A very short introduction to the AI behind agents, in sixteen slides: where language models sit in AI, a neuron, a network and its parameters, training, tokens, embeddings, attention, the transformer, generating one token at a time, other types of models, and what each of those facts means for running an LLM in production.
The slides
The tool is lesson 1 of
LLMOps for AI Agents, in sixteen slides:
- What is an LLM?: what a team that runs an LLM meets in production, and the mechanism that causes it.
- Artificial intelligence, machine learning and language models: four fields, each inside the one before.
- A neuron: a weighted sum and an activation: three inputs worked through to one output.
- A neural network: layers of neurons: every line a weight, and the parameters counted.
- Training: adjusting the weights from examples: the loss, backpropagation, and why we use models others trained.
- A token is a piece of a word: a sentence cut into tokens and their ids.
- An embedding: a token as a point in space: related words near each other.
- Attention: each token looks at the tokens before it: the word it finding the pod it refers to.
- The transformer: the architecture of every LLM: the blocks, and one probability for every next token.
- Generation: one token at a time: temperature and the stop token.
- Other types of models: seven types, each with what it accepts and what it produces.
- Multimodal LLMs: images, voice and video beside the text: a screenshot of a dashboard and a question about it.
- Speech models: voice to text, and text to voice: and a voice assistant made of three models.
- Image and video generation models: a description turned into a picture and a clip, and diffusion's steps.
- Embedding models: meaning as a semantic vector: three texts and a picture as points.
- What the model's nature means for operations: the chapter's map, one fact per lesson.
The buttons on the left and right edges move between slides. So do the arrow keys and a swipe.
The lab
The token ids, the positions of the embeddings and the attention weights are made up for the drawings. The neuron's
numbers and the network's parameter count are computed by the slides.
Sources