AI proves Aaronson-Ambainis quantum conjecture Lean 4 + Mistral Large 4 open weights 2026

An AI Proved an 18-Year-Old Quantum Computing Conjecture on Day 3. Mistral Released a 1T-Parameter Model for Free.

Physical Superintelligence's AI (unreleased Anthropic model) proved the Aaronson-Ambainis conjecture (open since 2008) on day 3 — Lean 4 verified, 9,454 lines, 750,881 declarations, two independent kernels. The result formally limits quantum exponential speedup on random inputs. PSI improved it twice more. Same week: Mistral released Large 4 (Le Chonk) open weights — 1T params, 49-52B active, $0.70/$2.10 per M tokens, MaaS restriction in license.

TL;DR

- Physical Superintelligence (PSI), an AI-native physics research lab, announced on November 6 that an internal research system running an unnamed, non-public Anthropic model proved the Aaronson-Ambainis conjecture — an open problem in quantum query complexity since 2008. The system generated a complete proof on day three of the run. PSI then directed it at related open problems; it improved the result twice. Lean 4 accepted the final proof: 84 files, 9,454 lines, 750,881 declarations checked across two independent kernels (Lean's own kernel + lean4lean). The proof artifacts and full formalization are public on GitHub. The paper has not completed peer review. Human mathematicians must still confirm that the Lean statement faithfully captures the original conjecture. - What the Aaronson-Ambainis conjecture means: when a quantum computer answers yes-or-no questions about random data with no hidden pattern, a classical computer can predict its answers for almost every input without needing exponentially more data reads. The result imposes a formal limit on when quantum computers can achieve exponential speedups over classical hardware on typical random inputs. It does not resolve the broader question of quantum advantage in all cases — it restricts the specific case of unstructured random inputs. - Mistral AI released the open weights of Mistral Large 4 on Hugging Face. Architecture: approximately 1T total parameters, 49-52B active per token, MoE, 1M-token context, multimodal (text + image in), reasoning modes (low/medium/high). Available via HuggingFace, Mistral API ($0.70/$2.10 per million input/output), Microsoft Azure AI Foundry, NVIDIA NIM, vLLM, SGLang. License: Mistral Ecosystem Open-Weight License — a modified MIT license. Free for most uses. MaaS restriction: companies that operate a Model-as-a-Service business must obtain a separate commercial license from Mistral.

Week 7 of the AI-proves-open-math series.

OpenAI released 722 manuscripts to GitHub in October. DeepMind's Gemini agent characterized 11 Busy Beaver Turing machines in Lean. PSI's unreleased-Anthropic-model system just proved the Aaronson-Ambainis conjecture.

The conjecture has been open since 2008. The proof took three days.

The Aaronson-Ambainis conjecture: what it is and why it matters

The conjecture is in quantum query complexity — the branch of theoretical computer science that studies how many times a quantum algorithm needs to query an input to solve a problem.

The specific question it answers: take a boolean function (a yes-or-no function) on N inputs. If a quantum algorithm can determine the answer to this function while reading only a few inputs — a low query complexity — does that mean the function's behavior is determined by a relatively small number of sensitive variables? The conjecture, originally posed by Scott Aaronson and Andris Ambainis, says yes: functions with low quantum query complexity must have high variable influence, meaning individual inputs matter.

In plain terms, the result says: when a quantum computer answers yes-or-no questions about random data with no hidden structure, an ordinary computer can correctly predict its answers for almost every possible input, without reading exponentially more data. It is a constraint on quantum advantage. Quantum computers cannot achieve exponential speedups over classical computers on random, unstructured inputs — there has to be hidden structure in the problem for the quantum advantage to appear.

The result does not prove that quantum computers lack advantage in all cases. Problems with hidden mathematical structure — like factoring integers (which has the Fourier structure that Shor's algorithm exploits) — are not ruled out. What the Aaronson-Ambainis result rules out is the more ambitious claim that quantum computers could be exponentially better on generic random inputs.

This matters for practical quantum computing in one specific way: it sets a formal boundary on what quantum advantage claims need to demonstrate. If a quantum application is claimed to offer exponential speedup, the Aaronson-Ambainis result means it must be exploiting non-random problem structure, and that structure must be identified and explained. Handwavy appeals to quantum advantage on generic data are now harder to defend.

The PSI proof: how it was produced

Physical Superintelligence describes itself as an AI-native physics laboratory, structured as a public benefit corporation. The lab ran instances of an unreleased Anthropic model on its own computing infrastructure.

The system worked for three days before producing a complete proof. PSI then gave the system two related open problems and says the system improved the result twice. Lean 4 accepted the final version.

The formal proof is 84 files, 9,454 lines (6,446 lines of code), built against Lean 4.33.0 and the matching Mathlib release. PSI validated it against two independent checking implementations: Lean's own kernel, and lean4lean — a separately implemented type checker. 750,881 declarations were checked. A sandboxed Comparator verified that the proved Lean statement matches a reference statement that imports only Mathlib, so a reader can confirm the statement without trusting any of PSI's custom definitions.

The proof contains no unproven placeholders and uses no additional axioms beyond what Mathlib already includes. This is the same standard of verification that the OpenAI 722 manuscript project used for its Lean-verified results, but applied to a single conjecture with the full complexity of an 18-year-old open problem rather than a batch of claimed results.

The caveats that PSI itself states: the mathematical paper describing the proof is not machine-checked — only the Lean formalization is. Human mathematicians must confirm that the Lean statement faithfully captures what Aaronson and Ambainis conjectured. No external party has yet independently run the full verification. The model used to produce the proof is not available for independent examination. Independent review by quantum complexity researchers is underway.

The pattern that the last source identified is worth repeating: Lean-verified results on named open mathematical problems now arrive from AI labs and AI-powered research systems within weeks of each other. The DeepMind Busy Beaver result (October 15), the PSI Aaronson-Ambainis result (November 6), OpenAI's math breakthroughs (October 7), and the ongoing AI-assisted formal verification work across multiple labs are establishing that formal proof checking is becoming a standard output of AI-assisted mathematical research.

Mistral Large 4: the largest open-weight model available for commercial use

Mistral's decision to release the weights of a 1-trillion-parameter model is a substantial event for the enterprise AI ecosystem, but the license requires careful reading before enterprise legal and procurement teams sign off on deployment.

The Mistral Ecosystem Open-Weight License (MEOSL) is a modified MIT license. MIT grants essentially unrestricted use, modification, and commercial distribution. MEOSL adds one condition: "businesses that operate a Model-as-a-Service business" must obtain a separate commercial license from Mistral before using the model commercially.

MEOSL defines Model-as-a-Service as giving third parties inference or post-training access in a way that lets them exercise "meaningful control over the inputs, parameters, or training data." The license explicitly exempts: internal use (using the model to build your own products and services, not reselling model access), and products that embed the model's capabilities in specific features (rather than exposing raw model access).

Practical implications: - An enterprise running Mistral Large 4 on its own infrastructure for internal workflows: free, no additional license required. - A startup deploying Mistral Large 4 as a hosted API, charging customers per token: MaaS restriction applies — requires a commercial license from Mistral. - A product that uses Mistral Large 4 to power a specific feature (e.g., a legal document summarizer, a code review tool): appears to fall under the product embedding exemption, but legal review is warranted.

This is meaningfully more restrictive than Apache 2.0, which Mistral used for Large 3. The practical effect: Mistral Large 4 can be deployed freely at scale for internal enterprise use but cannot be resold as an API endpoint without Mistral's consent. This keeps the model usable for everyone who is building with it, while protecting Mistral's commercial API business from being directly undercut by compute providers reselling the weights.

The model itself: three weeks of additional reinforcement learning were applied after the October preview (which is what dropped in Abu Dhabi). Mistral positions it as the strongest open-weight model for defensive cybersecurity and coding. It previously scored 93% on Cybench (40 CTF challenges from real competitions) and 82% on Artificial Analysis Cyber Index. Available in FP8 and NVFP4 quantization through Hugging Face, plus GGUF files from the llama.cpp community. Deployment stacks: vLLM, SGLang, llama.cpp, Unsloth.

The two stories and what they share

A formal proof of an open quantum computing conjecture. The largest open-weight frontier model released for commercial use.

Both are about expanding the accessible frontier — the set of problems that can be addressed with currently available tools. The Aaronson-Ambainis proof expands what AI systems can formally verify in mathematics. The Mistral Large 4 weights expand what AI systems any organization can deploy without ongoing licensing fees to a model provider.

The formal mathematics story has a caveat the model story doesn't: the conjecture was proved by an unreleased Anthropic model that no one outside PSI can examine. The proof is verifiable because it is in Lean. But the method — the actual mathematical ideas the system used to generate it — cannot be studied because the model is private. The Busy Beaver results from DeepMind had the same characteristic: the proofs were public and machine-checkable, but the reasoning process that generated them was inside a Gemini model.

This creates an interesting epistemic situation: the mathematical community can verify that the proof is correct, but cannot learn from it in the way that human proofs teach — by studying the approach, the intuitions, the failed attempts, and the eventual insight that made the proof possible. The mathematical community gains the result without gaining the mathematical understanding of how to find it.

For formal verification work, that is sufficient. For mathematical education and insight, it is a different kind of contribution than a proof by a human mathematician.