Physical Superintelligence (PSI) says it has new results on five open problems in quantum information theory, published on October 23, 2026. On its research page the company writes that "working with our AI research system, PSI researchers have obtained new results on five open problems." Each result ships with a paper, a Lean 4 formalization and verification code, and the release repository documents how AI was used in each.
The story has been spreading on X and in niche outlets, and it comes with a big framing: AI helping close long-standing physics questions. This post separates what PSI claims, what anyone can check today, and what nobody outside the company has confirmed. We are not quantum-information researchers, and we have not built the Lean projects ourselves, so treat our reading as a guide to the primary sources, not a verdict.
TL;DR: claimed vs. verified
| Question | Answer |
|---|---|
| Who announced it? | Physical Superintelligence PBC, on its research page and X account, on October 23, 2026 |
| What is claimed? | New results on five problems from the Open Quantum Problems list |
| What ships with each result? | A paper, a Lean 4 formalization, verification code |
| Was AI used? | Yes, PSI says so, and names models in the repository |
| Can I check the proofs? | Yes: the Lean projects build from pinned toolchains |
| Is it peer reviewed? | No sign of it yet; the papers are preprints |
| Independent expert confirmation? | None found as of October 27, 2026 |
| Are all five problems fully solved? | No: several results are partial or negative |
What exactly did PSI publish?
A sphere with two overlapping outlines and a green dot, illustrating open quantum information problems
According to the PSI research page, the five results address these questions from the Open Quantum Problems (OQP) list:
| OQP | Question | PSI's stated result |
|---|---|---|
| 2 | Can states with a negative partial transpose hide entanglement that cannot be distilled from any number of copies? | For Werner states in every local dimension d of at least 3, Bell pairs cannot be distilled from three copies when 1/d is below alpha and alpha is at most 37/80 (0.4625). The question for arbitrarily many copies stays open. |
| 12 | Can vacuum correlations violate the CHSH Bell inequality at arbitrarily large separations? | For two fixed-shape bounded regions in the vacuum of a massive free scalar field, no local observables violate CHSH beyond a finite separation, although the regions stay entangled at any separation. |
| 27 | Which d-outcome projective measurements on a maximally entangled state give the largest CGLMP Bell violation? | Fourier-phase measurement strategies are optimal, and every maximally violating strategy is unitarily equivalent to them. |
| 38 | How many PPT maps suffice to break entanglement? | Any 2(d-1) PPT maps on a d-dimensional system suffice. PSI notes that two do not suffice in general, as reported in a recent OpenAI preprint. |
| 43 | Can every extensibly causal process arise from a larger finite-dimensional reversible process? | A bipartite counterexample: there are extensibly causal processes with no such description. |
Read the wording closely. Only OQP 43 is presented as a clean yes-or-no answer (a counterexample). OQP 2 is a bound in one regime. OQP 12 is a finite-separation theorem for a specific setup. OQP 38 gives an upper bound of 2(d-1) maps, which is not the sharp answer the original PPT-squared conjecture asks about, since that concerns composing just two. So any headline claiming "solutions to five long-standing problems" is stronger than the research page's own language of "new results on five open problems."
Which AI models did the work?
The repository is unusually specific for the one result we read in full. In the OQP 38 result guide, PSI states that "the mathematics was developed by PSI researchers working with AI models (OpenAI's GPT-5.6-Sol, GPT-5.6 Pro and GPT-6-Astra and Anthropic's Claude Opus 5) in automated research workflows." It adds that the Lean formalization "was written by AI agents (Anthropic's Claude Opus 5.5 in Claude Code, and OpenAI's GPT-6-Astra in Codex) under PSI's direction."
Responsibility is assigned to people: the paper's named authors, David Turturean and Ning Bao, "are responsible for the correctness of the paper." A file called formalization.yaml records which agent did which part, and each paper has an AI disclosure. We read this provenance only for OQP 38; the other four result guides may differ, so check each one.
This is the same pattern we have covered elsewhere: a human-directed team, several frontier models, and agents writing the machine-checkable layer. Our post on Claude designing quantum circuits described a similar loop from Anthropic's side, and the TheoPhysEval result for GPT-6 Astra showed how far models now get on graded physics problems. Solving problems that researchers listed as open is a harder claim than scoring on a benchmark, which is why the verification question matters.
What does the Lean formalization verify, and what does it not?
A chain of nodes ending in a green check, showing how a Lean proof verifies PSI's quantum results
Lean is a proof assistant. If a theorem and its proof compile in Lean, the proof is logically valid relative to Lean's kernel and the stated definitions. That is real, strong evidence, and it removes the usual worry that an AI produced a plausible-looking but broken argument.
It does not remove the other worries.
- Statement fidelity. Lean checks that the proof proves the formal statement. It cannot check that the formal statement matches the open problem a physicist had in mind. Definitions can be subtly narrower or different. PSI's repository tackles this with a
Challenge.leanfile for statements, aSolution.leanfor proofs, and a paper-to-formalization map inlean/docs/FIDELITY.md, so reviewers can audit the match. That is the part that needs human experts. - Physics modeling. OQP 12 concerns a quantum field in a vacuum. Formalizing field theory is much harder than formalizing finite-dimensional linear algebra. How much of that result is formalized, and under which assumptions, is a question for specialists. The repository says the vacuum result also includes a longer manuscript alongside the shorter paper.
- Novelty. Whether each result is genuinely new, or follows from known work, is the sort of thing referees and domain experts judge. We could not find any outside assessment.
- Peer review. The papers are labeled preprints in the repository's citation notes, and no journal acceptance is mentioned.
So the accurate summary is: the proofs are checkable and PSI has made them easy to check, but no independent verification of significance or fidelity has been reported yet.
How can you check it yourself?
The repository's own guidance is a good checklist, and the steps do not need a quantum background to start.
- Open the oqp-batch-2026 repository and pick one result folder.
- In its
lean/folder, read the README for prerequisites and pinned toolchain files (lean-toolchain,lakefile.toml,lake-manifest.json), then run the build commands it lists. - Open
Challenge.leanand compare each theorem statement with the problem text on the Open Quantum Problems site, such as the PPT-squared conjecture page cited for OQP 38. - Read the paper's AI disclosure and
formalization.yamlfor who did what. - For the two results with earlier drafts (OQP 2 and OQP 27), the repo includes dated manuscripts and a
timestamps/verify.pyscript, run aspython3 -I timestamps/verify.py. This documents when drafts existed, which matters for priority claims.
If you want a general method for trusting AI-produced work you cannot judge on sight, our guide on making an agent prove it with differential testing lays out the same principle: require evidence you can run.
Who is PSI, and why does the framing matter?
We covered the company's launch in Physical Superintelligence raises $58M. It is a venture-funded company building AI systems for physics research. That context is worth keeping in mind: a company with a funding story has an incentive to present results at their strongest. That does not make the results wrong. It is a reason to read the papers rather than the headline, and to give experts time to respond.
The word "superintelligence" in the company name also invites confusion with the broader idea. If you want that background, see what superintelligence means. Solving partial cases of specialist problems is meaningful, but it is a long way from the general capability the term usually implies.
What would change our read?
Three things would move this from "claimed and checkable" to "confirmed":
- A named quantum-information researcher outside PSI who builds the Lean files, confirms the statements match the OQP problems, and says whether the results are new.
- Acceptance at a journal or posting to arXiv with referee-style feedback.
- Responses from the OQP list maintainers on whether the problems are now marked as resolved, partly resolved, or unchanged.
Until then, the safe wording is the one PSI itself used on its research page: new results on open problems, obtained with AI help, with formal proofs available for inspection.
What this means for builders
For people building with AI, the interesting part is the workflow, not only the physics. Several frontier models from two labs were used in automated research workflows, then agents in Claude Code and Codex wrote the formalization, and humans took responsibility for correctness. The deliverable was not only a paper but a runnable verifier. That is a template for any field where claims are cheap and checks are possible: ship the checker with the claim. It is also the direction public science funding is heading, as in Anthropic's credits for federal science agencies.
Related reading
- Physical Superintelligence raises $58M
- Claude designed quantum circuits that beat the best published ones
- TheoPhysEval: GPT-6 Astra on 49 of 50 physics problems
- What is superintelligence?
- Anthropic Genesis Mission credits for science agencies
- Make an agent prove it: differential testing
Details reflect PSI's research page and repository as read on October 27, 2026; the repository and any outside review may change.