Shield AI's CTO: When a Bad Model Means a Crashed Aircraft, Not a Bad Chat

A typo in a chatbot is annoying. A hallucination in a fighter jet’s AI pilot is a smoking crater. That brutal reality is driving a new conversation at TechCrunch Disrupt 2026, where the show is ditching a single AI stage for two. The new “Real World AI Stage” acknowledges that software demos and physical deployments are now completely different beasts, and the stakes have never been higher.

Shield AI CTO Nate Michael will lead a session bluntly titled “Building AI Systems When Failure Is Not an Option.” The company, which builds autonomous systems for defense, operates in a world where trust isn’t a nice-to-have; it’s the entire product. Michael’s talk promises to move beyond vague safety pledges and dig into the actual engineering rigor required to validate systems before they touch the real world. It’s the kind of grim calculus that consumer AI companies never have to make.

Colossal Biosciences CEO Ben Lamm is also on the docket, which guarantees fireworks. His billion-dollar bet on de-extinction—using AI to help bring back the woolly mammoth and other lost species—has been called both a conservation moonshot and a dangerous distraction. Lamm will have to defend whether his company is engineering nature’s comeback or just building a very expensive zoo when the planet’s current biodiversity is already collapsing.

The edge-computing panel might be the most technically substantive of the bunch. FieldAI’s Ali Agha, Medra’s Michelle Lee, and Eclipse Ventures’ Aidan Madigan-Curtis will grapple with a problem most AI labs ignore: how do you run complex models when you have no cloud, spotty connectivity, and a split-second latency budget? Their answers won’t just apply to drones and satellites. Every industrial AI deployment eventually slams into these same constraints, a lesson the hype cycle often forgets until the pilot project fails in the field.

A separate session on scaling prototypes to production features Bedrock Robotics’ Boris Sofman and Foxglove’s Adrian Macneil, both of whom have the scars to prove that a working prototype is the starting line, not the finish. Their advice will be tactical—what they’d do earlier, what supply-chain nightmares they’d avoid—because in deep tech, the lab is a comfortable lie. The factory floor tells the truth.