Interactive Illustrated AI History中文

Demo 09

LLMs And Modern AI Systems: Why do large models still need external systems?

Transformers let models learn general patterns from large-scale text, but real products are more than a single model call. Modern AI systems place the model inside an external structure of context, retrieval, tools, memory, and evaluation.

Why do large models still need external knowledge and tools?

Teaching interaction

LLM System Boundary Lab

Why does the same base model need different external system paths for different tasks?

Current policy Q&A

A newly updated private policy cannot come from parameters alone

This task needs current evidence and a verifiable citation, but no external action or persistent state.

User request
Under the internal policy updated this week, how many days of parental leave do I receive? Cite the source.
Required system path
Needs Retrieval, Context, and Eval; Memory and Tools remain optional on this path.
1 / 6

The task arrives with freshness and citation requirements

The user asks about an internal policy updated this week and requires a citation. The request states the goal but does not contain the answer.

What this boundary proves Task identified

A clear instruction can still lack the external facts required to complete it.

Learning goals
  • Distinguish model parameters, current context, external retrieval, and persistent task state.
  • Understand that tools turn generation into external action and therefore need separate authorization and result checks.
  • See systems select components for a task instead of stacking every module onto every request.
Simplification note

This demo calls no real model, retrieval store, memory service, tool, or evaluator. Requests, evidence, state, actions, and checks are scripted. The components are teaching boundaries; a real system may merge, split, or omit them.

Historical position

It connects Transformers, RAG, and agents

Earlier problemModel parameters are not the live world

Pretrained knowledge is powerful but becomes outdated, and it cannot directly read private documents, business databases, or tool results.

What it solvesPlace capability inside a system structure

Context windows, retrieval, tool use, memory, and evaluation turn a model into a composable software component.

Remaining problemThe system can still fail

Wrong retrieval, tool misuse, prompt injection, permission leaks, and evaluation blind spots still require engineering controls.

Simplification note

This is a system map, not a real orchestration engine

This chapter uses two scripted tasks to explain common LLM application components. It calls no real model, database, memory service, tool API, or evaluation service. The component relationships and outcomes are teaching abstractions that explain why modern AI applications need structure outside the model.

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.

Why is one box labeled “LLM” insufficient for a modern AI application?

Previous: Foundation Model Lifecycle Mark complete and continue Next: RAG Pipeline