AI Literacy Series: From Autocomplete to Agents - How AI Systems Build on Next-Word Prediction to Search Files, Use Tools and Decide What to Do Next

Autocomplete predicts the next word. An AI agent can search a file, use a tool, keep a result, and choose its next step. This course explains the changes in the model and the surrounding software that connect those two systems. 

Dr. Jason Dean starts with the short word windows of n-gram models, then explains learned representations, attention, Transformers, pretraining, and post-training. Animated diagrams show what each change adds and where its limits remain. The course then moves outside the model: files and retrieval supply relevant material, tools perform permitted actions, and a control loop uses results to choose later steps. 

You will learn to distinguish what a model learned during training from what an application supplies for a particular request. You will also compare a single response, a fixed workflow, and an agent loop, then examine the state, permissions, approvals, validation, and stopping rules around repeated actions. No programming background is required. 

By the end of the course, learners should be able to: 

- explain how an n-gram model predicts a next word and why longer contexts create sparse evidence;

- distinguish training from inference and trained parameters from current context;

- describe the roles of embeddings, attention, Transformers, softmax, and temperature;

- separate a model from the application, files, retrieval systems, and tools around it;

- distinguish additional reasoning within one response from a workflow or agent loop;

- identify the state, permissions, approvals, validation, and stopping rules needed around repeated model actions; and

- ask five concrete questions when evaluating an AI system: what was learned, what is in context, which sources and tools are connected, what actions are permitted, and what makes the process continue or stop.