Thinking machines laboratory wants to make artificial intelligence models more consistent

There was a great interest in what the Mira Murati Machines Machines laboratory adopts with 2 billion dollars in the financing of seeds And the team of the former researchers who joined the laboratory. in Blog post Posted on Wednesday, the MURATI Research Laboratory gave the first look at one of his projects: Create Amnesty International models with repetitive responses.

The search blog, titled “Nonderminism in LLM’s inference”, tries to empty the root cause of the randomness in the artificial intelligence model responses. For example, ask Chatgpt the same question several times, and you are likely to get a wide range of answers. This has been largely accepted in the artificial intelligence community as a fact-the models of artificial intelligence today are non-lively systems-but Machines thinking laboratories see this is a solution that can be solved.

This post, composed by a laboratory researcher composed by the Machines Machines Horace, argues that the root cause of the randomness of artificial intelligence models is the way the GPU nucleus – small programs that work inside computer chips in NVIDIA – together in the treatment of inference (everything that happens after pressing entering the chat). It suggests that by carefully controlling this layer of synchronization, artificial intelligence models can be made more inevitable.

In addition to creating more reliable responses to institutions and scientists, it is noted that obtaining the models of Amnesty International to generate repetitive responses can improve reinforcement learning training (RL). RL is the process of rewarding artificial intelligence models to obtain the correct answers, but if the answers are slightly different, the data becomes somewhat loud. Create more consistent responses of artificial intelligence can make the entire RL “more smooth”, according to He. Tell investors thinking laboratory that he is planning to use RL Allocating the artificial intelligence models for companiesInformation previously reported.

Moratti, the former chief technology employee at Openai, said in July that the first Machines Lab product would be The veil was revealed in the coming monthsAnd it will be “useful for researchers and startups that develop custom models.” It is still unclear what this product is, or whether it will use techniques from this research to generate more cloning responses.

Thinking machines laboratory also said he was planning Often the publication of blog publicationsThe code, and other information about its research in an attempt to “benefit from the audience, but also improve our own research culture.” This post seems to be the first in the company’s new blogs series called “Connectionism”, part of this effort. Openai also committed to opening research when it was founded, but the company has become more closed because it has become bigger. We will see whether the Moraati Research Laboratory remains the reality of this claim.

The research code offers a rare overview within one of the most emerging companies of artificial intelligence in Silicon Valley. Although it does not reveal exactly where technology is going, it indicates that the Machines Machines laboratory addresses some of the biggest questions on the boundaries of artificial intelligence research. The real test is whether the Machines Machines Laboratory has a solution to these problems, and to make products about its research to justify its evaluation of $ 12 billion.