“Mira Morat” unveiled the first Thinking Machines Lab product

The newly established Thinking Machines Lab, founded by some former Openai researchers, including “Mira Morat”, its first product called Tinker Introduced. According to its creators, this tool automates the Fine Tune process of custom and advanced artificial intelligence models.

Mira Morati, co -founder and CEO of the company, said in an interview with Wired:

“We believe that Tinker helps researchers and developers test models and make the advanced capabilities of artificial intelligence available to everyone.”

Using the new Tinker Artificial Intelligence Tool

Typically, the creation of specialized versions of the open source models (for tasks such as solving mathematical problems, writing legal contracts or answering medical questions) requires having large GPU clusters and advanced software tools for managing extensive training.

But the startup promises that Tinker makes this process much easier for companies, researchers and even ordinary users and performs many steps automatically.

In its current version, Tinker allows users to personalize two open source models, Llama (from Meta) and QWen (from Alibaba). Users can adjust the models in a variety of ways (such as Supervised Learning) with a few lines via the Tinker API and then download the optimized models and run on any platform they want.

Establishing Thinking Machines Lab by former members of Openai

Mira Morat was previously Openai’s chief technology chief, and after the temporary dismissal of Sam Altman in 2023, he was appointed as the temporary CEO of Openai for some time. About 10 months later, he was separated from the company and founded Thinking Machines Lab with a group of his former colleagues.

John Scholman, another Openai co -founder who designed the ChatGPT training process using the “Human Food Reinforcement Learning” (RLHF) technique, Bert Zef, former Vice President of Research Openai and Lillian Wong, a former Safety and Robotics researcher.

Finally, the report points out that several users have been experimenting with the ability to learn Tinker reinforcement to teach models that are capable of doing specific tasks that are not available through conventional APIs.