Technologies

Anthropic's open standard for connecting models to tools and data, released November 2024. Adopted by OpenAI, Google and Microsoft within months.

Models trained with reinforcement learning to think at length before answering. OpenAI o1 (September 2024), DeepSeek-R1 (January 2025). Spend compute at inference, not only in training.

Sequence models with linear cost in length. S4 2021, Mamba 2023. Used in hybrids with attention.

Have the model write intermediate steps before the answer. Prompting trick in 2022, trained behaviour in reasoning models by 2024.

A model in a loop with tools: think, act, observe. ReAct 2022, Toolformer 2023, coding agents 2024, computer use 2025.

Train a small set of extra weights instead of the whole model. LoRA 2021, QLoRA 2023. Why thousands of fine-tunes exist on Hugging Face.

Fine-tune on (instruction, response) pairs so a base model follows requests. FLAN 2021, InstructGPT 2022.

Loss falls predictably as a power law in compute, data and parameters. Kaplan 2020, Chinchilla 2022. The reason labs spend billions on training runs.

Retrieve documents, then generate with them in the prompt. The standard way to ground a model in private data.

Self-attention plus feed-forward blocks, no recurrence. Attention Is All You Need, 2017. The architecture of every large language model.

Train a reward model from human preferences, then optimise the policy against it. Christiano 2017, InstructGPT 2022. Turned GPT-3 into ChatGPT.

Google's matrix-multiply chip, announced May 2016. Trains Gemini.

Add the input of a block to its output. ResNet 2015. Makes very deep networks trainable. Inside every Transformer block.

Learn to reverse a noising process. Sohl-Dickstein 2015, DDPM 2020, latent diffusion 2021. Stable Diffusion, DALL-E 2, Sora.

A generator trained to fool a discriminator. StyleGAN faces. Replaced by diffusion for images after 2021.

Let the model weigh which inputs matter for each output. Bahdanau 2014 for translation. Self-attention made it the whole model in 2017.

Words as dense vectors learned from context. word2vec 2013, GloVe 2014. Meaning as geometry.

Neural networks as the function approximator in RL. DQN 2013, AlphaGo 2016.

Nvidia's CUDA made graphics chips programmable for general maths in 2007. AlexNet trained on two of them in 2012. The hardware story of the field since.

A recurrent cell with gates that keep gradients from vanishing. Ran speech recognition and translation until 2017.

Maximum-margin classifier with the kernel trick. The best general method from 1995 until deep learning.

Route each token to a few of many expert sub-networks. Jacobs and Hinton 1991, Shazeer 2017, Switch 2021. How frontier models grow parameters without growing cost.

Learn from reward rather than labels. TD learning 1988, Q-learning 1989, then deep RL and RLHF.

Networks with loops that carry state across a sequence. Hard to train over long ranges, which the LSTM fixed.

Compute the gradient of the loss with respect to every weight by the chain rule, layer by layer. Werbos 1974, Rumelhart, Hinton and Williams 1986. The algorithm that trains all deep networks.

Pearl's graphical models of probability. The probabilistic turn in AI of the late 1980s.

Shared local filters and pooling. Neocognitron 1980, LeNet 1989-1998, AlexNet 2012. The architecture of computer vision until the Vision Transformer.

Rules written down from human experts. DENDRAL (1965), MYCIN (1972), XCON (1980). The commercial AI of the 1980s, and the cause of the second winter when they proved brittle.

Programs that improve from data instead of rules. Named by Arthur Samuel in 1959.

Rosenblatt's single-layer learning machine. Learns a linear boundary from examples. Its limits, shown in 1969, caused the first neural network winter.

McCarthy's language. The language of AI research for thirty years.

Intelligence as symbol manipulation: search, logic and rules. From the Logic Theorist (1956) through expert systems. Dominant until the 1990s.

Can a machine hold a text conversation that a judge cannot tell from a human's? Turing's 1950 proposal.

Shannon's measure of information. Entropy, bits and cross-entropy, the loss that language models minimise.

Wiener's science of feedback and control in machines and animals. The intellectual parent of both AI and neural networks.

A unit that sums weighted inputs and fires past a threshold. McCulloch and Pitts, 1943. Every network is made of them.