No Priors Ep. 34 | With Ginkgo Bioworks Co-Founder and CEO Jason Kelly

While DNA resembles digital code, biological systems are noisy, non-deterministic, and physically embodied, so engineering them requires tolerating uncertainty and leveraging probabilistic and evolutionary approaches rather than strict logical debugging.

Ginkgo’s robotic ‘foundry’ turns manual experiments into high-throughput, standardized processes that generate large, consistent datasets—exactly what’s needed to train effective AI models for protein engineering and cell design.

Because humans did not evolve to read or design biological code, AI models trained across massive protein datasets may achieve superhuman performance in understanding and generating functional proteins much earlier than they do in human-language tasks.

Ginkgo is betting that many organism-engineering tasks share enough commonality to justify shared infrastructure and platforms—contrary to the prevailing belief in pharma and ag that every product requires its own bespoke tech stack.

Large models are now feasible, but high-quality, sufficiently rich biological and clinical datasets—especially those tied to outcomes like efficacy and safety—remain scarce and fragmented, slowing full-stack AI drug discovery.