李飞飞谈 ImageNet、世界模型，以及为什么每个人都该在 AI 里有位置

Li says she is not a utopian: AI will affect jobs and people, but she calls herself a humanist who believes that what AI does now and in the future is up to people. She sees humans as a fundamentally innovative species who keep improving their tools and lives, so she believes technology is a net positive. But every technology is a double-edged sword, and society can get it wrong.

She repeats a line from her testimony to Congress: there is nothing artificial about AI, because it is inspired by people, created by people and affects people. She reminds each graduating student of her lab of this. Her practical advice is for everyone, wherever they sit in AI development or use, to act as a responsible individual and care about how it affects their community and future generations.

Li traces the field from Turing's thinking-machine question and the 1956 Dartmouth workshop, where John McCarthy coined the term artificial intelligence, through logic and expert systems, to the machine-learning era of the late 1980s to early 2000s. The key shift was moving away from pure rules toward systems that learn patterns, so that after seeing three cats a machine can recognize the fourth, fifth and sixth. She began her PhD at Caltech in 2000, in the middle of what was called the AI winter.

She chose visual intelligence, with object recognition as a north-star problem, and noticed that the models she worked on all lacked data. Human learning and evolution are big-data processes, so she and her students conjectured that big data was a critically overlooked ingredient. Starting in 2006-2007, they curated 15 million images across a taxonomy of 22,000 concepts built on WordNet and released it to researchers, with an annual challenge.

In 2012 a group from Toronto led by Geoffrey Hinton entered the challenge, used ImageNet and two Nvidia GPUs, and made a huge step toward solving object recognition. Li calls big data, neural networks and GPUs the golden recipe for modern AI, and says ChatGPT still uses the same three ingredients, with internet-scale text, a more complex architecture and many more GPUs. She also notes that companies are still paying for labelled data today, and that Scale's founder told her ImageNet inspired his company.

Li recalls that in 2015-2016 some tech companies avoided the word AI because they were unsure whether it was a dirty word. She encouraged them to use it, calling it one of the most audacious questions humanity has ever asked. By her reading of Silicon Valley marketing, around 2017 companies began to call themselves AI companies.

She also cautions against crediting any single hero. AI is a 70-year-old field built by generations of researchers, and she says nobody could have got here alone. On AGI, she says she does not know that anyone has defined it, and that to her it is more a marketing label than a scientific one, since the north star is still the question of whether machines can think and act like humans.

Li insists more innovation is needed. Scaling laws still have room to run, but no scientific discipline has ever declared itself done, and AI is among the youngest. She points to a model that cannot count chairs in a video of office rooms, to the lack of Newton-like abstraction, and to the absence of emotional intelligence in a conversation between a student and a teacher.

Li says she was inspired by large language models, and as co-director of Stanford's human-centered AI institute she helped create a research center on foundation models. But she thinks much of human intelligence is spatial and beyond language, citing a chaotic first-responder scene where movement and situational awareness matter more than words. That led her to see spatial intelligence as the linchpin connecting visual intelligence with embodied AI and robotics.

She began formulating the idea in 2022, gave a TED talk on it in 2024 and founded World Labs. She defines a world model as one that lets anyone create worlds from an image or a sentence, interact with them by walking, picking up objects or changing things, and reason within them. For a robot consuming such a model's output, that means planning a path to tidy a kitchen.

Li stresses that the payoff is not only robots. Humans are embodied agents too, and could be augmented by spatial AI in games, creativity, design and science. She uses the discovery of DNA's structure as an example: Watson and Crick reasoned in three dimensions from a flat 2D X-ray photograph taken by Rosalind Franklin.

At Ben Horowitz's suggestion, Lenny asks about the bitter lesson, Richard Sutton's observation that simpler models with lots of data tend to beat complex ones with less. Li notes that the paper came years after ImageNet, and for her it was a sweet lesson. She says big data will still matter in robotics, but the field is in early experimentation and far less mature than language models.

Her core objection is a data mismatch. Language models are trained on words and output words, a perfect alignment, while robots must output actions in 3D worlds that web videos do not provide. Researchers have to supplement them with teleoperation and synthetic data, and the bitter lesson remains untested for robotics.

Robots are also physical systems, closer to self-driving cars than to chatbots, so they need bodies and application scenarios as well as brains. She points to Sebastian Thrun's 2005-2006 DARPA win and 20 years of progress to today's Waymo, and says cars are simpler than robots because they run on 2D surfaces and try not to touch anything. Reflecting on the brain's roughly 20 watts of power, she says the more she works in AI the more she respects humans.

Li presents Marble as one of the first products of World Labs, a frontier-model company she founded with Justin Johnson, Christoph Lassner and Ben Mildenhall. It is built on what she describes as the first generative model to output genuinely 3D worlds, after more than a year of work. Users can prompt with a sentence or one or more images, navigate the result, and even walk around in it with a headset.

When Lenny praises the dots that appear before a world renders, Li explains this was an intentional visualization that guides users in and is not part of the model, and many users have said they love it. She says Marble is early but is the first product that does "prompt to worlds".

Early uses include virtual production for film, where Li says a Sony collaboration with technical artists cut production time by a large factor on a one-month project, game developers exporting meshes, and robotics researchers creating diverse synthetic training environments. A psychology team also asked to use Marble to build immersive scenes of different types, for example messy or clean, to study how patients' brains respond.

Versus video generators like Veo, Li says World Labs' thesis is that spatial intelligence is not only about passively watching videos. Using Plato's cave, she describes vision as making sense of a 3D or 4D world from a 2D projection. Marble gives creators worlds with 3D structure they can use in their work, and can also export a camera path as video.

World Labs has around 30 people, mostly researchers and research engineers plus designers and product staff, and uses a lot of GPUs. Li says the company integrates deep tech with real products. She is the founder of a company at this scale for the first time, though she says she ran a dry cleaner at 19 and also founded Google Cloud AI and the Stanford institute.

What surprises her is how intensely competitive the AI landscape is, in both models and talent, including the sums certain talent can command. Asked about her career path, she credits curiosity and passion for AI and being intellectually fearless. She gave up the near-certainty of tenure at Princeton to restart the clock at Stanford, became the first female director of its AI lab, and went to Google to work with people like Jeff Dean and Geoffrey Hinton.

She says she does not dwell on everything that could go wrong. To young AI talent she says to focus on passion, mission, belief in the team and the impact they can have, instead of overanalysing every dimension of a job. Lenny agrees that the news and FOMO make this hard.

Li co-founded Stanford's Institute for Human-Centered AI in 2018 with faculty including John Etchemendy, James Landay and Chris Manning, after concluding at Google that AI would be a civilizational technology. She also wrote a New York Times piece that year arguing for a guiding framework anchored in human benevolence. It has grown into what she calls the world's largest such institute, with hundreds of faculty across all eight Stanford schools.

HAI works on policy because Silicon Valley was not talking to Washington or Brussels. She cites programs such as a congressional boot camp, the AI Index report and policy briefings, and advocacy for a national AI research cloud bill that passed during the first Trump administration. She remains a leader there though less involved operationally.

Her closing message answers the question she hears worldwide: whether a musician, teacher, nurse, accountant or farmer has a role in AI. Her answer is a resounding yes, and she criticizes Silicon Valley talk of infinite productivity or infinite leisure. No technology should take away human dignity, and dignity and agency should be at the heart of development, deployment and governance, she says, whether you use AI as an artist, take part in your community as a farmer, or are a nurse helped by AI to care for an ageing society.

Listeners can find World Labs at worldlabs.ai, which has research blogs, Marble and job postings for the San Francisco team.