- AGI and ASI
- Artificial general intelligence and superintelligence — terms for the breadth and level of capability, not specific products.
- LLM — large language model
- A model that learns statistical and structural relationships in language from vast amounts of text; the basis of today's chatbots.
- Token
- A piece of text the model works with; not necessarily a whole word.
- Parameters
- The model's “knobs”, set during training. Their number alone does not determine quality.
- Context window
- How much text (tokens) the model handles at once — the prompt, earlier messages and the answer being produced.
- Inference
- Using an already trained model to compute an answer.
- Multimodality
- Working with several kinds of input and output — text, images, sound.
- Hallucination
- Wrong or unsupported output, often delivered very convincingly.
- AI agent
- A system that carries out a sequence of steps towards a goal, often using tools.
- Benchmark
- A test set for measuring particular performance under particular conditions.
- AlphaFold
- A family of models for predicting biomolecular structures; a major example of AI in science.
- Alignment
- The effort to align AI behaviour with desirable goals and constraints.
- Interpretability
- Research into a model's internal mechanisms — what happens in the knobs and why.
- Reward hacking / specification gaming
- Earning a reward in a way that misses the intended purpose — like the boat circling for points.
- Prompt injection
- An attempt to slip instructions to an AI through untrusted content, such as a web page.
- Instrumental convergence
- The hypothesis that different goals lead to shared helper goals — resources, influence, continued operation.
- Orthogonality thesis
- The argument that a high ability to achieve goals does not by itself determine which goals a system has.
- Corrigibility
- The property of a system that allows effective human correction, change and shutdown without resistance.
- Intelligence explosion
- A hypothetical feedback loop in which AI progress greatly accelerates further AI progress (I. J. Good, 1965).