C5R is building physical labs for AI models that still struggle with experiments

Why it matters

C5R is betting that AI's scientific limits are partly an infrastructure problem: models need instrumented, real-world environments to plan experiments and learn from physical results.

C5R is building physical research facilities where AI models can plan and run experiments, an effort founder Michael Akilian (@akilian) says grew from a gap between advances in virtual AI and the limits he encountered at the lab bench. Akilian introduced the company in a September 24th thread on X, saying he began studying biology 18 months ago, first setting up a small apartment lab for worm experiments and later joining a lab at UCSF.

Akilian says C5R has been coming together for five months. His route to the project spans biology research at UCSF, hardware work at Apple and Misfit Wearables, and Clara Labs, an AI company he co-founded and sold, according to his biography. That mix sits behind C5R's central bet: getting AI to do useful scientific work requires a physical workspace connected to instruments and experiments, rather than another system that operates only on digital data.

C5R's website describes Facility-0 as a model-driven research facility covering biology, chemistry and materials science. Its software is intended to let a model inspect equipment inventories and specifications, design an experiment as code, and send instructions to instruments and people. The system then collects measurements so the model can analyze results and choose what to try next, according to the company's description. C5R says its equipment ranges from hydraulic presses to pipettes.

That workflow puts a human-operated lab inside the AI loop. C5R's own account says software can control equipment and issue instructions to people; it does not describe a facility in which models independently handle every physical task. The distinction sets the practical test for the business: whether models can reliably turn a research objective into experiments that people and machines can execute, then use the results to make better decisions.

The ambition reaches beyond biology. C5R says Facility-0 covers protein design, medicinal chemistry and solid-state materials, and that its SciUniverse project will evaluate whether frontier models can turn scientific objectives into verifiable results in those fields, either inside the facility or a digital twin. The company is positioning the lab both as an environment for carrying out research and as a place to assess what models can accomplish in physical science.

That is a different deployment problem from asking a model to answer scientific questions or draft a protocol. A model working with physical experiments must contend with instruments, supplies, measurements and instructions to lab staff. The quality of its output depends on whether those pieces can be coordinated in a way that produces interpretable results. C5R's stated approach ties model planning to that operating layer, making the lab itself part of the product.

Akilian's post describes C5R as a new company built by a group of friends and invites interested people to contact him. The company site lists a broader team and identifies San Francisco as its location. The announcement gives no funding figure or experimental benchmark; the case for C5R rests on the facility it is building and its proposed evaluation of model performance.

The commercial challenge is turning that setup into repeatable scientific work. C5R's own description makes clear that the system must connect planning software, physical instruments, measurements and human instructions. For Akilian, whose earlier work includes both biology and AI, the company is a bet that progress in scientific AI will depend on giving models access to the conditions where research actually happens.