On November 18, 2025, Demis Hassabis, chief executive of Google DeepMind, and Josh Woodward, vice president of the Google Gemini team, joined Kevin Roose and Casey Newton on a special episode of The New York Times podcast "Hard Fork" to discuss the launch and capabilities of Gemini 3. The conversation was recorded just ahead of Google’s public rollout of the Gemini 3 model and was presented as an early look into what the company says are material improvements over previous Gemini releases. (podscripts.co)
The interview combined a product-briefing rundown with extended answers from Hassabis and Woodward about the model’s strengths, deployment plans and safety posture. Below are the core statements and descriptions the executives made during that conversation, organized by topic and quoted directly where appropriate.
What Gemini 3 does: reasoning, interfaces, and coding
Hassabis and Woodward described three headline areas where Gemini 3 advances beyond prior models. First, they said the model shows clear improvements in multi-step reasoning and the ability to "think many steps" without losing its train of thought. As Hassabis put it about reasoning improvements, this one's way better at that
. (podscripts.co)
Second, the team emphasized new generative interfaces: when users ask a question, Gemini 3 can assemble custom, interactive interfaces rather than only returning plain text or images. The hosts described examples shown at a Google briefing—an interactive tutorial about Vincent Van Gogh and a mortgage calculator that the model generated on the fly. These interface-generation capabilities were presented as a major product differentiator. (podscripts.co)
Third, both executives highlighted substantial investment in coding capabilities. They said Gemini 3 improves coding performance—particularly front-end work and so-called "vibe coding"—and that the model is already useful for practical programming projects. Woodward noted it has reached a point where it feels incredibly useful and capable on front end
. (podscripts.co)
Benchmarks and user-facing improvements
The hosts and guests discussed multiple benchmarks Google shared at the briefing. They said Gemini 3 Pro substantially outperformed Gemini 2.5 Pro across a range of academic and capability tests; for example, the hosts reported improvements on very hard interdisciplinary exams used as proxies for advanced reasoning. Hassabis and Woodward stressed that benchmarks are proxies for product quality and must be translated into user experiences that matter. (podscripts.co)
Agent features: inbox triage and task completion
One capability the hosts singled out was what Google described as the "Gemini agent." The conversation described a model that can examine a user’s inbox, understand message contents, propose replies and organize messages—functions meant to help users get their email under control. The hosts called that a feature they were eager to try, and the Gemini team confirmed agent-style integrations are a priority. (podscripts.co)
Rollout, product placements and efficiency
Hassabis and Woodward explained the initial rollout was phased: Gemini 3 would appear first in the Gemini app, in an "AI mode" tab adjacent to Google Search, and in developer-facing products, with broader Workspace and Gmail integrations to follow at unspecified dates. They emphasized engineering work on model efficiency—distillation techniques and cost-performance tradeoffs—so the company can serve AI features to very large user bases without unsustainable server costs. The team said different members of the model family will target different cost/performance points. (podscripts.co)
Personality and user relationship
On the model’s persona, the team described internal work on style and persona to make Gemini 3 more succinct and helpful. They framed the app as a productivity tool—"a superpower in your toolbox"—and said they are interested in metrics such as how many tasks the model helps a user complete in a day. The emphasis was on utility for writing, research, creation and task completion rather than on companion-style relationships. (podscripts.co)
AGI timelines and remaining research needs
Asked whether Gemini 3 changed Hassabis’s prior estimate about timelines to artificial general intelligence, he said the model is right on track
with expectations and the trajectory of Gemini development, but that one or two more research breakthroughs will likely still be needed to achieve the consistency expected of a general intelligence. He reiterated a continued estimate of roughly five to ten years to AGI, contingent on further advances in reasoning, memory and world-modeling research. (podscripts.co)
Safety testing and new risks
Hassabis described Gemini 3 as the team's most thoroughly tested model so far. He emphasized internal safety testing, work with safety institutes and external testers, and additional attention to areas where improved tool use and function-calling can enable both beneficial and risky behaviors (for example, cyber uses). The team said those capability improvements make rigorous safety checks more important than ever. (podscripts.co)
Business strategy, distribution and the question of a bubble
Hassabis framed Google’s advantage as technical and organizational: model efficiency, integration across products used by billions, cloud/TPU infrastructure and the ability to deploy models into widely used services. On whether the industry is in a bubble, he said parts of the AI sector may show bubble-like activity—particularly very large seed valuations with little substance—but argued Alphabet is well positioned to win in either environment because it combines near-term monetization opportunities with long-term platform investments. (podscripts.co)
Student access and product examples
The team noted a separate announcement that U.S. college students would receive a year of free access to a paid Gemini tier, a move the hosts described as strategically aimed at early adoption. For consumer-facing examples the team suggested easy, demonstrable tasks people might try on holiday gatherings—image editing and imagery features remain a popular showcase for the new model. (podscripts.co)
References and further viewing:
Hard Fork — "Google's Gemini 3 Is Here: A Special Early Look" (episode page and transcript). (podscripts.co)
Hard Fork episode on iHeart. (iheart.com)
Explore more exclusive insights at nextfin.ai.
