Breaking: Mira Murati’s Covert AI Lab Unveils Its Initial Offering

Thinking Machines Lab, a well-funded startup co-founded by leading researchers from OpenAI, has announced its inaugural product—a tool named Tinker that streamlines the creation of customized frontier AI models.

“We believe [Tinker] will empower researchers and developers to experiment with models and significantly enhance accessibility to frontier capabilities for everyone,” stated Mira Murati, co-founder and CEO of Thinking Machines, in a WIRED interview prior to the announcement.

Large corporations and research institutions are already fine-tuning open-source AI models to generate new versions tailored for specific tasks, such as solving mathematical equations, drafting legal documents, or addressing medical inquiries.

Typically, this process involves acquiring and managing GPU clusters alongside utilizing various software tools to ensure stability and efficiency during extensive training runs. Tinker aims to enable a broader range of businesses, researchers, and even enthusiasts to adjust their own AI models by automating much of this process.

Essentially, the team is banking on the notion that facilitating fine-tuning of frontier models will represent the next significant advancement in AI. There’s reason to think they could be on the right path. Thinking Machines Lab is led by researchers instrumental in the development of ChatGPT. Moreover, beta testers I’ve consulted claim that Tinker is more robust and user-friendly than comparable market tools.

Murati emphasizes that Thinking Machines Lab aspires to demystify the complexities involved in calibrating the world’s most advanced AI models, making it feasible for more individuals to traverse the frontiers of AI. “We’re making a capability that otherwise remains at the frontier accessible to all, which is truly transformative,” she asserts. “There are numerous intelligent individuals out there, and we need as many of them as possible engaged in frontier AI research.”

Currently, Tinker enables users to fine-tune two open-source models: Meta’s Llama and Alibaba’s Qwen. Users can write a few lines of code to access the Tinker API and initiate fine-tuning through supervised learning, which adjusts the model with labeled data, or via reinforcement learning, a growing method that fine-tunes models by providing positive or negative feedback based on their outputs. Users can then download their customized model and deploy it as desired.

The AI industry is closely monitoring this launch, partly due to the high caliber of the team behind it.