Exploring AGI Using AI

I have long been interested in Artificial Intelligence (AI) and what is now called Artificial General Intelligence (AGI). How long? Check out this post. If I'd pursued my Master's in Computer Science, it would have been in AI. My bachelor's was in cognitive psychology.

First Use of LLM AI

As 2025 waned, our house was sold, leading to a subsequent influx of cash that needed to be invested. A portfolio existed that needed a major revamp. I’d not used the current AIs, specifically the Large Language Models (LLMs), although Grammarly counts as some use. I turned to LLMs to create the portfolio.

I wasn’t naive, letting them do the design and walk away. Instead, it became a collaborative effort among Claude, Gemini, ChatGPT, and me. Sharing the project status among the AIs brought information and insights that a single AI wouldn’t have found.

The project evolved from a simple list of Exchange-Traded Funds (ETFs) into a portfolio with an operating manual. The manual provides clear guidance on when specific funds should be transferred to other funds based on market rates, market volatility, and the spread between private and public funding. Other guidance indicates when to reinvest dividends rather than draw on them for income. The guidelines have delivered good results when backtested against problematic markets since 2000.

This project ran from the end of December 2025 to mid-February 2026. Two week-long vacations did intervene.

Today is 22 March 2026, and the volatility tracking is getting a workout due to the Iranian military activity. It brushed against the flag to swap funds but avoided the need.

Using AI for AGI

The experience was encouraging. I turned to my interest in AGI.

Previous experience showed the advantage of working with multiple AIs for ongoing project reviews. This project brought a new insight: the chats develop personalities. I’d noticed this in general on the last project, so I created a startup prompt to explain to a new chat how it should interact with me and how I work. The prompt is based on the AIs' own observations. One example is that I ‘pivot’ in discussions rather than following the linear, logical path that AIs assume. Especially with the AGI discussions, I recall a previous thought, or a new one is triggered. The first AGI attempt evolved to use three versions of Claude: the theoretician, the engineer, and the coder. These three personalities originated from the initial conversations with them. The coder less so since it is Claude Code, which is specifically meant for C++ development. The engineer chat started with creating an architectural system design based on a document from the theoretician. The theory discussion was the open question: how can we do this? The interesting part about the engineer AI is that it would propose a solution to an issue that would work, but didn’t follow the underlying theory. The AI often completes my general observation with details and implications. One of them generously suggested I would have completed it. In many cases, I didn’t have the deeper knowledge the AI is trained on, so no, I wouldn’t have completed the thought.

This work will continue despite the first experiment leading to an insurmountable challenge. That doesn’t discourage me. I subscribe to Edison’s position: “I have not failed. I've just found 10,000 ways that won't work.” Direct work on AGI may be interrupted by a challenge from DeepMind on how to measure progress toward AGI: "Measuring progress toward AGI: A cognitive framework." My exploration, in collaboration with an AI, provides insight into the problem and identifies areas for improvement in this measurement methodology.