Can AI Let Anyone Run a Real Lab Experiment? with Keith Robison

Keith Robinson describes Ginkgo’s autonomous laboratory as a Lego-like system built from reconfigurable automation carts. An agent converts a scientist’s intent into a protocol; the user loads reagents, launches the run, and monitors it from a phone or laptop. The system records execution data, including timings that may explain experimental variation, and agents also handle analysis. Humans and “agent scientists” can work together, and Keith says he has run an experiment from 30 miles away without being physically present.

Keith says Ginkgo Cloud Lab is intended for middle-school and high-school students, adult amateurs, and anyone interested in genetic engineering. Users can access free computational pipelines, design sequences for binders, biosensors, or enzymes, submit the needed compounds, sequences, and protocol, and receive experimental data run on the autonomous lab. He argues that people no longer need to be skilled at bench work: someone can design the experiment computationally and quickly switch into experimentation through the service.

The host recalls a teenage “nuclear Boy Scout” who built a working reactor in his parents’ shed, using materials he obtained by posing as a university professor, as evidence that young people can do more than adults expect. Keith agrees that teenagers could contribute to drug or disease research once access stops being the limiting factor. He cites a sequence-logo visualization developed by a high-school student working with an experienced researcher, and notes that family-driven rare-disease projects already motivate people to begin research before graduate school.

Keith contrasts traditional science-fair projects, where students with relatives in advanced laboratories often have an advantage, with Cloud Lab, which can be accessed anywhere on the internet. He wants serious protein-engineering projects to cost less than what many iGEM teams raise merely to travel, but says DNA synthesis remains slow and expensive. Open-source or free-for-noncommercial-use tools now support enzyme, binder, and biosensor design, including pH-sensitive biosensors, and could let students investigate organisms or fungi whose genera have not yet been sequenced.

The host asks whether younger scientists benefit from not knowing which ideas were previously considered impossible. Keith says that lack of preconceptions can help, but regulations often arose after real failures, so therapeutic research still needs a balance between experimentation and safeguards. He uses agents to get a rapid overview of unfamiliar fields and identify open questions, while crediting cross-disciplinary newcomers in genomics and bioinformatics with bringing in mathematical and computational ideas. He also points to biotech’s concrete effects, including Vertex’s cystic-fibrosis drug, which gives roughly three-quarters of patients normal lifespans and fertility.

Keith says Ginkgo has integrated more than 40 instrument types in Boston, but many instruments were designed for humans rather than robots. Protocols may be stored under names without version control or hashes, so a foreperson reviews changes to ensure a run invokes exactly the intended protocol—even though that prevents Keith from creating one spontaneously at night. Vendor APIs contain missing features, bugs, and poor documentation; one API described a sleep argument as milliseconds while the underlying C function used seconds. He also cites JPEG-only diagrams, unexplained error codes, and the lack of physical simulation models for robots such as Agilent’s Bravo.