In today’s venture capital market, “world models” are unquestionably the hottest of hot topics. We almost daily see new “world model” companies completing funding rounds, with valuations soaring and star-studded shareholder lists. And in the press releases announcing these financings, people repeatedly emphasize one key point: a qualified superintelligent agent should not rely solely on data feeding to gain capabilities, but must actively understand the physical world—just as humans do.
But Pete Florence wrote a long public letter after starting his company, beginning with: “Do not label my company as a world model.”
This is truly upside down. Pete Florence is far more than just a “startup founder.” Before founding his startup, Pete Florence worked at Google DeepMind, rising from a research scientist to a senior research scientist. He was one of the lead developers behind Gemini Robotics, DeepMind’s robotic control model released in 2025. However, his most influential contribution during this time was, alongside his colleagues, introducing the world in 2023 to a novel robotic architecture called “Vision-Language-Action Models.”
(Pete Florence, source: social media)
Yes, if today’s “world models” or “VLA” represent the most cutting-edge and widely accepted direction, then Pete Florence is undoubtedly a pioneer on this path. For someone like him to lead the charge in abandoning the “world model” label is truly shocking.
And now, the impact has doubled. Recently, Generalist AI, the embodied intelligence company founded by Pete Florence, completed a new funding round totaling $400 million (approximately RMB 2.7 billion), with a valuation of $2 billion (approximately RMB 13.55 billion). Investors in this round include NVentures, NVIDIA’s investment arm; NFDG, co-managed by renowned angel investors Nat Friedman and Daniel Gross; Bezos Expeditions, Jeff Bezos’s family office; Lin Bin, co-founder of Xiaomi; Yuan Zheng, founder of Zoom; and Fei-Fei Li, one of the most prominent scientists in the field of world models.
“Goal” is more important than “tag”
Why, as one of the primary founders of world models, does Peter Florence resist being labeled a “world model” researcher? Why does Li Fei-Fei, the most representative scholar in the field of world models, financially support someone who openly defies conventional norms? The story may begin in 2019.
At that time, Pete Florence was pursuing his Ph.D. in computer science at MIT, focusing on areas such as robotic manipulation, computer vision, and natural language processing—given this background, Pete Florence was thoroughly mainstream: his research direction was conventional, his academic credentials were solid, and he was not someone who needed to rely on being unconventional to secure resources. But the issue was that MIT assigned him a supervisor named Russ Tedrake.
Who is Russ Tedrake? First and foremost, he is undoubtedly a leading academic. In 2019, he served as a professor of electrical engineering and computer science at MIT and director of the Robotics Consortium at the Computer Science and Artificial Intelligence Laboratory. Each year, during the prestigious DARPA Robotics Challenge, he leads MIT’s team in competition. Outside of academia, he also holds the position of Vice President at the Toyota Research Institute’s Robotics Center. Russ Tedrake is widely regarded as one of the foremost scholars in robotics, with ample resources to help the young Pete Florence realize his academic aspirations.
However, in Lars Tiedeck’s self-perception, what fascinated him was not programming code, but “physics.” In a personal introduction, Lars Tiedeck recalled that his journey into computer science began when he encountered the “rich dynamic behaviors” while researching “bipedal robots,” sparking his deep interest in “complex fluid dynamics control.” Therefore, unlike other researchers who typically start by studying how to make robots pick up apples or make beds, his initial research focused on how to control “stalled aircraft or flapping-wing vehicles” and how to “navigate at high speed through dense obstacles.”
Given this background, Lars Tedrake places great emphasis on “understanding the physical world.” The MIT website describes his academic focus as follows: “His research centers on finding elegant control solutions for interesting (underactuated, stochastic, and/or hard-to-model) dynamical systems, and building experimental platforms to validate these solutions. He is particularly interested in the connections between mechanics—especially nonsmooth mechanics—and machine learning/optimization theory to enable robust control design for complex mechanical systems.”
Peter Florence, immersed in this environment, naturally became part of the "physics-oriented" school within computer science. For instance, his most representative academic achievement during his Ph.D. was a paper titled "Self-Supervised Correspondence in Visual Imitation Learning." This paper introduced a method that enables robots to accomplish challenging manipulation tasks with only 50 demonstrations, while also generalizing across different object categories and adapting to deformable object configurations. The paper was awarded the 2020 IEEE Robotics and Automation Best Paper Award.
What “school” he belongs to doesn’t matter; what matters is that, under the influence of this environment, Pete Florence developed a fundamentally different way of thinking. Most researchers are accustomed to starting with existing technologies, conducting experiments to explore their possibilities, and then determining practical applications. Pete Florence, however, believes the correct order should be: “first set a specific goal, then design the technological pathway.”
After joining Google DeepMind, Peter Florence pursued this direction, and his first major contribution was the Transporter Network, Google’s initial robotic model architecture introduced in 2021. In the paper announcing the model, Peter Florence stated that organizing items should be a fundamental skill, but for robots, performing this task requires “high-level and low-level perceptual reasoning”—they must determine where books should be placed, in what order to stack them, and ensure their edges are aligned to form a neat pile.
The Transporter Network is a model architecture designed specifically to "make simple actions simple," enabling robots to perform a variety of tasks based on vision alone, with faster training and reduced dependence on training environments.
In 2023, we jointly released the VLA architecture with the DeepMind team, building naturally upon this line of thinking. In the paper that launched today’s era of world models, the authors stated that they intended the VLA architecture to “significantly improve generalization to new objects, interpret instructions not seen in robot training data (such as placing objects on specific numbers or icons), and perform basic reasoning based on user commands (such as picking up the smallest or largest object, or the object closest to another object).”
Returning to the original question, why does Peter Florence, one of the primary founders of world models, resist being labeled a “world model”? The answer is this: Peter Florence believes that “goals” matter more than “labels.”
In his view, the current enthusiasm for world models is essentially “idea-driven,” with much of the excitement stemming from capital markets' thrill at identifying non-consensus opportunities in hot directions. And if the true goal is to genuinely integrate robots into our work and daily lives to create productivity, then building “world models” is clearly not the objective. The real goal should be enabling robots to accomplish a wide variety of unseen tasks with extremely high success rates and speed, entirely without any task-specific data.
This is precisely why Pete Florence decided to leave Google DeepMind and start his own company. At the 2025 NVIDIA GTC conference, Pete Florence made his first public appearance as co-founder and CEO of Generalist AI. He said, “We are determined to build robots that can do anything… Imagine what the world would look like if the marginal cost of physical labor dropped to zero.”
99% success rate
Beyond its technically unconventional philosophy, Pete Florence’s entrepreneurial journey has also been decidedly non-mainstream.
Theoretically, an entrepreneur with such a background would inevitably be highly sought after by VCs today. Yann LeCun, Ilya Sutskever, and Mira Murati are examples—all raised over $1 billion in seed funding before or immediately after incorporating their companies. But Peter Florence’s Generalist AI accepted investment from only a handful of institutions in its early stages, including NVIDIA, the Bezos Family Office, and NFDG. It wasn’t until NVIDIA’s venture arm, NVentures, hosted a “portfolio company roundtable” at the 2025 GTC conference that many even learned he had left to start his own company.
Why is this the case? The most likely answer is Peter Florence’s deliberate choice. As mentioned above, Peter Florence joined Google DeepMind right after graduating and worked there from 2019 until 2025, with no other employment history in between. This means that Generalist AI is his first entrepreneurial venture, and it is therefore essential to approach it with the utmost care.
In fact, at the 2025 NVIDIA GTC conference, where Peter Florence made his public debut as an entrepreneur, he clearly demonstrated his "caution." Beyond revealing that he was building "robots," he disclosed no specific business direction and explicitly stated, "We are still in stealth mode."
Until November 2025, the specific business of Generalist AI was not seen. In November 2025, Generalist AI released their first-generation embodied intelligence model, GEN-0. In their official announcement, Generalist AI stated that GEN-0 combines the strengths of vision and language models while also surpassing them—GEN-0 can capture human-level reflexes and physical common sense.
In simple terms, it can continuously improve its capabilities as model size and training data increase, overcoming the limitations of previous small models; it can think and act like a human, making rapid and natural responses in real physical environments; it is inherently compatible with different types of robots without requiring additional modifications; more importantly, it leverages vast amounts of real-world operational data, eliminating constraints due to data scarcity, and allows flexible adjustment of training data composition. Numerous tech media have noted that GEN-0 demonstrates that the mathematical “scaling laws” driving large language models like ChatGPT also apply to physical motion.
However, GEN-0 is not perfect. For example, GEN-0 also fails to address the dataset challenges plaguing the field of embodied AI. Therefore, by April 2026, Generalist AI had rapidly evolved into a completely new version: GEN-1.
("Robotic Arm," source: Generalist AI social media)
To address data set challenges, Generalist AI developed a wearable device capable of capturing subtle movements and visual information as humans perform manual tasks. Generalist AI stated that during the development of GEN-1, they collected over 500,000 hours of “petabyte-scale physical interaction data” using these robotic hands to train their physical model. After sufficient training, Generalist AI reported that GEN-1 achieves a 99% success rate in repetitive yet delicate mechanical tasks such as folding cardboard boxes, packing phones, and maintaining robotic vacuums—operating at approximately three times the speed of the previous GEN-0 model and achieving this performance in just about one hour.
Therefore, Generalist AI proudly announces that the physical model of GEN-1 is nearing a GPT-3-like inflection point, with performance on certain tasks beginning to “reach the level required for deployment in commercial practical environments,” and “we can expect each new generation of models to unlock a growing range of increasingly complex tasks that can all be mastered.”
On the official blog, Pete Florence noted that the development of GEN-1 best embodies his personal technical philosophy: first, he set a rational goal—enabling a robot to perform a wide variety of unseen tasks with extremely high success rates and speed, entirely without any task-specific data. He then defined a solution path that allows for the use of minimal robot data for specific tasks (referred to as X), achieving high-level performance on those tasks, and progressively reducing X while improving performance.
By now, the question we raised earlier has been answered. Whether or not the products developed by Generalist AI are called “world models” is no longer important; if you see potential in the embodied intelligence industry and believe robots will soon be widely adopted in actual production, then Generalist AI is indeed a worthwhile investment. Moreover, Generalist AI’s latest funding round was indeed finalized within two months after the release of GEN-1.
According to reports, existing shareholders NVIDIA, Bezos Expeditions, and NDFG all chose to reinvest, doubling their commitments. Additionally, new investors include Lin Bin, co-founder of Xiaomi; Yuan Zheng, founder of Zoom; Chinese-American scientist Fei-Fei Li; as well as institutional investors Radical Ventures, 8VC, Union Square Ventures, Hanabi Capital, and Norwest.
In other words, Pete Flores in June 2026 no longer needs to prove himself. At the very least, the claims he once made—such as when the newly founded Pete Flores, in 2025, said on a podcast: “General robotics isn’t about dabbling in everything, but about becoming proficient enough in real-world tasks to be genuinely useful”—are now on their way to being fulfilled one by one.
This article is from the WeChat public account "ChinaVenture," authored by Pu Fan.