“Nvidia is investing in Mira Murati's Thinking Machines Lab, part of a multiyear partnership in which the startup will deploy at least one gigawatt of cutting-edge chips to train and serve its frontier AI models.
The deal includes a collaboration to design artificial-intelligence training and serving systems using Nvidia technology. The size and structure of the investment couldn't be learned.
Thinking Machines, founded last year by Murati, OpenAI's former chief technology officer, and a number of her former colleagues, has so far been largely quiet about its intentions.
It released its first product, Tinker, a tool to help researchers train AI models, last year [1], and has said it aims to make AI systems that work with humans rather than operating autonomously.
The partnership is Nvidia's latest move to support the AI industry's up and comers. The deal offers the relatively young AI lab computing power to advance its research, as well as funding to pay its staff.
While Thinking Machines was already working with Nvidia chips to develop its models, the new partnership further cements the startup as a customer. Murati called Nvidia's technology "the foundation on which the entire field is built."
Thinking Machines has been competing fiercely for AI researcher talent as the valuations in its industry have skyrocketed. Today, it has about 120 employees from about 30 a year ago, according to people familiar with the matter.
Nvidia Chief Executive Officer Jensen Huang gave a nod to the staff assembled by Thinking Machines, which includes OpenAI co-founder John Schulman, as a reason for the deal, saying it "has brought together a world-class team to advance the frontier of AI." [2]
1. Tinker is a flexible training and fine-tuning Application Programming Interface (API) released by Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. It is designed to give developers and researchers "surgical control" over the fine-tuning of large language models (LLMs) without the burden of managing complex training infrastructure. The platform offers the opportunity to start using the services for free so that developers can experiment without initial costs.
Key Features of Tinker
Infrastructure Management: The service runs on Thinking Machines’ internal clusters, handling resource allocation, scheduling, and failure recovery so users can focus on their models.
Low-Rank Adaptation (LoRA): Tinker primarily uses LoRA to reduce the compute and cost required for customization.
Granular Control: Unlike typical high-level APIs, Tinker provides low-level primitives like forward_backward and sample, allowing users to implement custom training algorithms [3].
Model Support: It supports a wide range of open-weight models, including Mistral, Gemma, Qwen, and Kimi.
Tinker Cookbook: Thinking Machines provides an open-source library on GitHub containing modern implementations of post-training methods and "recipes" for specialized tasks like solving math problems or scientific reasoning.
Recent Developments (as of March 2026)
General Availability: Originally launched in private beta in October 2025, Tinker is now generally available and includes support for advanced vision input models.
Nvidia Partnership: In March 2026, Thinking Machines announced a massive compute deal with Nvidia to deploy at least one gigawatt of Vera Rubin systems to power future frontier model training.
Company Valuation: The startup reached a $12 billion valuation following a $2 billion seed round in July 2025.
2. Nvidia Invests in Thinking Machines. Keach Hagey. Wall Street Journal, Eastern edition; New York, N.Y.. 11 Mar 2026: B4.
3. Low-level primitives such as forward_backward and sample are foundational building blocks provided by Tinker, a managed fine-tuning API launched by Thinking Machines (founded by former OpenAI CTO Mira Murati). These primitives allow researchers to write custom, high-control training loops—ranging from supervised fine-tuning (SFT) to advanced reinforcement learning (RL)—while the API handles the complex infrastructure, such as distributed GPU scheduling and failure recovery.
Key Primitives in Tinker:
sample: Takes a prompt, generates a completion from the current model weights, and allows for scoring that completion for online reinforcement learning. A completion generated from the current model weights represents the model’s direct prediction of the next likely sequence of tokens based on the patterns it learned during training. The weights act as the "memory" or "knowledge" of the AI, controlling how input data is processed to generate output.
forward_backward: Sends a batch of data through the model to compute the loss and gradients.
optim_step: Updates the model weights based on the computed gradients.
save_state: Handles checkpointing. In the context of machine learning libraries like Hugging Face Accelerate, save_state is a method used to handle comprehensive checkpointing by saving the current state of a training session.
Key Aspects of these Primitives:
High Control & Flexibility: Unlike black-box solutions, these primitives enable users to define their own loss functions, training loops, and data workflows in standard Python, supporting complex methods like direct preference optimization (DPO) and online RL.
Performance Optimization: While they allow for granular control, they are designed to work with LoRA (Low-Rank Adaptation) to make fine-tuning efficient and cost-effective, even for large Mixture-of-Experts models like Qwen-235B-A22B.
Komentarų nėra:
Rašyti komentarą