Google AI subsidiary DeepMind is working on a new algorithm that will combine the power of natural language processing and reinforcement learning to create a more capable and responsible chatbot than ChatGPT, the popular AI system from OpenAI. The new algorithm, dubbed Gemini, is still in development and will be based on some of the techniques that helped DeepMind's AlphaGo defeat a world champion of the board game Go in 2016.
Gemini will also incorporate some new innovations that DeepMind's CEO and cofounder Demis Hassabis says are “pretty interesting”. According to an interview with WIRED, Gemini will aim to tap into the “amazing language capabilities” of large language models (LLMs) but also add some of the “strengths of AlphaGo-type systems”, such as planning and problem-solving.
“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 says. “We also have some new innovations that are going to be pretty interesting.”
The details of the new innovations are not disclosed, but they might involve using reinforcement learning to train Gemini on different tasks and domains, as well as using tree search to explore and remember possible actions and outcomes.
Reinforcement learning is a technique that allows software agents to learn from trial and error by receiving feedback on their performance. Tree search is a method that allows software agents to simulate and evaluate different scenarios and choose the best one.
AlphaGo used both techniques to master the complex game of Go, which has more possible moves than atoms in the observable universe. These techniques could help Gemini become more versatile and adaptable than ChatGPT, which relies mostly on statistical patterns learned from large amounts of text data.
Trying to Overcome ChatGPT From a Losing Position
ChatGPT has been widely praised for its ability to generate coherent and engaging texts on various topics, ranging from coding to poetry. However, it has also been criticized for its lack of factual accuracy, ethical awareness, and common sense. Even so, ChatGPT has been at the forefront of AI as the technology reaches the mainstream.
OpenAI, the developer of ChatGPT, has close ties to Microsoft and has been helping to spearhead the company's move to put AI on all its services. The GPT-4 LLM which is an updated version of the model that powers ChatGPT can now be found in Microsoft's Bing Chat, Microsoft 365 Copilot, GitHub Copilot, and Azure OpenAI Service.
Google has also launched its own chatbot, Bard, which is powered by a large language model called LaMDA. Bard is integrated with Google Search and can answer questions and have conversations on various topics. However, Bard is not as advanced or versatile as ChatGPT or Gemini, according to Hassabis.
Gemini will try to address some of these issues by providing more reliable and appropriate answers, as well as citing sources for its factual claims. Gemini will also be more transparent and accountable than ChatGPT, which sometimes produces harmful or misleading outputs. Gemini is still in the works and will take several months to complete. The project will cost tens or hundreds of millions of dollars, similar to the budget of GPT-4.
Hassabis hopes Gemini will be a game-changer for Google and for the field of artificial intelligence. He believes Gemini will be able to perform more tasks on the internet and on computers than any other chatbot.