If you are new to computers, you may have heard about this thing we call "AI" these days, specifically LLM-based AI chatbots such as ChatGPT and Google's Gemini. You may be wondering how do they work, how to use them, and if you can trust them. While the first two questions are bit complicated to answer, you absolutely can not trust them. Allow me to explain why.
The way LLM's work is fundamentally very simple. It literally just guesses the next word or phrase that comes after a piece of text. It's essentially an autocomplete, and it's actually a very good autocomplete in some cases, but just like any autocomplete, it's can get the completions completely wrong because it's just guessing what the next word is going to be.
While LLM's are genuinely impressive when they get things right, there is no guarantee that it will get things right. Most importantly, LLM's don't have any sort of reasoning or logic built into them. They're completely different from traditional algorithms. They don't take pieces of information and make a decision based on data available. Instead, they are just calculating through an absurdly complex mathematical operation what word is most likely going to come after another word in a text document.
When ChatGPT answers your question, ChatGPT has no idea what you just asked. It can not understand English. Its AI model was trained on a lot of English text, so when you ask ChatGPT "what is 2 + 2?" it's probably going to answer "4" because there is a lot of text on the Internet that says "2 + 2 equals 4." That's the only reason it knows what comes AFTER "2 + 2." Its ability to take the "4" and create a human-sounding sentence out of it to give the answer is rather impressive, but it's not actually understanding your question. It's just taking the keyword that it has to give you, "4," and transforming it in some way, e.g. it could say "it's four" or "the answer is 4" instead of just saying "4." That makes it look like it can have a conversation, but it's merely guessing what someone would answer you. Imagine, for example, a transcription of a conversation between two people, such as of a podcast. When you ask something, you are one person in the transcription. The chatbot simply guesses what the other person would tell you after that.
ChatGPT is not AGI
An AI that can truly understand what you tell it and reason about it is also known as an AGI (Artificial General Intelligence).
Experts in the field of AI have various opinions on whether LLM's can become AGI or not. Some believe it's already kind of AGI, some believe if you had infinite money and infinite GPU's you could make an LLM big enough that it would become AGI, and some believe it's simply fundamentally never going to work and a different approach is necessary.
Unfortunately, because some of these experts are working in companies that sell AI products, it's hard to tell if they're saying what they're saying because they believe their approach will actually work, or because they need investors to believe their approach will actually work. There's a lot of hype involved in AI.
However the future of AI looks like, what's important for us right now is that currently ChatGPT isn't AGI, nor is Gemini, nor is anything that we have in the market today. It's all just autocomplete.
Always Verify ChatGPT's Output
My recommendation is that if you are using ChatGPT, you must make sure that all output is verifiable in one way or another.
For example, if you want to learn about a subject, one trick is to make ChatGPT provide links to articles that talk about the subject rather than simply summarizing it. Just write in your prompt "give me links to articles that talk about X" and it will do that. This way you can use it more like a search engine than as a random text generator.