mgx

How I Use AI

Artificial Intelligence isn’t new, and my experience is nothing compared to the decades of research, papers, and real breakthroughs that came long before I ever touched an AI model. By the time I started playing around with it, AI had already been through multiple hype cycles, from the early days of symbolic reasoning to neural nets, to deep learning breakthroughs that were quietly transforming fields like vision and translation. Academic papers, government grants, startup pitches--it was all already happening, quietly and seriously, far from anything I was doing.

But around the edges of all that, something more scrappy was brewing. Before ChatGPT became a household name, there were open-source GPT models floating around, early versions that you can download and run locally, often with barely enough RAM and a lot of trial and error. That’s where I started.

There's a Facebook group called "The Bot Appreciation Society" where hobbyist builders shared ideas and helped each other troubleshoot. That’s where I first heard about GPTs. People were experimenting with generating content that was actually readable. Some were making bots that told jokes or spun strange little stories. The technology felt raw but full of potential, like holding a tool that was almost ready to be truly useful.

I built a UFO bot that used a dataset of UFO sightings, feeding the entries to GPT-2 to generate even wilder stories and lore. It would then post these enhanced tales on Facebook. It sounds simple now, but at the time, stitching together web scrapers, content filters, and posting schedules felt like genuinely complex work. The whole thing ran on a patchwork of services I’d cobbled together, probably held up by digital duct tape and sheer luck. I wasn’t doing anything cutting-edge. But I was learning, experimenting, and feeling my way through a world where language and code were starting to blur.

when everything changed: November 2022 happened. OpenAI released ChatGPT to the world, and suddenly everyone was playing with the beefed up version of the GPT I'd been experimenting with. Within months, they released the API, and that's when things got interesting for me. I didn't jump on the "AI will replace everything" bandwagon. Instead, I started integrating it into my actual workflow, treating it like any other tool in my toolkit. I began following developments from Anthropic, Google, and others; not because I wanted to be an AI influencer, but because I needed to understand which tools worked best for different problems.

How I Actually Use AI Today

Learning and relearning: I use AI primarily for learning. Yes, I know about hallucinations and reliability issues. But as a hyperactive person, I sometimes struggle with complex explanations or dense academic writing. LLMs become my patient tutor, always available, breaking down concepts until they click. When I'm reading something technical and hit a wall, I'll paste the confusing section into Claude or ChatGPT and ask for clarification. It's like having a knowledgeable friend who never gets tired of explaining things differently until you understand. This extends to subjective areas too, like art interpretation. I'll bounce ideas off AI: "I think this piece means X, what's your take?" I know it's not human perspective, but it offers different angles I might not consider. It's like having access to a diverse group of thoughtful people, even when I'm working alone.

Programming and problem-solving: Stack Overflow used to be my go-to for coding questions. Now? I barely visit it. When I'm building mini apps and want to optimize something quickly, I'll paste my code snippets and ask, "How can this be better?" The feedback is immediate and usually helpful. There was one exception recently; Apple's CLGeocoder coder issue. AI couldn't help there because I needed better context. I had to dig through Stack Overflow comments to understand what was happening. Sorry, pattern matching.

And it goes deeper than quick fixes. I use AI (v0.dev) to create skeleton UIs when starting new projects, instead of staring at a blank screen, I describe what I need and get a basic structure to build from. When exploring unfamiliar codebases, I'll feed it to cursor, or paste sections in Claude or ChatGPT and ask AI to explain what's happening, like having a patient code reviewer who never gets tired of questions.

My automation scripts have gotten significantly better since I started using AI to optimize them. I'll take something I wrote months ago and ask, "How can this be more efficient?" Often, I'll get back cleaner logic, better error handling, or suggestions for libraries I didn't know existed. It's like having a more experienced developer review my work, except they're available at 2 AM.

The iteration speed is what really changed my workflow. Instead of getting stuck researching the "right" way to implement something, I can start with a working solution and improve it. When working with technologies I'm not familiar with, I begin by having a conversation with AI, "I need to build X with Y technology or Z API, where do I start?"-- and suddenly the learning curve becomes manageable.

Real-world problem solving: The most memorable AI interaction wasn't mine, it was my wife's. One of the chicks at her parent's place fell into waterhole and was barely breathing, essentially drowning. In that moment of panic, she opened DeepSeek and asked what to do. The AI walked her through emergency care steps. She followed them exactly, and we managed to bring the chick back to life. It's completely fine now. Could we have called a vet? Sure, but this was a spontaneous crisis requiring immediate action. The AI was there, it was accurate, and it worked. I'm genuinely grateful for that moment.

Family technology bridge: I built a website for my mom that turns Google Sheets data into a proper site. She runs a small plant nursery but isn’t comfortable with English-heavy technical resources. I showed her how to use Gemini directly within Google Sheets to optimize her product descriptions and site content. Now, with Gemini’s improved vision (Live), she can also point her phone camera at a plant and ask questions, making it even easier to learn on the go.

Analytical work: I've built tools that analyze data in bulk, finding patterns and insights that would take forever manually. These aren't common use cases, but they're where AI really shines, handling tedious analysis so I can focus on decision-making.

The meta moment: There's something poetic about how I'm creating this blog post. I recorded my thoughts as a voice note, which gets transcribed and structured by AI (using Whisper and other models within voicenotes.com), then I edit it into final form. It's AI as a thinking partner, not a replacement for thinking. I have AI integrated into Telegram (my own bots) for quick questions and simple automation. It's become infrastructure rather than novelty.

Which AI for What

Different models excel at different tasks, so I've developed preferences:

I don’t keep active subscriptions to every model at once. For elevated usage, I top up my OpenRouter credits and use whichever AI model I need. When it comes to image generation, I mainly create quick icons or logos when I need placeholder visuals. I'm not particularly interested in any image-gen AI, but it's useful for rapid prototyping.

To me, the AI revolution isn't happening in boardrooms or venture capital presentations. It's happening in small moments when the right information appears at exactly the right time, when tedious work disappears so creative work can emerge, when language barriers crumble and learning accelerates. These tools have become invisible infrastructure in my daily work. I don't think about them much anymore, they just work.

Inspired by Nicholas Carlini's "How I Use AI." I've been wanting to write something like this about AI for a long time. I'll expand the post as I discover more use cases and reflect on how it impacts me in the future.