Tech & Future
Understand the machines rewriting everything
Novice
Subdomain
What an AI model actually is: training, tokens, inference, and why predictions are statistical guesses. The mental picture that makes everything else click.
Training vs inference
Training is the slow, expensive learning phase where the model's internal numbers are set. Inference is the fast usage phase that runs every time you ask it something.
Open
Tokens
Tokens are the sub-word chunks an AI reads and writes. About 4 English characters or 3/4 of a word per token.
Open
Neural networks and parameters
Billions of tiny tunable knobs that, together, store patterns.
Open
Supervised, unsupervised, and RL
Three different teachers.
Open
Embeddings
An embedding is a list of numbers (a vector) that represents the meaning of a chunk of text, image, or audio, so that similar things sit close together in that number space.
Open
Mixture of Experts (MoE)
An architecture where a large model contains many specialised sub-networks (experts) and a router decides which small subset of experts processes each token, so most parameters stay inactive on any one forward pass.
Open
Direct Preference Optimization (DPO)
A method that fine-tunes a language model directly on pairs of preferred and dispreferred responses, without first training a separate reward model as classic RLHF requires.
Open
Diffusion models for image generation
Generative models that learn to reverse a gradual noising process, producing images by denoising random Gaussian samples step by step.
Open
Parameter
A tunable weight inside a model adjusted during training.
RAG
Retrieval-Augmented Generation, giving a model external documents to ground its answers.
Overfitting
Memorising training data so well the model fails on new inputs.
Inference
Using a trained model to produce outputs.
Fine-tuning
Continuing to train a model on a smaller, more specific dataset to specialise it.
Embedding
A vector of numbers representing the meaning/features of a chunk of input.
RLHF
Reinforcement Learning from Human Feedback, alignment via human ratings.
Hallucination
A fluent, confident output that's factually wrong.
Token
The smallest input/output unit an LLM works with, typically 3 to 4 characters in English.
Context window
The maximum tokens a model can consider in a single forward pass.
Vector database
A store optimised for nearest-neighbour search over embeddings.
MoE
Mixture of Experts. Routes tokens to specialised sub-networks so active parameters per token stay small.