I started experimenting with ChatGPT at work in early 2023. At first, my use was pretty ordinary: summarizing meetings, drafting emails and researching topics. It was useful, but it still felt like a better assistant rather than a fundamentally different way of working.
That changed when I started giving it more context. I loaded mission and vision statements, goals, strategies, competitive research, product information and details about the team into ChatGPT and, later, Claude projects. The answers became much less generic. Now I could use AI to think through strategic decisions because it understood enough about the organization and market conditions to create meaningful, cross-referenced outputs that would have take a mere mortal days to compile.
As those experiments became more useful, my needs evolved. I wanted more than a simple chat window. I needed a system that could monitor information, perform research and work on projects over longer periods of time. I also wanted it to operate somewhere other than my primary computer, where I could give it more freedom without worrying about what it might break.
That was the spark that led me to build a home AI lab.
I happened to have a decent GPU (RTX 3090) left over from my crypto-mining days, along with enough spare parts to assemble another computer without spending much money. I installed Linux because Hermes, the agent framework I wanted to try, didn’t support Windows at the time. Then I used Claude Chat, Claude Code and Hermes to get everything working.
The process was slow and not especially elegant. Claude would tell me what to paste into Claude Code, which would sometimes tell me what to paste into Hermes. I spent a lot of time moving instructions between tools, but I also learned how the pieces worked together and what I actually needed from the system.
The first project that made the effort worthwhile was a competitive-intelligence monitor for my day job at a Pacific Northwest credit union. Every Sunday night, it reviewed competitors’ websites, social channels, press releases and hiring activity. It looked at the products they were promoting, how they were positioning themselves and what had changed since the previous week.
Before that, competitive intelligence was something I either did manually or paid a vendor to produce. Either way, the result was a snapshot that immediately started to age. Three or six months later, someone had to do the work again. The new system kept running, which helped me and other leaders stay informed without waiting for another research project. It continues to do that today.
That early success led me to try turning some of these ideas into products. I explored productizing the competitive-intelligence system and built a job finder for friends and family. The problem was that the major AI platforms kept adding the same capabilities. Agents, scheduled tasks, design tools and better web access quickly became native features. At the same time, everyone using AI to code could chase and deliver similar ideas.
I’ve written more about what happened to those product ideas, but the main lesson was that building software is no longer the difficult part. The new challenge is finding a problem the platforms won’t absorb, differentiating your solution from dozens of similar products and then out-marketing everyone else.
My own setup kept changing too. I eventually returned to ChatGPT and gave Codex the permissions it needed to work directly on the Linux box. That removed much of the back-and-forth. Now I can describe what I want in the morning, let it work and come home to a prototype. That was the original goal I wanted Hermes to do from the beginning.
ChatGPT and Codex are my primary tools today, although I still use Perplexity for research and haven’t completely given up on Claude. I jump between these platforms, depending on the task. One leads for a while, another catches up, and the workflow changes again. I’m less interested now in committing to one tool than I am in maintaining the ability to experiment with all of them.
I also don’t believe the tools eliminate the need for human experience. After roughly 25 years in marketing and business leadership, I can usually recognize when an answer sounds plausible but is incomplete or wrong. I know which assumptions to question and where business reality is more complicated than the output suggests. An earlier-career version of me would not be able to use these systems in the same way.
That's been the biggest change in how I think about the lab. I’m less focused on building software products and more focused on using AI to solve business problems. Sometimes that means creating an internal tool. Or it might mean helping a business leader understand what AI could do for a team or process. Over time, it may lead me into consulting, helping multiple organizations work through those questions.
I’m not blind to the risks AI creates for jobs, the economy or society. I also don’t think it's going away, and putting my head in the sand doesn’t seem like a good plan. The lab gives me a place to keep learning while the technology, the hype and the business models sort themselves out.
For now, I can learn about something in the morning and find out whether it is actually useful before dinner. That is enough reason to keep the machine running.