Thinking Machines Lab Aims to Solve Randomness in AI Models for More Reliable Responses and Smoother Reinforcement Learning

In a groundbreaking development, Mira Murati’s Thinking Machines Lab has offered insights into one of their projects with an aim to address a significant issue in the AI realm – creating AI models that produce consistent responses. The announcement was made in a blog post published on Wednesday, providing the first public look into the lab’s work.

Titled “Overcoming Randomness in AI Model Responses,” the post delves into the underlying causes of unpredictability in AI model answers. For instance, repeatedly querying ChatGPT with the same question may yield a variety of responses, which is generally accepted as a characteristic of current AI systems. However, Thinking Machines Lab believes this issue can be addressed and resolved.

The post, authored by researcher Horace He, suggests that the root cause of this inconsistency lies in the way GPU kernels – small programs embedded within Nvidia’s computer chips – are integrated during inference processing (the phase following user input in ChatGPT). By refining the orchestration at this level, He proposes, AI models could become more deterministic.

Beyond improving the reliability of responses for enterprises and scientists, such a development could also optimize reinforcement learning (RL) training. RL is the process of rewarding AI models for correct answers, but the slight variations in responses can lead to noisy data. Achieving greater consistency in AI model responses could streamline the entire RL process, according to He. Thinking Machines Lab has previously mentioned plans to employ RL for customizing AI models for businesses.

While details regarding the lab’s first product remain undisclosed, Murati, a former OpenAI chief technology officer, stated in July that it would be unveiled in the coming months and would prove beneficial for researchers and startups developing custom models. Whether this project will incorporate techniques from this research to generate more consistent responses is yet to be determined.

Thinking Machines Lab has also expressed its intent to regularly publish blog posts, code, and other research-related information as a means of benefiting the public and enhancing its own research culture. This post, the first in their new blog series titled “Connectionism,” appears to be part of this initiative. OpenAI, too, initially espoused open research principles but has since become more closed as it has grown in size. It remains to be seen if Thinking Machines Lab will adhere to its stated commitment.

This research blog offers a rare peek into one of Silicon Valley’s most mysterious AI startups. Although it does not reveal the exact direction of the technology, it underscores Thinking Machines Lab’s efforts to tackle some of the biggest questions in the forefront of AI research. The real test will be whether Thinking Machines Lab can successfully solve these problems and bring products to market that justify its $12 billion valuation.