Could AI Have Done That?

image source

While toiling away for more than a week going through every one of the 800 or so blog posts on my site and re-formatting them, I ran smack dab into Microsoft’s build Conference on June 2nd. This being 2026, it was all about A.I., agents, and all things meant to support this structure. In addition, Microsoft announced a huge slate of ARM-based computers hat would meant specifically to run Windows, agents, and NVIDIA CPUS. That entire presentation was something. While embarking on this insanely manual process, the thought came to mind whether an AI could have done this.

After the conference at a roundtable, Satyalla Ndealla said,

Let’s conceptualize this as an ecosystem play rather than a single model or a single tool.

My first instinct is no. There are too many small things that need to change, and too many places where code was put in places it didn’t need to be. The AI would have to know that it needed to insert a line in what’s called the front matter of my posts. In its way, I guess it’s possible I could teach an agent to do that. But, how could I teach an agent to know when a post looks right? Teach it my preferences. Could I could translate my fundamental likes and dislikes into an agent? Highly doubtful, but maybe.

What keeps ringing in my head are two of the buzzwords I heard over and over at the conference: “AI truth” and “cognitive load.” These appear to be the fundamental problems with AI systems and might be two hills that are impossible to climb when we think about usefulness of them.

First is AI truth and the lack thereof. I’m guided by the very long but worthwhile read from Kyle Kingsbury titled The Future of Everything is Lies, I Guess. The essay runs down many of the hazards of these systems but first an foremost is how they’ll all lie to us convincingly. The AI is going to provide a plausible answer to our requests and we’re not going to be able to verify what a “tork-spindle-rotor” is or why it’s so crucial to our car’s operation. Leaving aside that no such thing exists, the AI will build a convincing and plausible argument why it needs to be replaced and build legitimate, workable blueprints or 3MF to output to the 3D Printer. This sort of con man - esque problem is widespread and makes me not trust the AI would properly fix my post formatting problem. Kyle’s not alone, Cringley is all over this too. And then there are the ethical considerations in using generative AI. And, while you’re reading, take in the very excellent essay from Matthew Butterick entitled Extinction-Level Capitalism. Listen, there is no doubt this is world-changing stuff, but clarity of mind is essential (and we’re still staying here at the low level with my rudimentary problem).

Secondly, the concept of cognitive load, which is a fancy way of saying mental overload will be pervasive with AI systems. The agent might go and happily reformat my articles for markdown, come back and excitedly announce that it finished. But there will be no way for me to wrap my brain around verifying hundreds of thousands of words have been correctly laid out. It will just come down to trusting that it did what I asked. It will “convince” me it did too, thus bringing back the idea of AI truth.

When I think of how far these systems have come, I look to what Google put out with Google Age AI Eloquent on IOS. No account, no tokens, just dictate and it cleans up all the text for you and puts it right into the clipboard. I wanted to try this and see my writing made with my fingers and perhaps the same thing made with this app, but I knew both would influence each other. I settle on a short walk and talk to see what the app could produce, here it is:

I took a walk on a big street, on Yonge Street. I was curious to see if this would pick up my voice, but at the same time, I knew I had to keep track of my thoughts, and it’s different when you’re talking verbally.

Transcribing this is a solid utility within the AI sphere. I think this is something I could go to regularly if I spoke to my phone and wanted to speak out phrases and have the phone fix it for me. Although, there’s always this feeling of looking weird walking down the street, but I assume that people just think I’m talking on the phone. I think this is a very mature aspect of AI involvement.

It’s clear that this small app will not be free forever and given Google, it will probably be dead inside of a year¹. But if I were to simply dictate all of my old blog posts, AI could do that on this tool with some modifications.

But I come back to the complexities of my task. Though some 800 blog posts that were hosted on several sites, the approach to displaying information went through huge changes from Blogger’s <div> tag heavy html to WordPress’ classic editor and now Gutenberg’s need to add whitespace throughout the post. This meant no less than four unique ways to represent footnotes for example. The challenge here is that I discovered a different footnote style when I encountered it and had to fix a new way going forward. There was no time to scan all the 800 posts ahead of time to map these patterns for the AI systems to learn. With 300 posts left as I write this, there may yet be more curveballs in the source to uncover. As I saw these new patterns, then and only then they could be translated to markdown.

Would I have gone to AI for this? no. Do I even think an AI agent or tool could even do this? No. Maybe I’m wrong and that wrong-ness is what this new AI era is about me an other tech people need to learn. As I think of these systems, I’m thinking more about what Cory Doctorow said about how our focus should be on AI:

Far more important than what the gadget does is who it does it to and who it does it for."

Thoughts on that, coming soon.