How models work

someone in the meeting says AI and you have to work out whether they mean a model call, a classifier or a slide.

Artificial intelligence (AI)

you stop writing the rules, hand over the objective and the data, and let the training process find the parameters.

Machine learning (ML)

when someone calls it deep learning they are telling you the network has many layers, not that anything mystical is happening.

Deep learning

a giant chain of multiply-and-add operations whose constants were tuned automatically.

Neural network

when you offload part of a model, or read its config, or watch it load, the unit being counted is one layer of dozens.

Layer

the number in the model's name counts these, training is what changed them, and inference only ever reads them.

Parameters / weights

the model is fluent in the languages and frameworks that were plentiful in its corpus and vague about the ones that were not.

Training data / corpus

every mainstream model you call is one of these, which is why they all share the same context-window and attention economics.

Transformer

a learned relevance score between every pair of tokens in the context.

Attention

everything it does, the chat and the code and the reasoning, comes out of one trained habit of guessing the next token.

Large language model (LLM)

extremely sophisticated autocomplete, run one token at a time.

Next-token prediction

the answer arrives one word at a time, and a reply twice as long takes twice as long to finish.

Autoregressive

'the model learns from our conversations' is false for API usage; nothing you send updates the weights.

Training vs inference

this is the phase that gave the model its knowledge, and the same phase that gave it a cutoff date.

Pretraining

the model's manners, its refusals and its house style come from this phase, not from what it read.

Post-training

the model refuses something you consider perfectly reasonable, and what you have run into is the tuning, not a bug in your prompt.

Alignment

you load a model that never went through chat tuning, ask it a question, and it continues your sentence instead of answering it.

Base model

you never had to teach it how a conversation works; every model behind an API you call was already tuned to answer.

Instruct model / chat model

every API call you make is inference, and nothing you send in one of them changes the weights for the next.

Inference

someone tells you the model version moved and behaviour changed, without a line of your code changing.

Checkpoint

the number is billions of parameters, and the well-trained 8B beat the older 70B you assumed would win.

Model size naming (7B, 70B)

the industry kept betting on bigger training runs because quality improved predictably with parameters, data and compute together.

Scaling laws

the same prompt that failed on the small model just works on the bigger one, with nothing else changed.

Emergent capabilities

you asked about a library released last month and the model has never heard of it.

Knowledge cutoff

you hard-coded the model name and six months later the best one has a different name.

Frontier model

you can download these weights and run them on your own hardware, which is not the same as having the training data or the code.

Open weights vs closed models

every parameter runs for every token, so the memory you have to find is the whole model rather than a fraction of it.

Dense model

a big team where only two or three specialists look at each token, not the whole staff.

Mixture of experts (MoE)

the big expensive model becomes the teacher, and what you actually ship is the small one that learned to imitate it.

Distillation

use reasoning modes for hard problems, not for every call; a simple extraction task does not need thinking tokens.

Reasoning model / extended thinking

you get a better answer by letting it think longer or sample more, without training anything bigger.

Test-time compute

you paste a screenshot into the same request as your text and it answers about both of them.

Multimodal