DeepMind Paper Links AI Logic Errors to Statistical Learning

Four Google DeepMind researchers, András György, Tor Lattimore, Nevena Lazić and Csaba Szepesvári, argue that today’s AI systems make basic logic errors because of how they are trained. Their paper, ‘Beyond Statistical Learning: Exact Learning Is Essential for General Intelligence’, was posted on arXiv on 30 June 2025. The authors write: “We argue that their unsound behavior is a consequence of the statistical learning approach powering their development.”

The paper circulated on 23 October 2026 in a post on X by the account @crazyshyyt, which presented it as a new paper and said its authors argue that large language models will never reach AGI. The paper itself is an argument for a change of training goal, with a few mathematical results on simple problems. Its authors write that exact learning is “both essential and possible”.

Business Pill · BENCHMARKS

A short Business Engineer explainer on why a test score is not your own result. It teaches the general idea only and says nothing about the paper in this piece.

What the Paper Argues

• The problem. The authors write that “even the most advanced frontier systems regularly and consistently falter on easily-solvable deductive reasoning tasks.” Deductive reasoning means deriving new facts from existing facts and rules.
• The cause, as they see it. Models are trained to score well on average over a distribution of examples. That is statistical learning. A model can then learn “statistical shortcuts” that work on familiar inputs and fail on new ones, the paper says.
• The proposed goal. The authors want the field to aim for exact learning, “which demands correctness on all inputs”, rather than a high average score.
• The strong line. The abstract says current systems are “unfit to fulfill the dream of achieving artificial general intelligence capable of sound deductive reasoning.” That is a statement about today’s systems and today’s training goal, not a forecast that AGI cannot be built.

Sources: György, Lattimore, Lazić and Szepesvári (Google DeepMind), arXiv 2506.23908, version 1, 30 June 2025; post by @crazyshyyt on X, 23 October 2026. The 16-month gap is our arithmetic.

What Was Posted on X

In this 23 October 2026 post, the account @crazyshyyt describes the exact-learning paper as a new paper, says Google DeepMind researchers argue large language models will never reach AGI, and says more data or larger models cannot fix the problem. That is the poster’s framing; this piece compares it with what the paper says.

What the Math Actually Shows

The paper’s main result uses a toy problem: yes-or-no answers over inputs made of d binary variables. With randomly drawn examples, about d² of them are enough to make the average error small, the paper says. But to pin down the exact rule, a learner needs on the order of 2 to the power d examples, which the authors describe as essentially all of the possible inputs.

The authors give d=100 as a moderately large case, and write that “the number of possible inputs is an astronomical number.” By our arithmetic, at d=100 that is about 10,000 examples for a low average error against about 1.3 × 10³⁰ for the exact rule, ignoring constants and log factors.

A second result covers learners that treat their inputs symmetrically, such as neural networks trained with gradient descent. The authors show these can need at least 2 to the power d/2 examples to learn one simple comparison rule, and say a similar result applies to transformers, the networks behind large language models. They also show gradient descent can be slow to settle on the exact answer, and cite work in which next-token training destroyed the exactness of a transformer that had been set up to solve a logic problem exactly.

What the Authors Propose

The paper lists directions, most of them already being explored, it says. They include tests that actively hunt for failures instead of static benchmarks, ways to verify correctness, carefully chosen teaching examples, and learners and teachers that cooperate. Another is changing the task, for example training on step-by-step reasoning traces or adding a trained verifier that rejects wrong answers.

In one experiment, a 10-layer decoder-only transformer trained on simple propositional logic problems with reasoning traces reached 1.0 accuracy on one problem style and 0.999 on another, measured on 2,048 randomly generated problems. The paper says this is very high accuracy, but “exact learning is still not achieved”.

The paper also looks at systems that translate plain language into a formal language and hand it to a checker. Citing other work, it notes an on-distribution correctness of 76% and an off-distribution correctness of 46% for that translation step. In its conclusion the paper calls for future research on interactive learning and on “teaching algorithms step-by-step by explaining the process rather than simply providing input-output pairs.”

Why It Matters for Zero-Tolerance Work

The paper’s conclusion says the mismatch is most acute in engineering, mathematics and science, “where flawless formal reasoning is paramount.” In its section on alternative views, it adds that as models gain agency, “Even unlikely mistakes are unacceptable if they lead to a potentially catastrophic outcome.”

The viral post makes the business case in stronger terms. It says companies deploying these models into “zero-tolerance workflows, like financial analysis, legal compliance, or complex operations” face “a massive warning sign.” That framing is the poster’s, not the paper’s; the fields the paper names are engineering, mathematics and science.

Connecting the Coverage

Our formal verification piece reported on 23 October that Anthropic used Claude to check that 239 entry points across 123 open-source libraries do not crash on any input up to 64 MB, under stated assumptions, and that Theorem’s cost model starts at about $40 per kilobyte of machine code. Our verification-cost explainer defines verification cost as what it takes to check something, however cheap it was to make. What’s new: this paper argues for building correctness on all inputs into how models are trained, not only checking outputs afterwards.

The Structural Read

As we read it, the gap between average accuracy and exact correctness is a gap between two kinds of tests: a benchmark score measures performance over a distribution of examples, while the paper asks for correctness on all inputs.

As we read it, the authors’ own experiment shows how close and how far current methods are: a small transformer trained with reasoning traces reached 1.0 and 0.999 accuracy on two styles of logic problems, and the paper still says exact learning was not achieved.

As we read it, the remedies the paper lists, from adversarial testing to trained verifiers and formal checkers, move part of the burden from the model to the system around it; the paper cites 76% on-distribution and 46% off-distribution correctness for one translation step into a formal language.

Three Implications

FOR COMPANIES DEPLOYING AI The paper says the mismatch is most acute in engineering, mathematics and science, where it says flawless formal reasoning is paramount; as we read it, that is where a high average score and a guarantee on every input differ most.

FOR MODEL BUILDERS The paper proposes changing the training goal to correctness on all inputs, through chosen teaching sets, cooperative learning, reasoning traces and verifiers, and says most of these are already being explored.

FOR EVALUATORS The paper says static benchmarks have limited utility and calls for frameworks that actively challenge systems and for ways to verify correctness.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework lists four moats, one per data paradigm, including “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read it, this paper argues the same direction from the research side: it names trained verifiers and checks for correctness among the ways to get from average accuracy to correctness on all inputs. The framework was published in July 2026 and does not discuss this paper.

What Is Not Established

The paper is an argument with theory on simple problems, not a test of frontier models. Its sample-size results are for toy rules over binary inputs, and its one experiment uses a small transformer on synthetic logic problems. The authors themselves list the view that scaling current methods could reach exact learning or make models appear logically coherent for practical purposes. We read version 1 of the paper, dated 30 June 2025, and did not contact the authors or the poster. Nothing here is investment advice.

The Bottom Line

The DeepMind authors argue that training for average accuracy leaves AI systems open to logic errors on unfamiliar inputs, and that the goal should be correctness on every input. They write that this goal is “both essential and possible”, which is narrower than the viral post’s reading.