jakiro
Author: Wuji, special editor of Tencent Technology
On November 19, Beijing time, after Google released the Gemini 3 series of models, the technology podcast “Hard Fork” owned by the New York Times released a special program in which hosts Kevin Roose and Casey Newton conducted an exclusive interview with Google DeepMind CEO Demis Hassabis and Google Gemini team leader Josh Woodward.
This interview focuses on Google’s latest flagship AI model Gemini 3 (actually the Pro version of the Gemini 3.0 series). This is Google’s first milestone release widely considered by the industry to regain its technology and product leadership after the failure of Bard and the catching-up phase of Gemini 1.x and 2.x.
The two leaders elaborated on Gemini 3’s breakthroughs in multi-step reasoning, code generation (especially front-end and “atmosphere coding”), and dynamic generation of interactive interfaces. They emphasized that Google has quickly pushed the strongest models to billions of user products such as search, Gmail, and Workspace, reshaping competitive barriers.
Key points from the interview:
-
Gemini 3 is fully in line with the expected development trajectory,General artificial intelligence (AGI) still needs 5 to 10 years and 1 to 2 major research breakthroughs;
-
Google’s full-stack advantages in efficiency, cost, and distribution enable it to win in any market environment;
-
The AI bubble partially exists, but Google has the dual guarantees of short-term monetization and long-term trillion-level new track.
The following is a condensed version of the interview
Rhodes: Kathy, we are temporarily adding a special episode today, the theme is the release of Gemini 3.
newton: Yes, Kevin.This model has been long-awaited in Silicon Valley AI circles, and we finally get to experience the real finished product with our hands.
Rhodes: There are two main reasons why we broke the regular Friday release rhythm and recorded this issue specifically.First, we got an exclusive interview opportunity with two core AI leaders at Google (DeepMind CEO Hassabis and Gemini team vice president Woodward).
Secondly, the release of Gemini 3 has attracted strong attention in the industry.We heard internal sources from multiple laboratories saying that this model has achieved breakthroughs in some key areas and may pose a substantial threat to competitors.Google has been viewed as a chaser over the past two years, and now the question is: Have they returned to the lead?
newton: Before officially entering the interview, we will briefly introduce the known information.Google held a closed-door briefing before the release. The most eye-catching new capabilities of Gemini 3 include: greatly improved coding and “ambience coding” capabilities; and a new interactive interface generation function.
It no longer just outputs text, but directly generates customized interactive interfaces for users.For example, when a user asks about Van Gogh’s life, the model will instantly generate a complete learning page containing pictures, timelines, and interactive elements; another example is generating a mortgage calculator for properties worth more than a million dollars.These features mark the jump from “answering questions” to “building experiences.”
Rhodes: Gemini 3 significantly outperforms Gemini 2.5 Pro in all public benchmarks.For example, on an interdisciplinary doctoral-level problem set called “Humanity’s Last Exam”, the former scored only 21.6%, while the latter directly improved to 37.5%.Google’s overall stance is that any task you can do on ChatGPT, Claude or other older versions of Gemini can be done better on Gemini 3.
newton: They also showed an early demonstration of Gemini Agent: the model can deeply access the user’s mailbox, understand the content of all emails, automatically classify, formulate replies, and even help users completely clear their inboxes.
In addition, Gemini 3 will be available in Gemini App and Google Search AI Mode starting this week; American college students will receive free access to the premium version for one year.The keyword that Google repeatedly emphasizes is “Learn Anything”, which actually positions Gemini as the ultimate personalized education tool.
Rhodes: Demis, Josh, welcome to Hard Fork.Two years ago, Sundar Pichai compared the Bard to “a modified Honda Civic” racing around a track against more powerful rivals.So, what kind of car is Gemini 3?
Hassabis: I wish it was much faster than the Honda Civic.I’m not used to using the car analogy, maybe more like a professional drag racer (Drag Racer).It’s not designed for daily driving or circuit racing, it’s pure power focused on a specific purpose.It represents the perfect combination of our top research results and large-scale computing power. The goal is to show unparalleled instantaneous explosive power in this competition at the forefront of intelligence.
Rhodes: This is interesting.Compared with all previous AI models, what new things can Gemini 3 do on a specific level?Please give us some quantitative, practical examples.
Woodward: There are three points that stand out the most.First, in multi-step reasoning, it can think about more steps at the same time, and we have raised its reliability to a whole new level.Previous generation models often “lost their train of thought” or hallucinate when reaching the 5th and 6th steps of complex logical derivation, while Gemini 3 can reliably complete 10 to 15 steps of coherent reasoning tasks, such as complex tax planning, overall planning and booking of cross-border travel, or comprehensive debugging of a huge system with millions of lines of code.
Secondly, it will generate a new interactive interface on a large scale for the first time.What users need is no longer simple text answers, but customized software components.For example, if you ask it: “Help me design a dashboard that can track all my investment portfolios,” it will generate an interactive and operational dashboard interface in real time, instead of a bunch of text describing how to make a dashboard.
Third, we invest heavily in coding capabilities, especially the front end and “ambient coding,” which means it can generate fully functional and beautifully designed user interface code based on natural language prompts.Upcoming new products such as Google Antigravity will also fully demonstrate this, with models able to dynamically change the layout and functionality of the user interface based on context.
newton: Many people believe that for ordinary users, the use case of “chat” has been basically solved.They couldn’t even think of any new questions that would make Gemini 3’s answers qualitatively different from those of its predecessor.What do you think of this perception?
Woodward: I understand this point of view.On the surface, the accuracy rate of basic question and answer is already very high.But the real difference is reliability, integration and presentation of information.Gemini 3’s answers will be more concise, more expressive, and the information presented in a more understandable way. This is a change that most people will immediately perceive.
More importantly, the model begins to be deeply integrated with other users’ data sources, such as linking with other products in the Google ecosystem, truly transcending the simple question and answer model and becoming the user’s “digital steward.”It understands the context of your entire email so that when drafting a reply, it not only answers the question but also tailors the tone and content based on your past style and your relationship with the recipient.
Hassabis: I totally agree.Its reliability, style and personality have been refined to make it simpler and more to the point.In scenarios such as “atmosphere coding”, the threshold of practicality has been crossed.This is a transformation from “intelligent assistant” to “intelligent colleague”.I personally plan to use it to get back into game programming over the Christmas holidays, and it can now not only write functional code, but also provide architectural advice at the early stages of design.
Rhodes: Demis, when you were interviewed by us in May this year, you judged that AGI will still take 5 to 10 years and may require several major breakthroughs.Does Gemini 3 change this timeline?
Hassabis: Not at all.It fits perfectly with the trajectory we’ve set over the past two years.In fact, since the launch of the Gemini series, our progress has been the fastest in the industry.Gemini 3 is stunning, but still expected.
Before we can achieve true general artificial intelligence, we still need to make one or two key breakthroughs in consistency, reasoning depth, memory mechanism, and physical world modeling (such as the SIMA and Genie projects we are advancing).What we are doing now is “System 1 thinking” (fast, intuitive), but to achieve AGI, we must unlock “System 2 thinking” (slow, thoughtful, analytical).
In addition, models need to have long-term, selective memory mechanisms that can recall and apply specific interactions from weeks or months ago, rather than being limited to a limited contextual window.Therefore, the judgment of 5 to 10 years remains unchanged.
newton: Regarding the relationship between model personality and users, the industry is hotly discussing “AI companions”.What kind of relationship do you want users to have with Gemini 3?
Woodward: This is a very sensitive but important issue.We position it as a “super tool” rather than an emotional companion. Its core value is to help users complete daily tasks efficiently and improve productivity.We are paying more attention to a new indicator internally: How many tasks have we completed for you today?This is closer to the core value of the original Google search – efficiency.We believe that pushing models toward the position of emotional companions is both a security risk and a departure from Google’s core mission as a provider of information and tools.
Rhodes：You gave up the viral growth opportunity of “Erotic Companion”. Is this a major strategic mistake?
Woodward: No comment.Our security team has strict norms and guidelines for this.
Rhodes: Competitors have been noticeably nervous over the past few weeks.Do you think Google is currently leading the AI race?
Hassabis: The current environment is the most competitive in history.The only thing that really matters is the rate of progress, and we’re very happy with that.We never lost our research leadership, now it’s just product launches that have finally caught up.Competitors are excellent in research, but they cannot replicate our advantages in scale distribution and vertical integration.
We are injecting Gemini into billions of user products such as Maps, YouTube, Android, search, and Workspace. This distribution network and terminal data feedback loop is an insurmountable moat.In addition, our full-stack advantage on customized TPU chips makes our training cost and efficiency far higher than competitors who rely on external GPU resources.
newton: What do you think about the debate between the law of scale and diminishing returns?Some people believe that the larger the model, the lower the marginal benefit of performance improvement.
Hassabis: This is an ongoing debate.We are very satisfied with the improvement of Gemini 3 compared to 2.5, which is fully in line with expectations.The returns are not as exponential as they were in the early days, but the incremental utility and reliability improvements it brings are still far higher than our marginal costs, and are still worth our full investment.Until the 1 to 2 research breakthroughs required to reach AGI arrive, continuing to drive performance through the largest-scale basic model is still the most effective strategy at present.We believe that the law of scale still holds true.
Rhodes: Are we in an AI bubble?
Hassabis: This is too binary a question.There is indeed a bubble in some areas (such as billion-dollar seed rounds with no real products, just concept companies) where valuations are disproportionate to actual revenue.But Google has both short-term monetization (search, Workspace, cloud TPU) and long-term trillion-level new tracks (robotics, games, drug discovery, materials science, etc.).
For example, specialized models like our AlphaFold are creating real value in drug discovery, a trillion-dollar market that has nothing to do with consumer AI valuations.Whether short-term bubbles exist or not, we will win: seize opportunities when booms occur and be more resilient with full-stack advantages and deep cash flow during contractions.
newton: If it was a Thanksgiving party and someone wanted to change the subject of politics, what feature would you suggest they use Gemini 3 to show off to wow the crowd?
Woodward: I don’t know if it can save Thanksgiving, but it can bring laughter.Take out your phone and take a selfie, then let Gemini 3 edit the photos like crazy.
Our image model in Gemini remains the strongest globally.You can instantly transform a family photo into any comical scene, style, or period setting.It will definitely make the whole audience laugh.Later, when you show how it can help you write a well-written resignation letter or generate a customized holiday recipe calculator, they will naturally explore other new features.
