Headline & intro
The same recipe that crushed humans in chess and Go appears to fall apart on a children’s matchstick game. That’s not a curiosity—it’s a warning label for how we train modern AI.
A new study argues that AlphaZero‑style systems, which learn purely by playing themselves, hit a wall on simple impartial games like Nim. According to Ars Technica’s coverage of the work, the AI doesn’t just play badly; it basically stops learning. For anyone betting on “scale and self‑play” as the royal road to general intelligence, this is a deeply inconvenient datapoint.
This piece looks at what’s really going wrong, why it matters far beyond board games, and what it should change in how Europe and the wider industry think about AI.
The news in brief
According to Ars Technica, researchers Bei Zhou and Søren Riis published a paper in Machine Learning examining how an AlphaZero‑like training procedure performs on Nim, a very simple impartial game.
In Nim, players take turns removing matches from rows arranged in a pyramid; whoever is left without a legal move loses. Mathematically, there is a well‑defined procedure (a parity function) that lets you evaluate any position and know whether the current player can force a win.
Zhou and Riis built a system trained like AlphaZero: it knows only the rules, plays millions of self‑play games, and learns to predict which moves lead to victory. On small Nim boards (five rows), performance improved quickly. But once they increased the board to six and seven rows, learning essentially stalled: after extensive training, the “smart” move selector did no better than a version that chose moves at random.
The authors argue this shows that self‑play reinforcement learning struggles to discover the underlying parity rule, revealing a specific and severe failure mode.
Why this matters
On the surface, an AI tripping over Nim looks like a party trick. In reality, it slices right into the core mythology of modern AI: that if you give a big neural network enough compute, data, and self‑play, it can discover almost any structure hiding in the problem.
Zhou and Riis’ result says: not always. Nim is simple, fully observable, and solved. Yet an AlphaZero‑style learner flails as soon as the board gets modestly larger, because the winning strategy depends on a global mathematical relation (a parity/XOR‑like function), not on local patterns in the board.
That’s bad news for:
- The “just scale it” camp. If the learning signal doesn’t expose the structure in a usable way, more games and bigger models don’t help much.
- Anyone using pattern‑based models on symbolic domains. A lot of “AI for maths, code, or safety‑critical reasoning” quietly assumes the same recipe will generalise from Go to algebra. Nim suggests otherwise.
Who benefits?
- Hybrid and symbolic approaches. People working on neurosymbolic AI, program synthesis, and formal methods suddenly have a clean, widely understandable example showing why their tools matter.
- AI safety and evaluation teams. The paper hands them a crisp benchmark family where ultra‑strong game AIs can look competent but be systematically wrong.
The immediate implication: AlphaZero‑like systems are not “general game solvers.” They’re extremely strong pattern recognisers plus search. When the correct solution looks more like running an algorithm than spotting a shape, they can confidently fail.
The bigger picture
This result doesn’t come out of nowhere; it fits several recent threads in AI.
First, we’ve already seen brittleness in top Go engines. In 2023–24, hobbyists and researchers showed that systems like KataGo could be demolished by specially constructed positions, despite being far stronger than any human in normal play. The pattern learner hadn’t fully internalised the underlying combinatorial structure of the game.
Second, large language models are showing similar cracks. They can pass many math benchmarks, but often by memorising templates and surface patterns. Push them into slightly longer or more abstract puzzles—where you must latch onto invariants or parity arguments—and their performance collapses unless you bolt on tools like code execution or symbolic solvers.
Nim is the cleanest possible version of this problem. There are no distractions: no noisy language, no complex rules, no human annotation bias. Either your training setup discovers the parity rule, or it doesn’t. Zhou and Riis show that self‑play plus gradient descent basically doesn’t.
Historically, this re‑opens a debate that AI went through in the 1980s and 1990s: symbolic reasoning versus neural networks. The current fashion heavily favours deep learning and reinforcement learning. But parity‑like functions are a classic example from theory showing what vanilla neural nets struggle with unless you bake in the right inductive biases.
Compare this with Google DeepMind’s own later work such as AlphaTensor or AlphaDev, which search explicitly over programs or algorithmic structures instead of just learning “boards → values.” Those systems are far more deliberate about representing algorithms, not just heuristics. Nim suggests that if you want algorithmic generalisation, you must design for it; it doesn’t magically fall out of self‑play.
The European / regional angle
For Europe, this isn’t only an academic curiosity; it lands in the middle of regulatory and industrial debates.
The EU AI Act pushes providers of high‑risk AI systems to prove robustness and to document known limitations. Nim‑style games give European regulators and auditors a simple, transparent stress test: can your “general reasoning engine” handle a toy problem that demands a clear mathematical invariant? If not, how will it behave in a medical triage model, an air‑traffic optimisation system, or a bank’s risk engine where such invariants exist but aren’t as obvious?
European research is relatively strong in areas that this paper vindicates: formal verification (e.g., work at ETH Zürich or Inria), logic, and hybrid AI. Universities and labs in Berlin, Paris, Zürich, Ljubljana, Zagreb and beyond are already exploring neurosymbolic systems. Nim provides a showcase example those projects can use when arguing for funding or industrial adoption.
For European companies, especially in regulated sectors like finance, energy and mobility, the message is sobering: if you’re deploying black‑box reinforcement learning because “it works in games,” you should assume there are hidden Nim‑like zones in your problem space. Under GDPR and the AI Act, unexplained catastrophic failures are not just embarrassing; they’re legal liabilities.
Finally, Europe has an opportunity to lead in evaluation standards. Instead of accepting vendor demos on glossy benchmarks, regulators and public buyers could mandate batteries of impartial‑game‑style tests for any system claiming advanced reasoning.
Looking ahead
What might change after this?
- More algorithmic benchmarks. Expect a wave of papers proposing families of games and puzzles with known mathematical structure—beyond Nim—to probe where different model classes break.
- Architectural tweaks and hybrid systems. DeepMind, OpenAI and European labs are unlikely to accept that “self‑play just can’t do this.” We’ll see experiments mixing neural nets with small symbolic modules, parity checkers, or differentiable logic.
- Rethinking claims of “general game‑playing.” Marketing around “one algorithm to master any game” will face more scrutiny. Investors and policymakers should start asking which classes of games—and by analogy, which classes of real‑world tasks—are actually covered.
Timeline‑wise, this isn’t an overnight revolution. AlphaZero‑style training is deeply embedded in how the industry thinks about RL. But over the next 2–5 years, as more failures of this type accumulate in domains like theorem proving or program synthesis, it will get harder to pretend they’re edge cases.
Key open questions include:
- Can we coax neural networks to internalise parity‑like rules through clever curricula or representations, or do we need explicit symbolic components?
- How do we even detect Nim‑style blind spots in messy real‑world systems where we don’t know the ground‑truth rule?
For practitioners, the opportunity is clear: building tools, tests and hybrid architectures that can turn this kind of failure mode into a competitive advantage.
The bottom line
Nim exposes a structural weakness in one of AI’s most celebrated training paradigms. AlphaZero‑like systems are incredible at learning statistical associations but can be almost helpless when the right move depends on crisp symbolic rules.
That should temper the industry’s faith in self‑play and “scale solves everything,” and it strengthens the case for hybrid, verifiable approaches—an area where European research is well positioned. The practical question for readers is simple: where in your domain might a Nim‑like blind spot be hiding, and how would you know?
