##谷歌DeepMind：下一个Transformer，AlphaGo与Gemini强强联合
**机器之心报道**
谷歌DeepMind首席执行官兼联合创始人Demis Hassabis近日在接受采访时，透露了公司未来发展规划，并对当前人工智能领域的一些问题发表了自己的看法。
Hassabis表示，DeepMind和Google Brain的合并带来了很多创新机会，他们的目标是发明下一个能够推动AI前沿的架构，就像Google Brain发明了Transformer架构一样。他们希望将AlphaGo在规划和决策上的优势与Gemini等多模态模型结合，开发具备更强智能体行为的系统。
**AlphaGo与Gemini强强联合，打造更强大的AI系统**
Hassabis认为，现有的学术基准测试已经趋于饱和，无法区分顶尖模型之间的细微差异。他强调AI领域需要更好的基准测试，特别是在多模态理解、长期记忆和推理能力等方面。
他指出，现在很多模型都是从五、六年前发明的技术中产生的，因此这些模型仍然缺少很多东西，会产生幻觉、不擅长长期规划，无法主动完成复杂任务。为了解决这些问题，谷歌打算通过结合其在游戏智能体和大语言模型方面的专业知识，将AlphaGo在规划和决策上的优势与Gemini等多模态模型结合，开发具备更强智能体行为的系统。
**开源与前沿技术：谨慎与平衡**
在谈到开源时，Hassabis表示，他们已经开源了很多技术，如Transformer、AlphaFold。但他认为前沿模型需要经过更多的审核，在发布一到两年后才能开源，这种模式也是谷歌正在遵循的。
他解释说，开源的主要问题在于它就像是走过一扇单向门，一旦发布，就无法撤回。因此在开源之前需要非常谨慎。
**AI的未来：突破数学难题，创造全新理论**
Hassabis认为，AI可能会在一些复杂的数学问题上取得突破，例如帮助解决著名的数学猜想或在国际数学竞赛中表现出色。然而，目前的AI系统还无法自行提出新的数学假设或原创性理论。他认为，AGI的一个重要测试标准将是其是否能够自主生成像广义相对论那样的全新假设和理论。
**确保AGI的公平与安全**
关于如何确保AGI能够使每个人都受益，Hassabis认为不可能将所有偏好都包含在一个系统中，但是可以构建一套安全的架构，然后人们根据自己的偏好、使用目的、部署目的，决定AI系统可以用来做什么，不能用来做什么。
Hassabis的观点引发了业界广泛关注。他强调了AI发展中需要解决的挑战，也展现了谷歌DeepMind在未来发展方向上的思考和规划。相信随着技术的不断进步，AI将为人类社会带来更多益处，并推动人类文明迈向新的高度。
英语如下：
##Hassabis: Google Aims to Build “Super Transformer” with AlphaGoand Gemini Combined
**Keywords:** Hassabis, Transformer, Combined
**Content:**
**Google DeepMind: The Next Transformer, AlphaGo and Gemini Join Forces**
**Machine Intelligence Report**
Demis Hassabis,CEO and co-founder of Google DeepMind, recently revealed the company’s future development plans and shared his views on current issues in the field of artificial intelligenceduring an interview.
Hassabis stated that the merger of DeepMind and Google Brain has brought numerous opportunities for innovation. Their goal is to invent the next architecture that will push the boundaries of AI, much like Google Brain invented the Transformerarchitecture. They aim to combine the advantages of AlphaGo in planning and decision-making with multi-modal models like Gemini to develop systems with stronger agent behavior.
**AlphaGo and Gemini Join Forces to Create a More Powerful AI System**
Hassabis believes that existing academic benchmarks have become saturated and cannot distinguish subtle differences between top models. He emphasized the need for better benchmarks in the AI field, especially in areas such as multi-modal understanding, long-term memory, and reasoning abilities.
He pointed out that many current models are derived from technologiesinvented five or six years ago, which means they still lack many things. They can produce hallucinations, are not good at long-term planning, and cannot proactively complete complex tasks. To address these issues, Google intends to combine its expertise in game agents and large language models, integrating AlphaGo’s strengths in planning anddecision-making with multi-modal models like Gemini to develop systems with stronger agent behavior.
**Open Source and Cutting-Edge Technology: Caution and Balance**
Regarding open source, Hassabis stated that they have open-sourced many technologies, such as Transformer and AlphaFold. However, he believes that cutting-edge models require more scrutiny and should be open-sourced one to two years after release, a model that Google is currently following.
He explained that the main problem with open source is that it’s like walking through a one-way door. Once released, it cannot be taken back. Therefore, extreme caution isnecessary before open-sourcing.
**The Future of AI: Breaking Mathematical Barriers, Creating New Theories**
Hassabis believes that AI could make breakthroughs in complex mathematical problems, such as helping to solve famous mathematical conjectures or excelling in international mathematics competitions. However, current AI systems are not yet capable of independentlyproposing new mathematical hypotheses or original theories. He believes that a crucial test criterion for AGI will be whether it can autonomously generate entirely new hypotheses and theories like general relativity.
**Ensuring Fairness and Safety of AGI**
Regarding how to ensure that AGI benefits everyone, Hassabis believes that it’s impossible to include all preferences in one system. However, a safe architecture can be built, and then people can decide, based on their preferences, usage purposes, and deployment objectives, what the AI system can and cannot be used for.
Hassabis’s views have sparked widespread attention in the industry. Hehighlighted the challenges that need to be addressed in AI development and showcased Google DeepMind’s thinking and planning for future development directions. It is believed that with continuous technological advancements, AI will bring more benefits to human society and propel human civilization to new heights.
【来源】https://www.jiqizhixin.com/articles/2024-08-20-5
Views: 1