Baruch Sadogursky and Leonid Igolnik on Reliable AI — Interview with Alexy

Baruch Sadogursky, head of developer relations at TuxCare, and Leonid Igolnik on reliable AI at AI By the Bay 2025, after their talk on whether AI-generated code can be trusted. Working with AI is rewarding and frustrating in a cycle: the first vibe-coding results look amazing, then fall apart, then guardrails and shared context make them work, and the reward is being able to reason that the result is exactly what you wanted. AI is stochastic by design, so reliability is repeatability, and right and same are not the same thing. From Grace Hopper's compiler onward, every new abstraction needed new tooling; with English as the coding language and a non-deterministic compiler, that tooling is spec-driven development, Amazon's Kiro and GitHub's Spec Kit, and an intent integrity chain, on the way from the trough of disillusionment to the plateau of productivity.