Foundation models Concepts

Backpropagation is the procedure that lets a multilayer neural network learn by sending output errors backward to adjust its weights, using gradient descent. A 1986 Nature paper made it widely known, and it is standard in most neural networks today.

A residual connection adds a block's input to its output, so the block learns only the change it needs to make. This makes very deep networks trainable, and it is a standard part of today's transformers.

The Transformer is a neural network architecture introduced in 2017 that relies on self-attention instead of recurrence or convolution. Most of today's large language models are built on it.

Pre-training trains a large model on huge amounts of unlabeled text, and fine-tuning then adapts it to a specific task with a small labeled dataset. This two-stage recipe is the basis of the term foundation model.

Scaling laws are empirical trends showing that a language model's test loss falls smoothly, as a power law, when parameters, data and training compute grow. They help plan large training runs and explain the race for compute.

A vision-language-action model (VLA) is a robot policy that takes in camera images and a language instruction and outputs actions directly, usually by adapting a pretrained vision-language model to control.

An AI agent is a system that perceives its environment and acts autonomously toward goals. The term covers everything from a thermostat to LLM-driven software that directs its own tool use.

Quantization stores a language model's weights and activations at lower numerical precision, cutting the memory and bandwidth that inference requires. Methods such as LLM.int8() and GPTQ made very large models runnable on far smaller hardware.

AGI is the goal of AI systems that match or beat humans across a broad range of tasks. The term has no agreed definition; a framework by Morris et al. grades systems by performance level and generality.

Alignment is the problem of making a model's behaviour match what its users and designers actually intend, not just what its training objective literally rewards. Training on human preference comparisons is one practical route to it.

Hallucination is false or misleading information that a model presents as fact. It is a central failure mode of text-generating systems and a main obstacle to trusting their output.

A multimodal model handles more than one kind of input or output, such as images together with text, inside a single model. Work on CLIP and Flamingo showed how such models are trained and how they can be steered toward new tasks.

Tokenization splits text into the tokens a language model actually reads; most modern systems do it with byte pair encoding, a 1994 compression method adapted for translation in 2015.

A technique that re-centres each layer's inputs to zero and rescales them to unit variance over a mini-batch, making deep networks faster and more stable to train.

Knowledge distillation trains a small model to reproduce the outputs of a larger model or an ensemble, so that much of the predictive performance is kept at lower compute and deployment cost.

A component of sequence models, especially transformers, that encodes each token's position as a vector, giving order-blind attention a sense of sequence; its forms range from the original transformer's sinusoids to RoPE and ALiBi.

Model parallelism splits one model's parameters and computation across multiple accelerators, so networks too large for a single device can still be trained. NVIDIA's 2019 Megatron-LM paper showed an intra-layer version built in native PyTorch.

In-context learning is a large language model's ability to perform a new task from instructions and a few examples placed in the prompt, with no weight updates. Named and studied in the GPT-3 paper, it moved adaptation from fine-tuning to prompt writing.

A state space model represents a sequence as a hidden state that evolves with the input. S4 brought the idea into deep learning in 2021, and Mamba made its parameters input-dependent, enabling linear-time modeling of long sequences.

Instruction tuning is a supervised fine-tuning stage that trains a large language model to follow user instructions; the FLAN paper showed it substantially improves zero-shot performance on unseen tasks.

Constitutional AI is a training method introduced by Anthropic in December 2022, in which a model's behaviour is shaped by a written constitution of principles rather than by human labels of harmful outputs.

Direct preference optimization (DPO) is a method for fine-tuning language models on human preference data with a supervised-learning-style objective, removing the separate reward model and the reinforcement-learning loop of standard RLHF.

Test-time compute is the computation a model spends while answering a prompt, beyond what training already fixed into its weights. A 2024 study by Snell et al. showed how to spend it adaptively, and how it can stand in for a bigger model.

Prompt engineering is the practice of designing the text input that steers a generative AI system toward a desired output. Shared vocabulary and techniques such as chain-of-thought prompting make results easier to describe, compare and reproduce.

Long context is a model's capacity to take in and work over very large inputs — documents, images, recordings — in a single request. Systems now reach millions of tokens, though finding information inside that window is a separate skill.

A context window is the full span of text a model can refer to while producing a response, measured in tokens and including the response itself. Its size, and how evenly a model uses it, shape what the model can do.

A function applied to each node's weighted sum, giving neural networks their non-linearity. Sigmoid, tanh and ReLU shaped early deep learning; GELU now dominates transformers.

Synthetic data is training material written by a model rather than collected from people or the web. In the phi-1 experiments a small code model trained partly on model-generated "textbooks" reached competitive benchmark scores, suggesting data quality itself shapes capability.

Red teaming is the practice of actively probing a model for harmful or unwanted behaviour before deployment. A 2022 method by Perez and colleagues automated it by using one language model to generate test cases for another.

A jailbreak is an adversarial prompt that gets a safety-trained model to produce behaviour its training was meant to prevent. A 2023 analysis by Wei et al. traced these failures to competing objectives and mismatched generalization.

A regularisation technique for neural networks: during training it randomly zeroes the inputs or outputs of some neurons so that units cannot co-adapt, which reduces overfitting.