How do large language models actually work?
Start with how a language model generates output so you know which assumptions from ordinary software still hold.
Learning tracks · Track 4 of 7
For software engineers adding language models and agents to their existing engineering skills.
Start with step 1: How do large language models actually work?
After this track: You can explain an AI application’s components, test its behavior and review the code used to build it.
Follow these 12 steps in order, or return to the one you need for your current work. Reading times are estimates; listening times cover the full episode.
Start with how a language model generates output so you know which assumptions from ordinary software still hold.
Understand the input and output units before estimating request size or inference cost.
Add the request’s capacity limit to that model before designing conversation history or document inputs.
Practice specifying tasks and constraints before adding retrieval or orchestration to compensate for unclear instructions.
Follow the retrieve-then-generate flow before deciding how your application will supply facts the model lacks.
Learn what similarity search represents before choosing how to retrieve documents for that flow.
Connect embeddings to storage and search before treating a vector database as the whole retrieval system.
Compare a shared tool protocol with a custom integration once you know which data and actions the application needs.
Decide how much orchestration the task needs before taking on a framework’s abstractions.
Build checks for variable outputs before letting a successful example stand in for a test suite.
Examine a working AI-assisted coding process after establishing the tests you would need to trust its output.
Finish by deciding how generated changes will be reviewed when writing code is faster than checking it.
Take it into your work. You can explain an AI application’s components, test its behavior and review the code used to build it.
Choose another track or continue through the five-level pathway.