# Video script: the five disciplines of LLM optimization

Status: ready to record, no video has been published yet.

## 00:00, the problem

“LLM optimization” can refer to five different jobs. This video maps them so that SEO, application architecture, model compression, evaluation and agents are not confused.

## 00:35, AI answer visibility

The optimized object is a page or website. Work covers crawling, indexing, structure, entities, self-contained passages and observable mentions. Google points back to SEO fundamentals and requires no special markup for its AI features.

## 01:25, LLM applications

The optimized object is a product or pipeline. Task quality, cost and latency are measured together. Context, retrieval, routing, tools and caching need to remain connected to evals.

## 02:15, models and inference

The optimized object is the model. Fine-tuning, quantization, distillation and runtime selection change trade-offs between quality, memory and throughput. A theoretical estimate does not replace a benchmark on target hardware.

## 03:05, evaluation and reliability

An improvement only exists relative to a criterion. A versioned eval set retains inputs, expected results, graders, failures and regressions. An internal score is not general proof of safety.

## 03:55, agents and protocols

An agent orchestrates model, state and actions. MCP standardizes connections to data, tools and workflows. It certifies neither server security nor answer accuracy.

## 04:45, conclusion

Complete the sentence: “what are we optimizing, to improve what, measured how?”. The answer determines the discipline, team, protocol and evidence required.

Sources and full version: https://llmoptimisation.fr/en/five-disciplines-llm-optimization/