Test-Time Compute
Your AI isn't slow — it's thinking. Here's why that changes everything.
Grounded in the research on chain-of-thought prompting
Wait, it just thinks longer?
Yes. Test-time compute means the model spends more computation at inference — when it's answering you — instead of only during training. Instead of firing off the first plausible token, it generates intermediate reasoning steps, checks itself, and revises. More compute in = better answers out. It's not magic. It's the same math, run more times.
Who figured this out?
Google Brain showed it in 2022 with chain-of-thought prompting — Wei et al. demonstrated that prompting large models to reason step-by-step before answering dramatically improved accuracy on math and logic benchmarks. Then OpenAI operationalized it in 2024 with o1, a model trained specifically to spend extended thinking time before responding. DeepSeek-R1 followed in early 2025, open-source, matching o1 on reasoning benchmarks.