Building at the scale of a star takes more planning than people can do alone. This is a plain-language guide to artificial intelligence: what it is, how it learns, and where to learn more.
Each guide gives you the core idea in a minute, then something to try.
A neural network is layers of very simple units. Each one adds up its inputs, weighs them, and passes a number on. Nothing in it is programmed with rules. What it knows is stored in the weights.
Try it: sort ten photos into two piles and notice you cannot fully explain your own rule.
Training means showing the network an example, measuring how wrong its answer was, and nudging every weight slightly to be less wrong. Repeat that billions of times and skill appears.
Try it: guess a number, get told higher or lower, and count how fast you close in.
A language model does one thing: predict what comes next in a piece of text. Trained on enough writing, that one skill turns out to cover answering questions, translating and writing code.
Try it: cover the end of a sentence and predict the last word before you look.
The same method works on pictures and sound. An image is cut into small patches, sound into short slices, and the network learns patterns across them the way it does across words.
Try it: zoom into a photo until it is only coloured squares, then zoom back out.
An agent is a model given a goal and some tools. It works in a loop: decide a step, take it, look at the result, decide the next. That loop is what lets AI do tasks, not only answer.
Try it: write out every small step of making tea, including checking each one worked.
Large models are trained on thousands of specialised chips running for weeks, and the electricity that takes is now a real limit. Intelligence at scale is, in the end, an energy problem.
Try it: find the wattage on your laptop charger and compare it with a kettle.
One pick each, with the reason and the catch. These are ordinary links. We do not earn a commission from them. If that changes, we will say so next to the link and on our affiliate disclosure page.
Why this one: a working scientist explains what these systems can and cannot do, without hype.
The catch: it was published in 2019, before today's chatbots, so the examples are older.
Find it at Open LibraryWhy this one: the clearest account of why it is hard to make a machine do what we actually meant.
The catch: it is long and detailed. Expect a slow, rewarding read.
Find it at Open LibraryWhy this one: a physicist thinks through what intelligence could become over cosmic time, which is where this site is headed.
The catch: much of it is speculation, clearly labelled as such.
Find it at Open LibraryWe earn nothing from these. They are simply the best places to go deeper, including open papers and the agencies that measure the real world.
Plain-language pages that go further than the sketches on this home page.
A language model predicts what comes next in a piece of text. That one skill, trained on a lot of writing, covers many tasks.
Training is how a model gets its weights. Inference is how it uses them. The two have different costs, and both spend energy.
An agent is a model given a goal and some tools, then run in a loop: act, look at the result, act again. Here is what that loop is, and where it fails.
A neural network is layers of simple units that learn from examples. Here is what that means, and what it does not mean, in plain language.
Today's language models and agents are useful, and they have hard limits. Here is a clear list of things they do not reliably do.
A Type II civilization uses the full energy of its star. Getting there takes five things. This site covers one of them.
These are ordinary links, not a storefront. We do not earn a commission from them. If that changes, we will say so next to the link and on our affiliate disclosure page. These are shops, not reviews, and not a claim that we bought or tested the goods.