In the rapidly evolving world of generative AI chatbots, OpenAI’s ChatGPT is the AI bot to beat. That was the case even before OpenAI released GPT-4 — a multimodal large language model that significantly improves the reliability, creativity, and usefulness of ChatGPT. And yet, despite ChatGPT’s current dominance, Google DeepMind CEO Demis Hassabis told Wired in a recent interview that Google’s Gemini AI system will eventually outperform GPT-4.
Although Gemini will perform similarly to OpenAI’s model, responding to text inputs from users, Google also wants to combine the technology with techniques used in AlphaGo.
If you’re unfamiliar, AlphaGo is a computer program developed by DeepMind that defeated the human world champion of the notoriously difficult strategy board game known as Go. DeepMind trained AlphaGo with a technique called “reinforcement learning.” That’s when software receives feedback on its performance every time it makes a decision.
AlphaGo also used a Monte Carlo tree search (MCTS) algorithm to search for and decide which moves it would use in a game. By bringing these techniques to Gemini, you can start to see how the resulting consumer-facing product might give ChatGPT a run for its money.
“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 told Wired. “We also have some new innovations that are going to be pretty interesting.”
Hassabis says that Gemini is still “a number of months” away, and Wired notes that it will likely cost tens of millions of dollars, if not hundreds of millions. But given that Bard has failed to make any real impact, it’s clear that Google needs to try again.