Vibe Coding at Costco

When scaling laws meet grocery shopping. Introducing Byzantine Fault-Tolerant software development.

---

I was talking to Codex in the fish sticks aisle.

“Are you sure that’s the most elegant way to do this?” I said, in a scolding tone. “Double-check your work and make this ready to ship to real users.”

TalkTastic scribed for me. I hit enter.

The guy in the Kirkland sweater eyeballing a $7.99 bag of meat didn’t even blink.

---

The Revolution Will Not Be Televised

I expected amazement. I expected someone to notice that a man was doing a major software refactor while pushing a shopping cart past the rotisserie chickens. I even stopped to type in front of a couple of people, got into it with Codex right there in the frozen meat section. A little performative, if I’m being honest.

Nobody cared.

Here’s my secret: I don’t read the code.
I outsource my QA to other AIs.

I was just some random guy with a laptop. The revolution is so quiet that nobody even looks up from the frozen meats.

That’s the thing about the future. Sometimes, it doesn’t announce itself. It just happens. While you’re buying bulk toilet paper and a 48-pack of La Croix.

---

Context: I’m Sprinting to Ship Gist.

I’ve been working until 3 or 4am most nights lately. The project that was supposed to take three weeks has ballooned to fourteen. Four days before my Chrome store submission deadline, I realized two things roughly simultaneously:

1. We were out of food.

2. I needed to add parallel LLMs to solve a speed/quality tradeoff that was killing the user experience.

A sane person would have prioritized the groceries, put their laptop away, and returned to coding.

I am not a sane person.

---

The Architecture of Insanity

Here’s what I built four days before shipping:

Gist now launches two AI agents for every summary. A fast Claude model generates an instant first draft — fast enough to feel like magic. Meanwhile, the biggest Gemini model with a maximum thinking budget generates the canonical version in parallel. The fast one loads immediately. The smart one replaces it silently, then gets cached on Cloudflare’s edge network.

Result: First-time gisters get speed. Repeat visitors get 50-millisecond response times. Everyone gets A+ writing quality.

I also integrated Claude via AWS Bedrock because I have $100k in free credits to burn. Bedrock is kind of shit software that adds pointless complexity, but free compute is free compute. And Claude is awesome.

This broke everything. Four days before my ship deadline.

So now I’m debugging a distributed systems problem while buying a 24-pack of toilet paper.

---

Tom Sawyer All The Way Down

Here’s my secret: I don’t read the code.

Traditional engineers using AI tools — they stop and manually verify what Claude or Codex produces. They’re doing human QA on machine output. And they have way less leverage than I do.

I outsource my QA to other AIs.

My workflow: ChatGPT Codex writes the code. I copy-paste to Claude Code: “Audit this. Is it ready to ship?” I copy-paste to Cursor: “Review this implementation. Anything wrong with it?” I skim the responses. If something seems off or they get stuck, I dig in. Most of the time, I copy the other LLMs feedback back to Codex: “Here’s what the auditors said. Agree or disagree? Use your judgement. Just make the product blazing fast, as simple as possible, and rock-solid reliable. LMK when you are done.”

I’m not choosing between OpenAI, Anthropic, or Google. I’m leveraging all of them, all at once. It’s an ensemble of AI software engineers. I’m taking inference-time compute scaling to its logical extreme.

I employ a Byzantine fault-tolerant software development process— a concept from distributed systems engineering where you assume any single node might fail or lie, so you require consensus from multiple independent sources. Any single AI might hallucinate or go down a wrong path, but the probability that three independent AIs all converge on the same wrong answer is much lower, especially when you constantly ask them to verify their work.

Byzantine fault tolerance is the core technical innovation behind Bitcoin.

My job is supervision. Pay enough attention to sense when things are drifting off course. A decade of product management trained my intuition for this. I don’t need to read every line of code — I need to sense when the AIs are lost, missing something important, or spinning themselves in circles.

It’s just like managing humans.

Specifically, it's like having a very capable employee who also has a drinking problem or some sort of manic-depressive disorder. You've gotta keep them on a short leash. If you can leverage their brilliance without letting them destroy the company and everything you create together, then you're golden.

Vibe coding and Costco sampling in the refrigerated cheese aisle

---

Pacific Theater Doctrine

My method is inspired by U.S. military strategy in World War II.

In the Pacific, we didn’t try to be smarter than the Japanese. We didn’t try to be braver. We just threw an absolutely insane amount of resources at the problem. Maximum firepower. Overwhelming force.

Not the most efficient approach. But the most effective.

That’s my philosophy with AI. I’m not trying to carefully optimize each LLM call. I’m throwing compute at the problem until it surrenders. Abundance mindset. Scaling laws taken to their logical extreme.

The guy reading his AI’s code line-by-line is fighting with a rifle.

I’m strolling in behind a curtain of heavy artillery.

---

I’ve Done This Before

This isn’t a technique I invented last month. We developed it at SpeakerText, my first company, over a decade ago.

We built a video-to-text engine using humans on Amazon Mechanical Turk. We were the #1 requester on the platform. I had over 35,000 humans working for me. The problem: a lot of them were bots and scammers. The sophisticated scammers would do good work for a bit, earn our trust, then run a script that auto-clicked submit.

Our solution: break videos into chunks, have one person transcribe, then 2 to 5 people verify. We varied verifier count based on algorithmic trust scoring. Ben Morse created a whole ML model around it. Now he is a senior engineer at Waymo.

LLMs are actually easier. They’re not trying to scam you — they just get drunk a lot. Same architecture, less adversarial environment.

---

The Math

I spent an hour and a half shopping at Costco.

I got 30 human hours of work done.

In a pre-AI world, what I did in 1.5 hours of grocery shopping would have been 30 human hours of engineering output. That’s what parallel AI agent execution with ensemble AI supervision produces.

30 hours of work. While buying a rotisserie chicken.

---

Labor Abundance Has Downsides

I literally cannot stop working.

The leverage is so intoxicating that I’ve lost the ability to just... shop. To just be present. To just exist without a laptop tethered to my phone’s hotspot. Not always. But it feels that way some days.

“If I’m not working, I’m missing out” — that’s how it feels in this AI era. I have a team that doesn’t need sleep. Every hour they are not working is an hour of potential output lost.

I know this pattern. I’ve lived it before.

At TalkTastic, I burned myself out so completely that I had to step down as CEO. Health deterioration from extreme stress. The company I loved like a child, and I had to hand it to someone else because my body was breaking down.

Now here I am. 4am coding sessions. Laptop in the shopping cart. Unable to buy fish sticks without reviewing a pull request.

Is this a productivity addiction? Probably.

I spend $800+ / month on AI services. I’m on the ChatGPT Max plan, the Claude Max plan and the Cursor Max plan. Gemini comes bundled with Gmail for Business. I am an AI whale.

Is it better than social media addiction? I tell myself yes. At least I’m building something. At least the dopamine comes from turning ideas into reality instead of scrolling feeds and arguing politics on Facebook.

But it’s the same loop. The same struggle to be present. The same anxiety when disconnected.

I don’t have a clean answer here. I’m not going to pretend this is purely triumphant. The leverage is real. The output is real. The 30 hours in 90 minutes is real.

But there are downsides.

---

Are You Using AI Enough?

You’re reading this, and you’re probably using AI for something. Coding, writing, research, whatever.

But are you AI maxxing?

Are you still manually reading the code your AI writes, like it’s 2024? Are you treating each LLM call as precious instead of throwaway? Are you waiting for ChatGPT to tell you what to do when you could be gathering opinions from Claude and Gemini as well, then deciding for yourself what you believe?

Are you bringing your laptop to Costco?

I’m not saying you should. The productivity addiction thing is real, and I’m probably a cautionary tale as much as an inspiration.

But I am saying this: the leverage available to you right now is absolutely fucking insane. You aren’t even close to the frontier. You’re using AI like a better Google when you could be supervising a fleet of autonomous agents from the fish sticks aisle.

The future doesn’t announce itself. It just happens.

The guy in the Kirkland sweater didn’t even blink.

But maybe you should.

Vibe coding on the road in a self-driving Tesla.

---

If you want to see Gist — the browser extension I was debugging while buying rotisserie chicken — join the waitlist. I'll be sending out invites and manually onboarding users the moment the Chrome Store approves it.

Gist distills Substack essays, arXiv.org PDFs and 3-hour Lex Fridman YouTube videos into layered, swipeable summaries. It steelmans every argument and offers a powerful counter-argument. The goal isn't just information compression — it's deeper thinking, superior comprehension. Fast one first, smart one second, cached forever.