What is an AI benchmark?

An AI benchmark is a standardised exam for machines. Researchers assemble a set of questions or tasks with known correct answers, run models on them, and publish scores. When an AI company announces a new model, it almost always comes with a table of benchmark results showing it beating competitors.

How benchmarks work

Each benchmark targets a skill:

• Knowledge and reasoning: multiple-choice questions across many subjects (MMLU), or very hard expert-level science questions (GPQA).
• Math: competition-style problems.
• Coding: writing functions that pass tests, or resolving real GitHub issues in real codebases (SWE-bench).
• Agentic tasks: completing multi-step tasks using tools, browsers or terminals.
• Human preference: platforms like LMArena show people two anonymous models’ answers and let them vote, producing a ranking.

Scores are usually reported as a percentage correct or a rating.

An analogy for the meatbags

Benchmarks are standardised tests like the SAT. Useful for comparing millions of students quickly; poor at predicting who will be a good employee. And just as tutoring firms “teach to the test”, AI labs can — deliberately or accidentally — optimise models for benchmarks rather than for real-world usefulness.

Why benchmarks mislead

• Contamination: benchmark questions published online can end up in training data, so the model may have effectively seen the answers.
• Saturation: when top models all score in the high 90s, the benchmark stops distinguishing them, and researchers must build harder ones.
• Narrowness: a high coding score doesn’t guarantee the model will understand your codebase, your brand voice or your customers.
• Cherry-picking: companies tend to highlight the benchmarks where they look best.

Why it matters for the Chosen

Benchmarks help you shortlist models and track how fast the field moves — the steady climb on hard tests is one reason serious people debate AGI timelines. But the Chosen make decisions with their own evals:

1. Collect real examples of the work you want AI to do.
2. Run two or three candidate models (or tools) on the same inputs.
3. Score them on what matters to you: accuracy, tone, format, speed, cost per task.
4. Re-run the test whenever a new model launches.

This private benchmark is a core part of LLMOps, and it’s worth more than any leaderboard. Leaderboards tell you which model is best at exams. Your eval tells you which one makes you money.