More or Less · Topics · Episode #174
Vibe-Coding a Shotgun: The Last-Mile Problem, and Should AI Set Your Pay?
Clip from the full recording, 38:52–44:09 · download mp4 · watch at this point on YouTube
Parker says AI can write the last mile of software at install time but vibe-coding the core is a shotgun pointed at your own legs; Sam says AI is excellent at using software and wants it to negotiate pay.
Jess asks whether Rippling really needs AI for what it already automates. Parker says software has always wrestled with the last-mile problem: a customer’s needs are simple but every customer wants something different, which is solved by a zillion configuration knobs. AI lets software be written at installation time, downstream of company context and policy preferences. But he warns that having customers build their own creates a shotgun to blow their own legs off, with a maintenance burden and vibe-coded security holes. His answer: don’t vibe-code authentication, permissions, analytics or approvals; just the last mile.
Sam reframes it: the biggest problem with software is that people are bad at using it, while AI can both write software and excellently use it, reading the documentation and hitting the right button where humans get lost 87 menus deep in Photoshop.
Parker describes AI that surfaces problems and asks you to click approve. Sam asks why click approve, and why the AI shouldn’t negotiate employee pay. Jess says a comp negotiation with an AI wouldn’t go over well, and notes some states require a human in the loop. Sam says most people are bad at negotiating and an AI would be better.
Key points
- Parker: the last-mile problem is solved historically by endless configuration knobs; AI can write that last mile at installation time.
- Parker: customers vibe-coding their own software creates a multi-year maintenance burden and security holes; don't vibe-code auth, permissions or approvals.
- Sam: AI can write software and AI can use software, and humans are bad at using it.
- Sam: AI is excellent at reading all the documentation and hitting the right button, while humans get lost in menus.
- Parker: the ideal is AI that finds a problem and sets everything up so you just click approve.
- Sam: why click approve at all, and why shouldn't AI negotiate pay with employees?
- Jess: she doesn't think an AI comp negotiation would play well; Parker notes some states require a human in the loop.
Where they landed
Quotes
“now you've built this big old shotgun for customers to blow their own legs off, right?”— Parker
“software can be written at installation time, where the software that you build for a company is downstream of their particular company context, their policy preferences.”— Parker
“the biggest problem with software is that people are bad at using software as users.”— Sam
“AI can write software and AI can use software.”— Sam
“shouldn't the AI just negotiate with your employees for how much they get paid?”— Sam
Suggested tweets
Parker Conrad on customers vibe-coding their own software: "now you've built this big old shotgun for customers to blow their own legs off."
Tweet this →Sam Lessin: "the biggest problem with software is that people are bad at using software as users." AI can write software and AI can use software.
Tweet this →Sam: shouldn't the AI just negotiate with your employees for how much they get paid? Jess: I don't think it plays when I talk to an AI.
Tweet this →All topics · Full episode #174 + transcript · Subscribe on YouTube · Spotify