Chapter 1: What were we before AI?

This question must be answered before thinking about AI. What were we as developers? What was our purpose? What differentiated us from typists or coders? Were we paid for typing with ten fingers, or for something else?

Well, the most obvious answer (which applies to many, if not all, other professions) is that we are problem solvers. We are not just code editors with flesh. But if we are problem solvers, then what is AI?

The difference between human and AI

Let me ask you a question: How is it that an AI chatbot that knows so much more than each individual one of us still has some shortcomings? How is it that the same AI that might give me a very sophisticated and complex answer in any field I name still hallucinates, generates six fingers per hand, or straight-up lies? Humans are not like that. A professional artist doesn't just draw seven fingers per hand and say, "Oopsie, I made a mistake." It's either intentional or something's off.

The question is how AI and humans differ in the way they think, at least in the models we have today.

Let's take a human child, for example. A child learns very fast. It learns so much and does extraordinary things that no other animal can do. But it still can't remember anything. Most people don't remember anything at all from infancy up to age four or five. The opposite is true for LLMs. You can give an LLM a completely random piece of text that no human can memorize, and it will recite it back to you instantly. This is due to the memory-oriented nature of LLMs.

Memory-oriented beings

LLMs fundamentally work on a "big bucket of data." You might say, "Well, Hamid, humans work like that too. That's why children learn so fast: they digest so much data every day." While that is true, it does not necessarily mean it's the same. And in fact, this example is a perfect way to show why. A kid learns fast but forgets fast too. So what persists? What is stored that helps that kid grow instead of reverting back to being zero years old? If an LLM loses all of its parameters, it'll turn into the dumbest thing you know. But a human will forget five years of their life, and yet those five years could affect their whole life; traumas do that, for example.

This is due to the fact that a human brain is the most powerful pattern-recognition machine which has a very dedicated context-management system. Everything that you learn at those ages turns into patterns, logic, and ways of thinking. You don't memorize the exact same things. You only memorize how you felt about them or how you should treat them. This doesn't stop there; you carry it with you to your grave. When you're 10 years old, you might touch a hot pot and realize that you shouldn't touch things on the stove. You might never remember touching that pot, but you always know you shouldn't touch it. And since that's a pattern you saved, not just an example, you soon find different things that look hot and threatening.

Your brain is a massive library of patterns. But you do need to memorize more things as you age. You must memorize your native language, the people around you, your friends, your school, the laws of society, etc. You also try to turn everything into patterns; this helps you avoid potential harm instead of relying on explicit experiences. For example, if it's illegal to hit someone with a machete, it's also illegal to do it with a small kitchen knife.

There's a concept in programming called a harness. It got popular with the rise of AI agents. Fable 5, GPT Sol, and other models are just models, LLMs. What you run them with is the harness: Claude Code, Codex, Cursor, etc. It is not just the tool but also how you use the models. The way you optimize your use of these agents and models is part of your harness too. It has already been widely accepted that a harness is probably the most important factor in AI output quality.

Your brain is the most powerful harness. If you could somehow transfer the knowledge in current AI models, even GPT 3.5, into your brain, you would become the Mega Mind of our time, rather than just a new frontier model. Your dataset is very limited, yet you're able to do things no current LLM can do.

Current AI, on the other hand, works based on a prediction system. While it is true that right now it's not only a "next-word guesser" (though it roughly is), it's also true that it works based on "what's the most common chain of thoughts and answers that comes after this input." When you insert your input, it is combined with the billions of parameters in the LLM, and a pattern is formed based on how your input directs those "memories." So if you ask about house plants, for example, the AI sees the pattern and says, "OK, this part of the data from this section of the parameters matches perfectly with the input. What's the most probable thing to come after this question? ... This answer..." This is also why you receive the text word by word, or, more accurately, token by token. (You might want to give this paragraph to your favorite chatbot and ask it to explain it to you if you want a better understanding. I'm oversimplifying things so I can extract the core logic needed to do our philosophy thingy.)

To make it short: Humans have a much better prediction system but work with a much smaller dataset and have far fewer examples to give you. This also makes sense when you meet senior developers. They're mostly not able to recite every little step you have to take to build a large piece of software. They just know the tools, the stack you should choose, and the foundation you have to lay for your project. They use their massive set of patterns along the way to guide themselves and their teams toward the goal. They don't possess superhuman abilities.

This also explains why, when you ask AI to simply be creative, it doesn't work. Creative things tend to be less common. There is less information about them, so it's hard to verify their creativity objectively. Humans are opinionated individuals. A person might call something creative, while another person might not agree, or some people might not have thought about it at all. This is what subjective means. But AI's purpose is to work objectively. There isn't a perfect written manual for achieving creativity (or at least we've collectively rejected anything published about it). But the same AI model doesn't generate a very different answer every time you ask the same question, even though each chat session is different. I read the other day that if you ask an LLM to generate a list of 100 random words, AI detectors will say they're 100% sure that the list is AI-generated. It's not a coincidence; it's what LLMs are: memory-oriented (I wrote this part of the book before the whole "watermark" thing).