AI should only run as fast as we can catch up

AI should only run as fast as we can catch up

AI should only run as fast as we can catch up.

The story of Daniel and Eric

Recently I have spoke with two of my friends who all had fun playing with AI.

Last month, I met with Eric, a fearless PM at a medium size startup who recently got into vibe coding with Gemini.

After getting familiarized with Gemini, Eric was genuinely amazed by how AI quickly turns prompt into playable web applications. It served great purpose as a first prototype to communicate ideas to designers and engineers. But Eric really wanted to skip those steps and directly ship it to prod. But he couldn’t really understand that Gemini actually built a single-page HTML file that merely looks like a working app. Sadly, one cannot build a reliable enterprise product out of this. And there is really no effective way for Eric to catch up on these technical details and outpace the engineering team himself.

Last week, I had coffee with Daniel, a senior staff engineer who recently grew fond of AI coding and found it to be the true force multiplier.

Daniel was skeptical of AI at first, but lately he hasn’t wrote a single line of code for months already. What he does is just precisely prompt the AI to create new components in an existing framework (involving Kafka, postgres, AuthN/Z, and k8s infra stuff) and adhering to certain preexisting paradigms. He would just spot-check the correctness of AI’s work and quickly spin up local deployments to verify it’s indeed working. Later, he pushes the changes through code review process and lands those features. All without writing a single line of code and it’s production ready just as if he wrote them himself. To Daniel, building and shipping things fast and scalable is simpler than ever.

Interpolating between the two stories

After speaking with Eric and Daniel, I suddenly feel that there is an overarching theme around the use of AI that we can probably interpolate out of the stories here. And after pondering for a weekend, I think I can attempt to describe it now: it’s the problem of reliable engineering - how can we make AI work reliably.

... continue reading