From brain to algorithm. This atlas is built from specialists: a cortex to see, an area to speak, a hippocampus to remember, a cerebellum to move. Yet with those same modules you read, do algebra, play chess: things evolution never built anything for. One question ties all the regions together. How does a system of specialists generalize? And is an AI doing it?
The contested word
There is no agreement on what AGI, "artificial general intelligence", actually is. Not in academia, not in the labs. The definitions in circulation fall into three incompatible schools, plus one camp that rejects the term outright.
The first wants to match a human: AGI is a system with the cognitive versatility of a person across essentially all tasks. Demis Hassabis (DeepMind) uses scientific discovery as the test; Dan Hendrycks measures it across ten cognitive domains (A Definition of AGI, 2025); DeepMind's Levels of AGI, 2023 grades it on two axes, performance and generality. The second is universal and formal: Shane Legg and Marcus Hutter define intelligence as performance over all computable environments, weighted by simplicity (Universal Intelligence, 2007). A precise mathematical object and, by their own admission, uncomputable. The third is economic: OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" (OpenAI Charter). Intelligence measured by output, not by cognition.
Then there are those who want to abolish the word. Yann LeCun (Meta) argues that "general intelligence does not exist": humans are evolved specialists, and the term AGI should be retired. In February 2026 he formalized this in a paper with Goldfeder, Wyder and Shwartz-Ziv, AI Must Embrace Specialization via Superhuman Adaptable Intelligence (arXiv:2602.23643): aim not at AGI but at SAI, intelligence that can learn to exceed humans at what matters, measured by how fast it acquires new tasks.
The misunderstanding at the center of it all. Almost the entire dispute comes from two words mistaken for one. Universal intelligence (Legg-Hutter) is the ability to act well across every possible environment: a theoretical limit, uncomputable, and under the No Free Lunch theorem not even attainable, because no finite system can be good at everything at once. General intelligence, in the ordinary sense, is something else: broad competence over the structured, non-random distribution of our world. The "G" in AGI means "not task-specific", not "good over every possible world". The recurring error is to refute the first and present it as a refutation of the second. When François Chollet defines intelligence as "skill-acquisition efficiency over a scope of tasks, given priors and experience" (On the Measure of Intelligence, 2019), he is saying exactly this: the interesting generality is bounded, not universal. Hassabis caught the point precisely in December 2025, responding to LeCun's statements: "he is confusing general intelligence with universal intelligence; brains are in fact extremely general". Elon Musk backed him. The reply predates the paper by two months and answered his public remarks; only Ben Goertzel responded to the paper directly, calling SAI "nothing but AGI under specific assumptions".
The third term. Between universal and pure specialist there is a middle path, and that is where the generality we care about lives: general but bounded. Broad performance up to a level of difficulty, over a natural distribution of tasks. It is not a rhetorical invention: the false dichotomy of "domain-specific versus domain-general" has been dismantled both in cognitive science (Jacobs & Gärdenfors, 2017) and in the measurement of artificial intelligence (Hernández-Orallo et al., 2021). A Turing machine with finite memory also covers a minute fraction of all possible problems, yet we do not call it specialized. It is universal but limited, categorically different from a calculator. It is this middle term that the debate tends to erase. And it is exactly the shape of the brain.
Why the word stayed vague. The uncomfortable part is worth saying too. AGI's vagueness was not only an unsolved scientific problem: it was useful. A term in which everyone projects their own dream aggregates capital and talent better than a precise definition, which would instead make disappointment possible (Narayanan & Kapoor, AGI is not a milestone). According to reporting, the OpenAI-Microsoft contract would set AGI at the threshold of 100 billion dollars in profit (The Information, Dec 2024): a definition with nothing to do with cognition. The honest suspicion is that "SAI" or "world models" will inherit the same fate: the next north star, just as loaded with expectation.
How the brain generalizes
Here the question stops being terminological and becomes biology. The signature of human generality is not doing everything. It is doing things we were not designed for, with hardware designed for something else. The key idea: specialization and generalization are not opposites, they are the same fact seen from two sides. A circuit is specialized by its properties (recognizing shapes regardless of position and size), and it generalizes precisely because those same properties get reused for new tasks that need them. Generality is not inside any module. It emerges from reuse.
- Neuronal recycling. Reading is about 5,000 years old, too recent for evolution to have built a "reading area". Yet every literate brain grows one, in the same spot of the visual cortex, because that tissue already did invariant visual recognition, exactly what letters demand. Stanislas Dehaene calls it neuronal recycling (Dehaene & Cohen, Cultural Recycling of Cortical Maps, Neuron 2007). The parallel with AI is direct: a convolutional network trained to recognize objects, then fine-tuned on reading, grows a unit analogous to that area on its own (Hannagan et al., PNAS 2021).
- Neural reuse. Michael Anderson extended the idea into a general principle: brain regions are routinely multi-use, and more recent functions recruit more, and more scattered, regions, a sign that they are assembled from older parts (Anderson, Neural reuse, BBS 2010). The observation (regions are multi-use) is solid; that reuse is the single principle remains an open debate.
- Moravec's paradox. Making a machine play chess is easy, giving it a toddler's perception and movement is brutally hard. The standard reading: the sensorimotor is ancient and deeply specialized, abstract reasoning is recent, the thinnest veneer of thought running on top of and reusing that older machinery (Moravec's paradox).
- Generalizing structure. The brain reuses not only circuits but structures. Once a schema exists, a new fact that fits it consolidates in a day instead of weeks (Tse et al., Science 2007). The hippocampus's cognitive maps, born for space, generalize to abstract and relational structure: the same code that orients you in a city orients you in a space of concepts (Tolman-Eichenbaum Machine, Whittington et al., Cell 2020). And in the prefrontal cortex abstract variables are represented in a geometry that makes them recombinable, the substrate of compositionality (Bernardi et al., Cell 2020).
Magnus Carlsen at about 2880 Elo, starting from a brain shaped to gather food and avoid predators, is not proof that we are mediocre at chess. It is a demonstration of transfer. Whoever concludes the opposite measures how high we reach in one domain, the ceiling, and draws something about how many domains we cover, the breadth. They are two different axes.
How an AI generalizes
The same question, about neural networks: does gradient descent build machinery that generalizes, or a giant lookup table? The difficulty is that for a model trained on half the internet almost any behavior can be reread as interpolation over something already seen. It helps to order the evidence from cleanest to most ambiguous.
- Grokking, the cleanest proof. A small transformer trained on modular addition first memorizes (100% on training, about 0% on test), then, training well past that, suddenly generalizes (test toward 100%). The point is not the jump. It is that the model was opened up and the real algorithm it learned was found inside: not a table, but a trigonometric procedure in Fourier space (Nanda et al., 2023, after Power et al., 2022). Here the skeptic's escape "it just memorized" is not argued away, it is excluded by looking at the weights. And the generalizing solution wins because it is more efficient under the pressure of regularization (Varma et al., 2023). Honest about the limit: the clean result is on modular arithmetic; whether the same dynamic holds at LLM scale is contested.
- In-context learning. The model performs a task defined by a few examples in the prompt, without updating its weights. It rests on real circuits, the induction heads (Olsson et al., 2022). There is an irony: it is fast adaptation without weight updates, exactly the metric LeCun's paper proposes. Caveat: it is often "locating" a task already present in pretraining rather than learning a new one.
- Transfer and its limits. Instruction tuning generalizes to families of tasks never seen in training (FLAN, 2021); reasoning learned in English transfers to other languages. On the other side, memorization is real and measurable, and high scores are often inflated by benchmark contamination.
- ARC, the live frontier. ARC is built against memorization: every task is novel and requires inferring an abstract rule from a few examples. Its recent history is the answer to the question, watched in real time. ARC-AGI-2, at launch (March 2025), zeroed out frontier models while the human average sits at 66%. In fifteen months the climb: by June 2026 models pass that human baseline, GPT-5.5 at 85%, Gemini 3.1 Pro at 77% (ARC Prize leaderboard), and the climb requires adaptation at test time, not just more parameters. But ARC-AGI-3 (March 2026), centered on novel interactive problems, reset every frontier model below 1%. Three months on, the best score is 1.5% (Claude Opus 4.8), while untrained humans solve 100% of the environments.
There is also the bet on how to get there: world models, which predict in abstract latent space rather than over pixels, to capture structure and not surface (JEPA, Dreamer, Genie). It is a direct echo of the brain-as-predictive-model from the previous section: same intuition, two substrates.
How we read the verdict
There is a real middle path, general but bounded, on both sides of the parallel. In the brain: generality is not inside the regions, it is in their reuse. The specialists of this atlas generalize because they get redeployed onto tasks they were not born for. In AI: generalization is real, and we do not take it on faith, because in grokking the generalizing algorithm can be read in the weights. But every jump in scale repeats the same lesson: it turns out narrower than it looked. In-context learning is often location, not learning. Transfer numbers deflate under contamination control. ARC-AGI-2 falls, and ARC-AGI-3 reopens the chasm on the interactive. The boundary moves outward. It stays a boundary.
The real question is not whether networks generalize, grokking settles that, but how wide, and for now every honest measure says: wider than last year, still narrower than a human, and the next benchmark finds the edge again.
This is why the whole map makes sense read in two directions. Each region is a specialist with its AI counterpart and its maturity level. Generalization is what happens when the specialists talk to each other. It is the same story, biological and artificial, and here the circle between brain and AI closes. The discipline is in refusing the two opposite simplifications. Not the hype, for which capability means imminence. Not the dismissal, for which it is just a better autocomplete. In between is what is actually happening.