Demis Hassabis, the CEO of Google’s DeepMind, has announced the development of a new AI system called Gemini, which aims to surpass the capabilities of OpenAI’s ChatGPT. DeepMind’s Gemini, currently in the works, is a large language model similar to GPT-4, but it incorporates techniques used in AlphaGo, the program that defeated a champion player of the board game Go in 2016. By combining the strengths of language models with game-playing capabilities, Gemini seeks to offer new functionalities such as problem-solving and planning.
AlphaGo, a groundbreaking AI program, utilized reinforcement learning and tree search techniques to master complex games like Go. DeepMind plans to leverage this expertise to enhance the capabilities of language models, enabling them to perform a wider range of tasks on the internet and computers. While Gemini is still in the development stage, Demis Hassabis has indicated that it may take several months and cost tens or hundreds of millions of dollars to complete, underscoring the significance of the project.
Google’s response to the competitive threat
Once completed, Gemini could play a pivotal role in Google’s response to the competitive threat posed by ChatGPT and other generative AI technologies. Although Google pioneered numerous AI techniques, it opted for cautious development and deployment of products based on these innovations. But, since the debut of ChatGPT, Google has expedited its own chatbot, Bard, and integrated generative AI into various products, including its search engine. The merger of DeepMind and Google’s primary AI lab, Brain, has created Google DeepMind, a team that combines the strengths of both organizations. Demis Hassabis believes that 80 to 90 percent of recent AI innovations have originated from either DeepMind or Google, making this collaboration a powerhouse in the field.
Having previously experienced the AI gold rush, Demis Hassabis understands the dynamics and challenges of navigating technological advancements within tech giants. DeepMind’s acquisition by Google in 2014, following its remarkable achievements in using reinforcement learning to master video games, brought AI capabilities once considered decades away into the present. The stunning victory of AlphaGo over Go champion Lee Sedol in 2016 was a testament to the potential of AI to surpass human performance.
Expanding the capabilities of language models
Training a large language model like GPT-4 involves feeding vast amounts of curated text into machine learning software, allowing it to predict subsequent letters and words accurately. Reinforcement learning, based on human feedback, further refines the model’s performance. DeepMind’s extensive experience in reinforcement learning provides a unique advantage for the development of Gemini, potentially unlocking new capabilities beyond what traditional language models offer.
In addition to drawing from reinforcement learning, Hassabis and his team may explore ideas from other areas of AI to enhance large language models. DeepMind’s researchers work across diverse fields, such as robotics and neuroscience. Recently, the company showcased an algorithm capable of learning manipulation tasks using various robot arms. Integrating knowledge acquired from physical experiences, similar to how humans and animals learn, could be crucial to advancing AI capabilities. Overcoming the limitations of language models that primarily learn from text-based data is seen as a significant challenge in the field.
Balancing progress and risks
Demis Hassabis faces the challenge of accelerating Google’s AI efforts while mitigating potential risks. The rapid advancements in language models have raised concerns among AI experts about malevolent uses and the ability to control these technologies. Hassabis acknowledges these concerns but emphasizes that the tremendous benefits of AI, particularly in scientific discovery, necessitate its continued development. While some have called for a pause in the creation of more powerful algorithms, Hassabis remains committed to advancing AI responsibly and believes that comprehensive research, evaluation tests, and collaboration with external experts are essential to effectively addressing the risks.
