Thinking Machines Lab wants to make AI models more consistent | TechCrunch

Great interest in what Mira Murati’s thinking machine lab is building $2 billion in seed funding and an all-star team of former OpenAI researchers who joined the lab. exist Blog Posts Murati’s research lab published Wednesday, making the world its first introduction to one of its projects: creating AI models with repeatable responses.

The research blog post titled “Beat nondeterminism in LLM Inference” attempts to unravel the root causes of the introduction of randomness in AI model responses. For example, ask changpt the same question a few times and you may get a wide range of answers. In the AI ​​community, this is largely seen as a fact – today’s AI models are considered non-deterministic systems – but Thinking Machines Lab believes this is a problem that can be solved.

This post was written by Horace He He, a researcher at Thinking Machine Labs, who believes that the root cause of the randomness of AI models is the way the GPU core (the applet running inside NVIDIA’s computer chip) is sewn in the inference process (everything happens after you hit Enter Inter Inter in Chatgpt). He suggested that by carefully controlling this layer of orchestration, the AI ​​model can be made more certain.

He noted that in addition to providing more reliable responses to enterprises and scientists, it also pointed out that obtaining AI models to produce reproducible responses can also improve reinforcement learning (RL) training. RL is the process of rewarding AI models for the correct answer, but if the answers are all slightly different, the data will be a little noisy. According to him, creating a more consistent AI model response may make the entire RL process “smooth”. Thinking Machine Lab has told investors it plans to use RL Customize AI models for enterprisespreviously reported information.

Murati, former CTO of Openai, said in July that Thinkens Machines Lab’s first product will be Publicly available in the coming monthsand it will be “useful for researchers and startups who develop custom models.” It is unclear what the product is or whether it will use the technology in this study to produce a more reproducible response.

Thinking Machine Lab also said it plans Post blog posts frequentlycodes about its research, and other information to “benefit the public but also improve our own research culture.” This post is the first in the company’s new blog series called “Connectionism” and seems to be part of the work. Openai also promises open research when it is established, but as the company changes, the company becomes more closed. We will see if Murati’s research lab is true to this claim.

Research blogs provide a rare glimpse within one of Silicon Valley’s most secret AI startups. While it doesn’t fully reveal where the technology is going, it suggests that Thinking Machines Lab is solving some of the biggest problems on the AI ​​research boundaries. The real test is thinking about whether the Machine Lab can solve these problems and make products around its research to justify its $12 billion valuation.