After the release of Gemini3, the team spoke out: three major innovations and the scale law is still valid.

Author: Wuji, Special Correspondent for Tencent Technology

On November 19, Beijing time, following Google's release of the Gemini 3 series models, The New York Times' tech podcast “Hard Fork” aired a special episode, featuring hosts Kevin Roose and Casey Newton interviewing Google DeepMind CEO Demis Hassabis and Google Gemini team leader Josh Woodward.

This interview focuses on Google's latest flagship AI model, Gemini 3 (specifically the Pro version of the Gemini 3.0 series), which is widely regarded in the industry as a milestone release that marks Google's first return to a leading position in technology and products after the failure of Bard and the catch-up phases of Gemini 1.x and 2.x.

The two responsible parties elaborated on the breakthroughs of Gemini 3 in multi-step reasoning, code generation (especially in front-end and “atmosphere coding”), and dynamically generating interactive interfaces, emphasizing that Google has rapidly pushed its strongest models to billions of users across products like Search, Gmail, and Workspace, reshaping competitive barriers.

Core viewpoints of the interview:

• Gemini 3 is fully in line with the expected development trajectory, it will still take 5 to 10 years and 1 to 2 major research breakthroughs to achieve general artificial intelligence (AGI);
• Google's full-stack advantages in efficiency, cost, and distribution will allow it to succeed in any market environment;
• There is some AI bubble, but Google simultaneously has dual guarantees of short-term monetization and a long-term trillion-level new track.

Below is the condensed version of the interview content.

Rod: Casey, we are airing a special episode today on short notice, and the theme is the release of Gemini 3.

Newton: Yes, Kevin. This model has been long awaited in the Silicon Valley AI circle, and we are finally going to experience the real product ourselves.

Lodz: The reason we broke the usual Friday release rhythm to record this episode is mainly due to two reasons. First, we have the opportunity to conduct exclusive interviews with two core AI leaders from Google (DeepMind CEO Demis Hassabis and Gemini team Vice President Woodward).

Secondly, the release of Gemini 3 has sparked strong attention in the industry. We have heard from several labs' internal sources that this model has achieved breakthroughs in certain key areas, which could pose a substantial threat to competitors. Over the past two years, Google was seen as a follower, but the question now is: have they returned to the leading position?

Newton: Before formally starting the interview, let's briefly introduce the known information. Google held a closed-door briefing before the release, and the most striking new capabilities of Gemini 3 include: significantly enhanced coding and “atmospheric coding” abilities; as well as a brand new interactive interface generation feature.

It no longer just outputs text, but directly generates customized interactive interfaces for users. For example, when a user inquires about Van Gogh's biography, the model instantly creates a complete learning page that includes images, timelines, and interactive elements; or it can generate a mortgage calculator for properties over a million dollars. These features mark a transition from “answering questions” to “building experiences.”

Lodz: In all public benchmarks, Gemini 3 significantly outperformed Gemini 2.5 Pro. For example, in a cross-disciplinary doctoral-level problem set known as “Humanity's Last Exam,” the former scored only 21.6%, while the latter directly improved to 37.5%. Google's overall statement is: any task you can complete on ChatGPT, Claude, or other older versions of Gemini can be done better on Gemini 3.

Newton: They also showcased an early demonstration of the Gemini Agent: the model can deeply access users' email accounts, understand all email content, automatically categorize, draft replies, and even help users completely clear their inbox.

In addition, starting this week, Gemini 3 will be available on the Gemini App and in AI Mode on Google Search; American college students will receive one year of free access to the premium version. The keyword that Google repeatedly emphasizes is “Learn Anything,” which essentially positions Gemini as the ultimate personalized education tool.

Rodz: Demis, Josh, welcome to “Hard Fork.” Two years ago, Sundar Pichai compared Bard to “a modified Honda Civic,” racing on the track against stronger competitors. So, what kind of car is Gemini 3?

Hassabis: I hope it is much faster than a Honda Civic. I'm not very accustomed to using cars as metaphors, maybe more like a professional drag racer. It is not designed for daily driving or circuit racing; it has pure, immense power concentrated for a specific goal. It represents the perfect combination of our top research achievements and scalable computing power, with the aim of showcasing unparalleled instantaneous explosiveness in this race at the forefront of intelligence.

Rodz: That's interesting. What new things can Gemini 3 do on a concrete level compared to all previous AI models? Please provide us with some quantitative, practical examples.

Woodward: Three points stand out the most. First, in multi-step reasoning, it can think through more steps simultaneously, elevating its reliability to a whole new level. Previous models often “lost track” or generated hallucinations when reaching the 5th or 6th step of complex logical deductions, whereas Gemini 3 can reliably complete coherent reasoning tasks of 10 to 15 steps, such as complex tax planning, overall planning and booking for international travel, or conducting comprehensive debugging of a large system with millions of lines of code.

Secondly, it will generate a brand new interactive interface on a large scale for the first time. What users need is no longer simple text responses, but customized software components. For example, if you ask it: “Help me design a dashboard that can track all my portfolios,” it will generate an interactive, operable dashboard interface in real time, rather than a bunch of text describing how to create a dashboard.

Third, we have invested significant resources in coding capabilities, especially in 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 like Google Antigravity will also fully showcase this, with the model able to dynamically change the layout and functionality of the user interface based on context.

Newton: Many people believe that for the average user, the use case of “chat” has basically been solved. They can't even think of new problems that would make the answers from Gemini 3 qualitatively different from its predecessors. What do you think of this view?

Woodward: I understand this perspective. On the surface, the accuracy of basic Q&A has already reached a high level. But the real difference lies in reliability, integration, and the way information is presented. The answers from Gemini 3 will be more concise, more expressive, and the way information is presented will be easier to understand, which is a change that most people can perceive immediately.

Moreover, the model begins to deeply integrate with users' other data sources, such as interacting with other products within the Google ecosystem, truly transcending the simple question-and-answer model to become the user's “digital butler.” It can understand the context of your entire email, allowing it to not only answer questions when drafting replies but also to adjust the tone and content based on your past style and your relationship with the recipient.

Hassabis: I completely agree. Its reliability, style, and personality have been carefully honed, making it more concise and to the point. In scenarios like “atmospheric coding,” it has crossed the threshold of practicality. This represents a shift from “smart assistants” to “smart colleagues.” I personally plan to use it to get back into game programming during the Christmas holidays. It can now not only write functional code but also provide architectural suggestions early in the design phase.

Rodz: Demis, when you were interviewed by us in May this year, you assessed that AGI would still require 5 to 10 years and may need several significant breakthroughs. Has Gemini 3 changed this timeline?

Hassabis: Not at all. It aligns perfectly with the trajectory we set over the past two years. In fact, since the launch of the Gemini series, our pace of progress has been the fastest in the industry. Gemini 3 is stunning, but still within expectations.

There are still 1 to 2 key breakthroughs needed in consistency, depth of reasoning, memory mechanisms, and modeling of the physical world (such as the SIMA and Genie projects we are advancing) to achieve true general artificial intelligence. What we are currently doing is “System 1 thinking” (fast, intuitive), but to achieve AGI, we must unlock “System 2 thinking” (slow, deliberate, analytical).

In addition, the model needs to have a long-term, selective memory mechanism that can recall and apply specific interaction content from weeks or months ago, rather than being limited to a finite context window. Therefore, the judgment remains unchanged for 5 to 10 years.

Newton: There is a heated discussion in the industry about “AI companions” regarding model personality and user relationships. What kind of relationship do you hope users will establish with Gemini 3?

Woodward: This is a very sensitive but important issue. We position it as a “super tool” rather than an emotional companion, with the core value being to help users efficiently complete daily tasks and enhance productivity. Internally, we focus more on a new metric: how many tasks did we help you accomplish today? This is closer to the core value of the original Google Search—efficiency. We believe that positioning the model as an emotional companion poses safety risks and deviates from Google's core mission as a provider of information and tools.

Rodz: Have you given up the viral growth opportunity of “erotic partners”? Is this a major strategic mistake?

Woodward: No comment. Our security team has strict protocols and guidelines for this.

Lodz: In the past few weeks, competitors have clearly been on edge. Do you think Google is currently ahead in the AI race?

Hassabis: The current environment is the most intense competition in history. The only thing that truly matters is the speed of progress, and we are very satisfied with it. We have never lost our leading position in research, and now the product rollout has finally caught up. Our competitors excel in research, but they cannot replicate our advantages in scaling distribution and vertical integration.

We are injecting Gemini into products used by billions of users, including Maps, YouTube, Android, Search, Workspace, and more. This distribution network and terminal data feedback loop is an insurmountable moat. In addition, our full-stack advantage on customized TPU chips allows our training costs and efficiency to far exceed competitors relying on external GPU resources.

Newton: What do you think about the debate on scale laws and diminishing returns? Some believe that the larger the model, the lower the marginal benefit of performance improvement.

Hassabis: This is an ongoing debate. We are very pleased with the improvements of Gemini 3 compared to 2.5, which fully meets our expectations. The returns are not experiencing exponential growth as they did in the early days, but the incremental practicality and reliability it brings are still far above our marginal costs, making it worth our full investment. Until the 1 to 2 research breakthroughs needed to reach AGI arrive, continuously pushing performance through the largest foundational models remains the most effective strategy at this time. We believe that the law of scale still holds true.

Rodz: Are we in an AI bubble?

Hassabis: This is an overly binary question. There is indeed a bubble in certain areas (such as companies with billions in seed funding but no actual products, only talking concepts), where valuations are disproportionate to actual revenues. However, Google simultaneously has short-term monetization (search, Workspace, cloud TPU) and long-term trillion-dollar new tracks (robotics, gaming, drug discovery, materials science, etc.).

For example, our professional models like AlphaFold are creating real value in the field of drug discovery, which is a trillion-dollar market unrelated to consumer AI valuations. Whether there is a short-term bubble or not, we will come out on top: seizing opportunities during prosperity and being more resilient during contractions with our full-stack advantages and strong cash flow.

Newton: If it's Thanksgiving and someone wants to shift the political topic, what feature would you suggest they showcase with Gemini 3 to impress everyone?

Woodward: I don't know if it can save Thanksgiving, but it can bring laughter. Take out your phone for a selfie, then let Gemini 3 go wild editing the photos.

Our image model in Gemini is still the strongest worldwide. You can instantly turn a family photo into any funny scene, style, or era. It will definitely provoke laughter all around. Then, when you show how it can help you write a proper resignation letter or generate a customized holiday recipe calculator, they will naturally explore other new features.