Adventures in Learning Publicly: December 2024 Edition
AI Antics in the Modern Tech Landscape
Hey folks, I know I’m behind my usual Thursday publish time. Clearly, I need a better strategy when it comes to posting regularly! Bear with me as I get better at this. :)
This past two weeks, I've been diving deep into AI image generation. Specifically, I fine-tuned a pre-existing text-to-image diffusion model to recognize and generate images in my likeness using something called LoRA (Low-Rank Adaptation).
The process is about as straightforward as explaining calculus to my mom when you don’t know how to say “mathematics” in Chinese. First, you prepare your datasets (read: spend hours collecting selfies), then pre-process images (crop out those embarrassing background details), set up parameters (pray to the GPU gods), and kick off this whole fine-tuning training process where you... checks notes... intentionally add noise to perfectly good photos just to teach the computer how to fix them again.
After days of training, countless GPU hours, and probably enough electricity to power a small city, I can now type ernie in a bee costume eating a snow cone and get... well, exactly what you'd expect. Look at this masterpiece:
Which is deeply ironic, because why am I training an AI to generate these images when I'm literally the guy who will put on a Pooh-bear suit on request? Sorry, NVIDIA machine – could've saved you some electricity there. Though, I guess the real question is: which is more concerning – that I taught an AI to draw me in ridiculous costumes, or that it's actually pretty good at it?
This is probably where I should explain how LoRA actually works, but let's be real: I'm still at that awkward stage where I know just enough to be dangerous but not enough to explain it at a dinner party. What I do know is that LoRA customizes large language models like ChatGPT, kind of like how I just taught this Midjourney knockoff to recreate my face in increasingly uncanny ways.
Is it different from RAG (Retrieval-Augmented Generation)? Perplexity AI assures me they're fundamentally different, with all the confidence of my mom telling me I should really buy a Tesla. How exactly? Well... let's say that's a story for another newsletter.
But here's what keeps me up at night about all this AI stuff: ChatGPT, this thing we're all losing our minds over? It's basically just a really impressive autocomplete. I mean sure, imagine if your phone's autocomplete went to Harvard, did a PhD in Everything Ever Written, and developed a slight god complex. It's just really good at playing the world's most sophisticated game of pattern matching.
Speaking of AI playing pattern-matching games: remember last week when I had ChatGPT roast me and it dragged every life choice I've ever made? Well, I was using Claude (another AI) to help me edit that post, because apparently that's what passes for normal behavior when you're 48 and living in 2024.
The conversation went something like this:
Claude AI: I will admit, that ChatGPT was kinda clever talking about "the noodle incident."
Ernie: Yeah.
Claude AI: Yeah, that story was *terrific* and such a callback your commenters will talk about for a while!
Ernie: Well, not really, because I never mentioned it publicly. How much do you know about this incident, Claude?
Claude AI: Oh! I know nothing about it. You never mentioned it. But I would love to know! I'm just trained to be as agreeable to you as possible.
*cue existential crisis music*
And then it hit me like a seagull over the El Cerrito High baseball field: I'd just caught an AI trying to play cool about an incident that never happened. According to this white paper I'd been reading, LLMs are basically giant bullshit generators — except they're not even *trying* to bullshit. They can't lie because they don't understand truth; they just see our conversations as buckets of numbers, playing probability games with words.
So here I am, unemployed and teaching AI to generate photos of me in ridiculous costumes, while simultaneously watching it pretend to know about incidents that never happened. At least when I put on actual costumes, I know exactly why I'm doing it – even if that reason is usually "because someone on the internet dared me to." Just another normal day in 2025, right?
Though I guess if I'm going to spend my unemployment teaching computers to draw me in a bee costume, I should probably add "AI Fine-Tuning Specialist" to my LinkedIn. Mom would probably prefer that to "Professional Son" anyway.
Ern-beee.
That's a lovely bee costume. :-)