Google and UCB researchers detail “inference-time search”, which some call a fourth AI scaling law, though experts are skeptical of its usefulness in many cases
But Can It Deliver? Eric Zhao : Why We Can't Escape Brute-Force Search Bluesky: Dave Lee / @davelee.me : An AI that adds “but there's reason to be skeptical” to the end of every sentence in a story about AI [embedded post] X: Ethan Mollick / @emollick : So it looks like there's a third scaling law: you can make models better by training them with more compute, by having them “think” for longer about an answer, or by generating large numbers of answers in parallel and picking good ones. Each might be increased independently. Eric Zhao / @ericzhao28 : Thinking for longer (e.g. o1) is only one of many axes of test-time compute. In a new @Google_AI paper, we instead focus on scaling the search axis. By just randomly sampling 200x & self-verifying, Gemini 1.5 ➡️ o1 performance. The secret: self-verification is easier at scale! [image]
Context & Ripple Effects
Google and UCB’s paper lands amid an active debate over whether extra compute after training constitutes a durable new route to model improvement. A prior critique of inference-time scaling claims had already questioned whether such “laws” reliably predict performance, while Google was also testing an AI Mode for complex search questions.
The significance is less a settled breakthrough than a clearer articulation of a technique that could shift optimization from training runs to how a model explores and selects answers at serving time.
First-order effects
- Google and UCB give researchers and model builders a defined inference-time-search approach to evaluate alongside longer reasoning and conventional training-scale improvements.
- The paper intensifies scrutiny of when generating and selecting many candidate answers improves results enough to justify the added inference work; the reported expert skepticism limits any immediate claim of broad usefulness.
Second-order effects
- Model providers will need to compare answer-quality gains against serving cost and latency, making task-specific evaluation more important than treating inference-time search as a universal scaling lever.
- For Google, the research is adjacent to its effort to make search handle complex queries through its experimental AI Mode, where additional inference steps must fit a user-facing product experience.
Third-order effects
- If effective on valuable tasks, progress may increasingly come from allocating compute dynamically at inference rather than only from larger training runs—though the corpus does not establish that this method generalizes broadly.
- The debate shifts the competitive question from who trains the largest model to who can operationalize reliable, economically viable reasoning and search at serving time.
The trend: AI development is broadening from pre-training scale toward adaptive inference strategies, with practical economics and task fit determining which techniques endure.