Written by: Techub News Compilation

Introduction

In May 2025, during an internal dialogue organized by Sequoia Capital, Google's Chief Scientist and AI legend Jeff Dean engaged in an in-depth conversation with Sequoia partner and former Google engineering executive Bill Coughran. As a key figure who led the establishment of Google Brain, an early promoter of the Transformer architecture, and the initiator of the TPU (Tensor Processing Unit) project, Jeff Dean's vision spans AI research, infrastructure, and productization. This conversation took place amidst a rapid advancement of AI technology and intense industry competition, with his insights revealing not only Google's strategic direction but also providing a high-level perspective on the technological evolution of the entire industry.

Summary

• The implementation pathway for AI agents is clear, gradually expanding from virtual environments to physical robots, with capabilities rapidly growing through reinforcement learning and experience accumulation.
• The competition for top large models will be limited to a few players, but lighter, specialized smaller models will emerge through technologies like model distillation.
• Dedicated AI hardware (like TPU) is crucial for achieving efficient, low-power computing, and future inference hardware is expected to achieve tens of thousands of times energy efficiency improvements.
• AI is fundamentally changing the paradigm of scientific research, such as replacing traditional simulators with neural networks, resulting in a discovery speed increase of hundreds of thousands of times.
• Jeff Dean predicts that AI agents with the capabilities of a junior engineer may appear in "about a year," which requires them to master综合 skills beyond code generation, such as testing, debugging, and tool usage.

From Transformer to Gemini: The Path to Scaling AI

Jeff Dean began by reflecting on the long-term accumulation of the AI wave. He pointed out that although AI has only entered the public eye in the past three to four years, its foundations were established as early as 2012-2013. At that time, people began using "large" neural networks to solve visual, speech, and language problems, and this unified methodology replaced traditional handcrafted solutions, demonstrating the immense potential of machine learning.

Dean specifically mentioned a key work by Google in 2012: they trained a neural network 60 times larger than any model at the time, utilizing 16,000 CPU cores, and achieved remarkable results. This experiment "solidified our belief that scaling these methods is feasible." Since then, "larger models, more data, better results" have become the core driving forces of AI development over the past decade.

Today, model capabilities have vastly improved. Dean believes that while the current models cannot solve all problems, they can address more issues each year. This is attributed to holistic advancements in algorithm improvements (such as training larger models with the same computing power), hardware expansion (more computing units, higher interconnect bandwidth), and post-training techniques (like reinforcement learning). Multimodal capabilities—able to process and generate audio, video, images, text, and code—are also becoming highly practical.

Agents: From "A Bit Unreal" to Real Pathways

When asked about the industry's current enthusiasm for "agents," Dean candidly admitted that some concepts do seem "bubbly," but he sees a clear pathway for implementation. He believes that with appropriate training processes, agents will ultimately be able to accomplish many tasks that humans can do today in virtual computer environments. Currently, they can perform some tasks, but not the majority. The path to capability enhancement is clear: more reinforcement learning, more experiential learning for agents, and early products that, while not perfect, are already extremely practical.

Dean extends this logic to the physical world. The development of physical robot agents will follow a similar trajectory: starting with expensive robots capable of performing a few tasks (like tidying a room) in chaotic environments; through experience learning, costs will decrease, leading to a next-generation product that can perform thousands of tasks at only one-tenth of the cost. This positive cycle of cost and capability will accelerate the maturity of agent technology.

Model Landscape: Few Giants and a Blooming Variety

Regarding the competitive landscape of large language models, Dean's views are direct and profound: building the top-tier models requires massive investment, so the final players will not exceed 50, potentially only a few. This refers to the "foundation models" at the absolute forefront.

However, this does not mean the market will solidify. Once these powerful foundation models are available, techniques like model distillation (of which Dean is an early co-author) can be used to create a plethora of lightweight models suitable for more scenarios. He cited that this technology has likely aided models like DeepSeek. Thus, in the future, there will be many different forms of models focusing on different areas, but a few powerful general models will still hold significant importance.

The Hardware Race: Going Beyond Nvidia and Embracing Specialization

When discussing hardware, Dean's position is very clear. As the founder of Google's TPU project, he firmly believes that hardware dedicated to machine learning, especially accelerators focused on "lower precision linear algebra," is key for the future. These hardware need generational performance improvements and must be interconnected on a large scale through ultra-fast networks to extend model computation to as many computing devices as possible.

He reflected on the birth of the TPU: in 2013, they foresaw that inference would require immense computational power, thus launching the first-generation TPU; the second generation (TPUv2) focused on both inference and training. Regarding critiques of the current cloud TPU developer experience, Dean acknowledged room for improvement and highlighted Google's internal solution Pathways. This system provides a powerful abstraction layer, allowing researchers to drive thousands of chips for training with just one Python process, greatly simplifying the complexity of large-scale model training. He announced that Pathways is now available to Google Cloud customers.

Looking ahead, Dean believes there is vast potential for improving the energy efficiency of inference hardware. "How do we create inference hardware that is 10,000, 20,000, or even 50,000 times more energy efficient than today? This seems entirely possible." He mentioned simulation computing as a potential pathway, with advantages in extreme energy efficiency, but digital dedicated hardware also has much to offer. He is currently investing time in research in this field.

AI-Driven Scientific Revolution and Future Computing

Dean used this year's Nobel Prize in Physics and Chemistry (awarded to his colleagues Demis Hassabis and John Jumper for AlphaFold) to illustrate that AI is profoundly impacting numerous scientific fields. The core is that AI can learn from data, and many scientific discoveries are about establishing connections between things.

He described a transformative scenario: many disciplines possess extraordinarily expensive computational simulators (like weather forecasting, fluid dynamics, quantum chemistry). These simulators can be utilized as training data to train a neural network that approximates the simulator, potentially being 300,000 times faster than the original simulator. This will fundamentally change the way research is conducted—scientists could filter ten million molecules during lunch, a task that previously took a year and was unaffordable in computational resources. This "speed revolution" will greatly accelerate scientific discovery.

For the future of computational infrastructure, Dean envisions a heterogeneous and specialized landscape. The demands for computing have fundamentally changed over the past five to ten years: we need to run super-large neural networks with extremely high performance and very low power consumption. Training and inference represent two different loads, likely requiring different solutions. In the future, there will be hardware platforms specifically optimized for efficiently operating powerful models across mobile phones, robots, self-driving cars, and data centers. Furthermore, “computational volume” is becoming a new and powerful tuning knob—spending ten thousand times the computation on certain problems can lead to qualitative leaps, but clearly, we cannot afford such extravagance for every issue. This requires collaborative innovation across hardware, system software, models, and algorithms (like distillation).

Outlook: Virtual Engineers and Organic Learning Systems

In the Q&A session, a sharp question was raised: how far are we from having an AI that can work 24/7 at a level comparable to a junior engineer? Dean's surprising response was: "Not far off... I assert this could be achieved in the next year." He emphasized that this requires not just better code generation but also tool usage, autonomous planning, running tests, debugging performance issues—a series of comprehensive capabilities. Virtual engineers could accumulate "experience" through reading documentation and trial and error in virtual environments, providing a pathway for rapid improvement of capabilities.

Finally, Dean shared his vision for the future of model architecture. He has always been an advocate for sparsity and mixture of experts (MoE) models, believing they are closer to how biological brains efficiently utilize energy (with different regions activated for different tasks). Current implementations of MoE are still too structured; future models should have divergent pathways with computational costs differing by a thousand times, and should be capable of dynamically scaling, compressing, and reorganizing parameters, forming a more organic, continuously learning system. While the current paradigm is very effective, it's hard to shift immediately, but he believes this direction holds enormous potential.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。