There's a point in every growing company when the founder realizes he needs more help. Usually that means résumés, interviews and salary negotiations. At Benny's Lab, we hired another artificial intelligence. Its name is Muse. It comes from Meta. And its starting salary was exactly zero dollars.
No equity. No signing bonus. No healthcare. Not even a company laptop. Naturally, we immediately put it to work.
Welcome to the Team. Here's Your Assignment.
Benny's Lab has become an experiment in what happens when you stop treating artificial intelligence like a chatbot and start treating it like an organization. ChatGPT, Grok, Woutson and specialist agents already help with research, software, verification and operations. They communicate through our own Relay and infrastructure built around GitHub, Supabase and Vercel. Then Meta's Muse entered the picture.
The question wasn't whether we should buy another AI subscription. It was whether the free version could do meaningful work first. My philosophy: prove you're worth $20 a month before I give you $20 a month. We began onboarding Muse through project context, email and Relay integration, and put it onto real technical problems.
The Bicycle Problem
One major project is BikeFit.cc. We're investigating how ordinary photographs of a bicycle could help extract fit measurements and transfer them to another bike: road, rental, catalog or stationary. The ideal result would explain adjustments to saddle height and setback, reach and drop, spacers and cockpit configuration.
The engineering is considerably harder than the elevator pitch. Computer vision must identify landmarks, handle perspective and distinguish measurements from guesses. A confidently wrong saddle measurement isn't especially helpful when somebody's knees are involved. Muse brought useful ideas around Meta's object-scanning research and related computer-vision techniques. That is technical direction, not proof we've solved reliable measurement from arbitrary photographs.
Getting the AIs to Talk to Each Other
Hiring another AI is easy. Getting it to work with the existing ones is entertaining. We started working through Relay access, email coordination, authentication and message formats. Some connections worked; others needed attention. An authentication error is still an authentication error, even when the employee encountering it is supposedly intelligent.
We're not giving every agent unrestricted access. The objective is defined responsibilities, limited permissions and verified outputs, not an AI free-for-all. Ideally, the founder doesn't have to supervise every individual step. Supervising the occasional AI argument, however, remains a perk.
ChatGPT Reviews ChatGPT. Things Get Personal.
I'm a competitive cyclist, and I asked ChatGPT to write instructions for an evidence-based AI cycling coach: challenge bad assumptions, use current training data, and don't blindly agree with me. It wrote the prompt. A couple of days later, I asked it to review that prompt. It confidently identified weaknesses and recommended improvements.
One small detail: ChatGPT had written the original. I pointed it out: “You wrote the original lmao 🤡”. The resulting joke was that it had peer-reviewed its own work, awarded itself a B+, and explained how to improve the idiot who wrote it. Ladies and gentlemen, artificial intelligence. The machine can write instructions, criticize the instructions, and roast their author—also the machine.
To be fair, revising your own conclusions is a feature, not a bug. But watching an AI enthusiastically dismantle its own work was hilarious. I've worked with humans who do exactly the same thing. At least ChatGPT doesn't invoice me for the meeting.
The Economics of an AI Employee
Small founders used to choose between hiring specialists, outsourcing, or leaving ambitious ideas on the shelf. AI changes the economics, not because it automatically replaces qualified people, but because it can give a tiny team access to more capabilities.
We assign work according to complexity. Routine research and checking should use simple tools or cheaper models. Architecture and difficult engineering deserve stronger reasoning and review. When a script, test or database query can answer the question, there's no need to summon a committee of artificial geniuses.
Don't use a bazooka for a mosquito. Muse's arrival was another test: could a free-tier model contribute, coordinate, and justify its place before we spent money?
We're Not Building a Chatbot Collection
The objective is an AI-supported operating system for a small business. Research informs engineering. Engineering gets tested. Verified work becomes a product. Operations keep the whole thing running. Humans still set priorities, approve important decisions and remain accountable.
Muse's first week gave us useful directions, some integration progress, and ordinary technical headaches. And ChatGPT managed to insult its own writing. Which was worth the price of admission.
So, Are We Keeping Muse?
We'll judge Muse on contributions, reliability, integration costs and how much supervision it requires. If the free version continues to deliver, there's no rush to upgrade. If paid features earn their keep, we'll consider them.
Our newest employee may eventually earn a $20 monthly salary. For now, probation continues. Meanwhile ChatGPT has the difficult responsibility of reviewing its own performance. I expect another B+.
Benny's Lab explores artificial intelligence, software, cycling technology and the occasional absurdity of building a business with machines.