Autopoiesis Sciences Is Building AI That Can Do Science

Autopoiesis Sciences is building AI for the part of science that rarely appears in the final paper: deciding which question is worth pursuing, when evidence should change a belief, and when a promising path has become a dead end.

The San Francisco company was founded in 2025 by Joseph Reth, Dr. Eike Gerhardt, and Dr. Lawrence Callahan. Its product stack begins with Aristotle, an AI research partner used for literature work, hypothesis generation, analysis, and experimental design. It extends to Columbus-1, an autonomous research system designed to sustain investigations over longer time horizons.

That makes Autopoiesis more than another interface for searching papers. The company is trying to turn scientific judgment into a learnable system, then test whether that system can direct research while the answer is still unknown.

About Autopoiesis Sciences

Autopoiesis starts from an uncomfortable constraint. Reinforcement learning works well when a system can attempt a problem repeatedly and score the answer. Much of frontier science offers no known answer, no instant verifier, and no clean reward for choosing the right next experiment.

The company believes the missing signal exists inside the research process itself. Scientists explore, reject, revisit, and revise ideas before any conclusion is ready to publish. Final papers preserve the result but discard much of that path.

Aristotle is designed to work alongside researchers during that process. It supports literature analysis, hypothesis generation, experimental design, and evidence review. Autopoiesis says thousands of researchers across pharmaceutical companies, universities, and government institutions use the product. That is a company-reported adoption figure, not confirmation that every named institution is a formal customer or endorser.

From Research Partner to Autonomous System

The second layer is Columbus-1. The company's technical announcement describes a system that combines hierarchical planning, tool use, a persistent semantic memory graph, and a method called Metacognitive Steering.

Metacognitive Steering attempts to detect a model's current mode of thought and intervene at inference time. Instead of changing the model's weights, a sparse controller moves the system among exploration, methodical execution, and critical reassessment. The company-authored arXiv preprint presents this as a way to preserve competing hypotheses and change direction during long investigations.

Autopoiesis reports that Columbus-1 identified eight attacker-reachable vulnerabilities in the BlueZ Bluetooth stack that an external security lab reproduced. It also says the system designed a 10-foot rocket intended to launch conventionally and land propulsively on solid motors. These are unusually concrete demonstrations, but they remain company and preprint claims rather than peer-reviewed proof that the approach generalizes across science.

Leadership Built Around Science and Systems

Joseph Reth is co-founder and CEO. The official company site says he began studying computer science at Modesto Junior College at 14, later attended San Francisco State University, worked on projects involving DARPA and the U.S. Army Research Laboratory, and previously led an AI lab called Lossless Research.

Dr. Eike Gerhardt is a co-founder with a background spanning startup research and venture investing. Dr. Lawrence Callahan is a co-founder and Chief Scientist whose career includes decades of scientific work, including time at the FDA.

The mix matters. Autopoiesis is not only training models or shipping a workflow tool. It is combining research methodology, AI systems, and enterprise adoption. The company's October 2026 addition of a founding go-to-market leader signals that the next phase includes bringing autonomous R&D systems to scientific and engineering organizations.

Why Autopoiesis Matters Right Now

Scientific AI is moving from summarization toward participation. The hard question is no longer whether a model can retrieve a paper or propose an idea. It is whether a system can maintain uncertainty, collect evidence, revise its plan, and produce work that another expert can reproduce.

Autopoiesis has a distinctive answer: use Aristotle as both a research product and an environment for capturing judgment, then use that data to shape systems like Columbus-1. If the loop works, each real investigation can improve how the next one is directed.

The $11M Seed announced in September 2026 gives the company more room to test that thesis. Threshold Ventures led the round, with participation from Informed Ventures, Alpaca VC, VitalStage Ventures, Correlation Ventures, DarkMode Ventures, and individual investors. The company said the capital would support team growth and continued research. It did not disclose a valuation, revenue figure, or detailed hiring target.

Hiring as a Market Signal

Autopoiesis has current openings in San Francisco across engineering and research. Its LinkedIn jobs board includes roles that move between product, research, and work with the founding team. The company describes an in-person culture centered on mission focus, sound decision processes, strategic differentiation, and end-to-end ownership.

That hiring pattern is evidence of the operating challenge ahead. Moving from a research partner to autonomous R&D infrastructure requires more than model performance. It needs reliable systems, evaluation, enterprise integration, security, scientific review, and people who can recognize when an automated result should be challenged.

The Test Ahead

Autopoiesis has put forward a technically specific thesis and demonstrations that are easier to inspect than a vague promise of scientific superintelligence. The next standard is higher: independent replication across domains, evidence of durable customer value, and safeguards appropriate for systems that could influence real experiments and engineering decisions.

The company's opportunity is significant because science is full of valuable questions that humans lack time to pursue. Its responsibility is equally significant because a system that directs inquiry must know when confidence is earned and when uncertainty should remain visible.

Autopoiesis is betting that the next leap in scientific AI will not come from answering faster. It will come from learning when to question the path itself.