How AI Works

Eleven concepts, from neurons to Da Vinci · 8 October 2026 · Zachariah Roydel Robertson

Eleven concepts that power modern language models, in the order you need them: from a single artificial neuron to how Da Vinci learns. Each builds on the one before.

01. Neurons and weights

A neuron takes in numbers, multiplies each by a weight (a strength), adds them up and sends out a number. Weights are what gets adjusted during training. They are the model's memory.

02. Backpropagation

After the model makes a guess, it measures how wrong it was. Then it traces backward through every neuron, asking how much each contributed to that error. Each weight is nudged a tiny amount to be less wrong.

03. Tokens

Words and parts of words are broken into small pieces called tokens. Each token is mapped to a number, like a postcode for a chunk of language. The model works with these numbers, not letters.

04. Next-word prediction

The model takes everything said so far and predicts which token comes next, with a probability for each possibility. It picks one, and that new token becomes context for predicting the next.

05. Transformers and attention

The Transformer is the engine inside modern AI. Its key idea is attention: each word looks at every other word in the context and weighs how important it is for understanding. This is how the model tracks relationships across long passages.

06. Pretraining

Before a model ever talks to anyone, it is trained on huge amounts of text: books, websites, articles. It learns the patterns of language by guessing the next word billions of times, and absorbs knowledge of the world along the way.

07. Fine-tuning

After pretraining, the model is adapted to a specific task using smaller, curated datasets. A customer-support model sees examples of good support conversations. Fine-tuning adjusts the weights slightly to specialise the general knowledge.

08. Reinforcement learning

After the model generates a response, a person or an automated system scores it. The model learns to maximise that score by adjusting its weights toward what worked and away from what did not.

09. Test-time thinking

A model can spend extra computation working through a hard problem step by step, instead of answering immediately. More thinking gives more accurate answers on complex questions such as mathematics and code.

10. Open and closed weights

With open weights, a model's parameters are public: anyone can run and modify it. With closed weights, the model is kept private and reached through an API. Open is free and customisable; closed is usually more refined and supported.

11. Da Vinci: causal learning

Da Vinci, the assistant inside e1-4, combines three ideas: stochastic training (learning from randomness), causal reasoning (understanding what causes what) and self-improvement (learning from its own mistakes). It is built on open-weight foundations, trained on its own compute, and improves by thinking.