glossary

The words, in plain English.

Active parameters: The number of parameters a model actually uses to process each token. In MoE models, most weights are idle on any given token — so active params is much smaller than total params.go deeper →
Alignment: The process of training a model to behave in ways humans find helpful, harmless, and honest. Alignment techniques like RLHF and DPO shape what the model will and won't do after pretraining.go deeper →
Attention: A mechanism that lets each token look at every other token in the sequence and decide which ones are most relevant to its meaning right now. It's the core operation repeated throughout every transformer layer.go deeper →
Base model: A model that has only been pretrained — it completes text rather than answers questions. It knows a lot, but it hasn't learned to follow instructions or be helpful yet.
Chain-of-thought (CoT): A prompting technique where the model is asked to write out its reasoning step-by-step before giving a final answer. It improves accuracy on complex tasks, but only reliably at larger model scales.go deeper →
Context window: The maximum number of tokens a model can see at once — its working memory. Tokens outside the context window are invisible to the model while it generates a response.go deeper →
Corpus: The collection of text a model is trained on. Pretraining corpora are enormous — typically hundreds of billions to trillions of words scraped from the web, books, and code.
Decode: The second phase of answering: the model writes the reply one token at a time, each step waiting on memory. This is why long answers take a while.go deeper →
DPO (Direct Preference Optimization): An alignment method that skips the reward model entirely — it trains the language model directly on preference pairs (chosen vs rejected) using a clever loss function derived from the Bradley-Terry model.go deeper →
Embedding: A list of numbers (a vector) that represents a token or concept in a space where similar meanings sit close together. The model's first step is turning each token into its embedding vector.go deeper →
Fine-tuning: Continuing to train a pretrained model on a smaller, task-specific dataset. It's how you turn a general-purpose base model into a specialist — for tool calling, a medical domain, or a specific writing style.go deeper →
GRPO: A reinforcement-learning method that scores a group of answers to the same question and rewards the better-than-average ones — with no separate 'critic' network, which makes it cheaper. It's the algorithm behind DeepSeek-R1.go deeper →
Hallucination: When a model generates text that sounds confident but is factually wrong or made up. It happens because models predict likely-sounding text, not verified facts — any gap in training data can produce plausible-but-false output.go deeper →
HBM (High-Bandwidth Memory): The stacked memory chips on a GPU that hold model weights and the KV cache during inference. HBM is large and fast, but reading from it is still the main bottleneck during token generation.go deeper →
Inference: Running a trained model to generate output — as opposed to training, which updates the weights. When you send a message to a chatbot and get a reply, that's inference.
Instruct model: A base model that has been further trained (via SFT and usually RLHF or DPO) to respond helpfully to instructions rather than just continue text. Most chat models you use are instruct models.
KV cache: A running store of the model's intermediate work for every token it has already processed, so each new token doesn't redo that work. It grows as the text gets longer and is the main thing filling GPU memory during generation.go deeper →
Logits: The raw, unnormalized scores a model produces for each possible next token before any probability conversion. Higher logits mean the model thinks that token is more likely; softmax converts them to a proper probability distribution.go deeper →
LoRA (Low-Rank Adaptation): A fine-tuning technique that freezes the original model weights and trains only a tiny pair of matrices inserted at each layer. The result has a fraction of the trainable parameters, cutting memory and compute costs dramatically.go deeper →
MoE (Mixture of Experts): An architecture where only a subset of the model's 'expert' sub-networks activate for each token, keeping per-token compute low while total parameter count stays high. DeepSeek-V3 activates 37B parameters per token out of 671B total.go deeper →
Next-token prediction: The single thing a base language model does: given the text so far, estimate how likely each possible next token is, then pick one. Repeat to generate text.go deeper →
Parameter count: The total number of weights in a model, usually expressed in billions (e.g. '7B' means 7 billion parameters). Larger parameter counts generally mean more capacity, but also more memory and compute needed to run the model.
Prefill: The first phase of answering: the model reads your entire prompt at once and produces the first reply token. It's fast because the work happens in parallel.go deeper →
Pretraining: The first and largest training phase, where a model learns from trillions of tokens of text by predicting the next token over and over. This is where most of a model's knowledge comes from.go deeper →
Quantization: Reducing the number of bits used to store each weight — from 16 bits down to 8, 4, or fewer. Smaller weights mean less memory and faster loading, with a small cost to output quality.go deeper →
Reward model: A separate neural network trained on human preference data that scores how good a model's response is. In RLHF, it acts as a proxy for human judgment during the RL training loop.
RLHF (Reinforcement Learning from Human Feedback): A training method where human raters compare pairs of model outputs, those preferences train a reward model, and then the language model is tuned to produce outputs the reward model scores highly.go deeper →
RLVR (RL with Verifiable Rewards): A reinforcement learning approach where the reward signal comes from checking whether the model's answer is correct — no human raters, no reward model. Only works for tasks with objectively verifiable answers, like math or code.go deeper →
SFT (Supervised Fine-Tuning): A training step where the model learns from a curated dataset of high-quality instruction-and-response pairs. A few thousand good examples is often enough to teach a base model to follow instructions.go deeper →
Softmax: A function that converts a list of raw scores (logits) into a probability distribution that sums to 1. It amplifies high scores and shrinks low ones, so the model's confidence is reflected in the probabilities.go deeper →
Temperature: A knob that controls how random the model's token choices are. Low temperature makes the model pick the highest-probability token almost every time; high temperature spreads probability more evenly, producing more varied output.go deeper →
Test-time compute: Extra computation done at inference time to improve output quality — such as generating multiple answers and picking the best, or letting the model 'think' longer before responding. Reasoning models like o1 and DeepSeek-R1 trade inference cost for better answers.go deeper →
Token: A chunk of text — often a word or part of a word — that a model reads and writes one at a time. Models don't see letters or words directly; they see tokens.go deeper →
Tokenizer: The component that converts raw text into the sequence of token IDs a model actually processes. Most modern tokenizers use Byte-Pair Encoding (BPE), which merges frequent character pairs into single tokens.go deeper →
Transformer: The neural network architecture used by every major language model today. It processes tokens through a stack of identical layers, each containing an attention step and a feed-forward step.go deeper →
Weight (parameter): One of the billions of numbers inside a model that get adjusted during training. Together the weights are a compressed imprint of everything the model was trained on.

Active parameters

The number of parameters a model actually uses to process each token. In MoE models, most weights are idle on any given token — so active params is much smaller than total params.go deeper →

Alignment

The process of training a model to behave in ways humans find helpful, harmless, and honest. Alignment techniques like RLHF and DPO shape what the model will and won't do after pretraining.go deeper →

Attention

A mechanism that lets each token look at every other token in the sequence and decide which ones are most relevant to its meaning right now. It's the core operation repeated throughout every transformer layer.go deeper →

Base model

A model that has only been pretrained — it completes text rather than answers questions. It knows a lot, but it hasn't learned to follow instructions or be helpful yet.

Chain-of-thought (CoT)

A prompting technique where the model is asked to write out its reasoning step-by-step before giving a final answer. It improves accuracy on complex tasks, but only reliably at larger model scales.go deeper →

Context window

The maximum number of tokens a model can see at once — its working memory. Tokens outside the context window are invisible to the model while it generates a response.go deeper →

Corpus

The collection of text a model is trained on. Pretraining corpora are enormous — typically hundreds of billions to trillions of words scraped from the web, books, and code.

Decode

The second phase of answering: the model writes the reply one token at a time, each step waiting on memory. This is why long answers take a while.go deeper →

DPO (Direct Preference Optimization)

An alignment method that skips the reward model entirely — it trains the language model directly on preference pairs (chosen vs rejected) using a clever loss function derived from the Bradley-Terry model.go deeper →

Embedding

A list of numbers (a vector) that represents a token or concept in a space where similar meanings sit close together. The model's first step is turning each token into its embedding vector.go deeper →

Fine-tuning

Continuing to train a pretrained model on a smaller, task-specific dataset. It's how you turn a general-purpose base model into a specialist — for tool calling, a medical domain, or a specific writing style.go deeper →

GRPO

A reinforcement-learning method that scores a group of answers to the same question and rewards the better-than-average ones — with no separate 'critic' network, which makes it cheaper. It's the algorithm behind DeepSeek-R1.go deeper →

Hallucination

When a model generates text that sounds confident but is factually wrong or made up. It happens because models predict likely-sounding text, not verified facts — any gap in training data can produce plausible-but-false output.go deeper →

HBM (High-Bandwidth Memory)

The stacked memory chips on a GPU that hold model weights and the KV cache during inference. HBM is large and fast, but reading from it is still the main bottleneck during token generation.go deeper →

Inference

Running a trained model to generate output — as opposed to training, which updates the weights. When you send a message to a chatbot and get a reply, that's inference.

Instruct model

A base model that has been further trained (via SFT and usually RLHF or DPO) to respond helpfully to instructions rather than just continue text. Most chat models you use are instruct models.

KV cache

A running store of the model's intermediate work for every token it has already processed, so each new token doesn't redo that work. It grows as the text gets longer and is the main thing filling GPU memory during generation.go deeper →

Logits

The raw, unnormalized scores a model produces for each possible next token before any probability conversion. Higher logits mean the model thinks that token is more likely; softmax converts them to a proper probability distribution.go deeper →

LoRA (Low-Rank Adaptation)

A fine-tuning technique that freezes the original model weights and trains only a tiny pair of matrices inserted at each layer. The result has a fraction of the trainable parameters, cutting memory and compute costs dramatically.go deeper →

MoE (Mixture of Experts)

An architecture where only a subset of the model's 'expert' sub-networks activate for each token, keeping per-token compute low while total parameter count stays high. DeepSeek-V3 activates 37B parameters per token out of 671B total.go deeper →

Next-token prediction

The single thing a base language model does: given the text so far, estimate how likely each possible next token is, then pick one. Repeat to generate text.go deeper →

Parameter count

The total number of weights in a model, usually expressed in billions (e.g. '7B' means 7 billion parameters). Larger parameter counts generally mean more capacity, but also more memory and compute needed to run the model.

Prefill

The first phase of answering: the model reads your entire prompt at once and produces the first reply token. It's fast because the work happens in parallel.go deeper →

Pretraining

The first and largest training phase, where a model learns from trillions of tokens of text by predicting the next token over and over. This is where most of a model's knowledge comes from.go deeper →

Quantization

Reducing the number of bits used to store each weight — from 16 bits down to 8, 4, or fewer. Smaller weights mean less memory and faster loading, with a small cost to output quality.go deeper →

Reward model

A separate neural network trained on human preference data that scores how good a model's response is. In RLHF, it acts as a proxy for human judgment during the RL training loop.

RLHF (Reinforcement Learning from Human Feedback)

A training method where human raters compare pairs of model outputs, those preferences train a reward model, and then the language model is tuned to produce outputs the reward model scores highly.go deeper →

RLVR (RL with Verifiable Rewards)

A reinforcement learning approach where the reward signal comes from checking whether the model's answer is correct — no human raters, no reward model. Only works for tasks with objectively verifiable answers, like math or code.go deeper →

SFT (Supervised Fine-Tuning)

A training step where the model learns from a curated dataset of high-quality instruction-and-response pairs. A few thousand good examples is often enough to teach a base model to follow instructions.go deeper →

Softmax

A function that converts a list of raw scores (logits) into a probability distribution that sums to 1. It amplifies high scores and shrinks low ones, so the model's confidence is reflected in the probabilities.go deeper →

Temperature

A knob that controls how random the model's token choices are. Low temperature makes the model pick the highest-probability token almost every time; high temperature spreads probability more evenly, producing more varied output.go deeper →

Test-time compute

Extra computation done at inference time to improve output quality — such as generating multiple answers and picking the best, or letting the model 'think' longer before responding. Reasoning models like o1 and DeepSeek-R1 trade inference cost for better answers.go deeper →

Token

A chunk of text — often a word or part of a word — that a model reads and writes one at a time. Models don't see letters or words directly; they see tokens.go deeper →

Tokenizer

The component that converts raw text into the sequence of token IDs a model actually processes. Most modern tokenizers use Byte-Pair Encoding (BPE), which merges frequent character pairs into single tokens.go deeper →

Transformer

The neural network architecture used by every major language model today. It processes tokens through a stack of identical layers, each containing an attention step and a feed-forward step.go deeper →

Weight (parameter)

One of the billions of numbers inside a model that get adjusted during training. Together the weights are a compressed imprint of everything the model was trained on.