Sergey Brin, Google's co-founder and tech industry titan, who disappeared from the public eye four years ago, has been swept back into the spotlight by the winds of artificial intelligence (AI)!
According to information obtained from Google insiders by The Wall Street Journal on July 20, Brin, having already achieved immense success, has been visiting Google's Mountain View, California office three to four days a week. Brin collaborates with researchers to develop Google's AI model, Gemini.
It was reported that Brin attended Google's internal AI meetings last year and has been increasingly involved. He is working closely with a group of researchers on Gemini, a model eagerly anticipated by the public. Brin discusses technical issues such as "loss curves" with Google researchers and is also involved in some personnel matters, such as recruiting high-demand researchers.
Brin's presence has potentially influenced Gemini's progress. The Wall Street Journal, quoting insiders, noted that Google executive Demis Hassabis, who is in charge of the Gemini project, recently told employees at a company-wide meeting that Gemini could be ready as soon as this year.
Will the long-awaited Gemini, rumored to combine AlphaGo and GPT-4-like models, finally arrive?
Google's Tight-lipped Preparation
On December 4, 2019, Brin officially announced his retirement, passing control of the company to Google CEO Sundar Pichai.
He perhaps didn't foresee the AI revolution brought about by the sudden popularity of ChatGPT just three years later, which forced Google into direct confrontations time and again. Microsoft's announcement in February this year about integrating ChatGPT into Bing and creating its own generative search experience with GPT-4 sent a chill down Google's spine.
Multiple Google employees have reported that since the AI boom began in November last year, Brin has frequently attended Google's AI-related meetings. His influence at Google is growing, underscoring the intense competition in AI. No tech giant wants to admit defeat, reflecting Brin's enthusiasm for AI.
Notably, the renowned "Google Brain" team originated from Google X, a division Brin led in his early years. In a 2018 shareholder letter, Brin wrote, "The power and potential of AI, as well as the potential to solve significant problems through computation, have never been greater. AI's development will be the most significant development I've seen in the computing field in my lifetime."
Media reports suggest that Pichai was excited to learn of Brin's return to Google's AI research and encouraged his contributions.
Google has always been seen as a top-tier AI research center, but with the intensification of AI competition, this tech giant has shifted to a defensive mode to fend off a growing number of agile AI competitors, protect its core search business and share price, and mitigate potential future security threats.
The Wall Street Journal mentioned that in April of this year, Alphabet, Google's parent company, merged two top-tier AI teams, Google Brain and DeepMind, under the leadership of DeepMind founder Hassabis. Google claimed that this move would accelerate its AI development.
Hassabis expressed confidence in the new combined team, which he said brings together two forces that have been vital to recent AI advancements.
"If you look at our position in the AI field, you'd believe that 80% or 90% of future innovations will come from one of these teams. In the past decade, both teams have made incredibly outstanding achievements."
Google's Gemini, powered by this top-tier AI team, may well be the company's last stand in the AI battlefield.
How Powerful is the Rumored Gemini?
According to the latest disclosure by Google DeepMind CEO Hassabis, the new Gemini model will integrate elements from AlphaGo and large language models.
Combining an AI system that created history by using reinforcement learning to beat the human Go champion and a top-ranking multimodal model, what sparks could a combined model produce?
The Wall Street Journal previously mentioned that Gemini would integrate the language functions of large models like AlphaGo and GPT-4, significantly enhancing the system's problem-solving and planning abilities.
Some AI experts believe that learning indirectly through text is a major limitation in language model development. AlphaGo's advantage can address this issue. In 2016, DeepMind's AI system AlphaGo beat world Go champion Lee Sedol 4 to 1, becoming the first robot in history to defeat a Go world champion.
AlphaGo, based on DeepMind's pioneering reinforcement learning technology, learns to handle tricky decision-making problems by trying repeatedly and receiving performance feedback. AlphaGo also uses the Monte Carlo tree search method to explore and remember possible actions on the board.
The next leap for language models might involve executing more tasks on computers. An earlier article mentioned that Gemini's greatest strength is its multimodal capability. It can understand and generate not just text and code, but also images. In contrast, ChatGPT is a pure text model that only understands and generates text.
Moreover, an important step in creating a language model with capabilities similar to ChatGPT is to use human feedback to improve its performance through reinforcement learning. DeepMind's deep experience in reinforcement learning could endow Gemini with new abilities.
At the Google I/O Developer Conference in May, Google mentioned that from the start, Gemini's goal was to be a multimodal, efficient integration tool and API. Google hinted then, "Although it's still early days, we have already seen multimodal capabilities in Gemini that were never seen in previous models. This has been very impressive."
Where will the next major leap in language models be? Gemini may well point the way for the next generation of language models. Accelerating the release of Gemini could give Google a greater chance of winning the AI race.
