Google Deepmind is developing an AI system called Gemini that will combine the capabilities of AlphaGo with large language models such as GPT-4, the technology behind OpenAI’s ChatGPT.
Deepmind CEO Demis Hassabis tells Wired that the company’s expertise in reinforcement learning could give Gemini distinctive features. Using AlphaGo’s techniques such as reinforcement learning and tree search, Gemini aims to provide new capabilities such as problem-solving and planning.
“At a high level you can think of Gemini as combining some of the strengths of AlphaGo-type systems with the amazing language capabilities of the large models. We also have some new innovations that are going to be pretty interesting.”
Gemini was officially unveiled in May and, according to Hassabis, will be in development and training for several more months. Gemini is said to have multimodal capabilities not seen in previous models, be highly efficient at integrating tools and APIs.
The system could cost tens or hundreds of millions of dollars to develop and could play a critical role in Google’s response to ChatGPT and other generative AI technologies, Wired reports.
Gemini will be available in multiple sizes and designed to support future innovations such as memory and planning. Rumors in March said that the model will have a trillion parameters like GPT-4 reportedly does. The project is said to use tens of thousands of Google’s TPU AI chips for training.
Meanwhile, according to OpenAI CEO Sam Altman, GPT-5 is still a long way from launch and won’t begin training for at least six months. A launch in 2024 seems likely. Google Deepmind hasn’t said when Gemini will be released.