Former OpenAI Chief Technology Officer Mira Murati today launched a new artificial intelligence startup, Thinking Machines Lab Inc., that will develop multimodal models.

The company is entering the crowded AI market four months after Reuters reported that Murati (pictured) was in talks to raise more than $100 million for a new venture. Thinking Machines Lab didn’t address the funding reports in its launch blog post today. However, the company did confirm another previously reported detail: that its initial team includes former OpenAI research executive Barret Zoph.

Murati and Zoph left OpenAI last September shortly after John Schulman, one of the ChatGPT developer’s co-founders, stepped down as well. Murati is Thinking Machines Lab’s chief executive, while Zoph is its CTO. Schulman is the startup’s chief research officer.

Thinking Machines Lab plans to train multimodal models that can process not only text but also multimedia files such as images. As OpenAI’s CTO, Murati oversaw the development of ChatGPT and DALL-E, a series of image generation models. She also played a key role in forging OpenAI’s partnership with Microsoft Corp., which has provided much of the cloud infrastructure that powers the AI developer’s research. 

The startup will approach “infrastructure quality as a top priority,” it detailed in its launch blog post. “Research productivity is paramount and heavily depends on the reliability, efficiency, and ease of use of infrastructure. We aim to build things correctly for the long haul.”

The company also shared other details about its development roadmap. Its models won’t be specifically tailored to do programming and math tasks like OpenAI’s o1. Instead, they will feature the ability to “adapt to the full spectrum of human expertise and enable a broader spectrum of applications,” the company stated. It’s unclear whether those applications might include consumer use cases.

Customizability will be another focus of Thinking Machines Lab’s engineering efforts. The company plans to make it simple for customers to customize its AI models to their requirements. Today, companies customize neural networks using prompts and by fine-tuning them on proprietary training datasets.

Last year, it was reported that Thinking Machines Lab’s products will be based on proprietary models. However, it appears that the company plans to open-source at least some components of its AI stack. 

“We believe that we’ll most effectively advance humanity’s understanding of AI by collaborating with the wider community of researchers and builders,” the company stated on its website. “We plan to frequently publish technical blog posts, papers, and code.”

The open-source program will encompass, among others, AI safety. The company plans to publicly release code and other technical assets produced as part of its efforts to avoid harmful AI output. It intends to test its algorithms’ safety using established methods such as red-teaming, or the practice using simulated cyberattacks to find weak points.

Thinking Machines Lab currently has 29 employees. Besides former OpenAI staffers, the initial team includes researchers from Google LLC, Meta Platforms Inc., Mistral and other major players in the AI ecosystem.

Photo: OpenAI