In a riveting statement, Demis Hassabis, CEO of Google DeepMind, excitedly revealed plans for their next brainchild, Gemini, an algorithm set to outshine OpenAI’s ChatGPT. Hassabis hopes to integrate the groundbreaking techniques which propelled AlphaGo to victory against a Go champion back in 2016 into Gemini.
Not long ago, the AI world was abuzz when Google DeepMind’s AlphaGo, an artificial intelligence system, defeated a reigning champion at the intricate board game, Go. The event marked a historic moment, illustrating the potential of AI in problem-solving and strategic thinking. Now, Hassabis and his team aim to harness the capabilities that powered AlphaGo to develop a new AI system, Gemini, which is expected to outperform OpenAI’s ChatGPT.
Still under construction, Gemini represents a large language model closely related to GPT-4, which is the engine behind ChatGPT. However, the team intends to merge this technology with the game-changing techniques used in AlphaGo to enhance Gemini’s abilities, potentially enabling it to solve complex problems and execute planned actions. Hassabis describes Gemini as a confluence of AlphaGo’s strengths and the linguistic prowess of large models, with a sprinkle of unique innovations that promise to be intriguing.
DeepMind’s pioneering work in reinforcement learning, where software gradually learns to tackle challenging problems through trial and error and constructive feedback, was crucial to AlphaGo’s success. A similar methodology, known as tree search, was employed to envisage and remember possible moves during the game. The next major leap for language models like Gemini could involve more dynamic roles on the internet and in computer systems.
Development of Gemini, as Hassabis admits, will be a long journey that might incur tens or even hundreds of millions of dollars. It’s worth noting that OpenAI’s CEO, Sam Altman, previously stated that GPT-4’s creation exceeded a staggering $100 million.
Once Gemini is completed, it may serve as a significant piece in Google’s countermove against the increasing threats from ChatGPT and similar generative AI technologies. While Google was instrumental in many innovations that led to the recent surge of AI concepts, they’ve been deliberate and cautious in their product development and deployment based on these techniques.
Following the launch of ChatGPT, Google expedited its own chatbot, Bard, and integrated generative AI into its search engine and a host of other products. To catalyze AI research, Google merged the DeepMind team with its principal AI lab, Brain, to form Google DeepMind in April. This merger combines two titans that have been pivotal in recent AI advancements.
Hassabis, no stranger to the tumultuous race of AI technology, was a key figure in the frenzy after DeepMind was acquired by Google in 2014. The organization demonstrated impressive results through software employing reinforcement learning to master simple video games. In the subsequent years, DeepMind showcased how this technique could surpass human capabilities, stunning the AI community when AlphaGo triumphed over Go champion Lee Sedol.
Despite its ambitious goals, Gemini’s future remains uncertain, with Hassabis tasked with both accelerating Google’s AI initiatives and managing potential risks. The rapid advancement of language models has raised concerns about misuse or loss of control over the technology. Some experts have called for a temporary halt to the development of more powerful algorithms to avoid potential hazards.
Regardless of these apprehensions, Hassabis firmly believes in the tremendous potential benefits of AI in fields such as health and climate science. He argues against halting development, stating the impracticality of enforcing such a pause. According to Hassabis, if handled correctly, AI could be the most beneficial technology humanity has ever encountered.
This doesn’t mean AI development should proceed recklessly. DeepMind has been analyzing potential AI risks for years, with a specialized “AI safety” group led by company co-founder Shane Legg. Last month, Hassabis joined other prominent AI figures in signing a statement warning of possible catastrophic risks posed by AI, similar to those of nuclear warfare or pandemics.
Hassabis understands that determining the potential risks of increasingly capable AI is a paramount challenge at present. He emphasizes the need for urgent research in fields like evaluation tests to understand the capabilities and control of new AI models. Moreover, he suggests that DeepMind might provide more accessible systems for external scientists, which could mitigate concerns that experts outside major corporations are being excluded from cutting-edge AI research.
The question remains: how concerned should we be? Hassabis admits that the actual danger posed by AI remains uncertain. However, if advancements continue at their current speed, it is crucial to establish safeguards swiftly. “I can see the kinds of things we’re building into the Gemini series right now, and we have no reason to believe that they won’t work,” says Hassabis, signalling an exciting yet challenging future for AI development.
