Thinking Machines Lab, led by Mira Murati, has shed light on one of the projects launched with an initial $2 billion funding. According to a post on the company’s blog, the team is working to make artificial intelligence (AI) models produce more consistent and reproducible responses.
Currently, it is common for AI models like ChatGPT to provide different answers when asked the same question multiple times. This is widely accepted in the AI field because these models are considered "non-deterministic systems." However, Thinking Machines Lab sees this as a problem that can be solved.
In an article by Thinking Machines Lab researcher Horace He, the main cause of this inconsistency is attributed to the way GPU (graphics processor) cores—specifically, the small programs running within Nvidia chips—coordinate with each other. According to He, careful control of this management layer can make AI models more consistent.
This improvement could not only produce more reliable responses for businesses and researchers but also enhance reinforcement learning (RL) processes. RL rewards AI models for correct answers, but inconsistent responses reduce the quality of training data. More consistent outputs could make this process smoother.
Mira Murati, former CTO of OpenAI, noted that the first product from Thinking Machines Lab is expected to launch in the coming months and will be “useful for researchers and startups developing custom models.” It is not yet clear what the product will be or whether it will directly utilize the results of this research.
The company also plans to regularly publish blog posts, code, and other data about its research. This is intended to “benefit the public and foster a culture of open research.” The recently published blog post appears to be part of this strategy. While OpenAI once followed a similar open research policy, it gradually became a more closed company. Whether Mira Murati’s lab will stick to this promise remains to be seen.
