Google has introduced the “AI co-scientist” (or virtual scientific collaborator) by adapting Gemini 2.0 to help scientists propose hypotheses and research plans based on assigned research topics.
The AI assistant for scientists uses a multi-agent system to create multiple agents that play different roles, such as idea generation, idea evaluation, and idea ranking. These agents can search for information on the web or utilize other specialized models to evaluate the plausibility of hypotheses.
In addition to brainstorming high-quality ideas, the AI assistant for scientists can also create research plans, allocate resources, and schedule tasks for us.
Google claims that using a specialized multi-agent system allows the AI co-scientist to achieve better evaluation scores compared to general market models. For example, using Elo rating evaluations and comparing them to Gemini 2.0 Flash Thinking, Gemini 2.0 Pro, and OpenAI o1.
Beyond evaluating model capabilities, Google has tested hypotheses generated by the AI co-scientist in real science labs, with satisfactory results in research areas such as pharmaceuticals, medicine, and biology.
Google has also conducted closed trials of the AI co-scientist through the Trusted Tester Program, where participants must register and be selected to participate.
TLDR: Google has introduced an AI co-scientist to assist scientists in proposing hypotheses and research plans, using a multi-agent system for idea generation and evaluation. The AI co-scientist has shown promising results in real science labs and is currently being tested in closed trials through the Trusted Tester Program.
Leave a Comment