophical fork — “is intelligence following rules, or learning from experience?” — and that fork is still alive today.
In 1969, Minsky and his colleague Seymour Papert proved the mathematical limits of the perceptron, and funding for connectionist research collapsed. Strictly speaking, the result applied to single-layer perceptrons, not to neural networks as a whole — but the field treated it as a death sentence. Labs shut down, money flowed en masse toward symbolism, and the period earned the name “AI winter.” Then, in the 1980s, the back-propagation algorithm and multilayer neural networks turned the tables. What Minsky had demolished was a one-story building; the answer, it turned out, was to stack many floors. In 2012, deep learning crushed an image-recognition competition, and connectionism completed its comeback. ChatGPT and Claude today are both descendants of that lineage.
Lately, though, deep learning — the apparent winner — has been showing some uncomfortably familiar weaknesses. Ask it why it judged something the way it did, and it cannot really explain. “Hallucination” — confidently inventing facts it doesn’t actually know — appears to be structural. When Anthropic ran a detailed analysis of its own AI model’s internals in 2024, it found millions of features tangled together within a single layer — and even that turned out to be a fraction of the model’s total knowledge. In domains like medical diagnosis or legal judgment, where you must show your reasoning, that’s a fatal limit. Symbolism collapsed because it “reasoned well but could not perceive the world”; connectionism is now hitting the exact mirror-image wall — “perceives well, but cannot explain its reasoning.” Relying on a single paradigm reliably exposes the other axis as a blind spot. It’s a pattern that has repeated for seventy years of AI history.
This is where the two long-separated roots finally meet again. The name is “neuro-symbolic AI.” Within a single system, combine the pattern-recognition strength of neural networks (the “neuro”) with the reasoning ability of symbolic logic (the “symbolic”). The mental shift is from “which side wins?” to “let’s add the strengths together.” The U.S. Defense Advanced Research Projects Agency (DARPA) launched the ANSR program in 2023 to pursue exactly this, with participants including UCLA, Carnegie Mellon, and SRI International.
Results are already appearing. Google DeepMind’s AlphaGeometry is a hybrid in which a neural network proposes auxiliary lines from geometric intuition, and a symbolic engine then carries out the rigorous mathematical proof. With that structure, it solved 25 of 30 International Mathematical Olympiad geometry problems (Nature, 2024). The average human gold medalist solves about 25.9 — putting AlphaGeometry essentially at gold-medal level. Neither a neural network alone nor a symbolic engine alone could have produced that result. An MIT-IBM team’s system, with no additional labels, hit 98.9 percent accuracy on visual question answering while also being able to lay out its reasoning step by step. Research that grafts structured fact databases — knowledge graphs — onto large language models has reported benchmark accuracy improvements of several multiples in efforts to reduce hallucination.
There are real technical obstacles still to clear. Cleanly combining the continuous learning of neural networks with the discrete operation of logic remains an open problem, and large-scale real-world deployments are at an early stage. But the very fact that the two roads of AI, after walking apart for seventy years, are starting to converge points clearly toward the direction of the next generation of artificial intelligence.