Talks that carry receipts.

Every production AI workflow has the same bottleneck, and it is not the model: the output looks right, nobody can promise it is right, and the checking costs roughly what the AI saved. This talk makes the case that the durable layer of AI value is guarantee, not generation — and shows what owning the transformation seam looks like in practice: hard contracts around probabilistic cores, outputs that fail closed, and V = Δ, the discipline of measuring the reliability delta instead of claiming quality. Drawn from systems running in production and a doctrine published at zachshallbetter.com/gray.

A tabletop game funded at 330% on Kickstarter becomes 25,000 delivered units, a ~$500K seed, and a just-in-time factory that ran largely on its own — through re-engineered machines, custom equipment, and invented processes. This is the factory story as systems doctrine: why the leverage is never in the thing, it is in the machine that makes the thing, and how the same discipline transfers from a physical line to software platforms. Honest about the failures, priced in the lessons.

Building a nano-ceramic game board teaches you something software forgets: physics does not read your spec. Materials enforce their parameter envelopes, fail fast, and never drift out of date — the strictest contract most builders ever meet. This talk carries that lesson back into software: characterization over assumption, fail-fast validation at the boundary, and living documentation as the input-side contract that keeps generative tools honest.