GPT: Generally Pretending To (understand)

For two years I've been confidently telling people that GPT stood for "General Purpose Technology." I said it in meetings. I put it in presentations. I don’t recall where I learned it, but I didn’t make it up. It made sense and I adopted it.

But no, I was wrong, GPT stands for Generative Pre-trained Transformer.

After looking that up (yes, on ChatGPT), I stared at my screen for a solid minute, wondering how many people knew I was wrong and said nothing. Then I started wondering what else I was getting wrong.

So here today I’d like to bust some jargon and explain this stuff - in case you are interested and want to avoid getting things wrong too.

Let's break down GPT, because once I understood what each word actually meant, everything clicked:

Generative: It makes stuff up. Not "retrieves" or “finds”, it literally generates new combinations of words based on patterns. Like that colleague who sounds brilliant but is basically freestyling based on every meeting they've ever attended.

Pre-trained: All its learning happened before you met it. Your prompts aren't teaching it anything. It's like hiring someone who read the entire internet up to some point but hasn't learned anything since (you can of course instruct it to go learn something contemporary)

Transformer: This is the architecture bit. It means the AI can pay attention to relationships between all words in a sentence simultaneously, not just adjacent ones. It's why it can remember what you asked about three paragraphs ago. Also why changing one word at the beginning of your prompt can completely change the output, something I learned after wondering why my prompts were so inconsistent.

When you type "Write me a marketing strategy" and hit enter (and I hope that your prompts are far more sophisticated than this…), this is what actually happens:

First: Your words get chopped up into tokens (chunks smaller than words). "Marketing" becomes "Market" and "ing." This is why character counts don't matter, token counts do. That 4,000 token limit? It's roughly 3,000 words.

Second: Each token gets turned into math. Literally. "Marketing" becomes a point in mathematical space, hanging out near "advertising" but far from "engineering." The AI doesn't know what marketing IS, it knows what words usually hang around with marketing. It's like navigating by social proximity rather than actual understanding. "Oh, marketing? That usually stands near ROI and campaigns and budget."

Third: it looks at all your tokens simultaneously and figures out which relationships matter most. This is why "urgent" at the beginning of your prompt hits different than "urgent" at the end. It's calculating the mathematical weight of every word against every other word. Mind-bending stuff.

Fourth: Your prompt rockets through 96-176 layers of processing (depending on the model). Early layers figure out grammar. Middle layers understand concepts. Final layers add nuance. This all happens on massive computers, probably in Iowa, burning through enough electricity to power a small house, costing about a penny each time. Every. Single. Query.

Fifth: It generates one token at a time, calculating probabilities for what should come next. Not retrieving, but generating. It's literally making it up as it goes, which explains why it can be brilliant one sentence and bonkers the next. It's not looking up answers; it's playing a hyper-sophisticated game of "what word probably comes next?"

Sixth: Before you see anything, safety filters check if the AI is about to say something problematic. This is why it sometimes refuses reasonable requests as the filter isn't as smart as the main model. Like having a paranoid legal team review every email before it sends.

The whole thing takes about 3 seconds and happens thousands of miles away from your laptop, in a data centre that probably has better cooling than your office.

Context Window: How much the AI can "remember" in one conversation. Once you exceed it, it starts forgetting the beginning.

Temperature: Controls how creative vs. conservative the output is. Low temperature (0.3) = boring but safe, like your legal team wrote it. High temperature (0.9) = creative but might go rogue, like your intern after an energy drink. Most of us never touch this setting and wonder why outputs are inconsistent.

Hallucination: When it confidently makes stuff up. Not a bug, it’s literally how the system works. It's generating plausible-sounding text, not retrieving facts. Never trust it with numbers or citations without checking.

Fine-tuning: Training a model specifically on your stuff by adding domain-specific knowledge files to create a more direct context for your prompt and how the LLM responds to your query.

Embeddings, vectors, attention mechanisms? Unless you're building AI, you can safely ignore it all. It's like understanding engine mechanics. This is useful if you're a mechanic, irrelevant if you just need to drive to work.

I expect that this topic could easily filly 800 pages instead of just 800 words, so I’ll stop here. So here's my challenge: Take a moment to figure out one AI thing you've been pretending to understand. Google it. Ask ChatGPT to explain it like you're five (ironic, I know, but this prompt works). Watch a YouTube video.

Then Monday, kill one AI initiative that doesn't make sense now that you actually understand how this stuff works.

Your budget will thank you. Your team will thank you. And you'll finally stop nodding along when vendors throw around terms like "semantic understanding" and "cognitive reasoning."

We're all just doing our best here. Might as well be honest about it.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 13 September 2025 in the Brandflow newsletter on LinkedIn.