All-In Podcast - Sergey Brin, Google Co-Founder | All-In Live from Miami
发布时间:2025-05-20 15:08:17
原节目
以下是对视频文字稿的总结,重点关注讨论的关键话题:
对话内容主要围绕着谢尔盖·布林在半退休状态后,重新回归谷歌,特别是参与人工智能领域的工作。他强调了人工智能发展的变革性和指数级增长的特性,将其比作互联网的早期阶段,但强调其发展速度更快、影响更深远。与互联网的发展相比,他觉得人工智能的创新速度“令人震惊”。
布林描述了他回归编码,对系统进行小修改并进行实验,以更深入地了解人工智能的各个方面。他对预训练和后训练(特别是思维模型)都表示兴奋,并承认已经取得了巨大的进步。
讨论深入探讨了提示工程和人工智能深度研究的力量。布林解释说,人工智能的超能力在于它能够以人类无法比拟的量和速度执行任务。他用处理数千个搜索结果并进行后续搜索的例子来说明这一点,这对于个人来说在合理的时间范围内是无法实现的。人工智能进行深度研究的能力,比如讨论中涉及的计算每英里死亡率(F1)的例子,展示了人工智能超越初始数据并构建理论的能力,这类似于通常分配给本科生完成的任务。
对话还探讨了人工智能对教育和未来工作的影响。布林承认,人工智能在某些认知任务(如数学和编码)方面已经超越了人类,并对传统教育路径(如大学)的相关性提出了质疑。人工智能快速进行高级计算的能力表明,大学对家长来说可能不再像过去那样重要。布林强调了社交智能和心理韧性的重要性,认为在先进人工智能时代,这些技能可能比死记硬背的知识更有价值。
布林分享了他对机器人技术和硬件的看法,借鉴了谷歌之前与机器人公司合作的经验。他对人形机器人表示怀疑,认为人工智能可以在不同情况下进行调整,而不一定需要模仿人类的形态。他认为,人工智能可以利用模拟和真实世界的数据进行学习,并在各种环境中有效地运行,从而可能使人形形态变得不必要。
对话涉及了人工智能对编程的影响。布林幽默地讲述了谷歌内部关于使用Gemini进行编码的争议,突出了最初禁止在编码任务中使用人工智能的限制。他提倡广泛采用人工智能工具来提高开发人员的生产力,并认为人工智能可以识别有前途的员工。
讨论转向了基础模型的问题。布林认为,存在一种趋同的趋势,即通用模型变得更强大和更通用。虽然专用模型可能对特定任务或研究目的有用,但他认为,长期趋势将是更少、更强大的通用模型。
对话讨论了开源与闭源人工智能。布林承认了开源模型取得的进展,特别是DeepSeek发布的模型,并表示谷歌同时追求开源和专有模型。
对话讨论了人工智能时代的人机交互,布林回顾了搜索框的演变。他谈到了增强现实眼镜的潜力。他坦言,谷歌眼镜发布得太早了,技术还没有准备好。他指出,电池续航问题是需要解决的问题。
布林还分享了一个有趣的轶事,讲述了他在内部聊天中使用人工智能来总结对话、分配任务,甚至识别潜在的晋升候选人。他强调了人工智能能够检测到人类管理者可能忽略的贡献和见解,从而展示了人工智能在增强管理和决策过程方面的潜力。
提到了无限上下文窗口和具有准无限上下文的Gemini构建的用例。布林说,这些领域可能存在内部发展,但他表示总是有发展,问题在于它们的工作效果如何。
在硬件方面,布林指出,谷歌为其Gemini使用了自己的TPU,尽管他们支持Nvidia。他说抽象层还不可用,还有很多事情需要解决。
布林建议采用一种响应时间机制,使语音交互真正值得进行,并且速度与去年相比大幅提高。
Here's a summarization of the video transcript, focusing on the key topics discussed:
The conversation features Sergey Brin, discussing his re-engagement with Google, particularly within the realm of AI, after a period of semi-retirement. He emphasizes the transformative and exponential nature of AI development, likening it to the early days of the web but highlighting its more rapid and profound evolution. He finds the pace of innovation "astonishing" compared to the web's development.
Brin describes his return to coding, contributing minor changes to the system and experimenting to gain deeper insights into various aspects of AI. He expresses excitement about both pre-training and post-training (particularly with thinking models), acknowledging the huge advancements that have come.
The discussion delves into the power of prompt engineering and deep research in AI. Brin explains that AI's superpower lies in its ability to perform tasks at a volume and speed that humans cannot match. He illustrates this with the example of processing thousands of search results and conducting follow-up searches, which would be impossible for an individual to achieve within a reasonable timeframe. The ability to perform deep research, like the discussed example involving calculating F1 death rates per mile, showcases the ability of the AI to go beyond initial data and construct theories, mirroring tasks typically assigned to undergraduate students.
The implications of AI on education and the future of work are also explored. Brin acknowledges that AI is already surpassing humans in certain cognitive tasks, such as math and coding, and raises questions about the relevance of traditional education paths, like college. The ability of the AI to quickly perform advanced calculations suggests college may not be as high of a priority as it has been to parents in the past. Brin emphasizes the importance of social intelligence and psychological resilience, suggesting that these skills may be more valuable than rote knowledge in an era of advanced AI.
Brin shares his thoughts on robotics and hardware, drawing from Google's previous experiences with robotics companies. He expresses skepticism about humanoid robots, arguing that AI can adapt to different situations without necessarily mimicking the human form factor. He believes that AI can leverage simulations and real-world data to learn and operate effectively in various environments, potentially rendering the humanoid form unnecessary.
The conversation touches upon the impact of AI on programming. Brin humorously recounts a dispute within Google regarding the use of Gemini for coding, highlighting the initial restriction against using AI in coding tasks. He advocates for the widespread adoption of AI tools to enhance developer productivity and suggests that AI could identify promising employees.
The discussion shifts to the question of foundational models. Brin suggests a trend towards convergence, with general models becoming more capable and versatile. While specialized models may be useful for specific tasks or research purposes, he believes that the long-term trend will be towards fewer, more powerful general models.
The topic of open source versus closed source AI is addressed. Brin acknowledges the progress made by open-source models, particularly those released by DeepSeek, and states that Google pursues both open-source and proprietary models.
Human-computer interaction in the age of AI is discussed, with Brin reflecting on the evolution of the search box. He addresses the potential for augmented reality glasses. He confesses that Google Glass was released too early and the technology was not ready. He notes that battery life issues are something that needs to be addressed.
Brin also makes a humorous anecdote about using AI in internal chats to summarize conversations, assign tasks, and even identify potential candidates for promotion. He highlights the AI's ability to detect contributions and insights that might be overlooked by human managers, showcasing the potential for AI to enhance management and decision-making processes.
The use cases for infinite context windows and a Gemini build with quasi-infinite context is mentioned. Brin says there may be internal developments in those fields but he says there are always developments and the question is how well do they work.
On the topic of hardware Brin notes that Google used its own TPUs for Gemini, although they support Nvidia. He says the abstraction is not available yet and there are a lot of things that have to be addressed.
Brin suggests a response time thing where it is actually worth doing voice and that the speed has increased drastically from just last year.