These AI Leaders All Studied Physics

11/19 2026 449

Back then, I felt I had been "born a hundred years too late," having missed the era of explosive physics discoveries. Now I have found a golden age in AI.

Behind what is set to become the largest IPO in history stands a group of physicists who "changed careers."

On November 19, US Eastern Time, Anthropic will list on Nasdaq under the ticker "ANT." In addition to the public offering, the company will simultaneously conduct a private placement, with the two parts raising up to approximately $90.4 billion in total; based on the upper end of the offering price range, the fully diluted valuation is approximately $1.4 trillion. This scale exceeds the record set by SpaceX in June this year and is expected to become the largest IPO in history.

A look through the resumes of the company's core founders reveals an interesting detail: this large-model team has a striking number of people with physics backgrounds. CEO Dario Amodei majored in physics as an undergraduate and studied neural circuits during his PhD; Chief Science Officer Jared Kaplan was once a theoretical physicist; Chief Architect Sam McCandlish has a PhD background in theoretical physics.

This is somewhat different from people's conventional image of a large-model company. Usually, people think of computer scientists and machine learning researchers, but at Anthropic, physicists not only appear in the founding team but also constitute an important part of the company's talent.

This phenomenon even has a certain universality. From large-model laboratories in Silicon Valley to AI startups in China, more and more people with physics backgrounds are standing at the core of the AI wave. According to Digital Intelligence Frontline, this year even the founder of a domestic internet company specifically wanted to interview people with physics backgrounds when recruiting AI talent.

What exactly does physics training bring them? Why have physicists become important participants in this round of the AI wave?

01 The Trillion-Dollar Anthropic Is Packed with Physicists

In early 2021, seven researchers left OpenAI and decided to found a new AI company. It was during the pandemic, and the small group could only meet in one person's backyard, six feet apart and wearing masks.

A year later, when two co-founders of Anthropic recalled this experience, they added an interesting detail: besides having worked together at OpenAI and some members having collaborated at Google Brain, this group also had one thing in common—"many of us are current or former physicists." Dario said everyone had already known each other in the physics community.

Dario's own experience is a microcosm.

In the spring of 2007, 24-year-old Dario was still a physics PhD student at Princeton University. In an office in the basement of Guyot Hall, he had a conversation with neuroscience professor Michael Berry. After that, he seriously considered switching to biophysics for the first time. In his second year of the PhD, he joined Berry's lab to study how nerve cells in animal retinas process and transmit visual information.

The new direction soon presented a difficult problem: the lab needed to record electrical signals from a large number of nerve cells simultaneously but had no suitable equipment.

So Dario simply built a multi-electrode array himself. Berry later called it "a rather ambitious technical development project." Dario stayed in the cleanroom making devices and adjusting equipment, and this work was eventually published as a paper in 2012. This may be exactly the underlying quality left by physics training: the habit of understanding problems from first principles, and when tools are lacking, building them yourself.

Why switch from physics to neuroscience? Dario later explained that he had "always been interested in two questions: how intelligence actually works, and whether science and medicine can improve human life," that is, seeking the "laws of physics" of intelligence.

After his PhD, Dario went to Stanford University for a postdoc under professor Parag Mallick, studying cancer detection and its mechanisms. Mallick later recalled that Dario had almost no background in cancer research, but he would use weekends to read through all the published papers on a given topic and quickly catch up.

It was also during this postdoc period, around 2012 to 2013, that research results in deep learning from teams such as Google Brain emerged one after another, especially AlexNet's performance in image recognition, which made him see that this technology was beginning to show real potential.

At the end of 2014, Dario joined Baidu and entered the team led by Andrew Ng. The first project he participated in was the development of the end-to-end speech recognition system Deep Speech 2. In less than a year, he moved to Google Brain as a senior research scientist and joined OpenAI in 2016.

Another key figure at Anthropic, Chief Science Officer Jared Kaplan, initially focused on the universe itself.

Kaplan's mother was a science fiction writer. From childhood, he was fascinated by one question: could humanity possibly build a faster-than-light spacecraft?

To figure out this question, he chose physics. For many years afterward, he followed a typical academic path in theoretical physics: undergraduate at Stanford, PhD at Harvard, postdoc at the SLAC National Accelerator Laboratory operated by Stanford University, and finally teaching at Johns Hopkins University as a theoretical physics professor.

But after years of moving among fields such as Large Hadron Collider physics, particle physics, cosmology, and string theory, he gradually felt frustrated and burned out. "Progress in these fields was not as fast as I had hoped. Many friends told me that AI was becoming a very important field," Kaplan later recalled. In early 2018, Kaplan began spending part of his time learning and conducting machine learning research.

Similar in background to Dario and Kaplan was Chief Architect Sam McCandlish. He pursued a PhD in theoretical physics at Stanford, studying quantum gravity and holography. In 2018, he also joined OpenAI and began participating in machine learning research.

In January 2020, Kaplan, McCandlish, Dario, and other researchers jointly published the paper "Scaling Laws for Neural Language Models." Through a series of experiments, they found that the performance improvement of language models was not entirely unpredictable but had a relatively stable power-law relationship with model size, training data volume, and computational resources.

This is what later became famous as the "Scaling Laws."

"The most helpful thing about physics training is that it got me used to looking for the broadest and most important trends, and then describing them as precisely as possible," Kaplan said years later when reviewing this research experience.

This entanglement with physics also entered the company with the founding of Anthropic, influencing its team building and talent selection.

Anthropic co-founder and President Daniela Amodei once revealed in an interview that in the company's early days, two-thirds of the first roughly 15 employees had PhDs in physics. The company even developed an internal talent development mechanism called the "Physicist Pipeline," attracting people who are very smart, good at quantitative analysis, and capable of macro-level thinking, and then training them to conduct machine learning research.

In an August 2023 program, Dario said: "Physicists learn new things very quickly. We usually find that if we hire a physics PhD, he can often quickly learn machine learning and make contributions." He mentioned that at that time there might have been 30 to 40 physicists on the Anthropic team.

Of course, AI is still a relatively young field. Anthropic's team includes software engineers and machine learning researchers, as well as researchers with backgrounds in physics, ethics, and neuroscience, along with policy and operations staff. Everyone is not confined by professional boundaries: researchers write code, and engineers also participate in research; the company also organizes small training sessions, with people familiar with Transformers teaching courses according to different levels, and operations and policy teams can also attend introductory learning.

02 Physicists Have Never Been Absent from AI

The 2024 Nobel Prize in Physics was somewhat unexpected: it was awarded to two scientists studying artificial neural networks—John Hopfield and Geoffrey Hinton—for "foundational discoveries and inventions that enable machine learning with artificial neural networks." The Nobel Committee for Physics specifically mentioned that physics has provided important tools for the development of machine learning.

In fact, the connection between physicists and AI is far earlier than this wave of large models. When neural network theory took shape, they were there.

Hopfield himself was a condensed matter physicist. In 1982, he proposed the "Hopfield network," bringing concepts from physics such as energy functions and stable states into neural network research. In 1985, Hinton, David Ackley, Terry Sejnowski, and others went further, applying methods from statistical physics to neural network learning and developing the Boltzmann machine, a name that itself comes from statistical physics.

Sejnowski's experience is particularly interesting. He pursued a physics PhD at Princeton University, initially studying general relativity, and his advisor was none other than Hopfield. Later, he audited a neuroscience course and became fascinated by the complex questions behind the brain and behavior, gradually turning to neural network research, believing that "the mysteries of the universe pale in comparison to the mysteries of behavior."

A group of researchers trained in physics began applying tools from probability, statistics, and information theory to machine learning problems. For example, in the 1980s, physicists began using statistical physics to study the memory mechanisms and storage capacity of neural networks; by the 1990s, this intersection further expanded into areas such as Bayesian inference, statistical learning, pattern recognition, and kernel methods.

Entering the new century, people with physics backgrounds still appeared at key moments. Fei-Fei Li, who studied physics as an undergraduate at Princeton, began promoting the establishment of the ImageNet dataset around 2007, giving machines, for the first time, massive amounts of labeled images to "see."

But what really made a group of young physicists unable to sit still was a game of Go in 2016.

In March of that year, AlphaGo played Lee Sedol, and the 37th move in the second game greatly surprised many professional players. For the first time, people intuitively felt that machines might be able to find solutions humans had not thought of through learning.

Misha Laskin, then pursuing a theoretical physics PhD at the University of Chicago, was also moved by this match. As he approached graduation, he began to reconsider his career choice. He later recalled that this was not because he no longer liked physics. It was just that he gradually realized he had become an expert in a very narrow field, and even if his research succeeded, it might take decades to know whether it had truly had an impact. "Maybe I'm just impatient," he said.

He began trying various things. Then he saw AlphaGo. He dropped everything and plunged into deep learning and reinforcement learning: first a postdoc at Berkeley, then joining Google DeepMind to participate in research on the Gemini reward model direction. In 2024, he and colleagues founded the AI company Reflection AI, aiming to be an open-source alternative to "closed frontier labs" such as OpenAI and Anthropic, and was also called by outsiders the "American version of DeepSeek."

A similar choice also appeared with Yao Shunyu. But this "Yu" is not that "Yu": he and Tencent's Yao Shunyu differ by only one character in their names, both come from Tsinghua, and both work on large models. To outsiders, a Tsinghua physics undergraduate degree and a Stanford theoretical physics PhD are quite an outstanding resume, but he was not satisfied with his research achievements: "No one would say the articles I published during my PhD were bad, but in all honesty, how much impact did they have on the world? I think almost none."

"I want to find a direction with more opportunities for young people." Quantum computing and AI were both on the candidate list, but the main bottleneck in quantum computing lies in experimental realization, which was not the part he was best at or most interested in. AI instead showed him a familiar way of doing research: "More often, you have an idea and can use some numerical values to verify it. In AI, these numerical values might mean training a model."

In his view, today's AI is very much like 18th-century physics, where theory and experiment had not yet clearly separated. "You can do experiments yourself, and you can also make theoretical speculations."

So, in 2024, after staying only two weeks in a theoretical physics postdoc position at Berkeley, he turned and joined Anthropic to participate in Claude model research. At the end of September 2025, he joined Google DeepMind to continue large-model research.

The opportunities brought by AI are increasingly extending beyond foundation models.

In 2022, several physicists who had worked together at a fintech company decided to bring AI into the music industry. Initially, they just wanted machines to understand and analyze sound, but an experiment after work one day led them to discover that AI could even write songs directly. After they turned the model into a Discord bot, they unexpectedly gained a group of enthusiastic users. Thus the AI music company Suno was born.

A few years later, Suno became the world's first unicorn in AI music, valued at $5.4 billion. In a May 2026 podcast episode, the host asked CEO Mikey Shulman why a physics PhD who had studied quantum computing would go on to found a consumer-facing music company. Shulman said one of the most important lessons physics taught him is that opportunities often lie at the intersection of different fields. For Suno, that intersection is music and technology.

Jiang Xu spotted another opportunity. He studied physics for both his undergraduate and doctoral degrees and explored several research directions during his PhD. "I feel that in life, you have to ride a rocket once," he said.

But the deeper he went into his studies, the more he felt he had been "born about a hundred years too late." The golden age of physics came in the first few decades of the 20th century, when many fundamental laws had yet to be discovered—when "even second-rate physicists could do first-rate work." Today, most of the simple laws have been uncovered, and what remains are the hard bones.

In 2016, AlphaGo gave Jiang Xu his first encounter with deep learning. He began teaching himself Python and machine learning, took Andrew Ng's Coursera course, and worked on some open-source projects. He landed his first AI job in Silicon Valley quite smoothly, and a year later joined OpenAI.

After some time working on large-model R&D, Jiang Xu set his sights on embodied intelligence, returned to China, and founded Liangyuan Xinchuang to research foundation models for embodied intelligence. In his view, AI's capabilities will gradually expand from the text domain—which has the most data and is the easiest to obtain—to other domains. Today's AI is like physics a hundred years ago: a huge number of problems await exploration, and "gold is everywhere."

03 After a long detour, physicists have returned to the physical world

In November 2026, Anthropic released Claude Fable 5.2, and the most eye-catching demo showed the model starting to wrestle with a bottle cap—directly controlling a robot to solve a Rubik's cube and unscrew a bottle cap.

Previously, OpenAI had also demonstrated GPT-6 Astra's involvement in electronic device design: from function definition and enclosure design to circuit board layout and software integration, the model could get its hands on everything—only prototyping and validation still had to be handed to engineers and manufacturers.

The capabilities of large models are reaching from inside the screen to outside it.

Industry enthusiasm is heating up as well. "We are in an era where the physical AI cycle and the AI super-cycle resonate at dual frequencies," Wang Jiping, IDC's global and China vice president, said at IDC FutureScape 2027 on November 18, citing data: 16.7% of enterprises have made physical AI their second-highest AI investment priority for the next 24 months, and 35% of CEOs have listed it among their top three emerging technology investments.

For physicists, this carries a certain irony of fate. Back then they left physics and rushed toward AI; now AI is about to enter the physical world, and old friends—gravity, friction, collision, space, and time—are coming knocking one by one.

If AI is to act in the real world, the first step is to understand that world.

Li Fei-Fei founded World Labs, betting on "spatial intelligence." She hopes AI will not only recognize that there is a cup in an image, but also know its position in three-dimensional space, its relationship to the table, and how it will change over time.

In 2025, World Labs launched its first product, Marble, with which users can generate explorable, editable 3D scenes from text, images, or video. In 2026 it launched the world model Atlas, which can reconstruct 3D space from images and simulate scene changes; it also acquired the robot simulation team SceniX to explore training robots in virtual environments and then validating them back in the real world. In September, AMD announced the acquisition of World Labs for about $8.2 billion, and Li Fei-Fei will serve as AMD's executive vice president and chief scientist.

Jing Li, who holds a bachelor's degree in physics from Peking University and a PhD in physics from MIT, has worked on models such as DALL·E 3, Sora, and GPT-4o, and has now joined AMI Labs, founded by Yann LeCun, as co-founder and head of pretraining.

What AMI Labs wants to do is let models learn the laws of the world from video and real-world observations. Its approach is based on JEPA: rather than painting the next second pixel by pixel like a painter, it learns abstract world representations and, combined with memory and planning, predicts the consequences of actions. Just like physicists: not memorizing every detail, but grasping the laws that govern change.

After understanding, one still has to act.

Jiang Xu regards robots as an important direction for achieving Physical AGI: both input and output can be handled with the data representations of multimodal models, and how to train and scale multimodal models is exactly what he knows best. Turning an unfamiliar problem into one already solved is almost an instinct for physicists.

In September 2026, his startup Liangyuan Xinchuang open-sourced the general navigation model LightNav-0. To train it, the team first collected more than 2,000 real indoor and outdoor scenes and converted them into simulation environments, synthesized more than 4,000 hours of training data, and then put the trained model back on real robots for validation. In his view, embodied intelligence cannot rely solely on collecting data one by one on real machines; it also needs to find a way to produce training data at scale.

Wang Hao, a PhD in computational physics and co-founder and CTO of the company Zhi Bian Liang, is pondering a more "physical" question: what exactly is the relationship between actions and results? For a robot, picking up a cup is a complete mechanics problem—how much force to use, how to grasp, how the cup will move once subjected to force, gravity, friction, collision; if any one of these is miscalculated, the cup may fall to the floor. The force analysis from high school textbooks has become one of the most cutting-edge research problems.

Traditional approaches often first train a video generation model and then connect it to an action model; Wang Hao wants to put vision and action into the same framework for joint modeling, so that the model not only predicts what will happen next in the environment, but also understands how its own actions will change the environment. In April 2026, Zhi Bian Liang released the embodied intelligence foundation model WALL-B, adopting what it calls a "world unified model" architecture.

Brian Ichter, co-founder of Physical Intelligence (PI), is trying to push general capabilities even further: can the same model learn to complete as many tasks as possible and adapt to robots of different forms?

He studied both aerospace engineering and physics as an undergraduate, then earned a PhD in aerospace at Stanford, and has worked at both Google Brain and DeepMind. In 2024, he co-founded PI with several researchers. In October of the same year, they launched the first general robot foundation model, π0, combining internet-scale vision-language pretraining with operational data from multiple robot platforms; the same model can fold clothes, clear tables, and pack groceries. The later π0.5 tried to let robots adapt to kitchens and bedrooms they had never seen before.

Robots are not the only entry point for AI into the physical world. On real roads, autonomous driving has long closed the loop of data collection, model training, and vehicle validation.

The AI path of Cao Xudong, founder of Momenta, began with a statistical physics course. He studied in the Department of Engineering Mechanics at Tsinghua University as an undergraduate; in his sophomore year, that course sparked his interest in statistics, which led him to statistical learning. During his PhD, he gradually realized that artificial intelligence was the direction he wanted to devote himself to long-term, so he left school and joined Microsoft Research Asia to work on computer vision. In 2016, he founded Momenta to bring AI into autonomous driving. In his view, the car is only a carrier; what really needs to be solved is enabling machines to perceive the environment, understand the environment, and make decisions.

Now he is also setting his sights on robots, and believes autonomous driving is the prologue to physical AI: this field has gradually formed scaled data loops and commercial loops; for robots to grasp, move, and manipulate in the real world, data acquisition and validation are far more difficult.

The combination of AI and the real world goes beyond this: in scientific research, AI is combining with automated equipment, attempting to connect computing, design, and experiments into a closed loop.

In 2014, Lai Lipeng and two other researchers with physics backgrounds co-founded XtalPi to explore combining AI, quantum physics calculations, and automated experiments for drug and materials R&D. AI assists in designing molecules and predicting reactions; automated equipment performs weighing, pipetting, and sample analysis; experimental results are fed back to the model to adjust the next round of plans. This precisely echoes Yao Shunyu's remark: in AI, theory and experiment are no longer separate. Only this time, besides humans, machines are also doing the work.

A hundred years ago, physicists pried open the laws of the world bit by bit in laboratories. A hundred years later, those who left physics back then have returned to the real world with AI. Whether gold is everywhere remains to be seen, but at least they have swapped in a new shovel.

Reference links:

1. "Daniela and Dario Amodei on Anthropic," Future of Life Institute, March 4, 2022

2. "'Very sharp, very ambitious': Inside Anthropic CEO Dario Amodei's Ph.D. at Princeton," The Daily Princetonian, October 16, 2026

3. "Generative SF: How Anthropic is building better, safer AI models," December 11, 2024

4. "Dario Amodei (Anthropic CEO) — The hidden pattern behind every AI breakthrough," Dwarkesh Patel, August 8, 2023

5. "Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan," June 16, 2025

6. "Suno's Mikey Shulman: Everyone Can Make Music Now," May 13, 2026

7. "A 4-Hour Interview with Yao Shunyu: Please Allow Me to Go a Little Crazy! Training Models at Anthropic and Gemini, Technology Forecasting, Heroism Is Over," Zhang Xiaojun Business Interviews, May 11, 2026

8. "Reflection co-founders Ioannis Antonoglou and Misha Laskin," Manifold, March 13, 2025

9. "ChatGPT Core Contributor Returns to China to Start a Business: Bringing LLM Scaling to Embodied Intelligence," DeepTech, May 22, 2026

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.