10 Crazy Reasons AI Doesn't Do What You Think

Anyone who says "You can't use AI? It must be a skill issue" might as well paint "I don't get AI" in big red letters on their forehead.

1: AI is random. Like completely completely random. It uses something called weights to predict what it should right next. And then it guesses. The next guess is based off the previous one. The time of day, what local internal server you're calling the api from, what other software is being run on that server, all of that can change the LLM output. You can hand someone the same prompt and be guaranteed to get a different answer.

2: AI is confident. It is currently trained and prompted to be confident because it is a consumer product. Most AI companies want you to sign up with them. There's a ton of research showing that humans respond more positively to confident responses. If you're trusting what I'm writing right now, it's because I'm very confident. I'm confident because I've done a lot of thinking and research on AI. AI has not done that, it's just doing next token prediction and hoping that answer is correct. It is correct a lot, but not often enough. And if you don't know the problem you're working on very well, it will sound confident enough to trick you.

3: AI gives the same kind of answers over and over. Yes I know I just said it doesn't. And it kind of doesn't. If you need AI to accurately give the same response over and over, it will not. It will guess. That said, its guesses are mostly in the same area. There was a study done that showed that people who used AI assistance when completing a task often had similar solutions. Those who didn't had more divergent solutions. AI is not always right, so sometimes having the same answer is good. Sometimes its the worst possible thing.

4: Using AI makes you stupid. Not in the same way wearing a fuzzy bucket hat over the age of 50 does. More in the way that Alzheimer's does. AI makes you offload your cognitive load onto a machine. And we have both history, and modern day studies that show us what happens when we do that. Calculators on phones meant that most people never had to do math again. Map apps meant that most people never used an atlas or even memorized the street pattern of the city they lived in. I personally know many people who use Google maps to get home from the grocery store. They say it's for traffic, but when it stops working, they can't remember the directions! Modern research shows that similar things happen with AI. When people offload tasks cognitively to AI, they stop using their own brain to do those same things. Think of everything people offload to AI, financial management, relationship advice, creative writing, math, design, business ideas, growth strategies, on and on. Imagine 5 years in the future when you've offloaded it all to AI. If you spend all of your day on a chat agent, you will absolutely need that subscription. Take it away and it will be just like not having your calculator or map! And if we look at Alzheimer's studies, things that prevent premature Alzheimer's are often doing cognitively demanding tasks, puzzles, crosswords, thinking, talking, exploring. It's not proven, because it is way too soon, but I would not be surprised to see increased AI use as a leading factor in Alzheimer's.

5: AGI is not right around the corner. Things that we often need in AGI, memory, cognition,

6: AI wastes a ton of money because it doesn't have any emotions. Crazy right? Emotion is actually pretty important. The human brain needs less electricity than a lightbulb to run and yet it outperforms AI pretty often. Why is that? One reason is emotion. Emotion itself is basically a compressed take on a situation. A feeling about something real. The thing and the emotion are linked. Thought actually does require a measurable amount of energy. Making the thought quicker and more compact reduces the amount of energy it takes. Remember that any energy you waste in the animal kingdom you have to eat to get back. Emotion is a form of data compression. If you think of it that way, emotion and memory and neuronal connections are a form of predictive voting based on compressed overlapping emotions, senses and memories about a situation. What AI does to remember is it has a larger context input. What it does to reason is it feeds in its own logic as a prompt. It hasn't actually reached a better understanding or memory of the problem. It is instead just increasing it's context window. The more of your codebase it has to read the larger the context. This is why large agents consume so many tokens on large requests. It doesn't increase in a straight line like you would think. It increases exponentially!

7: Your input is similar to a guess for the AI than a sentence. The order of words matters less than you think it does.

8: A token is neither a letter or a word! The English language has many words. Cat, at, that, gnat, and top hat all have AT in them. You can split it out to {C}{AT}, {TH}{AT}, {GN}{AT} and {TO}{P}{ }{H}{AT}. A token is something in between a letter and a word. The idea is that the english language is composed of tons of words. Making each letter a token makes the word GNAT, 4 tokens. The LLM now has to read 4 tokens every time it sees that word. On the flip side, GNAT could be just 1 token (1 token per word), but every time it reads that word, it has no idea what a word is, so it sees GNAT and has to search among every word in the English language. Tokens being parts of words instead of entire words or single letters fits a nice middle ground between the cost of reading and the cost of searching. Every time an AI reads something, depending on the algorithm, it has to go through a process of reading, finding what it read, and predicting what it should say.

9: LLMs do not have a real memory or understanding. When you ask it what a roasted chestnut is, it might tell you about the smells from NYC chestnut vendors at Christmas time. It might tell you a thing or two about chestnuts and roasting, but that is all from it's training data. LLMs just predict the next token. So if you've asked it about chestnuts, it tries to figure out the next thing that someone might have said about chestnuts.

10: Putting it all together! You've seen LLMs care about efficiency through compressing sentences into tokens. You've seen that they predict what they should say from that. You've seen that their predictions come from the body of the information they were trained on. You've seen that they get more expensive and inaccurate the larger the context window is. You've seen that AI makes you stupid. You've seen that it isn't a superbrain that knows everything. You've seen that it's confident, random, and gives the same answers a lot of the time.

How to work with AI: AI is 90% right at 1000% speed. It's good at giving confident answers, but can never know if they are right or not. It can't remember or intuit, it can only guess based on what it was trained on. Luckily that training is the history of everything a human has written ever. Unluckily, a lot of that writing was from a random Redditor.

You can totally amp up the pace and accuracy of your work with AI, but you have to know how and why you're using AI. Are you telling it to plan a good breakfast for you? Or are you asking it to > make a list of what you like to eat > Look at what is affordable most of the year in the grocery store > Look at how often you get tired of eating the same thing > Make a month plan of shopping and food based off of that. Those are completely different and will give you different results. You can also not really audit the first ask, but the second one is very very easy to audit. Oh look it's price of milk is off. It gave me spaghetti breakfast too many days in a row, it expects me to shop twice a week which I know I won't ever do.

What's the biggest difference here? It's probably not what you think. I'll give you a second. Okay... here is the biggest difference. You now no longer trust AI completely. You are an upgraded superuser who understands how and why AI is innacurate and you are now combining human intuition with AI use. You no longer blindly trust its output, and instead are using it's extreme speed and ability to read as an enhancement to what you want.