Learn AI · Level 4 · Frontier view · step 2 of 15 · word 1 of 4
What is test-time compute?
Test-time compute is the computing power a model spends while answering, for example by thinking longer before it replies. More of it often means better answers to hard questions.
In plain English
For years, models got better mainly by spending more on training: bigger models and more data (see scaling laws). Test-time compute is the other lever: letting a finished model spend more effort on each question, by working through steps, trying several answers and checking them.
OpenAI’s o1, released in September 2024, showed that answers to hard maths and coding questions improve steadily the longer a model is allowed to think.
An example
Many chatbots now offer a “think longer” or “extended thinking” mode. A tricky planning question might take a minute instead of a second, and come back with a more careful answer.
Watch out
Thinking longer costs more and takes more time, and it doesn’t help much with simple questions or facts the model doesn’t know. It is also a big reason AI’s demand for electricity keeps growing.
Learn AI path · Level 4 · Frontier view · step 2 of 15
Next wordMixture of experts →See all of level 4