Abstract:
Google is close to releasing its latest flagship model, Gemini 4, Google DeepMind chief Koray Kavukcuoglu said on Wednesday, a long-awaited development for the tech giant after falling behind rivals Anthropic and OpenAI in the artificial intelligence model race.
Kavukcuoglu said at an event that Gemini 4 is in the early stages of a development phase known as "post-training," in which the underlying AI model will be refined to make it perform reliably. He said he expected the model to be released "well before the end of the year."
“Our intention is to roll out an early version of the post-training phase as quickly as possible because we’re seeing the results and we’re excited about it,” he said. Subsequently, Google will "continue to iterate rapidly."
Gemini 4 will be Google's first next-generation flagship model since the Gemini 3 series was launched late last year - although the company released the Gemini 3.1 update in February and introduced a series of smaller, lower-cost Flash models that were weaker in complex reasoning capabilities. Google CEO Sundar Pichai said at the company's annual I/O conference in May that a major update, Gemini 3.5 Pro, would be released in June, but the version was not released as scheduled.
Kavukcuoglu said that after releasing the Gemini 3 and 3.1 models, the company did not launch the Gemini 3.5 Pro, but "slowed down a little bit" to focus on launching the Flash series of models.
“The most important thing for us at the time was to learn as fast as possible,” he said. He did not specify whether Google has officially abandoned Gemini 3.5 Pro, but made it clear that the company's current focus is on releasing Gemini 4.
Competitor Anthropic launched a powerful new generation of its Mythos model line this spring and received an upgrade this month, while OpenAI began rolling out its equally capable GPT-6 model in early September. Both companies’ models outperform Google’s current state-of-the-art models.
When asked if he was worried that Google had fallen behind its peers, Kavukcuoglu said he had "full confidence" in his team.
"We have an extremely strong team and I have full confidence in this team," he said. "In my opinion, we will always be at the forefront of technology."
Kavukcuoglu said that Google is developing protection mechanisms and conducting security tests for Gemini 4; currently, engineers are using the model internally to run its Antigravity AI programming tool.
Kavukcuoglu took over Google's AI business in August and moved from chief technology officer to senior vice president of Google DeepMind. Previously, Demis Hassabis resigned as CEO of Google DeepMind and became chairman of the department and chief scientist of Google's parent company Alphabet. Kavukcuoglu is also the chief AI architect at Google.
Google has a history of starting behind in the AI model race and then catching up. Its Gemini 3 model represents a huge technological leap forward, prompting OpenAI to sound a “red alert” in an all-out effort to catch up to the search giant.
Kavukcuoglu said that unlike his peers in the artificial intelligence industry, he does not have a strong view on whether the industry has achieved AGI (artificial general intelligence).
“The discussion about whether we have achieved AGI is not really the important topic,” he said. “It is more important to discuss: Can we build a trustworthy agent?”
Kavukcuoglu said Google's progress on cutting-edge models will also help the tech giant boost its fast-growing hardware business, which produces AI chips called TPUs (tensor processing units). To step up competition with Nvidia, Google this year began selling TPUs directly to customers, rather than just providing access to the hardware through its cloud business. Kavukcuoglu pointed out that Google DeepMind's research and development work on new models provides important reference for hardware engineers and helps them plan the design direction of "the next two to three generations" of chips.
Comments