Google DeepMind, renowned for its groundbreaking artificial intelligence (AI) research, is forging ahead in its development of an innovative AI system, named Gemini, that may outshine OpenAI’s ChatGPT, according to CEO Demis Hassabis.

DeepMind, which grabbed global attention in 2016 when its AI program AlphaGo defeated a Go champion, plans to integrate the techniques that made AlphaGo a success into Gemini. As a large language model similar to GPT-4, which powers ChatGPT, Gemini will not only work with text but will also incorporate AI capabilities such as planning and problem-solving.

Lessons from AlphaGo

Hassabis compares Gemini to a combination of the strengths of AlphaGo-type systems with the superb language capabilities of large models, plus some intriguing new innovations. Key to AlphaGo’s victory was reinforcement learning, a technique allowing software to tackle complex problems through repeated attempts and performance feedback. This method, along with a tree search for exploring and remembering possible moves, is expected to propel language models into performing more tasks online and on computers.

Catching Up with OpenAI’s ChatGPT

Gemini’s development could significantly shape Google’s strategy in the AI arena, particularly against competitors like ChatGPT and other generative AI technology. Google, despite its pioneering role in AI, has exercised caution in deploying AI-based products. In contrast, since ChatGPT’s debut, Google has launched its chatbot, Bard, and incorporated generative AI into its search engine and other products.

The consolidation of DeepMind with Google’s primary AI lab, Brain, to create Google DeepMind, was a strategic move aimed at enhancing AI research. Hassabis believes the synergy of these two AI powerhouses will be instrumental in driving future AI innovations.

The Promise of Reinforcement Learning

Training a large language model, like OpenAI’s GPT-4, involves feeding vast amounts of curated text into machine learning software known as a transformer. This forms the basis for the model’s proficiency in predicting subsequent text based on given input, a mechanism that proves astonishingly efficient in answering questions and generating text or code. The introduction of reinforcement learning, where AI models are fine-tuned based on human feedback, could be a game-changer for Gemini, given DeepMind’s extensive expertise in this area.

Navigating the Risks

As AI development accelerates, Hassabis acknowledges the risks and uncertainties associated with it. The concern among AI experts is that technology might be exploited for malicious purposes or become challenging to control. Some have even called for a halt in the development of more powerful algorithms to avoid creating potentially dangerous entities.

However, Hassabis argues that the enormous potential benefits of AI, such as advancements in areas like health or climate science, necessitate the continuation of AI development. Despite acknowledging the risks, he firmly believes in progressing with AI technology, describing it as potentially “the most beneficial technology for humanity ever,” if done correctly.

DeepMind has been cognizant of the potential risks of AI since its inception, with an internal “AI safety” group led by one of the company’s co-founders, Shane Legg. Hassabis also joined other high-profile AI figures in issuing a statement warning of risks associated with AI that could be comparable to nuclear war or a pandemic.

Towards More Transparent Research

One of Hassabis’s immediate challenges is identifying the likely risks of more capable AI. In an effort to promote transparency and open up AI research, Hassabis suggested that DeepMind may grant early access to frontier models for academia. This move would counter the growing concern that experts outside of large companies are being excluded from the newest AI research.

While the threat posed by AI remains uncertain, Hassabis urges the development of safeguards considering the pace of AI advancements. As the Gemini project continues to unfold, it symbolizes both the promise and challenges that lie ahead in AI development. The world will undoubtedly be watching how Google DeepMind navigates this journey.