Insights from Google AI Executive Omar Shams on the Future of AI

On July 3rd, Omar Shams, head of Google's AI division, shared insights on the cutting edge of the AI industry in an interview with Steve Hsu, a professor of computational mathematics at Michigan State University.

Key Points:

The scarcity and value of theoretical physics talent. Omar Shams explained the importance of physical intuition in AI research, drawing from his experience transitioning from string theory research to AI entrepreneurship. The optimization process of loss functions is similar to a sphere rolling on an energy manifold, and there is a direct correspondence between KL divergence in information theory and Hamiltonians in physics.

While chips are important, energy supply is the key constraint for the long-term development of AI. He believes that the expansion of the U.S. power grid is slow, while China's annual increase in power capacity has surpassed the total of the UK and France, highlighting the growing energy gap in the U.S.-China AI infrastructure competition. He even proposed the idea of deploying solar power stations on the moon or in space to support AI computing power.

There are no secrets in the AI field, but there is invaluable tacit knowledge. Through social interactions and job changes, companies' research directions and methodologies have become largely similar. However, the core competency of top AI talent lies in their intuition, experience, and judgment in building and training large-scale models, facing countless variables and subtle issues.

Restructuring talent needs in the software development industry. AI tools are expected to lead to a 30% unemployment rate among programmers within two years, with junior engineering positions at risk of being replaced by AI agents, fundamentally changing companies' hiring logic.

Commercial breakthroughs in AI agent technology. AI agent technology is moving from proof-of-concept to practical application. In the software development field, AI agents can now autonomously complete complex multi-step tasks. Similar breakthroughs are occurring in the legal services field, with companies like Harvey generating significant revenue.

Highlights:

Beyond chips, the U.S.-China AI race is about energy.

Host: Talking about AI development, the competition is very intense nowadays. The stories of DeepMind and OpenAI are examples. What do you think is the biggest bottleneck in the current AI race?

Omar Shams: There are two. The first is the well-known chip issue. The second, increasingly prominent, is energy. Power supply for data centers is becoming a hard constraint.

Host: When I talk about the U.S.-China AI race, there are two issues. One is the showdown between Nvidia and Huawei, but the other is, how will you power these data centers? In the U.S., increasing power supply to the grid is very difficult, while China's annual increase in power generation is equivalent to the total of the UK or France.

Omar Shams: And the U.S. is every seven years.

Host: Exactly. And they are now twice our size. So, if power ultimately becomes the fundamental element transforming into intelligence, how will you compete with them?

Omar Shams: I don't want to go too far, but I wrote a speculative article called "The Moon Should Be a Computer." If Earth's energy consumption increases by two orders of magnitude, it will have a thermal effect on the atmosphere. The real issue is the slow increase in the supply of base load power to the U.S. grid, possibly due to regulatory constraints or lack of construction capacity.

Omar Shams: So I speculated, could this problem be solved in space or even on the moon? This idea came up while chatting with friends. Although crazy, some smart people think it's a good idea. I also found that Eric Schmidt is now the CEO of Relativity Space, and he wants to put computers in space, partly because of energy limitations on Earth.

Host: Is the energy for his space project from solar panels or orbital nuclear reactors?

Omar Shams: I guess solar, because space nuclear energy would violate too many treaties, and if a rocket launch fails, it would be a dirty bomb. I've calculated that to obtain one gigawatt of power, you might need one square kilometer or more of solar panels.

Host: Putting so much stuff into orbit requires huge carrying capacity, and it can't be in low orbit, or the debris would be dangerous after disintegration.

Omar Shams: Yes, it must be at the Lagrange point. Fortunately, there's enough space in space. (The so-called Lagrange point is a special position within the solar system or between any two celestial bodies where an object can maintain a stable orbit relative to these two bodies.)

Talent Competition in the AI Race

Host: This leads to an interesting question: since everyone's technical routes and information are relatively transparent, why is Meta willing to pay a fortune to poach a top talent? If there are "no secrets," what is this money buying?

Omar Shams: The value of a top talent lies in their precise judgment based on deep experience, which can save a lot of trial-and-error costs and win valuable time in the race towards AGI. It's like a team can do without wings, but not without an engine. Top talent is that engine.

Host: So, the normal distribution of individual abilities, amplified through industry effects, ultimately manifests as a power-law distribution of company output.

Omar Shams: Exactly.

Host: But did Zuckerberg rely on intuition when assembling a super-intelligent team?

Omar Shams: I wouldn't dare to comment on that, but I must admit, Zuckerberg is indeed an excellent founder. Regarding his decision-making, I think it's a very bold adventure—such a gamble only founders and CEOs with super voting rights like him would dare to make. After all, Meta's cash flow is very abundant, and compared to some other money-burning projects, investing in AGI (Artificial General Intelligence) is a relatively wise choice. I think it's too early to judge now; we can wait a while to see the results.

Host: If I were Zuckerberg, with his resources, I would also wonder why not use our idle cash flow to assemble the best team we can? Why not put them here? So I'm not questioning that strategic decision. I'm questioning whether it's the right strategy to spend $100 million to get the so-called "best talent." Maybe you have to do it, you can argue, because there are only so many people who truly understand. But the counterargument is, no, there are many who understand.

Omar Shams: That's a good question. I think what they're buying isn't "secrets," but "tacit knowledge" and "taste." When building large-scale AI systems, there are countless subtle engineering and theoretical choices.

The value of these talents lies in their judgment and intuition accumulated in actual work, helping companies avoid common mistakes and take fewer detours. For example, Zuckerberg might have learned from Meta's Llama project. Developing AI is like building an airplane; even if you have all the theory, you need someone to guide "which screw to turn first."

With the era of AGI approaching, Zuckerberg would rather spend more money than miss the opportunity. After all, Meta can afford the cost, and the potential return could be enormous.

The Reality and Future of AI Agents, Layoffs Are Coming

Host: As the head of AI agents, what do you think of the current "hype and reality" in this field? It's clear that AI tools have brought productivity improvements. But I want to distinguish between AI tools and agents. I see sending queries to ChatGPT or having ChatGPT modify or write a draft as "tools."

I don't see that as an agent. I think an agent is something more autonomous, capable of taking multiple steps without human supervision, rather than a tool for single or few interactions with human oversight at each step of output. So, is there an example where, for instance, I want to write a function in my codebase, I let the agent do it, and it does a bunch of complex things and returns the result? Is that a reality now?

Omar Shams: In the software development field, agents have become a reality. In projects I'm involved in, tools like Cursor and GitHub Copilot have completely changed the way programmers work. Nowadays, even startups have significantly raised the standard for software quality, and low-quality code can no longer easily pass. This pressure has driven progress across the industry.

In the legal field, AI companies like Harvey have already started generating considerable revenue. Although progress may be slower in other industries, the introduction of AI assistants in white-collar work has become an inevitable trend. While I can't determine the specific impact of this trend on the job market, it's certain that workflows will undergo significant changes—AI assistants will either assist human work or directly replace some jobs. This has also led to elevated standards in the software industry. Junior software engineer positions are facing challenges because AI can already handle most basic tasks, and the future engineer role is more like a "technical supervisor" managing a team of AI agents.

Host: This isn't good news for computer science graduates.

Omar Shams: This is indeed a structural change. A few years ago, almost anyone who knew a bit of programming could get an offer, but this bubble is clearly unsustainable.

From a more fundamental perspective, the disconnect between computer education systems and AI development is also a major issue. Most university courses still focus on traditional content such as discrete mathematics and algorithm theory, neglecting the cultivation of practical software development skills. I believe this will force education and personal development to focus more on "agency" and "practical ability." Someone with rich project experience who can solve real-world problems will be more valuable than a graduate with only a degree.

Omar Shams: Regarding AI's impact on employment, I also want to discuss Anthropic CEO Dario Amodei's prediction, which suggests that with the development of AI, large-scale layoffs will occur within the next two years. He predicts that the layoff rate could reach 30% in two years.

He believes companies like Tesla, even though they are relatively streamlined, may face layoffs in the future. But personally, I think a 30% layoff rate might be somewhat high, but even so, industry insiders like Amodei believe AI's impact is much larger than we anticipated.

From String Theory to AI Entrepreneurship: Physical Intuition as a Key Driver

Host: Omar, from your physics/math major at Carnegie Mellon, to studying string theory, and finally diving into AI, what initially ignited your passion for physics?

Omar Shams: My first love was physics. At 15, I saw the "twin paradox" in a physics textbook—a twin traveling in space returns younger than his brother. I thought it must be made up, but the teacher told me it was true. At that moment, it felt like discovering that "magic" really exists. I decided then that I had to learn this magic. For the next ten years, I immersed myself in the world of physics, studying deep theories like holography and noncommutative geometry.

Host: I can relate to that story. The beauty of physics is that with simple algebra, you can derive conclusions like Lorentz transformations that overturn perceptions. I've always wondered why every smart person isn't passionate about physics. The physical intuition you mentioned, that feeling of playing a movie in your mind, is the key difference between physicists and pure mathematicians.

Omar Shams: Absolutely. For me, physics problems aren't cold formulas but an action movie in my mind. This visual, intuition-based thinking has had a profound impact on my later AI research.

Host: When did you seriously consider transitioning from physics to AI?

Omar Shams: It was during my later graduate years. I started dabbling in genomics, where I attempted to use principal component analysis (PCA) to process human mitochondrial DNA data, and for the first time, I felt the powerful ability of machine learning technology to transform data and reveal patterns. This, combined with my previous summer project experience in lattice quantum chromodynamics (Lattice QCD), showed me a new, full-of-possibilities field. My first formal job was building a music recommendation engine.

Host: So, you didn't completely abandon your physics way of thinking but brought it into a new battlefield.

Omar Shams: Yes, especially in my startup project Mutable. We developed a tool called "Auto-Wiki," which can automatically generate Wikipedia-style documentation for a large codebase. The inspiration for this idea partly came from the "renormalization group" in physics—through continuous "coarse-graining" (summarization) operations, extracting macro, key structures and information from micro details. This process not only helps humans understand code but also provides excellent context for large language models (LLM), greatly enhancing the efficiency of code Q&A systems.

Host: There are many physicists in the AI field, from Hinton to Karpathy. What "superpowers" do you think a physics background gives you?

Omar Shams: I think there are three points. First is physical intuition; we're accustomed to visualizing and systematizing abstract problems. The optimization process of AI's loss function is like a small ball rolling on an energy manifold, and physicists can intuitively "see" and understand this process.

Secondly, mastery of continuous mathematics; physics training makes us proficient in handling continuous systems, approximations, and probability distributions with mathematical tools like path integrals and partition functions, which are highly compatible with the mathematical nature of large-scale neural networks.

Lastly, experience in dealing with "emergent" phenomena. Physics is full of examples where complex phenomena emerge from simple rules, such as phase transitions. AI's "emergent ability" is similar. We're accustomed to finding patterns at different scales and have a deep understanding of this "qualitative change from quantitative change" phenomenon.

Host: What about the weaknesses of physicists?

Omar Shams: Possibly a lack of sensitivity to discrete algorithms and engineering details. But overall, when the problem scale reaches a certain level, continuous physical thinking often becomes more effective.

Host: So, if you were to give Zuckerberg a suggestion, it would be to hire more theoretical physicists?

Omar Shams: (Laughs) I think that would be a very wise investment.