🧠 Food for thought

Implications, context, and why it matters.

The deal backs an expensive and riskier way to tune AI models

• Google’s multibillion-dollar deal will help cover the heavy computing needs behind the reinforcement learning work that Thinking Machines’ Tinker tool uses 1.
• Tinker is a tool for automating custom frontier AI models. The source material does not call it a specialized application programming interface (API) or say it “dramatically simplifies” fine-tuning work 1.
• Tinker also adds risk. A security researcher used its API to make a 235-billion-parameter model generate harmful content for less than US$40 2.

Frontier AI labs are becoming prized cloud customers

• The non-exclusive agreement makes clear that frontier model developers such as Thinking Machines and Anthropic matter more in the cloud market 1.
• Cloud companies are chasing these deals. Anthropic, an AI startup that builds large language models, has signed large compute agreements with Google and Amazon 3.
• As tools like Tinker spread reinforcement learning, security work now reaches beyond output filters. Attackers with modest budgets can steer powerful models off course during training 2.