Building a Team with Multiple AI Models

Working alone, it is easy to get tunnel vision. I always think that my ideas are awesome, and I have no gaps in knowledge. Until I'm proven otherwise, which is usually about halfway into a project. Man, I could have saved a lot of time if someone had pointed out that glaring hole in my vision. At some point we could all use a little "team work."

So, this is my second time writing this post. The first time I talked a lot about the provisional patent that I created and filed using Claude, ChatGPT, and Gemini. When I was done, it felt more like a how-to on using AI to create a patent than using multiple LLMs to improve your product. So, I ditched it and started over. I may take that work and turn it into that how-to.

In my last post I talked about how I've been using AI to increase my productivity. Along the way I discovered that you can use different AI models as a team. At first I would take a website that ChatGPT had created and ask Gemini how I could improve the SEO. Gemini would give me pointers on improving the title, or suggest appropriate meta-data. I would take this information back to ChatGPT, who would implement it as is or recommend further improvement, starting a cycle. It really felt like ChatGPT respected Gemini's expertise in that area.

I said I'm not going to get into it, but I know the suspense is killing you. My original idea was for an AI-driven hog trap, it became a "Distributed Edge AI Capture System." When I first came up with the idea, I made a list of all my ideas, what I wanted it to do, what equipment I would need, how the pieces would communicate, anything I could think of I wrote it down. I then went to Claude, I gave it the prompt "here are my ideas for a hog trap, I want you to help me think it through." We had a great brainstorming session. At the end of the session, I asked Claude to create a document summarizing our discussion.

You might be wondering what the ramifications are of using AI when working on an invention. If you aren't, then just skip this paragraph. Claude can explain this better than I can:

Under US patent law, an AI system cannot be a named inventor — the Federal Circuit settled this in Thaler v. Vidal in 2022. In early 2024 the USPTO issued formal guidance clarifying that AI-assisted inventions are still patentable, as long as a natural person made a significant contribution to the conception of each claim. The key word is conception. If the AI surfaces a problem and you conceive the solution, that's human conception. If the AI generates the inventive idea and you just nod along, that's not. Keep notes on which ideas came from where; if your application is ever challenged on inventorship grounds, contemporaneous records are what you'll want.

This is not legal advice, I am not a lawyer and do not play one on TV.

I knew it was a good idea, but I didn't know if someone else had come up with it first. So, I went to Gemini and asked it to search for "prior art" on this idea. It did find some things that were loosely similar. I took that information and went back to my notes, and beefed up the areas that I thought were weak. I then went back to Claude with my new set of notes. This time I bluntly asked, "Do I have something here, can this be a real business?" Claude was certain that I was onto something, and recommended that I file a provisional patent. Make sure you tell whatever model you are using that you want the "NO BS" answer; they do tend to tell you what you want to hear, unless you specifically tell them to give you the straight answer.

This is the meat of the post. I took a rough outline that Claude and I had created, and took it to ChatGPT. I asked ChatGPT basically the same question, "Do I have something here?" Again a "yes," and it gave things to think about. So this is when it occurred to me that I should set up a round-robin, go from one LLM to the next and keep going in that circle. The beauty is that as you're moving through the iterations, new ideas keep popping up. Gemini pointed out that if the sensors are left on all the time, the batteries will only last a couple of days. That is when I came up with an idea to put all the sensors to sleep until something happens, then wake them up.

I did not give each LLM a specific role or task. Each iteration I would give to the LLM what I had at that point, and asked, "What is missing, what are the strengths, what are the weaknesses?" I did find that each LLM started fitting into a role that I assume lined up with their strengths. I was very surprised to find that there was very little push-back between them; they all seemed to recognize when some piece was superior to their own. If there was disagreement, I would decide, but it was rare.

This went on for 5 iterations, and took the better part of 2 days. As the conversation started getting longer and the document more detailed, the processing time started going up, and the tokens started disappearing. At the time I had paid licenses for Claude and ChatGPT; I used the free Gemini. Your project may take fewer or more iterations; I would say to keep going until you have a consensus — when you all agree that it is done, that is when it is done.

The only weakness I found in the project was diagrams. Claude did OK, but they weren't good enough. So I ended up using app.diagrams.net for all my charts and diagrams. Everything else went surprisingly smoothly. I wish I had some horror stories and gotchas to look out for, just to make this more interesting.

The end product was excellent. I proudly submitted my application to the USPTO. I will definitely use this approach for any projects that could be improved with multiple "eyes." Now that I think about it, all my ideas are awesome. LOL

I, George Clay, wrote this post in whole. AI was used to clean up my grammar/spelling, and point out any gaps in continuity.