I Used to Cook Dinner for 40 Interns Every Night to Bootstrap My Startup. Now They’re All Named GPT.

I’m a serial entrepreneur, angel investor, and startup advisor. When I started my first company, which I cofounded in college with my roommate, one of our biggest limiting factors early on—because we were bootstrapping the business and raised no venture capital—was being able to afford enough technical talent to build the initial version of our product.

Without money, we had to take a nonconventional approach. One of my first steps was to hire four interns from my network in college to help me build the app. Over the next year, four turned into eight, then eight turned into sixteen, and I made the somewhat crazy decision to not only hire all of these college-aged interns, but actually have them live with me while we worked on the product.

Over the next two years, that one original house with a bunch of interns and a couple of recent grads grew from one house to two and eventually four houses and two apartments—complete with dozens of folding tables and cheap OfficeMax chairs, and an industrial-scale food production operation where I was feeding 40 people a night and provisioning over 800 meals a week. The Washingtonian wrote a whole article about it- it was quite the experience.

While we eventually outgrew the “houses” era of my first company, our need for technical talent to build all of the tools and features that we needed only kept growing. By 10 years in, we had 150 software engineers, product managers, designers, data collectors, and more across four continents working to build and maintain our products, and 400 employees total across the business.

But that was before AI.

Now, as I build my latest startup—SmarterDumbphone—I’m realizing just how completely AI has changed the early-stage playbook. For me, as a bootstrapper, there’s no intern house, no folding tables, and no group dinners (unless you count my three kids). Just me and an army of AI agents.

I started building SmarterDumbphone just weeks ago. Today, we’re launching with dozens of features, integrations across multiple platforms, a polished marketing site, robust test coverage, a scalable serverless backend, and a solid legal and security foundation.

And I built all of it solo—with ChatGPT and a handful of other AI tools by my side.

Here are a few things I’ve learned building this company with AI:

1. AI works best when you treat it like an extremely fast but inexperienced intern.

AI tools can do a lot, and both the model and AI companies want you to believe that they can write an entire application for you in one go. Once in a while, ChatGPT or Claude occasionally generate something novel with a broad prompt (make my code do X), but most of the time, I need to know exactly what I want it to do, provide clear scaffolding or examples, and expect to spend most of your time testing, reviewing, and cleaning up the output. That part hasn’t changed from my intern days (or, in many cases, my 100+ engineer days).

What has changed is speed and scale. I can work on six things in parallel, hand off well-scoped tasks to the model, and then integrate them all later. Or, I can ask it to write tests for a file, and it will finish them in literally seconds, a task that would take an engineer hours to complete. That kind of velocity was unimaginable before AI.

2. AI can only handle the middle 80% of most technical tasks

It struggles with the first 10%, which is figuring out what structure, systems, framework, and overall approach is going to be sufficiently stable, testable, extensible, and performant. This is what makes “vibe coding” hard for people that don’t have significant technical experience.

When I asked ChatGPT’s best reasoning model to design my initial routing framework with plugins to handle different tasks (that first 10%)—its proposal was bloated, too abstract in many places, and not sufficiently extensible in others, hard to test, and generally something that looked good and sounded impressive, but failed to meet many of my requirements.

By contrast, once I had designed the extensible framework that I needed myself, I could give the model an example of one of the tools or plugins that I needed to build, provide it with a few additional requirements, and it could successfully build a full plugin by itself, which made the linear scaling of building many of these dramatically faster.

Modern AI also struggles with the last 10% of many coding projects—it is still far too confident and self-congratulatory when it’s fixing bugs. (I can’t tell you how many times I’ve had ChatGPT tell me “here’s the final, perfect, performant, tested solution to your problem!” and then I find three more bugs.)

However, that middle 80% still represents a lot of work, and many human engineers aren’t good at the first 10% and last 10% anyway. From my perspective, AI doesn’t replace all software engineers yet—and I’m not sure that it ever will—but it can certainly automate a truly significant amount of the work that goes on within an engineering team’s day-to-day.

In other words: you still need a real and experienced engineer to build the plan and make sure that the final product is high quality. But AI can handle a lot of the execution.

3. AI excels at testing.

Writing tests is by far the worst part of being a software engineer. Testing complex systems requires paying a lot of attention to mocking and writing enough test cases to actually cover all of the logic that you need. It also requires testing at many different layers—unit testing, integration testing, end-to-end testing, smoke testing, performance testing.

All of this used to require entire teams or significant levels of effort by software engineers, and I have found that by far the largest percentage speed-up in my time (with respect to quantity of work produced) has been in writing countless test suites and then using them to detect problems. The model doesn’t necessarily get all the tests working out of the box (I’ve found that for best results I need to phase my requests and first tell it to generate the cases to test, then build the test scaffolding and mocking, and then finally write all the tests), but because it iterates so quickly you can often just paste in the test output directly and it will be able to fix the problems long before I’d even have read the output (and agents like ChatGPT’s Codex are starting to do that iteration all on their own).

4. AI makes learning new systems and frameworks dramatically easier.

While some of the technologies that I’ve used in building SmarterDumbphone were very familiar to me, one of my core requirements in building this system was that everything run serverlessly with near-infinite horizontal scalability—because one thing that doesn’t scale well without engineers is on-call time.

This approach has required me to learn a lot of new technologies within AWS that support this approach, and I’ve had to create a bunch of abstraction layers, local tooling, and deployment infrastructure that is different than what I’ve previously used.

In my former life, whenever I needed to learn how to use a new framework or tool, I had to sit down and read the documentation, then build myself a dummy system to play with, then run into infinite bugs and spend untold hours perusing StackOverflow and Google and GitHub issues to try and figure out the many exceedingly long tracebacks that get generated whenever you try to run or install a new dependency.

With AI, I don’t have to do any of that. I can ask it for a comparison of a certain technology to something that I’m familiar with. I can ask it to help me write an abstraction layer so that I can speak to DynamoDB like I used to query Django. Or I can ask it to write me the skeleton of a function or a framework or a system so that I can understand how all the pieces fit together.

Better yet, I don’t even have to read error messages anymore—I can just copy them directly into ChatGPT, say nothing else, and ask it to figure out why my directory structure is wrong, or the version of a particular package in my system is incorrect, or something else. It’s a huge time saver and makes it possible for me to use more tools and frameworks that are outside of my comfort zone.

As my wife Kendall said while she was working through some bugs the other day in a language she had never coded in before—AI doesn’t replace coding experience, but with it an experienced engineer can write code immediately in a language or framework they’ve never used before.

5. AI is helping technical founders operate more easily on the “business” side of the company.

AI has been instrumental in helping me make quick progress on the non-technical parts of the company. For example, I’ve always found long-form writing and generating polished marketing copy mentally taxing, and as a result I’ve historically had to dedicate my most productive deep work time to those kinds of activities.

With AI, I can dictate a blog post, have the model format it and clean it up, and get it published without a ton of effort. And I’ve found the same thing to be true for reviewing legal documents, setting up the basic structure of the business, emailing initial contacts for the company, finding and supporting beta testers, etc.

For both my first company and many of the companies that I’ve mentored, technical founders could do many of those tasks—but they couldn’t do them easily enough, and wanted to dedicate brain space to actually creating their products.

I think we’ll see more companies in the AI era that are single-founder now that it is much easier to work on both sides of the cliched hacker/hustler divide.

6. AI is changing the calculus for founders—and for investors.

I bootstrapped my last company. It was hard, and I’ve always told other founders that bootstrapping only works when upfront costs and customer acquisition are both low, and if you need to hire engineers to build your initial product, upfront costs will be high.

But AI has eliminated half of that equation. If you can build the product with your own time and a $20/month ChatGPT subscription, capital requirements drop dramatically. You don’t need to give away 25% of your company just to build a working MVP anymore.

Capital will still be necessary for certain products that are particularly technically complex or expensive to create, and go-to-market costs will continue to demand the majority of investor dollars for Series A and beyond. But for many founders, the era of giving up 20–25% of your business and half of your time for the first few years of your company to fundraise for your basic fixed costs pre-revenue has dramatically changed.

I never thought I could build a production-grade tech company solo in a bit over a month. But here we are, and I’m really excited both for this company and for what this means for the industry at large.

The next few years are going to be wild. I can’t wait to see what happens when the next generation of founders realize they don’t need a team—or a seed round—to ship something real.

And I definitely don’t miss cooking dinner for 40.