Today, July 16, is National AI Day. One of the truths of working in this field is that the vocabulary never sits still. You pick up a new term more or less every day, and the moment you feel fluent someone drops one you have to go and look up. So to mark the occasion we lined up a handful of AI terms in the order most people tend to meet them, starting with the words that have gone fully mainstream and ending with the ones that still make researchers lean in.

The ones everyone knows now

LLM. Large language model. The thing most people now mean when they say “AI”, trained to predict the next chunk of text across an enormous amount of writing.

Hallucination. When a model states something false with total confidence. Confidence is the dangerous part, because it looks identical to being right.

Prompt. The instruction you give a model, an entire craft has grown up around wording it well.

The ones you pick up quickly

Token. The unit model actually reads and writes, closer to a chunk of characters than a whole word. Models think in tokens, not language.

Context window. How much text a model can hold in mind at once. Anything outside it may as well not exist.

RAG. Retrieval-augmented generation. Give the model a search step so it answers from real documents rather than memory, which is one way to keep it honest.

The ones that mark you as a regular

Fine-tuning. Taking a general model and training it further on your own data to specialise it for a task.

Chain-of-thought. Coaxing a model to show its working step by step. It often improves the answer, though the steps it shows are not always the steps it actually took.

Mixture-of-experts. A model built from many specialised sub-networks that routes each input to only the few it needs, so it can be enormous while staying affordable to run.

Distillation. Training a small model to imitate a much larger one, keeping most of the ability at a fraction of the cost.

The ones that make researchers lean in

Grokking. The strange moment when a model suddenly generalises long after it looked like it had simply memorised the answer. Nobody fully agrees on why it happens.

Speculative decoding. A small model drafts the next few tokens and a large model checks them, which turns out to be a neat trick for going faster without losing quality.

Attention sinks. A few tokens, often the very first ones, that a model quietly parks its attention on. Remove them and things break in ways the field took a while to understand. This one is a favourite of ours.

Superposition. The finding that networks pack in more concepts than they have neurons by letting those concepts overlap, which is a large part of why they are so hard to read from the inside.

And the one underneath all of it

Go far enough down the list and the terms start circling the same question: can you actually tell what the system is doing, and why. That question is the whole reason we build the way we do. Neurosymbolic AI, the last term for today, combines the pattern recognition of neural networks with the precision of symbolic logic, so a system can read the messy real world and reason about it in a way you can inspect, trace and check.

On a day for appreciating how far AI has come, that feels like the right note to end on. The vocabulary will keep growing. What matters is whether we can still understand what the words describe.

Want to learn more about the technical terms and the latest research in AI? Check UnlikelyAI Lab blogs.

Follow us on LinkedIn to hear more from the team.