Thinking Machines Lab Unveils Tinker AI Tool

Thinking Machines Lab Launches Tinker

Thinking Machines Lab, a well-funded startup established by former OpenAI researchers, has launched its first product: Tinker. This tool is designed to automate the creation of custom frontier AI models. Mira Murati, cofounder and CEO, stated that Tinker aims to empower researchers and developers, making advanced AI capabilities more accessible.

Fine-Tuning AI Models

Currently, large companies and academic labs adjust open source AI models to produce versions suited for specific tasks like math, legal work, or medical inquiries. This usually requires managing GPU clusters and using software to ensure stable and efficient large-scale training. Tinker seeks to broaden AI model fine-tuning to more businesses, researchers, and hobbyists through automation.

The team is banking on fine-tuning models becoming a major trend in AI, a bet that seems plausible given the team's background with ChatGPT. According to beta testers, Tinker is more user-friendly and powerful than similar tools. Thinking Machines Lab intends to simplify tuning powerful AI models, enabling more individuals to explore AI's potential. Murati emphasizes that making frontier capabilities accessible is revolutionary, encouraging more smart individuals to engage in AI research.

Tinker's Capabilities

Tinker currently supports fine-tuning Meta’s Llama and Alibaba’s Qwen. Users can fine-tune models through supervised learning by writing code to use the Tinker API, or they can use reinforcement learning by providing feedback based on the model's outputs. The fine-tuned model can then be downloaded and run anywhere.

The AI industry is keenly observing this launch, partly due to the team's expertise. Murati, formerly OpenAI's CTO and briefly CEO, cofounded Thinking Machines Lab with other OpenAI veterans, including John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz. The startup had previously garnered attention by raising $2 billion in seed funding, valuing it at $12 billion.

Schulman, who led fine-tuning efforts for ChatGPT, explains that Tinker will simplify extracting new abilities from models using reinforcement learning. He notes that while they abstract away distributed training details, users retain control over the data and algorithms.

Access to Tinker will be available via application. While the API is currently free, the company plans to charge for it in the future. Eric Gan, a researcher at Redwood Research, is using Tinker’s reinforcement learning to tune models for inserting backdoors in code. He notes that Tinker enables capabilities that wouldn’t be accessible through an API and simplifies adjustments to fine-tuning. Robert Nishihara, CEO of Anyscale, highlights Tinker's unique combination of abstraction and tunability compared to other fine-tuning tools.

Open Source and Future Plans

There are concerns about open source models being misused. Thinking Machines currently vets API access but intends to implement automated safeguards. The company has also been publishing research on model training, including ways to improve neural network performance and fine-tune language models more efficiently, which is used in Tinker.

The company aims to promote openness, contrasting with the trend of US AI companies keeping their models closed. Murati hopes Tinker will help reverse the increasing closure of commercial AI models, bridging the gap between frontier labs and academia.