Thinking Machines, an AI startup established by former OpenAI CTO Mira Murati, has introduced Tinker, a Python-based API aimed at facilitating large language model fine-tuning. Currently in private beta, Tinker provides developers and researchers with control over their training pipelines while managing the complexities of distributed computing and infrastructure.
Murati described Tinker as a tool that simplifies the process of designing experiments and training models, allowing researchers to focus on innovation without the burden of traditional training complexities. The launch of Tinker follows Thinking Machines’ successful $2 billion funding round earlier this year, with backing from notable investors like a16z and NVIDIA.
Tinker distinguishes itself from typical drag-and-drop interfaces by offering a low-level but accessible API. This allows researchers to manipulate loss functions, training loops, and data workflows using standard Python code while leveraging Thinking Machines’ managed infrastructure for efficient distributed execution. Key features of Tinker include Python-native functions, support for diverse model architectures, integration with LoRA-based tuning for cost efficiency, and an open-source companion library.
Before its public announcement, Tinker was already being utilized in various research labs, including universities like Princeton and Stanford, addressing different model training challenges. For example, Princeton’s Goedel Team successfully fine-tuned models for formal theorem proving, while Stanford’s Rotskoff Lab improved accuracy in chemical reasoning tasks. Other institutions also reported enhanced flexibility and capabilities in their research projects thanks to Tinker.
The introduction of Tinker has generated positive feedback within the AI research community, with notable figures acknowledging its design and potential. Currently, Tinker is available for free during its private beta phase, with plans for a pay-as-you-go pricing model to be implemented shortly.
Thinking Machines, established amid a series of organizational changes at OpenAI, focuses on enhancing AI systems through customization and fostering open science. As the AI landscape becomes increasingly competitive, the company aims to carve out a niche by supporting adaptable models and ensuring safety standards.
Source: https://venturebeat.com/ai/thinking-machines-first-official-product-is-here-meet-tinker-an-api-for