A Year of AI Experiments, Failures, and Surprises

What I've learned using AI over the last year or so.

AI is powerful, but it’s not magic. It’s a tool that rewards clarity, structure, and iteration. Over the last year I’ve used it for coding, cooking, writing, and even building autonomous agents — and the pattern is always the same: the more intentional the workflow, the better the results. AI saves time, but only after you invest time. And that’s the real lesson.

Over the past year I’ve pulled AI into almost every corner of my work and hobbies — coding and system reviews at the office (including factoring out legacy third‑party controls and replacing them with standard .NET/ASPX controls), blog and book drafting at home, image and video generation, and personal builds like my fitness tracker website, the My Dog Whistle Android app, and a Wordpress site for the preschool where my wife is the director. I’ve even used it for capacitor logging and troubleshooting during an old tube‑radio restoration project. At this point, AI has become my go-to when I want to get something done quickly. Or even when I'm bored on a long drive in a car with no radio.

I've even built an entire website, AI for Busy Humans, primarily with AI (see the "how it was made" here), to help "busy humans" learn how to utilize AI as a tool to save time in their daily routines (at work, home, school, etc.), and managed to cook a phenomenal dinner with AI input (including Gemini yelling at me!). I even built an AI-based automated "AI security events" system (in make.com, utilizing the Google Gemini API as "the thing which evaluates multiple concatenated RSS feeds") which updates a spreadsheet so I can quickly review and mark "active" any new threats (and they show up on the live "Security Watchlist" on the website).

AI is a time saver (or it can be, if used correctly).

But it's not magic. It's a generative, predictive tool. It's only as good as its data (which can be annoying at times when it's looking at "current" events from a couple of years ago). And it's only as good as the prompts you give it. And, I'm finding out, sometimes it's only as good as the model it's running. (And sometimes it goes off the rails and does something totally unexpected, but cool.)

Recently I've been working on a few local-AI automated generation projects. That is, running a LLM locally on my own computer hardware (either my Nvidia GPU on my Windows tower or on my MacBook Air's integrated CPU/GPU with unified memory), feeding it with Python scripts. First I was attempting to build a "fully automated video generation platform" - which (eventually) it would decide its own genre, write a script, then animate the script through AI video generation. That generated some "interesting" results - and I was running into limitations based on my GPU (a 2060 Super 8GB card). (Or at least that was the presumption.)

I switched gears, pivoting to a "book writing" system, trying to turn my "AI for Busy Humans" into a book (using RAG - Retrieval Augmented Generation - and giving it access to the full source of the website). This has been an extremely frustrating project as I try to get the system aligned; the output drafts, which I'd feed back to AI or Gemini or Claude for review, were underwhelming. (Originally I wanted to create a "book-agnostic" set of agent files - python scripts - and have all the book-specific information in a "design" file that's fed in, so swapping the design doc gives a whole new book - but eventually I gave up on this and allowed the python agents to include book-specific logic just to try to get this book finished. This project is currently on hold.)

I subsequently "accidentally" created a book series topic - while working on "Bitty the Algorithm" (a kid's book, currently live on Amazon), my wife misunderstood the title as "Beating the Algorithm" - and I thought, "Hey, that's a good book title!" - and expanded it into a three-book series (none currently written). Now, on my MacBook, I'm working through a series of (much simpler) Python scripts to write the chapters. And it's getting better, but still a work in progress.

The interesting thing on these projects - I'm having Gemini and Copilot write the Python scripts to do the actual prompting of the local LLM. This is "sort of working" - but also a very iterative, repetitive process. Write the script. Have the script write the book. Review the output. Revise the script. Repeat.

But, it's interesting. It's something. It's a tool. At the moment, it's costing me time... but, once I get things ironed out, it should save time in terms of drafting, and - if I can get the "autonomous movie generation agent" running - that will be really, really cool. (Hopefully.)

I’m still ironing out the rough edges — the book pipeline, the autonomous video agent, the local‑LLM scripts — but that’s the point. AI isn’t a finished product; it’s a moving target. Every iteration gets a little better. Every workflow gets a little tighter. And somewhere in that loop of “build → test → revise,” the time savings start to appear. I’m not there yet on every project, but I can see the shape of it. And that’s what keeps me experimenting.

And after a year of pushing AI into every corner of my work and hobbies — coding, writing, cooking, automating, experimenting — the pattern is obvious: AI doesn’t replace the work. It reshapes it. You still have to think, design, revise, and iterate. But if you’re willing to do that, AI becomes a multiplier. Not magic. Not a shortcut. A force‑amplifier. And the more intentional you are, the more it gives back.

Stay tuned. More to come...