Google's DeepMind is developing a cutting-edge artificial intelligence (AI) system, Gemini, expected to outshine OpenAI's ChatGPT, according to CEO Demis Hassabis.
Speaking to Wired, Hassabis said that Deepmind is enhancing the capabilities of its large language models leveraging on techniques that helped its AlphaGo defeat a Go champion in 2016.
As a large language model, Gemini resembles GPT-4, the foundation of ChatGPT, but with additional attributes inspired by AlphaGo, such as advanced problem-solving and planning abilities.
"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," Hassabis explained.
Demis Hassabis, DeepMind’s cofounder and CEO, says his engineers are using techniques from AlphaGo to make an AI system dubbed Gemini that will be more capable than that behind OpenAI’s ChatGPT. https://t.co/KL9cJMREUP
— WIRED (@WIRED) June 26, 2023
Competitive threat for ChatGPT
Creating Gemini could cost tens to hundreds of millions, comparable to the cost of GPT-4, as confirmed by OpenAI CEO Sam Altman in April.
The finished product is anticipated to play a significant role in Google's response to the competitive threat posed by ChatGPT and other generative AI technologies.
DeepMind's legacy in reinforcement learning, showcased through its domination of simple video games, could enable Gemini to possess novel capabilities.
Health and climate change benefits
DeepMind's wide-ranging research in areas from robotics to neuroscience also provides a unique foundation for the development of this large language model.
Despite growing concerns about the potential misuse or uncontrollability of advanced AI systems, Hassabis defends the need for further development.
He asserts the immense potential benefits, particularly in scientific discoveries related to health or climate change.
However, he emphasises the importance of developing safeguards against potential risks.
As AI development accelerates, the challenge lies in determining what risks more capable AI might pose.
Hassabis has expressed the hope of providing academia with early access to frontier models like Gemini, helping address concerns about the concentration of AI research within big tech companies.