##谷歌DeepMind：下一个Transformer，AlphaGo与Gemini强强联合
**机器之心报道**
在与伦敦大学学院高级空间分析中心城市数学副教授汉娜·弗莱的最近一次对话中，谷歌DeepMind首席执行官兼联合创始人戴密斯·哈萨比斯透露了公司的一些规划，并就当前AI领域的一些问题发表了自己的看法。
哈萨比斯认为，尽管AI在短期内被过度炒作，但从长期来看，它却被低估了。他强调，判断AI领域哪些是炒作，哪些是可以实现的，除了做调研，还需要关注发表言论者的背景和技术水平。他指出，那些仅仅跟风的人，贡献好点子的概率很低。
2023年，谷歌将Google Brain和DeepMind合并成立了Google DeepMind，这一举措带来了许多创新机会。哈萨比斯表示，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。
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。哈萨比斯认为，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，目前许多模型都源于五、六年前的技术，因此仍然存在许多缺陷，例如产生幻觉、不擅长长期规划、无法主动完成复杂任务等。
为了解决这些问题，谷歌打算结合其在游戏智能体和大语言模型方面的专业知识，比如将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。
在谈到开源时，哈萨比斯表示，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。
哈萨比斯进一步谈到，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。
哈萨比斯还认为，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。
关于如何确保AGI能够使每个人都受益，哈萨比斯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。
此次采访让许多人感到振奋，因为哈萨比斯听起来更像是一个计算机科学家，而不是推销员。许多人认为，收购DeepMind并让他们自由发展是谷歌做出的最好的人工智能决策，希望谷歌能让他们继续自己的工作，尽可能不要打扰。
谷歌DeepMind的未来发展方向备受关注，其在AI领域将取得怎样的突破，值得期待。
英语如下：
##Hassabis: Google Aims to Build a “Super Transformer,” Combining AlphaGo and Gemini
**Keywords:** Hassabis, Transformer, Collaboration
**News Content:**
## Google DeepMind: The Next Transformer, AlphaGo and Gemini Join Forces
**Machine Intelligence Report**
In a recent conversation with HannahFry, Associate Professor of the Mathematics of Cities at the UCL Centre for Advanced Spatial Analysis, Demis Hassabis, CEO and co-founder of Google DeepMind, revealed some of the company’s plans and shared his views on current issues in the AI field.
Hassabis believes that while AI is currently overhyped in the short term, it is underestimated in the long run.He emphasizes that judging what is hype and what is achievable in the AI field requires not only research but also attention to the background and technical level of the speakers. He points out that those who simply follow the trend have a low probability of contributinggood ideas.
In 2023, Google merged Google Brain and DeepMind to form Google DeepMind, a move that brought numerous opportunities for innovation. Hassabis says their goal is to invent the next architecture that will push the frontiers of AI, just as Google Brain invented the Transformer architecture.
Existingacademic benchmarks have become saturated, unable to distinguish subtle differences between top models. Hassabis believes the AI field needs better benchmarks, particularly in areas like multimodal understanding, long-term memory, and reasoning abilities. He points out that many current models are based on technology from five or six years ago, hence they still havemany flaws, such as hallucination, poor long-term planning, and inability to proactively complete complex tasks.
To address these issues, Google intends to combine its expertise in game agents and large language models, such as integrating AlphaGo’s strengths in planning and decision-making with multimodal models like Gemini, to developsystems with stronger agent behavior.
Regarding open-sourcing, Hassabis says they have open-sourced many technologies, such as Transformer and AlphaFold. However, he believes cutting-edge models require more scrutiny and should be open-sourced one to two years after release, a model Google is currently following.
Hassabis further argues that open-sourcing is like walking through a one-way door, once released, it cannot be retracted. Therefore, caution is needed before open-sourcing.
Hassabis also believes that AI could make breakthroughs in complex mathematical problems, such as helping solve famous mathematical conjectures or excelling in international mathematics competitions. However, current AI systems are not yet capable of independently proposing new mathematical hypotheses or original theories. He believes a crucial test for AGI will be whether it can autonomously generate entirely new hypotheses and theories like general relativity.
On ensuring that AGI benefits everyone, Hassabis believes it’s impossible to incorporate all preferences into one system. However, a safe architecture can be built, and then people can decide, based on their own preferences, usage purposes, and deployment goals, what the AI system can and cannot be used for.
This interview has energized many, as Hassabis sounds more likea computer scientist than a salesman. Many believe that acquiring DeepMind and allowing them to develop freely is the best AI decision Google has made. They hope Google will let them continue their work and not interfere as much as possible.
The future direction of Google DeepMind is attracting much attention. It remains to be seen whatbreakthroughs they will achieve in the AI field.
【来源】https://www.jiqizhixin.com/articles/2024-08-20-5
Views: 10