Today’s stories are taking longer to arrive.
Browse earlier stories in Timeline →
A few ideas that make everything else click.
Software that learns patterns or uses other methods to do tasks we associate with intelligence—like recognizing a picture, understanding a question, or making a plan. It can be very good at a task without understanding the world the way a person does.
A system of connected calculations. An example goes in, an answer comes out, and during training the connections are adjusted to reduce mistakes. Think of a huge mixing desk: training keeps turning the knobs until the output gets better.
A software toolbox for building and training machine-learning models. Like a set of building blocks, it gives developers useful pieces so they do not have to make every part themselves.
A model design that uses attention: it weighs which parts of its input are relevant to each other. In “the dog chased the ball because it was bouncing,” those connections help work out what “it” refers to.
A model trained on lots of text to learn patterns in language. It generates text one small piece at a time. That can produce useful explanations and reasoning, but also convincing mistakes.
The calculation work computers can do. When an AI lab wants “more compute,” it usually means more powerful chips, more time using them, or both. Those chips also need memory, electricity, cooling, and fast connections.
A processor that can do many calculations at once. Originally built for graphics, GPUs are also useful for AI because training models involves enormous numbers of similar mathematical operations.
The learning stage. A model processes examples, measures its mistakes, and adjusts its internal numbers. This may take a huge amount of computing work before anyone starts chatting with the finished model.
Using a trained model to produce an answer, image, or other result. Training is like practice before a performance. Inference is the performance itself—and it still uses computing power.
An AI system that can take steps toward a goal, often using tools. Instead of only suggesting a trip, an agent might search flights and compare options. Its reliability depends on the model, tools, permissions, and checks around it.
The idea of an AI helping improve the next version of itself, which then helps improve the next one, and so on. An AI helping write code does not by itself prove that this loop will keep accelerating. Its speed and limits remain open questions.
A small piece of text that a model processes, sometimes a whole word and sometimes part of one. Models read and produce sequences of these pieces. Token counts help measure how much text a model handles.
Learning through feedback on actions. A system tries something, gets a reward signal, and adjusts to earn more reward. Getting the reward does not always mean doing what a person really intended.
The internal numbers a model learns during training. Sharing the weights lets others run or adapt the model, subject to its license. It does not necessarily share the training data or all the code used to build it.
An answer that sounds plausible but is wrong or made up. A model can produce fluent language without having checked whether its claims are true.
A way to generate images by learning to remove noise. Starting with random static, a model repeatedly makes small changes until a picture appears, guided by a description or other input.