Notes From a Former GenAI Sceptic
I am a millennial and a digital native. I've been on the Internet since I was maybe ten years old, if not earlier. In the late 90s and 2000s I built websites for fun. I know how to use scripting, macros and formulas to build really complicated Excel spreadsheets. I've even done some basic coding. But outside of the websites, which were never all that fancy anyway, I never built something that I would be comfortable sharing with other users. It was always personal projects.
Last summer I spent a month trying to get ChatGPT to become a personal agent to help me manage parts of my life; daily tasks, birthdays, calendar management, basic planning, things like that. I was always impressed by the quality of the writing and the back and forth chat interface but it was honestly a waste of time, half the time the AI came up with wrong or completely made up info. I could never figure out ways to actually extract value out of the tool.
A fundamental truth of large language models is that they aren't figuring things out. They're making probabilistic guesses about responses based on prompts that are being provided. And that truth was kind of central to the results I saw when trying to use the tools. I was getting different results when I would ask the same questions.
It's why you shouldn't use LLMs to do math. A guess, even a really good guess, is not the same thing as math. Math is truth. On any kind of task where I need truth, generative AI tools that leverage LLMs just aren't up to the task.
I figured maybe ChatGPT (and by extension other GenAI tools) was just another toy that I would play around with, but not a real phase shift in how I work.
But then in the last four to six months, I noticed that the quality of the publicly available tools, including the ones I get through work at IBM, had dramatically improved to the point where I was finding that I was able to do all kinds of things that didn't work previously; document summarization and manipulation, even legitimate research.
These improvements led me down the path last week to seeing if I could actually produce something with these tools to provide real value, and what I realized was that GenAI tools are very good at coding applications. This entirely bypasses the math problem. Use GenAI to do the things it's good at, and create applications that handle the math deterministically. Best of both worlds!
I spent half a day doing a back and forth of "this is what I'm looking for, what comes next?". I didn't say I want a web app. I said I want to create something that looks like this. The tool said, these are some things that you could do to create something that looks like this. I got a framework, a structure, and direction. I just started following that direction.
By the end of the next day, I had an app that did basically exactly what I wanted: a "Career Command Centre" to help me structure and track my career development, built to my exact specifications based on years of working within the consulting industry.
Over the next few weeks, I'm going to write up a series of posts that talk about the Command Centre, what lessons I've learned while building it, and how I see myself using these tools moving forward now that I've turned from sceptic to believer.