The researchers say they discovered a new method of “climb” ai, but there is no reason to be skeptical

They discovered the researchers discovered a new Ai “Saldy Dight”? That’s what some buzz on social media Suggest – but experts are skeptic.

You are back the laws, a bit of an informal concept, describes the average patterns improved as the size of rosecoses and computer resources. Up to about a year ago, scaling “duty dance models – it was the dominant law – it was the dominant law in the face of the sense that most of the sense.

Pre-training has not been parade but both of additional scaled laws, post-training scaling and Typing time scalingare emerged to complement it. The post-training scarce is essentially the behavior of a modes of time to the nearest to the nursing to the inference – “robe” (see: models ” R1).

Google and UC Berkley researchers have recently proposed in a card What some online commenters described as a fourth law: “Search inference.”

The pursuit of the inference has a model generate many possible answers to a question in parallel and then select the “best” of the handle. The researchers that require can you see the performance of a model of an elder, as Google 1.5 Pro Geminito a level that exceeds opening o1 preview “Reasoning” Model on science and math benchmarks.

Our card focuses in this search assets and their scaling trends. For example, for simply randomly 200 answers, Gemini 1.5 (an old pattern of 2024!) Beats O1-Preview and approach O1. This is endless, rl, or truth veritors of the truth. Pic.TWitter.com/hb5fo7ifnh

– Eric Zhao (@ errzehao28) On 17th March 2025

“(B) y randomly randomly 200 responses, gemini 1.5 – an old model of 2024 – beat o1,” eric zhao, a document, written Places series on X. I am “Magic is that auto-verification becomes naturally to scale! Will be expected that a correct solution becomes harder solutions is, but the contrary is the case!”

Many experts say the results are not surprising, however, and the search of inference can not be useful in many scenarios.

Matthew, Average Professor Professor at Alberta University, then gather the good “- In other word” when the best answer can be facility. But most questions are not so cut and dry.

“(I) F we cannot write the code to define what we want, we cannot use (inference search),” he said. “For something as the general language interaction, we can’t do this (…) do not usually have a great app to truly approach.”

Mike Cook, a Fellow London University of London to King, adding the Guzdial assessment, adding the sense of “Reason, to our own trials of thinking.

“(Inference search) does not” lift the processing process, “(s) to work around a matters to make a mistake if your template must make your issue must make that easier more easy.”

That research may be insured they are secure to be news indeased to scale the “reasoning” model “as the co-authors of the paper, the reasoning models today may rack up thousands of computation dollars on a single math problem.

It seems that the pursuit of new scale techniques will continue.