a capability evaluation

Before you hire someone or trust a tool, you want to know what it can actually do — can this person code, can this calculator handle big numbers? AI systems are similar but harder to judge, because a single model can attempt almost anything you ask in words, and its abilities are written down nowhere — they emerged from training. A capability evaluation is a structured test built to measure what an AI model can do: not whether it is nice or safe, but how skilled it is at some task.

In practice an evaluation is a set of tasks plus a way of scoring the answers — for example thousands of exam-style questions, programming problems with automatic checks, or multi-step tasks in a simulated environment. Researchers run the model on the tasks and record a score (say, 'solves 62% of these coding problems'). Good capability evals try hard to get the model to perform at its best — giving it tools, multiple attempts, and well-crafted prompts — because the goal is to find the ceiling of what the model can do, not a lazy lower bound.

Capability evals are the foundation of much AI safety work: you cannot decide whether a model is dangerous until you know what it can do. But a measured score is only a lower bound on true capability — a better prompt, more attempts, or fine-tuning might reveal abilities the test missed (the elicitation problem). Scores can also be inflated if test questions leaked into training data (benchmark contamination). So a capability number is informative but should be read as 'at least this capable', not 'exactly this capable'.