Demis Hassabis, CEO of Google DeepMind reveals plans for “Gemini,” a system harnessing techniques employed by AlphaGo to defeat a Go champion in 2016.
In a groundbreaking achievement in 2016, Google’s DeepMind AI lab stunned the world with AlphaGo’s victory over a champion Go player. Now, Demis Hassabis, DeepMind’s CEO and co-founder, announces the development of Gemini—an AI system leveraging techniques from AlphaGo to surpass the capabilities of OpenAI’s ChatGPT.
Gemini, currently in the developmental phase, is a sophisticated language model akin to GPT-4, which powers ChatGPT. However, Hassabis explains that his team will infuse Gemini with the strategic methods utilized in AlphaGo, enabling the system to exhibit novel capabilities like problem-solving and planning.
Hassabis describes Gemini as an amalgamation of AlphaGo’s strengths and the language prowess of large-scale models. With additional exciting innovations, Gemini was first teased at Google’s recent developer conference, where a host of new AI projects were unveiled.
AlphaGo’s foundation lay in Google DeepMind’s pioneering technique known as reinforcement learning, where software tackles intricate challenges by iteratively attempting actions and receiving feedback. The implementation of tree search further facilitated exploration and memory of potential moves on the Go board. Looking ahead, language models are expected to make significant strides by handling diverse tasks on the internet and computers.
Gemini’s development is projected to span several months and could incur a substantial cost in the range of tens or hundreds of millions of dollars. Sam Altman, CEO of OpenAI, had previously stated that the creation of GPT-4 exceeded $100 million in expenses.
Closing the Gap
Upon its completion, Gemini is poised to play a significant role in Google’s response to the competitive landscape, including the emergence of ChatGPT and other generative AI technologies. While Google pioneered numerous techniques that fueled the recent surge in AI advancements, the company adopted a cautious approach in developing and deploying AI-based products.
Since the introduction of ChatGPT, Google has swiftly launched its own chatbot named Bard and integrated generative AI into its search engine and various other products. To further accelerate AI research, the company combined DeepMind, led by Hassabis, with Google’s primary AI lab, Brain, in April, creating Google DeepMind. Hassabis emphasizes that this collaboration brings together two influential powerhouses that have been instrumental in recent AI progress. “If you look at where we are in AI, I would argue that 80 or 90 percent of the innovations come from one or the other,” says Hassabis. “Both organizations have made remarkable advancements over the past decade.”
Hassabis has firsthand experience navigating the AI gold rushes that disrupt tech giants, having previously triggered one himself. In 2014, Google acquired DeepMind after the company demonstrated impressive results with software utilizing reinforcement learning to master simple video games. In the following years, Google DeepMind showcased how this technique could achieve feats that were once considered uniquely human, often surpassing human-level performance. When AlphaGo defeated Go champion Lee Sedol in 2016, it astonished many AI experts who had believed it would take decades for machines to reach such proficiency in a game of such intricate complexity.
Embracing Fresh Perspectives
The process of training advanced language models, such as OpenAI’s GPT-4, involves feeding vast amounts of carefully curated text from various sources into a transformer-based machine learning system. This training enables the model to grasp patterns in the data and become proficient at predicting the subsequent letters and words, resulting in its impressive ability to generate text or code and answer questions.
However, refining the performance of language models like ChatGPT requires an additional step: reinforcement learning. By incorporating feedback from human evaluators on the model’s responses, researchers can fine-tune its performance. Google DeepMind’s extensive expertise in reinforcement learning positions them well to equip Gemini with unique capabilities.
Furthermore, Google DeepMind’s researchers may explore the integration of ideas from diverse AI domains to enhance large language model technology. With their involvement in fields such as robotics and neuroscience, Google DeepMind showcased an algorithm recently that can learn to manipulate various robot arms to perform tasks.
The incorporation of physical experiences, akin to how humans and animals learn, is considered a crucial element in advancing AI capabilities. The limitation of language models relying solely on textual information for understanding the world is recognized by some experts in the field. Therefore, Google DeepMind’s multidisciplinary approach and exploration of alternative learning methods hold promise for addressing this challenge.
A Clearer Path Ahead
Hassabis faces the challenging task of propelling Google’s AI endeavors forward while also addressing unknown and potentially serious risks. The swift advancements in language models have raised concerns among AI experts, including those involved in algorithm development, about potential malevolent uses and the challenges of control. Some insiders have even called for a temporary halt in the development of more powerful algorithms to prevent the creation of potentially dangerous technologies.
Hassabis recognizes the extraordinary benefits that AI can bring, particularly in scientific discoveries related to areas like health and climate. He emphasizes the importance of not halting the development of this technology, stating that if implemented correctly, AI could become the most advantageous tool humanity has ever had. He believes mandating a pause would be impractical, as enforcement would be extremely challenging. Instead, he advocates for a bold and courageous pursuit of AI’s potential.
However, Hassabis emphasizes that this does not imply a reckless rush in AI development. Google DeepMind has long been exploring the risks associated with AI, even before ChatGPT emerged. Shane Legg, one of the company’s co-founders, has been leading an “AI safety” group within DeepMind for years. Hassabis recently joined other prominent figures in AI by signing a statement warning about the potential risks of AI, which could rival those posed by nuclear war or pandemics.
Assessing and understanding the risks associated with increasingly capable AI models pose a significant challenge. Hassabis emphasizes the need for urgent research in areas such as evaluation tests to determine the capabilities and controllability of new AI models. In an effort to address concerns about access to the latest AI research being limited to large companies, DeepMind is considering making its systems more accessible to external scientists. Hassabis expresses his desire for academia to have early access to these cutting-edge models.
When it comes to concerns about AI becoming a major danger, Hassabis acknowledges that certainty is lacking. However, given the current pace of progress, he believes there is limited time to develop adequate safeguards. Hassabis is confident in the measures being incorporated into the Gemini series and sees them as effective solutions, instilling confidence that responsible AI development can be achieved.