Additional Side: The Mix lesson shows suspicion of the correct AI and 34%

The researcher from the meta The appropriate team with University of Jerusalem They have found that to force large languages ​​that “thinking” also helps to do their work in the discussion.

A decode The release today found to be shortcuts in Ai Syspes leads to many corrects and reduce the amount of money to have problems.

“In this activity, we condemn the idea that chains are thinking of good skills,” write their records in their papers “Never fail. Love of a looks looking goodly due to a positive attitude. “

Research agrees to the most interesting person, where companies produce too much in the production of the supply of allowing colors to achieve more than everlastingThe imaginary chains“- the passage of the passage used to solve difficult problems.

Ai’s accuracy of AI jumps 34% when colors use the tiny chains

The researchers realized that in a job with the same viewpoint, “the applicant has the correct answers – to 34.5% of the most accurate than the highest form.” This finding effect occurred between AI’s advanced AI and Benchmarks.

“Though he shows a fascinating result, results in expenditure and less time,” authors of authors tell how much description had been given.

Based on this, the team made a process called “short-m @ k“It’s a number of ways of suspending the same but Halts that make up a number of the first few methods. The last answer is selected through the hand-in hand.

A new ‘@ k

Instructions show the largest alternatives, meaning can be large. The researchers found their course of action to reduce the costs to 40% while keeping the same session of the functioning of the work as a proper way.

“3 @ k, whereas a little bit more than the short-1 @ k, all the time more than the low costs of all the money, up to 33% of time to reduce time),” do.

Michael Hassid, a writer on this sheet of paper, and his organization recognized that they taught the final discussion – to condemn another large idea in Ai’s civilization.

“A short teaching brings out a good job,” investigators wrote. “Similarly, struggle in a long time on S1-tight time without a helpful value.”

Tech Giants can save millions of installation “not failed” method

What is available will come during the most difficult time of ai industrial, if companies will break the most powerful varieties that use many useful items.

“Our discovery also indicates that the projects of testing time is discussed in LLM, which means ‘thinking’ does not mean to be better and can change, meaningless,” the researchers did.

‘The study is different from other popular methods. Popular training in the past, plus the opening work “Ables of a signal” and “Isolation“Methods, which often represents many routes about. They make a recent job as princeton and Google OURGemind’s”The value of the mind“Frame and carnegie Mellon”Self-self“The process, which has changed various ways to do not to reason.

In order to make a technical idea monitoring Ai, the surveys show that the main and more gentle is not always better. The lesson points that can be maintained with money that can cost money and change how to achieve better instead of using raw energy.

In the busy companies and growth, it is found to teach Ai being a shorter will not me to save the power to make – it makes the opposite machine. In some cases, despite tactful wisdom from intestinal intelligence: never fail.