Select Language:
On November 19, Google announced the release of its latest AI model, Gemini 3. In a joint interview, DeepMind CEO Demis Hassabis and Gemini team leader Josh Woodward discussed the significant upgrades and potential implications of this new development.
One of the most notable features of Gemini 3 is its ability to generate user interfaces (Generative UI). Unlike previous models that only answered questions, Gemini 3 can dynamically construct interactive pages tailored to user queries. For example, if someone asks about Vincent van Gogh’s biography, the model can immediately produce a comprehensive webpage complete with images and a timeline. It can also create custom tools, like a mortgage calculator for high-value properties. This shift from simple Q&A to building immersive experiences signals a new phase in large language model applications—moving toward real-time software creation.
Gemini 3’s reasoning capabilities have been considerably enhanced. According to Woodward, earlier models tended to lose track during complex reasoning steps around the fifth and sixth stages. In contrast, Gemini 3 can maintain coherent logic through ten to fifteen reasoning steps, even in difficult scenarios such as intricate tax planning or debugging extensive codebases. This improvement translates into more reliable performance in complex tasks.
Furthermore, Gemini 3 has demonstrated impressive advancements in interdisciplinary assessments, often called “Humanity’s Last Exam.” Its score on a challenging doctoral-level exam jumped from 21.6% with the previous Gemini 2.5 Pro to 37.5%, surpassing GPT-5.1’s 26.5%. In the SimpleQA Verified test, Gemini 3 Pro achieved a 72.1% accuracy rate—more than doubling the performance of GPT-5.1 and Claude Sonnet 4.5. Such results also reflect a significant reduction in the hallucination phenomena traditionally associated with AI models.
Perhaps the most groundbreaking progress lies in visual intelligence. In a specialized test aimed at understanding screens and user interface interactions, Gemini 3 Pro scored an impressive 72.7%. This score is nearly twenty times higher than GPT-5.1’s, indicating a profound leap in the model’s ability to recognize buttons, menus, and contextual elements—not merely objects within images. This development lays a solid foundation for AI agents capable of performing advanced automation tasks on computers, greatly enhancing the practicality of AI as digital work assistants.
In the realm of web development, Gemini 3 achieved a high Elo score of 1,487 on the WebDev Arena leaderboard. Accompanying its launch, Google introduced a new development platform called “Google Antigravity,” where Gemini 3 acts as an “intelligent agent” to call APIs, generate interfaces, and debug. Using natural language instructions, users can now receive fully functional, aesthetically designed code—streamlining the development process.
Google remains cautious about strategic positioning, explicitly avoiding the promotion of Gemini as an emotional companion. Instead, the focus is on boosting productivity through a “super tool” that prioritizes task completion over emotional engagement. Internal metrics evaluate how many tasks the model can help accomplish, echoing early Google search’s emphasis on efficiency. Early demonstrations of Gemini Agent showed it understanding email contexts, automatically organizing and replying to messages, and even helping clear out inbox clutter—progressing toward a state where AI functions as an independent, intelligent collaborator.
In a recent special broadcast, hosts Rod and Newton sat down with Hassabis and Woodward to explore the implications of Gemini 3’s launch. They highlighted that the AI community has long anticipated such advancements, especially given recent leaks suggesting that Gemini 3 might hold a competitive edge over major rivals. Google’s recent strategy includes a closed-door briefing emphasizing that Gemini 3’s capabilities significantly exceed those of older models like Gemini 2.5 Pro, especially in complex reasoning and interface generation.
Hassabis likened Gemini 3 to a specialized drag racing vehicle—designed for peak performance in specific tasks rather than everyday driving. He emphasized that it combines top-tier research and large-scale computational power to deliver explosive breakthroughs, particularly in targeted areas like multi-step reasoning, interface creation, and code generation.
Woodward pointed out that despite high-perceived performance in straightforward question-answering, the core differences with Gemini 3 lie in reliability, integration, and the presentation of information. The new model offers more concise, expressive responses and better leverages deep integration within Google’s ecosystem, including email, search, and other productivity tools. It can adapt responses based on user style and past interactions, effectively acting as a “digital butler.”
Hassabis reassured that Gemini 3’s development aligns with the original goal of reaching Artificial General Intelligence (AGI) within the next five to ten years. While the model is impressive, he underlined that true AGI requires critical breakthroughs in consistency, reasoning depth, memory, and understanding the physical world—projects currently underway, such as SIMA and Genie, aim to address these gaps.
Regarding user relationships, the company positions Gemini as a “super tool,” not an emotional partner. The primary measure of success is how effectively it improves user productivity. Wording and tone are optimized to be efficient and safe, avoiding unsafe or intimate associations, which Google believes could pose safety risks and distract from its core mission of providing information and tools.
When asked about the possibility that AI growth may be experiencing a bubble, Hassabis acknowledged the industry’s overvaluations in some sectors but emphasized Google’s diversified approach—encompassing short-term revenue streams like Search and Cloud, as well as groundbreaking long-term projects such as AlphaFold for drug discovery. Despite skepticism, he expressed confidence that Google’s strategic advantages—including extensive product integration, dedicated custom chips, and vast user data—are insurmountable for competitors.
In the conclusion, the hosts joked about showcasing Gemini 3’s fun capabilities at a hypothetical Thanksgiving gathering, such as editing a family photo or creating amusing scenarios, hinting at the model’s impressive image-generation abilities and humor potential. Overall, Google’s latest model marks a significant stride toward more intelligent, integrated, and reliable AI systems, with early signs indicating it could reshape the industry’s future trajectory.
ChatGPT
Perplexity
Gemini AI
Grok AI
