Thinking Machines Lab wants to make AI models more consistent

There is a lot of interest in what Mira Murarati’s Thinking Machine Lab is making with Seed funds are $ 2 billion And the former Open Researchers All -Star Team that joined the lab. A Blog post Published on Wednesday, Murarati’s Research Lab gave the world the first look of a project: creating AI model with reproductive response.

The research blog post titled “LLM defeats Nonaditaminism in LLM infection” attempts to unpack the root cause of what randomly introduces in response to the AI ​​model. For example, ask Chatzipt the same question a few times and you can probably get a broad answer. It is taken as a fact in the AI ​​community — today’s AI models are considered as non-determined systems-but the thought machine lab sees it as a solved problem.

Thinking Machine Lab researcher Horace Hey argues that the root cause of the random of the AI ​​models is GPU kernels – small programs driven inside NVIDIA computer chips – infection processing is stitched together (whatever happened after you enter the chatter). He suggests that the AI ​​models can be further determined by cautiously controlling this level of orchestration.

Beyond creating a more reliable response for initiatives and scientists, he notes that AI models can improve training (RL) training found in AI models to create reproductive response. The RL is the process of rewarding AI models for the correct answer, but if the answers are a bit different, the data becomes a bit noise. Creating a response from a more series AI model can make the entire RL process “smooth”, according to it. Thoughts machine labs have told investors that it has planned to use it from RL Customize AI models for businessThe data has been previously reported.

OpenAI’s former Chief Technology Officer Murarati said in July that thoughtful machine would be the first product of the machine lab Uncovered in the next monthsAnd it “will be useful for” researchers and custom models developed startups. ” It is not yet clear whether the product is what, or whether it will use techniques from this study to create a more reproductive response.

Thoughts machine lab also said that it plans Often publish blog posts“Public benefits, however, in the attempt to improve our own research culture, this post seems to be part of the new blog series of the company known as” Connection “. The OPNA also promised to open the research at the time of establishment, but as the company was even greater, the company will see that the research lab is more closed.

The research blog gives a rare glimpse of one of the most confidential AI startups in Silicon Valley. It does not reveal where the technology is going on, indicating that the thought machine lab is dealing with some of the largest questions on the AI ​​research border. The original test is that the thought machine lab can solve these problems and make products around its research to equate $ 12 billion evaluation.