Autopoiesis Sciences Raises $11M for Scientific AI

Scientific discovery does not come with unit tests. A proof can be checked and code can be run, but a frontier research question may have no known answer, no cheap experiment, and no clean signal telling a scientist when to keep going or walk away.

Autopoiesis Sciences has raised an $11M Seed round around that missing control problem. Threshold Ventures led the financing, with Informed Ventures, Alpaca VC, VitalStage Ventures, Correlation Ventures, DarkMode Ventures, Mike Mahlkow, Cross Atlantic Angels, Max Keller, JD Montgomery, Adam Grosser, and others participating.

The San Francisco company is building AI systems designed to support and eventually conduct long-horizon scientific inquiry. Its thesis is unusually specific: the published record preserves results more reliably than it preserves the judgment that produced them. Autopoiesis wants to capture those abandoned hypotheses, challenged assumptions, failed searches, and changes of direction, then use that process to help models decide how to reason while an investigation is still moving.

What Autopoiesis Sciences Announced

Autopoiesis announced the Seed on September 30, 2026. The company said the capital will help grow the team, expand its research, and advance AI systems capable of autonomous scientific discovery. It did not disclose a valuation, audited revenue, a detailed hiring target, or a verified total-funding figure beyond the current $11M round.

The founding team spans several sides of the problem. Joseph Reth is co-founder and CEO. Dr. Eike Gerhardt is a co-founder with experience across academic research, venture capital, and company building. Dr. Lawrence Callahan is a co-founder and Chief Scientist whose public record includes decades of scientific work involving the FDA and NIH.

Threshold's Lisa Xu framed the investment around a gap between scientific judgment and the data available to train AI. Papers show the result that survived. They rarely preserve every question a researcher rejected, the evidence that changed a mind, or the point when a plausible path stopped being worth another experiment. That missing record is expensive because it contains the decisions that keep science from becoming an endless search through attractive ideas.

Aristotle Captures Research in Motion

Autopoiesis first built Aristotle, an AI research partner for literature review, hypothesis generation, analysis, and experimental design. The company says thousands of researchers across institutions including Pfizer, the FDA, Harvard, GSK, Merck, MIT, AstraZeneca, Stanford, Johnson & Johnson, the NIH, Bristol Myers Squibb, the European Medicines Agency, Oxford, Yale, and Johns Hopkins use the system.

Those institutional names require careful accounting. They describe company-reported usage by researchers affiliated with the named organizations. They do not establish that every institution is a customer, procurement partner, or endorser. The distinction matters because trust in scientific software is earned through validated work, not borrowed from a long list of recognizable affiliations.

The more interesting asset may be the interaction record Aristotle creates. A normal paper compresses the research path into a legible result. Aristotle can retain the discarded branches, weak evidence, revisions, and timing of decisions that occurred before the result. The company says it collected 3.5M reasoning hours of scientific-judgment data through that work.

Metacognitive Steering Changes the Reasoning Mode

Autopoiesis used those traces to develop Metacognitive Steering. The company's September 2026 arXiv preprint describes an inference-time controller that reads a model's represented cognitive state and applies layer-specific interventions. The goal is to shift among exploration, methodical execution, and critical reassessment without changing the underlying model weights.

That approach targets a different problem from simply giving a model more context or compute. A research system can keep generating ideas and still fail because it commits too early, protects a weak assumption, or explores forever. Metacognitive Steering attempts to control when the model should broaden its search, execute a procedure, challenge its own reasoning, or abandon work after new evidence arrives.

The paper is detailed, public, and authored by the Autopoiesis team. It is also a preprint. The reported findings should be treated as company research awaiting wider independent scrutiny, especially when the ambition extends from changing a model's reasoning trajectory to producing reliable discoveries across medicine, engineering, materials, and other scientific domains.

Columbus-1 Moves From Answers to Investigations

Columbus-1 is the company's first autonomous research system built around Metacognitive Steering. It combines a main reasoning model, a smaller policy that provides strategic direction, tool use, and a persistent semantic memory graph that retains hypotheses, evidence, decisions, methods, and rejected paths.

The company reports that Columbus-1 identified eight independently reproduced, attacker-reachable vulnerabilities in BlueZ, the Linux Bluetooth stack. It also reports that the system directed the design, simulation, and fabrication of a 10-foot rocket intended to land propulsively with non-throttleable solid motors. Those demonstrations span software security and physical engineering, which is useful evidence that the architecture is not confined to one narrow scientific workflow. They remain company-authored preprint results rather than proof of broad autonomous discovery.

Autopoiesis enters a market where investors are funding several routes into AI for science. Lila Sciences is connecting AI with autonomous laboratories, Chai Discovery is building molecular-design models, and Basecamp Research is moving biological foundation models toward therapeutics. Autopoiesis is concentrating on the reasoning control and memory required before a research system can choose which experiment, hypothesis, or path deserves the next unit of time and money.

Why the $11M Seed Matters

The round gives Autopoiesis resources to hire and deepen its research while the AI market is learning that scientific fluency and scientific reliability are different products. Generating a plausible hypothesis can be valuable. Sustaining an investigation requires the system to preserve uncertainty, expose its assumptions, use tools correctly, change direction when evidence weakens, and leave a record another researcher can inspect.

That creates a demanding commercialization path. Autopoiesis must show that its control techniques improve research outcomes across tasks, that the system's work can be reproduced outside company-selected demonstrations, and that researchers can understand why it made consequential choices. In medicine and regulated science, a faster path to a wrong conclusion can create more work rather than less.

The opportunity is equally substantial. Much of science is limited by the number of qualified people who can examine a question, design a useful test, interpret an ambiguous result, and remember why an earlier branch was rejected. If Autopoiesis can make that judgment legible and reusable, its product becomes more than an interface for asking scientific questions. It becomes part of the operating record that allows people and machines to pursue them together.

The $11M finances that transition from a provocative technical thesis to a system other researchers can repeatedly trust. Autopoiesis has shown enough to attract serious capital and a growing user base. The work now moves into the harder territory its own technology is designed to navigate: uncertainty that cannot be resolved by sounding confident, and discovery that only becomes valuable after somebody else can follow the path.