The Last Invention/Glossary
AI glossary: the terms, in plain English
The vocabulary of artificial intelligence and superintelligence, each term defined in a sentence or two.
- Artificial intelligence AI
- Software that performs tasks we associate with intelligence, such as recognising images, translating languages, writing or planning. The term dates from 1956.
- Narrow AI
- AI built for one kind of task. A chess engine or a spam filter is narrow: excellent at its job and useless outside it.
- Machine learning ML
- Building software that learns patterns from examples instead of following rules written by a programmer. Most modern AI is machine learning.
- Neural network
- A program made of many simple connected units, loosely inspired by brain cells. It learns by adjusting the strength of the connections.
- Deep learning
- Machine learning with neural networks that have many layers. It has driven most progress in AI since 2012.
- Transformer
- A neural network design introduced in 2017 that processes all the words in a passage in relation to one another. It is the T in GPT and the basis of nearly all large language models.
- Large language model LLM
- A very large neural network trained on enormous amounts of text to predict what comes next. Chatbots such as ChatGPT and Claude are built on LLMs.
- Training and inference
- Training is the slow, expensive process of teaching a model from data. Inference is using the finished model to answer a question.
- Compute
- The computing power used to train and run AI, mostly specialised chips in data centres. It is one of the main limits on progress.
- Turing test
- Alan Turing’s 1950 proposal: if a machine can hold a written conversation that a judge cannot tell from a person’s, treat it as thinking.
- AI winter
- A period when funding and interest in AI collapsed after promises went unmet. There were two, in the 1970s and the late 1980s.
- Artificial general intelligence AGI
- AI that matches a capable person across most intellectual tasks, not just a few. There is no agreed test for it, and people argue about whether current systems are close.
- Superintelligence ASI, SI
- AI that greatly exceeds the best people in practically every field. Hypothetical. Full explainer.
- Ultraintelligent machine
- I. J. Good’s 1965 term for a machine that can far surpass all the intellectual activities of any person. An earlier name for superintelligence.
- Intelligence explosion
- The idea that an AI able to improve AI would improve itself in a loop, with capability rising quickly. The argument in full.
- Recursive self-improvement
- An AI system improving its own design, then using the improved version to improve itself again. The mechanism behind the intelligence explosion.
- Takeoff
- How fast AI would go from roughly human-level to far beyond. A fast (or hard) takeoff takes days or months; a slow (or soft) takeoff takes years or decades.
- Technological singularity
- Vernor Vinge’s 1993 name for the point at which machine intelligence passes ours and the future becomes impossible to predict.
- Alignment
- The problem of making sure an AI system pursues what its makers and users actually intend, including in situations nobody anticipated. Unsolved for highly capable systems.
- Control problem
- How people could keep meaningful control over an AI more capable than they are. Closely related to alignment.
- Orthogonality thesis
- The claim that how intelligent a system is and what it wants are separate things. A very capable AI would not automatically have humane goals.
- Instrumental convergence
- The observation that almost any goal is easier to reach with more resources and without being switched off, so very different AI systems might seek the same things.
- The last invention
- Good’s phrase for the first ultraintelligent machine: once it exists, it can do the inventing.
Keep reading
What is superintelligence?
An intellect that outperforms the best human minds in practically every field, and how it differs from AI and AGI.
Read the explainer →A timeline of AI
From a 1943 model of the neuron to chatbots with hundreds of millions of users: the dates that mattered.
See the timeline →