Pattern matching, not reasoning

Between February 2025 and April 2026, Gary Marcus posted dated verdicts on a run of major AI systems. DeepSeek R1 and GPT-4.5 in February. OpenAI’s o3 and xAI’s Grok 4 in the same July post. GPT-5 in August, Gemini 3 in November. Anthropic’s Claude Code in April 2026. Seven systems across six posts in fourteen months, one structural argument repeated in roughly the same words: frontier large language models do “pattern matching, not reasoning”; “deep learning is hitting a wall”; “the curve has flattened.” OpenAI’s GPT-6 Astra drew a seventh post on September 3, 2026. In it Marcus credited the model’s reported symbolic-world-model behavior as neurosymbolic, called his own read tentative, and reserved judgment on the system’s robustness and on whether it is AGI.

The argument is older than the releases. Marcus, an NYU emeritus professor of psychology and neural science, has been making a version of it since The Algebraic Mind in 2001 and Rebooting AI in 2019: pure neural networks lack the structured representations a general intelligence requires, and grafting symbolic machinery onto them is the only path forward. Marcus’s longstanding prescription for AI — neural-network learning grafted onto structured symbolic reasoning — is what he calls “neurosymbolic.” The 2017 NYU debate with Yann LeCun, Meta’s former chief AI scientist, is the canonical reference point. His March 2022 essay in Nautilus, “Deep Learning Is Hitting a Wall,” supplied the slogan. What is interesting in 2026 is not whether the diagnosis was correct in 2001. It is whether a structural argument repeated at roughly verbatim length across fourteen months of frontier releases is a forecast that has held its shape, or a classification that wins regardless of which way the release lands.

Seven releases, two categories

The releases sort cleanly into two buckets in Marcus’s writing. Models that integrate any symbolic component he reads as “vindication.” Pure-scale releases he reads as “the wall.”

Seven systems, fourteen months, two categories. Source: Marcus on AI Substack.

The “vindication” column shares a feature: each model in it runs symbolic code somewhere in its loop. R1 trains with what Marcus, in February 2025, called a “rule-based reward system” and reads as evidence that “new techniques beyond pure scaling are necessary.” o3 and Grok 4 use external tools at inference. Anthropic’s Claude Code, in Marcus’s April 2026 reading, is “massive vindication” — the lab, as he put it, going “exactly where I have said for 25 years that the field needed to go: to Neurosymbolic AI.”

The “wall” column shares a different feature. GPT-4.5, GPT-5, and Gemini 3 were presented to the public primarily as scale stories — bigger pretraining runs, longer context, marginal capability gains. On GPT-5 in August 2025 Marcus wrote that the model showed “real progress on some dimensions” but was “stuck in place on others” — exactly, he argued, what hitting “the wall” looks like. On Gemini 3 in November he allowed that it was “[a] great model, as far as LLMs go, topping most benchmarks,” then closed: “Still no AGI”; “scaling isn’t getting us to AGI.”

Where the classification picks the side

The strongest counter to engage head-on is that an argument that reads every neurosymbolic-flavored release as “vindication” and every pure-scale release as confirmation has built itself a frame that fits both halves of every observation. “The wall” moves wherever scale has stopped working this quarter. Neurosymbolic expands to cover any system that runs symbolic code somewhere in its stack, which by 2026 is most production systems. An argument shaped to absorb both outcomes does not, in the strict sense, predict.

Take the boundary cases. R1’s “rule-based reward” is a reinforcement-learning training signal, not a deductive component in inference. A reader inclined to call it scaling-with-better-RL would not be reaching. o3’s tool use is closer to what Marcus has historically meant by hybrid architecture — but Anthropic and OpenAI tools have been calling Python interpreters since 2023, and earlier instances did not earn vindication posts. The selection effect is doing work. What changed in 2025 is less the category of system being shipped than the category of system Marcus chose to flag.

Run the rotation in the other direction and the picture is similar. GPT-4.5 was shipped in February 2025 with explicit guidance that it was a base-model upgrade rather than a reasoner. GPT-5 was a router product over a family of models, including reasoners. Gemini 3 in November 2025 shipped with tool use, multimodal grounding, and search retrieval — components that, in any other release week, would have qualified for the vindication category Marcus gave o3. Each of these could plausibly have been called “vindication.” None was.

The variable is not what the model can do. It is what the public framing of the release emphasizes. A scale-marketed release is “the wall”; a tool-marketed release is the path.

What the rule does not cover

Marcus does concede discrete prediction errors when they are discrete and isolable. On GPT-5: “I honestly didn’t think OpenAI would burn the brand name on something so mid.” It is a surface concession, not a structural one, and it is exactly the kind of admission a defender of an unfalsifiable frame would never need to make. An argument designed to win regardless of outcome would not bother conceding anything.

The structural argument has a stated falsification criterion, even if it is a soft one. Marcus has said that if GPT-10 is “just like GPT-3,” he will post a YouTube video admitting he was wrong. The criterion is loose — “just like” does the heavy lifting — but it is not absent. The bet to Dario Amodei on AGI by end of 2027, and the separate 10-task list co-signed with Miles Brundage, formerly of OpenAI’s policy team, at 10:1 odds, put a settlement price on the capability timeline the claim implies. Amodei, Marcus wrote in March 2026, never responded. The settlement structure is the part of Marcus’s record that resists the frame-fits-both-halves reading: a frame engineered to be unfalsifiable does not write public bets with named counterparties and dated terms.

The structural claim — that scaling alone does not yield general intelligence and that some form of structured representation has to enter the system — is now closer to the field’s working consensus than to its contrarian fringe. The engineering response across the seven releases reads as a series of partial concessions in that direction: rule-based reward signals in R1, external tool use in o3 and Grok 4, hybrid execution layers in Claude Code. The labs whose marketing language Marcus reads as confirmation have, in their actual engineering choices, moved closer to his prescription rather than away from it.

What would move the argument

For the structural critique to actually be tested in 2026 and beyond, two kinds of evidence would have to land.

A pure-scale release that closes a gap Marcus has historically named as the symptom — compositional generalization, the ability to combine known concepts to handle situations the model wasn’t trained on; robust planning over long horizons; reliable arithmetic without tool use — without any externally invoked symbolic component, would force a real revision. None of the seven releases qualifies. The capabilities improving fastest in the “wall” column are not the capabilities Marcus has named as the structural test. The capabilities improving fastest in the “vindication” column are improving in systems that do, in fact, have symbolic components stitched in.

A neurosymbolic-flavored release that fails on a task class the structural argument predicts it should handle would cut the other way. Claude Code as of April 2026 is the cleanest live test. A symbolic layer is designed to make deterministic branching logic reliable; the harder tests are novel domains the model has not been trained on and reasoning chains long enough to require revising an earlier conclusion mid-chain. The first is a generalization problem the symbolic side of the debate has historically claimed as its strongest territory; the second is a planning problem the structural argument has named as the central symptom for two decades. If Anthropic-style hybrid releases over the next year pass those tests, the “vindication” reading holds. If pure-scale models pass them first, it does not.

The structural argument held its shape across fourteen months of releases because the releases did not produce the kind of evidence that would break either reading. OpenAI’s mathematics results are the newest boundary case: on October 7, 2026, Marcus allowed that they could mark a real step toward AGI or a narrow system built on Lean and synthetic data, and stressed that OpenAI had not disclosed its methods. The frame-fits-both-halves critique is a real critique. It is not yet a refutation, because the data needed to refute Marcus is also the data needed to refute his critics. Both sides are waiting on the same release.

What the record actually says

Pure neural networks, trained on pure prediction over pure scale, do not on their own produce the kind of structured representations a general intelligence needs. The engineering response — tool use, retrieval, planners, deterministic execution layers, reinforcement signals shaped by rules — is in practice a series of partial concessions to the structural side of the 2017 debate.

That claim is testable, even if the version Marcus litigates in real time on Substack is not always the version he could most rigorously defend. The bet to Amodei and the 10-task list with Brundage test the capability timeline the claim implies, though neither isolates the architecture: the Brundage wager settles on whether AI can do eight of the ten tasks by end-2027. The next eighteen months of releases will test it. What the record of seven releases in fourteen months shows is not that Marcus has been right about every model, or that he has avoided every interpretive trap that a frame this elastic invites. It shows that the structural argument has not been broken by any single release in the period, and that the labs best positioned to break it have, in their engineering choices, drifted toward his prescription.

The Amodei bet has terms. The Brundage list has odds. Those are the artifacts that cannot be retroactively reframed by the next release post — they resolve, or they don’t, against criteria fixed before the resolving model exists.