PSI Reports AI-Assisted Aaronson–Ambainis Proof as Agents Build Scientific Models

Quantum Complexity: PSI Reports an AI-Assisted Proof

On November 6, Physical Superintelligence (PSI) announced that an internal research system, using instances of a non-public Anthropic model, found a proof of the Aaronson–Ambainis conjecture, open since 2008. PSI says the system had a complete proof on the third day of the run, that Lean accepted it, and that the system later improved the result twice, with Lean accepting the final version. David Turturean developed and ran the system, scoped the problem, reviewed a model-produced paper draft, and takes responsibility for correctness; Jim Halverson audited the proof, and Phiala Shanahan commented on the Lean formalization. [1]

On PSI’s account of the reported proof, an ordinary computer can estimate a quantum algorithm’s probability of answering yes, up to a small error, on all but a small fraction of uniformly random inputs, using a number of input queries that grows at most polynomially with the quantum query count rather than exponentially. An input query, in this setting, means reading one bit of the input. That claim is about query complexity for approximating acceptance probability on most uniformly random Boolean inputs, not a general classical running-time result for every quantum computation. [1]

This Watch reports PSI’s announcement. The underlying manuscript and Lean code were not independently reviewed here, the model is unnamed, and peer review is not established in the sources used for this issue.

Science: Agents That Build Models, Not Only Fit Parameters

Also on November 6, Anthropic published a guest article by Jakob Macke, Daniel Gedon, and Stefan Wahl of the ELLIS Institute Tübingen and the University of Tübingen on teaching AI to build scientific models. [2]

The authors separate three steps: building a model’s form, constraining its parameters to data, and interpreting what the model says about the real system. They report that large-language-model agents can write, run, fit, compare, and revise model code. They favor distributions of plausible models over a single best fit so uncertainty and the next experiment stay visible. [2]

On constraining models, they describe simulation-based inference: many forward simulations train a neural network that, given new data, returns ranges of parameters that could have produced it. They say the simulations can take weeks while trained inference can take seconds. [2]

They state a hard limit clearly. So far, such systems have mostly been tested on rediscovering models scientists already know. Evidence on problems with no known answer is sparse. [2]

Interpretation—turning a fit into understanding of the real system—remains the hardest step to automate. Their remark that an AI system could plausibly find a Hodgkin–Huxley-like equation in an afternoon is an illustration, not a measured discovery-speed result. [2]

Software Factories: Definitions and Live Workflows

Zach Lloyd defined software-factory criteria that include versioned, code-defined workflows; cloud execution; different agents; stored agent data; and a measurable, testable, self-improving process. He also separates agents, cloud agents, and triggers. That framing is his definition and product positioning, not an industry consensus or a proven claim of recursive self-improvement. [3]

Eno Reyes described a concrete pipeline: a signal enters human or workstream triage, then a specification and pull request, then agent and security review, an author-blind risk judge, continuous integration, live-preview quality checks, shipping, and production failures linked back to the pull request. He notes that most work still starts with a person and that most changes still receive human approval. [4]

Dex Horthy argued that stack-specific tests, linters, and previews resist one-size-fits-all automation, and he favors composable building blocks instead. Ryan X. Charles quoted factory and workflow posts and urged builders to automate what can be automated. Those are attributed opinions about practice. [5][6][7][8]

Security: Firecracker Escape and Sandbox Configuration

PWN Research published a technical report on escaping a Firecracker microVM to host execution while the guest virtual machine remained available. The reported target was Firecracker v1.16.1 on x86-64, with host kernel 7.0.0-1011-aws and guest kernel 6.18.44+. [9]

A critical detail in the report is configuration: Firecracker was launched directly as root, without the jailer, so the root outcome reflects that setup. The authors state that production privilege dropping, namespaces, and seccomp need separate assessment. The report depends on an earlier host-memory primitive and does not, by itself, establish that an AI system discovered the technique. [9]

Agents: Persistent Cloud Machines

boat’s founder (@AniC_dev) reported more than 2,000 active paying users and said that figure had doubled in less than two weeks. That is an unaudited company adoption count. [10]

The boat product site advertises persistent Ubuntu virtual machines with SSH and SCP, Docker inside the VM, a virtual desktop, snapshot forks, and CLI, API, and SDK access. Those are product claims; the service was not tested for this issue. [11]

Hardware: Software Systems for Engineering Design

Blake Scholl, founder of Boom, described reusable software systems for engineering calculations that combine software architecture with hardware physics expertise. In a note originally posted November 5 and newly surfaced in this coverage window, he claims a turbomachinery tool lets one engineer do in real time what previously took 100 engineers six months, and that another system auto-generates about 65% of aircraft structure. He also says an aircraft was built by a team of 50 and a full-scale turbine by 150. [12]

Those figures are founder claims without a reproducible public benchmark or audit in the sources used here. They describe software-driven design productivity at Boom, not a newly certified airliner or a completed manufacturing transformation, and not an exclusively AI result to generalize across the industry. [12]

Coverage note: This issue reviewed 51 public quote posts from @rxcastrohacker and 99 public posts from @AstrohackerLabs from after November 6, 4:41 AM CT through November 7, 4:00 AM CT. Most Labs entries were reposts of older material rather than fresh developments. Likes were not accessible for this day, so there is no complete likes coverage.

Sources: