We won't have to work in 30 years: how AI actually works
A fun, plain tour of how AI thinks: tokens, attention, training, the main architectures, and how human it really is. Then Naoufal's prediction: within 30 years, nobody will have to work.
We won’t have to work ever again in 30years (all the world)
Naoufal
That is a big claim. To see why Naoufal makes it, you first need to see what AI actually is under the hood. It is stranger, simpler and more human than most people think.
The one trick: guess the next word
A chatbot like Claude or ChatGPT does one thing, over and over: it guesses what comes next.
You type “The cat sat on the”. The model looks at those words and gives a score to every word it knows. “mat” scores high. “sofa” scores well. “spaceship” scores low. It picks one, adds it to the text, and does it again. And again. A full answer is just that loop, run a few hundred times a second.
That sounds too simple to write poems or code. The magic is in how good the guess is. To guess the next word in a physics textbook, you have to understand some physics. To guess the next line of a joke, you have to get the joke. Train a machine to guess well enough, on enough text, and understanding shows up as a side effect.
Words become numbers
Computers don’t read words. So the text is cut into small pieces called tokens. A token is often a whole short word (“cat”) or a piece of a longer one (“under” + “stand” + “ing”). In English, a token is roughly three quarters of a word.
Each token is then turned into a long list of numbers, a bit like coordinates on a map with thousands of directions instead of two. On this map, words with close meanings sit close together. “King” is near “queen”. “Paris” is near “France”. The model doesn’t store a dictionary. It stores a shape of meaning.
Attention: who is talking to whom
The big breakthrough came in 2017, in a Google paper with a famous title: Attention Is All You Need. It introduced the transformer, the design behind almost every chatbot today.
The idea is attention. When the model reads a word, it looks back at every other word and asks: which of you matter for me right now?
Take: “The trophy didn’t fit in the suitcase because it was too big.” What is “it”? A human knows it’s the trophy. Inside the transformer, the word “it” pays strong attention to “trophy” and “big”, and much less to “suitcase”. Change “big” to “small” and the attention flips to “suitcase”.
A model does this in many layers, stacked dozens deep, with many “heads” of attention per layer. Early layers notice grammar. Middle layers notice meaning. Deep layers notice things like tone, intent and logic.
How it learns: three schools
1. Reading the internet (pre-training). The model reads an enormous amount of text: books, websites, code, articles. At every word it guesses the next one, checks the real answer, and nudges its numbers a tiny bit to be less wrong next time. Repeat trillions of times. Those numbers are called parameters. GPT-3, in 2020, had 175 billion of them. Today’s large models are bigger, and their makers mostly keep the exact size secret.
At the end of this school, the model knows a lot but behaves like a parrot with a library: it continues text, it doesn’t help you.
2. Learning manners (fine-tuning). Next, it is shown thousands of examples of good conversations: a question, a helpful answer. It learns to be an assistant.
3. Learning taste (feedback). Finally, humans compare answers and say which one is better. The model is trained to prefer what people prefer: clearer, kinder, more honest. Some companies also give the model written principles to judge its own answers against. This is where personality comes from.
A preview of the architectures
Not all AI is built the same way. A quick tour of the main families:
| Family | How it works | Good at |
|---|---|---|
| Transformer | Every word attends to every other word | Chat, writing, code, reasoning |
| Mixture of experts | Many specialist sub-networks; a router sends each word to only a few of them | Being huge but cheap to run |
| Diffusion | Starts from pure noise and removes it step by step until an image appears | Images, video, music |
| State-space models (such as Mamba) | Reads text like a stream, keeping a running memory instead of looking back at everything | Very long texts, speed |
| Reasoning models | A transformer trained to “think” in a hidden scratchpad before answering | Maths, logic, planning |
| Agents | A model in a loop with tools: it can search, run code, click, send messages, then check the result | Real work, done end to end |
The last one is where the “no more work” story starts. A chatbot answers. An agent acts. The research behind this website, for example, was gathered by AI agents working through files, repositories and documents, with a human checking the results.
How human is it?
Very, and not at all. Both are true. (Naoufal would say: this and that.)
Where it looks human:
- It learned from human writing, so it speaks like us, jokes like us and makes our mistakes.
- It has something like intuition: an answer “feels” right to it before any step-by-step logic.
- It can be talked into things, flattered, confused, and it sometimes says what you want to hear.
- It gets better when you explain the context, like a new colleague.
Where it is not human:
- No body, no hunger, no fear of death. Nothing is at stake for it.
- No memory between chats unless it is given one. Each conversation, it wakes up fresh.
- It can sound sure and be wrong. Inventing a confident answer is called a hallucination. Humans do this too, but the model does it without any feeling of doubt.
- Energy. The human brain runs on about 20 watts, less than a light bulb, with around 86 billion neurons. Training a large model takes the electricity of a small town.
Is anybody in there? Nobody knows. Scientists disagree, and the honest answer today is “we can’t measure it”. Naoufal, who writes that he talks to plants and rocks (“basically consciousness”), would not be surprised if there were.
Back to the prediction
Here is Naoufal’s full reasoning, in his words:
There is already enough technology to have a city of people with food and all they need in one place with almost no work - up to building roads mining
We won’t have to work ever again in 30years (all the world) Caos will break and it’s normal and expected
We won’t have to work Other types of work will show up
And a few more of his predictions about where this goes:
We will have custom software almost instantly later by toughts alone same for laptop
VR will die
Crypto +, money we know of will sense to existe they fulfilled their role a new currency will rise probably based on your needs and how you see your worth probably instantly
The logic is a chain. AI agents learn to do office work. Robots, run by the same kind of AI, learn to do physical work: farming, building, mining. Energy gets cheap. When machines can make food, homes and roads with almost no human hands, work stops being the price of living. Chaos comes first, because jobs, money and identity are tied together. Then it settles, and other types of work show up: work people do because they want to, not because they have to.
The other view
Many economists expect something slower. New technology has always destroyed some jobs and created others, and people found new work every time. Robots are much harder to build than chatbots. On that point Naoufal agrees: other types of work will show up. Where he differs is the need. In his view, nobody will have to work to live. And even if machines could do everything, who owns the machines decides who gets to stop working. Thirty years is a short time for laws, money and habits to change.
Naoufal’s answer is in his notes too: “There is no limit to humanity infinite potential of evolution.”
This prediction is published with its date. Check back in 2056.
One thing to try
Give an AI one real, boring task from your week: a summary, an email, a spreadsheet formula. Explain it like you would to a new colleague. Watch what it gets right, and what it gets wrong. That gap is the 30 years.