Demo 08

Foundation Model Lifecycle: How does token prediction become instruction following?

The Transformer supplied a scalable architecture, but moving from a base model to a cooperative assistant also requires choices about data, objectives, and post-training. This chapter focuses on which stages update weights and which information affects only one inference call.

Teaching interaction

Foundation Model Lifecycle Lab

What changes during pretraining, instruction tuning, preference feedback, and runtime context?

Weights can change Weights stay fixed
1 / 6

Pretraining turns a large corpus into a prediction task

The model repeatedly predicts later tokens from earlier text. More data and compute can reduce training loss, but the corpus is not a fact-checked database.

What changes at this stage Pretraining objective

The prediction objective supplies broad language patterns; it does not directly define the most helpful response to a user.

Scripted output preview Training signal

Given: The history of AI… → predict the next token

Learning goals
  • Separate broad patterns learned in pretraining from interaction behavior shaped in post-training.
  • Understand that instruction data and preference comparisons can update weights while runtime context changes only the current generation.
  • Recognize that better instruction following does not provide live knowledge, permissions, or a reliability guarantee.
Simplification note

This demo uses one historically important post-training path to explain responsibility boundaries. It runs no real training, labeling, reward model, or inference. Models use different data, objectives, and optimization methods, and preference feedback does not guarantee factuality, safety, or complete alignment with user intent.

Historical Position And Boundary

From scalable architecture to cooperative model, then into a complete system

Pretraining learns conditional-generation patterns from a large corpus. Instruction tuning and preference feedback continue updating weights and shaping behavior. Runtime context participates only in the current inference and does not retrain the model on the spot. Post-training can improve instruction following, but it does not automatically provide live knowledge, external permissions, factuality, or complete safety, so the next chapter still places the model inside retrieval, tools, memory, and evaluation.

References

Concept Check

Test The Core Intuition With One Question

This check supports reflection and never blocks the next chapter. Records stay on this device.

When new material enters Runtime Context, what changes in this demo?