Mira Murati’s Thinking Machines Lab has garnered significant attention with its $2 billion seed funding and the recruitment of former OpenAI researchers. A recent blog post by the lab delves into their project of developing AI models with reproducible responses, aiming to address the issue of randomness in AI model outputs.
In the world of artificial intelligence, Mira Murati’s Thinking Machines Lab has been making waves with its substantial seed funding of $2 billion and the recruitment of top-tier former OpenAI researchers. A recent blog post published by the lab offers a sneak peek into one of their innovative projects: creating AI models with predictable and reproducible responses.
The blog post, titled “Defeating Nondeterminism in LLM Inference,” aims to unravel the mystery behind the randomness often observed in AI model responses. For instance, posing the same question multiple times to ChatGPT can yield a wide range of answers, highlighting the non-deterministic nature of current AI models. However, Thinking Machines Lab sees this as a solvable challenge.
According to Horace He, a researcher at Thinking Machines Lab and the author of the post, the root cause of AI models’ randomness lies in the orchestration of GPU kernels during inference processing. By meticulously controlling this layer of operation, He believes that it is feasible to enhance the determinism of AI models.
Beyond the implications for enterprises and researchers seeking more reliable AI responses, He points out that achieving reproducible AI model outputs could also benefit reinforcement learning (RL) training. RL involves rewarding AI models for correct responses, but if the answers vary slightly each time, the training data becomes noisy. By ensuring more consistent AI model outputs, the entire RL process could be streamlined, making it “smoother.”
Thinking Machines Lab has disclosed its intention to leverage RL to customize AI models for businesses, as reported by The Information. Mira Murati, the former chief technology officer at OpenAI, hinted in July that the lab’s inaugural product will be unveiled soon, catering to researchers and startups developing customized models. The specifics of this product remain undisclosed, raising questions about its alignment with the lab’s research on reproducible responses.
Moreover, Thinking Machines Lab has pledged to regularly share blog posts, code, and research insights to benefit the public and enhance its research culture. The debut post in the lab’s new blog series, “Connectionism,” signals a commitment to transparency and knowledge dissemination. This approach contrasts with the trajectory of OpenAI, which has transitioned towards a more closed-off stance as it expanded. Time will tell if Murati’s lab stays true to its promise of open research practices.
The blog post offers a rare glimpse into the workings of one of Silicon Valley’s most enigmatic AI startups. While it doesn’t divulge the exact direction of the technology, it hints at Thinking Machines Lab’s engagement with fundamental questions in AI research. The ultimate test lies in the lab’s ability to tackle these challenges and develop products that validate its substantial $12 billion valuation.
