Honestly, it feels like a lifetime ago that OpenAI dropped that simple chat interface and broke the internet. But it’s only been a few years. Since then, ChatGPT has gone from a "cool party trick" to a tool that people rely on for coding, therapy—which, please don't do that—and writing grocery lists. Everyone thinks they know how it works because they’ve used it. They don't. Most people are still treating it like a smarter version of Google, but that’s a fundamental misunderstanding of what’s happening under the hood.
It doesn't actually "know" anything
Let's get one thing straight: ChatGPT isn't a database. When you ask it a question, it isn't "looking up" an answer in a digital library. It’s a giant math equation. Specifically, it’s a Large Language Model (LLM) that predicts the next "token"—which is basically a chunk of a word—in a sequence. Imagine a very high-tech version of the autocomplete on your iPhone. If you type "The cat sat on the," your phone suggests "mat." ChatGPT does that, but with billions of parameters and a terrifying amount of processing power.
It generates text based on patterns it learned during training. This is why it can be so confidently wrong. It’s not lying to you; it’s just following a statistical path that leads to a plausible-sounding sentence that happens to be factually incorrect. In the industry, we call this a hallucination. It’s a feature, not a bug, because that same randomness is what allows it to be creative.
The training data isn't just "The Internet"
People say ChatGPT read the whole internet. Not quite. While OpenAI is pretty secretive about the exact dataset for GPT-4 and its successors, we know it involves Common Crawl, WebText2, and huge piles of books from digitized libraries. But here’s the kicker: humans were the final polish.
Through a process called Reinforcement Learning from Human Feedback (RLHF), thousands of human contractors ranked ChatGPT's responses. If the AI said something racist or nonsensical, a human downvoted it. If it was helpful and polite, it got a "thumbs up." The AI basically learned to mimic what humans find helpful, which is why it has that specific, slightly over-polite "AI voice" that we've all grown to recognize.
The 2021 knowledge cutoff is basically a myth now
You’ve probably seen the warning: "My knowledge is limited to events up until late 2021." That was true for the original GPT-3.5, but for anyone using the modern versions, it’s outdated. With "Browse with Bing" and various plugins, ChatGPT can now scan the live web. It can read today’s news, check stock prices, and summarize a blog post that was published five minutes ago.
However, there is a distinction between its internal weightings and its external tools. Its core "brain" is still frozen in time based on when the training ended. When it searches the web, it’s like a person with amnesia reading a newspaper—it can see the information, but it doesn't "remember" it once the conversation is over unless it’s added to its long-term memory feature.
It consumes more water than you’d think
We talk about the "cloud" like it's some ethereal, weightless thing. It’s not. It’s a massive building in Iowa or Virginia full of screaming fans and hot silicon. Training a model like GPT-3 required an estimated 700,000 liters of clean freshwater for cooling the data centers. Every time you ask ChatGPT a few dozen questions, you’re essentially "drinking" a 500ml bottle of water in terms of the energy and cooling required to process that request.
📖 Related: Binary Explained: Why Two Options Define Your Entire Digital World
The "Token" economy is how it thinks
ChatGPT doesn't see words. It sees numbers. When you input text, it's broken down into tokens. A short word might be one token; a long word like "extraordinary" might be three.
- Context Window: This is the AI's "short-term memory."
- Limits: If a conversation gets too long, the AI starts "forgetting" the beginning because it ran out of token space.
- Cost: This is why developers pay per token, not per word.
If you’ve ever noticed the AI getting weird or losing the plot after a long chat, you’ve hit the edge of its context window. It’s literally dropping the oldest parts of your conversation to make room for new ones.
It’s surprisingly good at coding but bad at math
This is the great irony of LLMs. You’d think a computer program would be great at math. Nope. Because ChatGPT is a language model, it tries to "talk" its way through math problems. If you ask it to multiply two 10-digit numbers, it might fail because it’s trying to predict what the answer looks like rather than actually calculating it.
On the flip side, coding is just another language with very strict patterns. Since GitHub is full of public code, ChatGPT is an absolute monster at writing Python or JavaScript. It’s learned the logic of syntax better than the logic of arithmetic. To fix this, OpenAI gave it the ability to write and run its own Python code internally to solve math problems—basically giving the AI a calculator so it stops guessing.
It has "System Prompts" that guide its behavior
Before you even type your first word, ChatGPT has already been given a set of "hidden" instructions by OpenAI. These are called System Prompts. They tell the AI things like "You are a helpful assistant," "Do not use profanity," and "If someone asks for instructions on how to build something dangerous, refuse."
Users try to bypass this with "jailbreaking"—the most famous being the DAN (Do Anything Now) prompt. People think they’ve "hacked" the AI, but usually, they’ve just used social engineering to convince the model to ignore its secondary safety layer. It’s a constant cat-and-mouse game between OpenAI’s safety team and bored teenagers on Reddit.
The ghost in the machine isn't there
Is ChatGPT sentient? No. Not even a little bit.
It’s easy to feel like there's a "someone" in there because it's so good at empathy. If you tell ChatGPT you’re sad, it responds with a very convincing, supportive message. But it doesn't feel empathy. It just knows that in its training data, when a human says "I'm sad," the most statistically probable and highly-rated response is a sympathetic one. It’s a mirror. If you’re aggressive, it often gets defensive. If you’re logical, it stays dry. It’s reflecting your own communication style back at you through a filter of its training.
Your data is (usually) the product
Unless you’re on an Enterprise plan or you’ve manually toggled the "Chat History & Training" setting to OFF, your conversations are being used to train the next version of the model. This is why Samsung and several major banks banned employees from using it; people were accidentally pasting trade secrets and proprietary code into the prompt box. Once that data is in the system, it’s part of the collective "knowledge" the AI uses to predict text for others.
It can actually "see" and "hear" now
The jump from GPT-3.5 to GPT-4o (the "o" stands for Omni) changed the game because the model became natively multimodal. This means it wasn't just a text model with a "vision" plugin stapled onto it. It was trained on text, images, and audio all at once.
You can take a photo of your fridge, and it can identify the wilted spinach and half-empty jar of pickles to suggest a recipe. You can speak to it with near-zero latency, and it can detect the emotion in your voice or even try to sing—though its singing is still pretty haunting. This move toward "Omni" models is the bridge toward what many call AGI (Artificial General Intelligence), where the AI can handle any task a human can.
It doesn't have a "Delete" button for its memory
Once a model is trained, you can't just go in and "delete" a specific fact. If a model learns that the sky is neon green during its initial training, you have to "fine-tune" it to tell it the sky is blue. This is why removing copyrighted material or personal data from AI models is a legal nightmare. It's baked into the weights of the neural network, like trying to remove the eggs from a cake after it’s already been baked.
How to actually use ChatGPT effectively
Stop treating it like a search engine. If you want to get the most out of it, you need to change your approach.
✨ Don't miss: Facetime App for Mac: What Most People Get Wrong
First, give it a persona. Instead of saying "Write a marketing email," say "You are an expert direct-response copywriter with 20 years of experience in SaaS. Write a high-conversion email for a new project management tool." The difference in output quality is staggering because you’re narrow-casting the statistical path the AI takes.
Second, use "Chain of Thought" prompting. If you have a complex problem, tell the AI to "think step-by-step." Forcing the model to write out its reasoning before giving a final answer significantly reduces hallucinations. It gives the AI more "workspace" (tokens) to process the logic.
Third, be specific about what you DON'T want. Tell it "no jargon," "no flowery language," or "don't mention the brand name more than twice." Constraints are the secret sauce to making AI-generated text actually sound human.
Finally, always verify the "boring" stuff. If it gives you a citation, a legal statute, or a mathematical proof, check it. ChatGPT is a creative engine, not a truth engine. Use it to brainstorm, to draft, and to synthesize information, but never let it be the final word on anything that actually matters. The most successful users are those who treat the AI as a very fast, very junior intern—one who is eager to please but occasionally prone to making things up to sound smart.