Summary in English:
This article delves into the development logic and industry trends of the Gemini 3 model through an interview with Sébastien Bubeck, the pre-training lead for the Gemini 3 project at Google. Bubeck emphasized that the success of Gemini 3 was not due to a single technological breakthrough but the result of team collaboration. Hundreds of people continuously optimized the model, data, and infrastructure across multiple dimensions, achieving cross-generational progress. He also pointed out that the industry is shifting from a “data abundance” to a “data scarcity” model, requiring research to focus more on architectural innovation and data efficiency rather than relying solely on scaling. Additionally, Bubeck shared DeepMind’s practices in evaluation systems, alignment issues, and continual learning, and reflected on his personal career trajectory, highlighting the critical role of teamwork and research quality in AI development.

Key Points:

  • Gemini 3’s Success Logic
    • Model breakthroughs stem from team collaboration rather than singular innovations, achieved through incremental improvements across multiple dimensions by hundreds of people.
    • The pre-training team consists of about 200 people, coordinating complex tasks such as data curation, architectural optimization, and evaluation system refinement.
  • Industry Trends and Research Direction Shifts
    • The industry is transitioning from “data abundance” to “data scarcity,” necessitating research to enhance intelligence through architectural optimization with limited data, rather than relying solely on scaling.
    • Projects like Chinchilla and Retro validated that data expansion should outpace model scaling, and architectural innovations (e.g., retrieval-based learning) can enhance performance.
  • Research Quality and Team Collaboration
    • Research must prioritize integration, avoiding isolated progress (e.g., a modification that increases usage difficulty, even if improving a single metric, may slow overall progress).
    • Balance short-term goals (addressing current performance bottlenecks) with long-term exploration (e.g., long-context processing, attention mechanisms), with DeepMind focusing more on long-term value.
  • Evaluation and Alignment Challenges
    • Evaluations must predict large model performance and reflect subsequent training outcomes; DeepMind relies on internally built evaluation systems to avoid external benchmark contamination.
    • Alignment issues require models to encounter harmful information to identify and avoid generating it, rather than directly excluding it.
  • Technical Details and Future Directions
    • Gemini 3 adopts a native multimodal architecture, processing image, text, and other multimodal data, albeit at high cost but with significant efficiency gains.
    • Long-context processing, continual learning (e.g., updating models via search tools), and agent interaction capabilities (e.g., screen understanding) are future priorities.
  • Personal Growth and Career Development
    • Bubeck began with programming interests, pursued education in multiple countries, joined the University of Cambridge, and later joined DeepMind for large model research, gradually becoming the pre-training team lead.
    • He emphasized the importance of teamwork and cross-domain integration skills for AI research and expressed optimism about the AI industry’s future.

Reference Documents & Links Mentioned:

  1. Gemini 3 Project (core discussion topic, no external links).
  2. Chinchilla Project (DeepMind’s model training research).
  3. Retro Project (architectural innovation, improving model performance via retrieval).
  4. DeepThink Model (Gemini 3’s released thinking model).
  5. Gravity Project (agent execution and visual perception research).
  6. Gopher Project (DeepMind’s early large language model research).

(Note: No specific external links were mentioned in the text; only project names are listed as references.)

Translation

Summary in Chinese:
本文通过访谈Google Gemini 3项目预训练负责人塞巴斯蒂安·布尔若，深入探讨了Gemini 3模型的开发逻辑与行业趋势。布尔若强调，Gemini 3的成功并非源于单一技术突破，而是团队协作的成果，通过数百人对模型、数据、基础设施等多维度的持续优化实现跨代级进步。同时，他指出行业正从“数据无限”转向“数据有限”模式，研究需更注重架构创新与数据效率，而非单纯依赖规模扩展。此外，布尔若分享了DeepMind在评估体系、对齐问题、持续学习等领域的实践，并回顾了其个人职业发展轨迹，强调团队协作与研究品味对AI发展的关键作用。

Key Points:

  • Gemini 3的成功逻辑
    • 模型突破源于团队协作而非单一突破，通过数百人对模型、数据、基础设施等多维度的微小改进积累实现。
    • 预训练团队规模约200人，需协调复杂任务如数据筛选、架构优化、评估体系完善等。
  • 行业趋势与研究方向转变
    • 行业从“数据无限”转向“数据有限”，研究需在有限数据中通过架构优化提升智能，而非单纯依赖规模扩展。
    • 项目如Chinchilla和Retro验证了数据扩展速度应快于模型规模扩展，且架构创新（如检索式学习）能提升性能。
  • 研究品味与团队协作
    • 研究需注重整合性，避免孤立推进（如某改进若增加使用难度，即使提升单一指标也可能拖慢整体进度）。
    • 平衡短期目标（解决当前性能瓶颈）与长期探索（如长上下文处理、注意力机制等），DeepMind更关注长期价值。
  • 评估与对齐挑战
    • 评估需预测大模型性能并反映后续训练表现，DeepMind依赖内部构建的评估体系以避免外部基准污染。
    • 对齐问题需模型接触不良信息以识别并避免生成，而非直接排除。
  • 技术细节与未来方向
    • Gemini 3采用原生多模态架构，处理图像、文本等多模态数据，虽成本高但效率提升显著。
    • 长上下文处理、持续学习（如通过搜索工具更新模型）和Agent交互能力（如屏幕理解）是未来重点。
  • 个人成长与职业发展
    • 布尔若从编程兴趣起步，经历多国教育后进入剑桥大学，后加入DeepMind参与大模型研究，逐步成长为预训练团队负责人。
    • 他强调团队协作与跨领域整合能力对AI研究的重要性，并对AI行业未来充满期待。

Reference Documents & Links Mentioned:

  1. Gemini 3项目（核心讨论对象，无外部链接）。
  2. Chinchilla项目（DeepMind的模型训练研究）。
  3. Retro项目（架构创新，通过检索提升模型性能）。
  4. DeepThink模型（Gemini 3发布的思考模型）。
  5. Gravity项目（Agent执行与视觉感知相关研究）。
  6. Gopher项目（DeepMind早期大语言模型研究）。

（注：文中未提及具体外部链接，仅列出项目名称作为参考。）

Reference:

https://www.youtube.com/watch?v=cNGDAqFXvew

<
Previous Post
Sam Altman: OpenAI 2026 townhall
>
Next Post
Boris Cherny: Claude Code