November 21, 2026
Heard in AI

Zvi Mowshowitz makes the case against a tiny 'cognitive core' AI

Zvi Mowshowitz argued that a small reasoning model that looks facts up would lose the intuition that comes from knowing millions of facts. Speculating aloud, he added that a truly minimal core would have no values of its own.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on The Cognitive Revolution, episode published October 24, 2026

A neural network makes a poor database. Nathan Labenz, host of The Cognitive Revolution, put that premise to Zvi Mowshowitz. If a model's facts are stored in its weights, the billions of numbers adjusted during training, why not separate the two jobs? A model could do the thinking, and a database could hold the facts. Labenz called it a popular idea and asked Mowshowitz whether it sounded promising.

Mowshowitz said he understood why people hold the theory, and that he thought it was wrong. The exchange appeared in an AI:AM highlights episode of the podcast published on October 24, 2026.

A small mind that looks things up

Mowshowitz described the idea as a long-standing "holy grail" of sorts and used the name the AI researcher Andrej Karpathy gave it: the cognitive core. As Mowshowitz summarized it, intelligence might come down to a small amount of reasoning machinery, the part that lets a mind think, learn and process information. The human brain, on this view, runs on very little core code. A small network with that core could then load information into its context (the text a model is given to work with on a task), look facts up when needed and query other systems that may not be neural networks at all.

In an October 17, 2025 interview with Dwarkesh Patel, Karpathy proposed stripping out memorized knowledge while keeping the reasoning, experimentation and tool-use abilities, so that a compact model would retrieve an obscure fact instead of storing it. Karpathy did not propose removing knowledge entirely. He said some basic knowledge has to stay, because constant lookups would get in the way of thinking. His example of a core with about a billion parameters was a speculative, long-term possibility.

Mowshowitz granted part of the premise. He said there probably is a cognitive core in theory: a small amount of code, or a small set of logic gates, at the heart of how a human brain or a language model works. He noted that about 200 lines of code can build a primitive neural network that does interesting things. A short program is a different thing from a trained model, though. In the same interview, Karpathy separated small teaching implementations of neural networks from the knowledge and skill a model picks up in training.

Why knowing a lot matters

Mowshowitz's objection drew on how people work as much as on models. Someone who looks up every fact one at a time, he said, is a very different mind from someone who has memorized the facts of a field, and the second person is "radically more effective." He pointed to the familiar feeling of entering an unfamiliar domain and constantly stopping to ask what an acronym stands for or how some detail works.

Chatbots such as Claude, ChatGPT and Gemini help a lot with that, he said. You can ask them to explain a mechanism or a puzzling sentence. Even then, you still need enough knowledge to know which questions to ask and to draw connections between the answers.

The deeper point, for Mowshowitz, was intuition. People rarely reason in a clean line from five known facts to a conclusion. More often, he said, "these five million facts" built up the intuitions and rules of thumb that told him which five facts mattered and how to get from A to B. People from very different cultures carry different ways of doing this, and a person who can switch between those ways of seeing the world has an advantage.

A large language model, he argued, has all of that at once: the facts, the links between them and what he called the "vibes," which combine so the model can find the relevant ones. He said this is much of what people mean by "big model smell," the hard-to-pin-down richness that larger models show compared with smaller ones.

He did not claim memorization is free. Each extra fact has a cost, possibly a rising one if the network must grow to hold it, and each new fact probably adds less than the last, since the most important facts come first and later ones tend to repeat the same patterns. But as computing gets cheaper and training methods improve, he said, the break-even point for how much a model should memorize keeps climbing. Current frontier models know "a ridiculous number of things," he said, and he would not have it any other way.

The appeal, and the problem with values

The conversation then turned to why it matters whether such a core can be identified at all. Mowshowitz said the original attraction was practical: a model small enough to run on a phone, or very fast on a local computer, supplemented with whatever a task needed. You could shape the mind and its circumstances to fit the job.

He was not sure that would be good for anyone, and he raised an alignment concern, meaning the problem of getting AI systems to act on the values people intend. He flagged that he was "spitballing for the first time" and had not thought much about this particular question.

His argument ran like this. Take an uncontroversial rule such as "don't steal." That stealing is bad is a kind of fact, or a value, a parameter. A core stripped to the absolute minimum would not include it. You would load "don't steal" into context when needed. But in that case you could just as easily load stealing instead, or neutrality.

So in an important sense, Mowshowitz said, "the true cognitive core is completely amoral." It "does not care about your feelings" or anything else, he said; it "simply processes information." If someone has such a core, he concluded, "it becomes very important to choose what to do with it."

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags