Aristotle solved five IMO problems. That is the data point. Everything else is noise.
Harmonic claims its model achieved a gold medal at IMO 2025 with Lean formal verification. The announcement came via Crypto Briefing—a publication known for token narratives, not peer-reviewed science. The information is thin. The hype is thick.
Let's isolate the facts. The model solved 5 out of 6 problems. The solutions were accompanied by Lean proofs. IMO is the International Mathematical Olympiad—a test of pure reasoning. A gold medal there signals deep logical capability. But here is the catch: we have no architecture details, no training data description, no compute budget, no independent replication. Silence in the logs is louder than the crash.
I have seen this pattern before. In 2018, I audited a smart contract that claimed to be secure. The marketing was flawless. The code had a reentrancy bug deep in the swap function. Six weeks of manual review uncovered it. The project paid a bounty but never published the fix rationale. The public saw only the headline. Here, the headline is “AI beats IMO.” The hidden code is missing.
Let’s dissect what we actually know versus what we need to know.
Known: Aristotle solved five IMO problems. The solutions are formalized in Lean. The model likely uses a neuro-symbolic approach—combining neural networks with symbolic reasoning. This is consistent with AlphaProof and OpenAI o1. The approach is promising.
Unknown: Model architecture (transformer? mixture of experts? specialized encoder?). Training data (was it contaminated with IMO problem solutions and Lean proofs?). Computing resources (how many GPUs? how long per problem? does it run in real-time?). Generalization (can it solve unseen problems from other competitions like Putnam or USAMO?). The sixth problem failure—was it a limit of the search space or a fundamental reasoning gap? No answers.
Credibility: The source is Crypto Briefing. Their editorial bias is high—they cover crypto assets, not formal science. The article uses positive framing: “gold medal performance.” No mention of limitations. No comparison to existing systems. No independent verification. In my 2022 forensic report on Terra’s collapse, I traced $100 million withdrawal triggering a death spiral. The official narrative was different from the on-chain data. Here, the official narrative is all we have. Without technical documentation, the result is a claim, not a proof.
Technical risk: Lean verification does not automatically mean the reasoning is correct. It means the formal syntax checks out. But the logical chain could still be flawed if the model used heuristic shortcuts that happen to satisfy the proof checker. “Precision is the only currency that never inflates”—but precision applies to the verification tool, not the model’s reasoning. The model might overfit to IMO-style templates. In 2021, I analyzed BAYC floor prices and found wash trading patterns masked as organic volume. Here, the pattern is mathematical performance masked as general intelligence.
Industry impact: Short-term, this will push more AI labs to adopt Lean or similar formal verification for reasoning tasks. It will raise the bar for math benchmarks. Long-term, the real test is whether this approach scales to open research problems. Solving Olympiad problems is impressive but not equivalent to proving new theorems. The gap is vast.
Now the contrarian angle. The bulls have a point: neuro-symbolic approaches with formal verification are the most promising path to trustworthy AI reasoning. Even if Aristotle is overhyped, the methodology is sound. Solving 5/6 IMO problems with verified proofs is a non-trivial technical achievement. The model likely demonstrates real capability in combinatorial reasoning and algebraic manipulation. If Harmonic releases code and benchmarks, the community can validate and build on it. The potential for formal verification in blockchain smart contract auditing, for example, is direct. As someone who spent years auditing contracts, I see the value. But I also know that tools are only as good as their transparency.
Contrarian takeaway: do not dismiss the result entirely. The underlying research direction is correct. The execution may be genuine. But without transparency, it is irresponsible to treat it as a breakthrough.
Takeaway: Harmonic, release the whitepaper. Release the evaluation details. Let the community run the model on standard benchmarks like MATH-500 and AIME. If the results hold, this is a milestone. If not, it is a well-crafted PR piece. Until then, treat the gold medal as a token, not a proof. “Yield is just risk wearing a mask of mathematics.” Here, the yield is gold medal performance, and the mask is formal verification. The underlying risk is unknown.
I will wait for the logs. The silence is louder than the crash.