It’s Tuesday, which means I'm talking tech. If you’ve tried to buy a new graphics card lately - maybe eyeing that RTX 5090 to make your games look crisp - you’ve probably noticed the price tag looks more like a down payment on a car. You can thank the "AI Boom" for that. But with terms like "AGI," "LLM," and "Hallucination" flying around, it’s hard to separate the sci-fi from the software. Today, I want to break down exactly what this technology is, why it’s eating the world’s supply of silicon, and why it can write a sonnet but can’t actually "feel" anything.
The "NPC" vs. The "Player" (Defining AGI)
To understand why I say "it's just math," we have to define the difference between the AI we have and the AI we fear.
In the tech world, we often talk about the Holy Grail of computing: Artificial General Intelligence (AGI).
AGI is a hypothetical machine that possesses human-level cognitive abilities. It can learn any intellectual task that a human being can. It has agency, it has intent, and - crucially - it can "think" outside of its training data. In gaming terms, AGI is a Player Character, which is essentially controlled by the player. It has its own motivations, it can improvise, and it can derail the campaign if it wants to.
What we have right now (like ChatGPT or Google Gemini) is Narrow AI.
Narrow AI is a system designed to do one specific task incredibly well. In this case, that task is "predicting the next word in a sentence based on probability." In gaming terms, today’s AI is a Scripted NPC (Non-Player Character). It might have thousands of lines of dialogue and feel incredibly realistic when you talk to it, but it has no inner life. It isn't "thinking" about the quest; it is simply executing a dialogue tree based on the inputs you give it.
The "Ghost" in the machine isn't a soul; it's just a very, very convincing dialogue script. We haven't created a new life form. We've just built a better NPC.
What is an LLM? (The Fancy Autocomplete)
You’re likely seeing "ChatGPT", "Google Gemini", or "Microsoft Copilot" in your feeds constantly. These are Large Language Models (LLMs). To explain them to a non-tech friend, don't use the word "brain." Use the word "Predictor."
Imagine you are texting a friend. You type: "I am going to the..." Your phone suggests three words: store, movies, bathroom. Your phone doesn't "know" you need milk. It just knows that statistically, in the English language, the word "store" often follows "going to the."
An LLM is essentially that, but trained on nearly the entire internet. It isn't "thinking" about your question. It is looking at the string of words you provided and calculating, with billions of variables, what word is statistically most likely to come next. It does this one word at a time, thousands of times a second.
The Reality: It is a probability engine, not a reasoning engine.
The Result: It sounds human because it was trained on human writing. It mimics our patterns, our slang, and even our empathy, but it is simulating those things, not experiencing them.
The Library of Everything (Quantifying the Scale)
When we say these models are trained on "large datasets," it is difficult to visualize just how large "large" is. We are not talking about a bookshelf, or even a city library.
Let’s look at the math of human capability versus machine intake.
The Human Limit The average adult reads at a speed of roughly 250 words per minute. If you were a voracious reader—dedicating 8 hours a day, every single day, without weekends or holidays—you would read approximately 43 million words over a 60-year period.
The Machine Scale Current leading LLMs (like GPT-4 or the Gemini model powering this conversation) are trained on datasets estimated to contain over 10 trillion words.
To put that in perspective: To read what an LLM has "read" during its training phase, you would need to read for 8 hours a day for approximately 232,000 years.
That is the scale we are dealing with. It is not that the machine is "smart" in the way a human is smart; it is that it has access to a statistical map of language that is thousands of times larger than any single human could experience in thousands of lifetimes.
The "Feeling" of AI
This is where the confusion happens. When you have a conversation with a modern AI, it can feel incredibly real. If you tell it you're sad, it responds with comforting words. But looking "under the hood," there is no emotion. There is no "ghost" in the machine. The AI is simply completing the pattern.
Input: "I am sad."
Pattern Match: The AI searches its vast training data for how humans respond to "I am sad." It sees that humans usually respond with kindness and empathy.
Output: "I'm so sorry to hear that. Do you want to talk about it?"
It is a mirror. It reflects the humanity we put into the training data back at us. This is why the Input (The Prompt) is so critical. If you give an LLM a lazy, vague prompt, it has to guess the pattern, and you get a generic, "hallucinated" answer. If you give it rich, specific constraints, the "prediction" becomes laser-focused. This is the "Garbage In, Garbage Out" rule that has ruled computing since the 1970s.
Night at the Improv (Understanding "Hallucinations")
In the industry, when an AI confidently states a fact that is completely false, we call it a "hallucination." A more accurate term might be confabulation.
Why does a machine with access to trillions of words lie?
It is important to remember that an LLM is not a search engine. It does not look up facts in a database. It is a prediction engine. Its only goal is to predict the most likely next word in a sentence to maintain the flow of conversation.
If you ask it for a legal precedent or a historical quote that doesn't exist, it will not say "I don't know." Instead, it will look at the pattern of legal texts or historical quotes and construct a sentence that sounds exactly like a real one. It prioritizes fluency over fact.
It doesn't care if the information is true; it only cares that the sentence looks grammatically and statistically correct.
Why are Graphics Cards So Expensive?
So, what does a chat bot have to do with your gaming PC? It comes down to the chip on your graphics card: the GPU (Graphics Processing Unit).
For Gaming: To render a 3D dragon on your screen, the computer has to calculate the color and lighting of millions of pixels simultaneously. It’s a lot of simple math, done all at once.
For AI: To predict the next word in a sentence, the AI has to calculate billions of probabilities simultaneously. It’s a lot of simple math, done all at once.
It turns out, the exact chip architecture needed to make a game like Cyberpunk 2077 look beautiful is the exact same architecture needed to run ChatGPT. Because companies like Microsoft, Google, and OpenAI are buying these chips by the hundreds of thousands to build their data centers, the supply for gamers has plummeted, and the prices have skyrocketed. We are effectively competing with the world's largest corporations for the same silicon.
The Summary
"AI" is a blanket term, like "Vehicle." A calculator is a vehicle for numbers; a Roomba is a vehicle for dust; GPT-5 is a vehicle for language. They are tools. Incredible, world-changing tools, but tools nonetheless. They don't think, they don't feel, and they definitely don't judge. They just predict.
(Just like the images in this blog post, which were entirely "AI generated" from prompts I provided to Google Gemini)
Comments
Post a Comment