Researchers say they have found a new method to “magnify the AI”, but there is a reason to take a doubt

Researchers have discovered a new AI “Scale Law”? That is Some noise on social media Suggests – but experts are skeptical.

Describe how AI scale laws, a slightly informal concept, the performance of the performance of AI models, the size of the data used for their cultivation and the size of calculation sources. About a year ago, “Pre-preparation” – constantly educating larger models on a larger data – at least, at least the reason for the fact that AI laboratories were a dominant law.

There was no preparation in advance, but two additional scale laws, post-education measurements and Test time scaleappeared to complete it. The test-training scale adjusts the behavior of a model, testing time scale is a test-time scale to apply more calculation to make more computation – IE running models – “Information” management (see: Models) R1).

Recently Google and UC Berkeley researchers a paper Some commentators described the online fourth law: “Search-time search”.

The result-time search creates a model of a model, parallel creates many possible answers and then select the “best” of the gang. Researchers claim that one year of a year can increase performance Google’s Twins 1.5 ProTo a level that exceeds Openai O1-preview “Justification” model for science and math criteria.

Our paper focuses on this search arrow and its scale trends. For example, random answers and self-affirmation, twins 1.5 (ancient early 2024 models are approaching in advance and approaches. This is without the truth of Finetuning, RL or ground truth. pic.twitter.com/hb5fo7ifnh

– Eric Zhao (@ ericzhao28) March 17, 2025

“(B) Y, only 200 answers and self-affirmation, Gemini 1.5 – ancient Early 2024 model – O1, and approaches” Messages series in X. “Spell, checking the peculiarity, you will be able to choose a proper solution, and the solutions would expect the solutions to be larger, but the opposite is!”

Several experts may not be surprising, and the failure of this may not be useful in many scenarios.

Matthew, the EU researcher and associate professor of the University of Alberta, TechCrunch, when approaching is a good “evaluation function”, it was reported that the best answer could be easily identified. But most of the questions are cut and dry.

“(I) f cannot write the code to determine what we want, we cannot use (result-in-time).” “We cannot do this for something like common language interaction (…) Generally a great approach to solving most problems.”

A researcher Mike Cook with London’s specialization in the EU has agreed with the assessment of the Genter of London, and stressed the gap between the “regulation” in the AI ​​part of our word and our thinking processes.

“(Subsequent time searching) does not lift the model’s justification process,” he said. “(I) t, if your model is wrong, a way of being a very thin-supported error (…) a way of a tender technology is a way of a tender technology that will make it easier for 200 attempts to check these errors.”

This infertility may be confident that there are restrictions on the “thinking” model “thinking” to an AI industry “thinking” to a scale of “thinking” to an EI industry. As co-authors of the paper note, it can collect justification models today Thousands of dollars calculating in a single math problem.

As can be seen, the search for new scale methods will continue.