News
Google DeepMind pilots double-blind AI evaluation
Google DeepMind is piloting double-blind AI evaluation with several external organizations, preventing the evaluators from viewing the model weights and the model providers from reading confidential test prompts.
Google DeepMind is collaborating with the Singapore Institute for AI Security, OpenMined, AVERI, and MLCommons to pilot a double-blind AI evaluation.
The solution utilizes Google Cloud Confidential Space, preventing the evaluators from viewing the Gemini model weights, while also preventing Google from accessing confidential test prompts from external organizations, thereby reducing benchmark contamination and protecting the intellectual property of both parties.
refer to:Google DeepMind official statement