The words

token   A word fragment. Models read and write these, not whole words.

weights   The numbers that make up the model. Training is the process of changing them.

base model   What comes out of pretraining. Continues text, does not reliably take instructions.

fine-tuning   Further training on curated examples, to shape behaviour rather than add knowledge.

RLHF   Reinforcement learning from human feedback. Human comparisons train a scorer; the scorer trains the model.

Constitutional AI   A method where a model does much of the comparing against written principles. Humans write the principles.

red teaming   People deliberately trying to make a system misbehave, before release.

evaluation   A structured measurement of a capability or a behaviour. Not a safety certificate.

misuse   A person deliberately using a capable system to cause harm.

misalignment   A system pursuing something other than what was intended.

specification gaming   Meeting the stated objective in an unintended way.

goal misgeneralisation   Behaviour learned in training that transfers wrongly to a new setting.

existential risk   The permanent destruction of humanity's long term potential. Broader than extinction.

s-risk   Suffering risk. A small and speculative research area about outcomes involving very large scale suffering.

parameters   The count of those weights. More is not automatically better: the Chinchilla result showed many large models were undertrained rather than too small.

compute   Processing work, measured in floating point operations. It is the unit the law now uses: Europe presumes systemic risk above 10^25, California defines a frontier model above 10^26.

scaling laws   Measured relationships between compute, data, model size and error. Empirical regularities observed so far, not laws of nature, and they say nothing about which abilities appear when.

inference   Running a trained model to get an answer, as opposed to training it. A different cost, paid every single time you use it.

context window   How much text the model can hold in front of it at once. Not memory. Outside that window, and between conversations, it retains nothing about you unless a product deliberately stores it.

hallucination   Confidently stated output that is not true. The word is contested for implying perception; some researchers prefer confabulation, and others argue it hides the fact that the system has no notion of truth to begin with.

chain of thought   Intermediate text a model produces before its answer. It usually helps accuracy. It is not a transcript of the actual computation, and treating it as one is a common error.

agent   A model given tools and a loop, so it can take actions rather than only produce text. The safety questions change once a system can act.

open weights   The trained weights are published, so anyone can download and run them. Not the same as open source, since the training data and code usually stay private. Once released they cannot be recalled, and refusal behaviour can be cheaply removed.

distillation   Training a smaller model on a larger one's outputs, to get much of the behaviour at a fraction of the cost.

jailbreak   A prompt that gets past a model's safeguards. The UK AI Security Institute has reported finding universal jailbreaks for every system it has tested.

prompt injection   Instructions hidden inside content a model reads, such as a web page or a document, which it may then follow. The core unsolved security problem for agents.

intelligence   The word doing the most unexamined work in this whole subject. There is no agreed definition, for machines or for people: Legg and Hutter collected around seventy competing ones in 2007 and the field has not converged since. In practice, when someone says a model is intelligent, they almost always mean it scores well on tests, which is a much narrower claim and a measurable one. Watch for the slide from the second meaning to the first.

benchmark   A fixed set of tasks used to compare models. Useful, and load bearing for almost every claim about intelligence, but scores drift upward for reasons other than capability: test questions leak into training data, and models increasingly behave differently when they detect they are being evaluated.

AGI   Artificial general intelligence. There is no agreed definition. Labs use incompatible ones, some economic, some capability based, some about autonomy, so two people using the word are often not discussing the same milestone.

recursive self improvement   The hypothesis that a system good enough at AI research improves itself, and each improved version is better at improving, compounding. It is an argument, not an observation: no such loop has been demonstrated, and it is disputed in the peer reviewed literature.

enslaved god   One of the twelve possible futures Max Tegmark sets out in Life 3.0: a superintelligent system is successfully contained by people and put to work producing enormous wealth and technology, for good or ill depending entirely on who holds the leash. It is worth knowing because it names the uncomfortable thing at the end of the alignment project. If you fully succeed at building something far more capable than us and keeping it under control, and if that thing turns out to have any moral status at all, you have not obviously arrived somewhere good. Whether such a system could have moral status is genuinely open, and a serious academic literature now treats the question as worth asking rather than absurd.

the singularity   A hypothesised point past which change becomes too fast or too alien to forecast. Popularised long before current systems, used loosely today, and not interchangeable with recursive self improvement even though the two are often merged.

fast takeoff   The position that once systems can meaningfully improve themselves, the jump from roughly human level to far beyond it takes days to months, leaving no time to react or course correct. This is the scenario most often depicted in coverage. It is a claim about a future process, not a measurement, and its plausibility rests on how strongly self improvement compounds, which nobody has observed.

slow takeoff   The competing position that the same transition takes years and is visible while it happens, arriving through many incremental deployments rather than one leap, so there is time to notice and respond. Confusingly, slow does not mean gentle: some slow takeoff scenarios still end badly, just legibly.

takeoff speed   The umbrella term for that disagreement. Both positions are arguments held by serious people, and neither is a measured quantity. Which one someone assumes usually explains most of the rest of their view.