Skip to content

AdvancedLesson 5 of 59 min read

Evaluating an AI: benchmarks, limits, and critical thinking

A course from the Academy: Under the hood of LLMs

Go to the quiz ↓

"The new model X crushes GPT on such-and-such benchmark!" Every week brings its triumphant announcement. Latest lesson: learning to read these figures, and to be just wary enough of them.

What is a benchmark?

A benchmark is a standardised set of questions: general knowledge and reasoning (MMLU), competition maths (AIME), real code (SWE-bench, where the model has to fix real bugs in real software), etc. Same test for all models → comparison is possible. It's essential… and tricky.

The three classic pitfalls

  • Contamination. Public benchmarks end up in training data: the model has potentially "seen the answers". A score can measure memory more than intelligence.
  • The target becomes the goal. Labs optimise for trending benchmarks; one more point on a score doesn't necessarily translate into real-world usefulness.
  • The average hides everything. A model "better on average" can be worse on your specific task: legal French, your code framework, your business domain.

The essential complements

Rankings based on blind human votes (two anonymous responses, you pick the better one) measure real preference. Safety evaluations test risky behaviours before deployment. And the best benchmark in the world remains yours: keep a small list of your typical tasks and run it on every new model.

Analogy. Benchmarks are like exam grades: useful for comparing at scale, but you wouldn't hire someone on their grades alone. You'd interview them on your own problems. Do the same with models.

End of lesson

Transformer, RLHF, RAG, MoE, benchmarks: you now have the vocabulary and concepts to read any AI announcement, separating signal from marketing. Alyriaz's article feed is your ongoing training.

Key concepts

Complete the lesson

A 4-question quiz: 3 correct answers are enough, plus an exercise to practise. It opens in the Academy, on this same lesson.

Take the quiz and complete the lesson Free. With an account, your progress is saved and leads to the course certificate.

The course: Under the hood of LLMs

  1. The Transformer: the architecture that changed everything10 min
  2. From raw text to assistant: fine-tuning and RLHF10 min
  3. Embeddings and RAG: connecting AI to your data10 min
  4. Mixture of Experts, distillation, quantization: the engineering of efficiency9 min
  5. Evaluating an AI: benchmarks, limits, and critical thinking9 min

The whole Academy →The AI glossary