Summary of the Key Points in Plain Language
This article breaks down the core information from an interview with Li Feifei, widely recognized as a pioneer in the field of AI, conducted by Bloomberg in August. Twenty years ago, she used a dataset of 14 million manually annotated images called ImageNet to transform deep learning from an obscure technology that no one took seriously into the cornerstone of the AI industry today. In this interview, she not only revealed her current bets on the next generation of AI—world models that can understand the 3D physical world—but she also pointed out the inherent limitations of large language models like ChatGPT. She criticized the prevailing trend in Silicon Valley, which uses apocalyptic sci-fi narratives about AGI (Artificial General Intelligence) to hide the fact that the power of AI, its data, and its development are controlled by a few giants. This concentration of resources aims to turn AI into a private asset controlled by a few, rather than a public technology that should serve all of humanity.
---
Detailed Interpretation of the Key Points
1. What Most People Don’t Know: The Origin of All Current AI
Most people think the sudden popularity of deep learning in 2012 was due to the excellence of the Hinton team’s algorithms. However, that’s not the case. At that time, the entire AI community was focused on tweaking code and adjusting parameters, much like people trying to make car engines more efficient, without realizing that they didn’t even have a proper testing ground. The available annotated images were limited to a few hundred thousand, which was far from enough to develop reliable image recognition capabilities. Li Feifei, on the other hand, took a different approach. She used crowdsourcing to collect 14.19 million images across more than 20,000 categories, essentially building an eight-lane highway for the AI industry. This database transformed a philosophical question of whether machines could understand the world into a practical engineering problem that could be tested and improved. All subsequent AI advancements in image generation, facial recognition, and autonomous driving owe their success to this dataset. What was initially dismissed as a trivial data project ultimately raised the bar for the entire AI industry.
2. Why Li Feifei’s 3D World Models Represent the Next Real “Turning Point” for AI
Current AI news often focuses on large language models achieving new milestones, but Li Feifei’s argument is straightforward: “Can text put out fires? Can text make pancakes?” While such models can generate recipes, they have no understanding of the real world—eggs breaking when dropped from a height, door handles requiring a specific angle to open, or the heat from hot coffee. The technology she is working on aims to give AI the “common sense of space,” enabling it to generate and simulate 3D environments and understand physical laws. Once this is achieved, robots will be able to perform real tasks in kitchens, construction sites, and hospitals. She emphasizes that this technology is still in its early stages, far from the so-called “universal AI” that some Silicon Valley companies claim to have achieved.
3. The Hypocrisy of Silicon Valley’s “AI Will Destroy Humanity” Narrative
Li Feifei criticizes the Silicon Valley giants for using this apocalyptic narrative as a cover for their control over AI resources. They claim AI will be dangerous, yet they hold all the high-end GPUs, data, and talent, preventing smaller companies, universities, and public institutions from accessing them. This strategy is similar to running a chemical plant and claiming it’s dangerous while preventing anyone from interfering, only to profit from it. By focusing on a distant threat, they avoid addressing real issues such as the low wages of AI workers in Kenya, the exploitation of resources by tech companies, and the lack of funding for AI education in schools.
4. The Public’s Fear of AI
A Pew survey from 2025 shows that only 16% of people are excited about AI, while 34% are concerned. The fear is not about a lack of understanding of the technology; rather, it’s about the practical consequences: jobs being lost to AI, the lack of benefits for the general public, and the lack of control over AI’s direction. Li Feifei points out that the power of AI is in the hands of a few, and the technology will primarily serve their interests, not the broader community.
5. Li Feifei’s Unchanged Logic for 20 Years: A Third Path for AI
There are two extreme views on AI: either to immediately halt all research or to accelerate its development at all costs. Li Feifei advocates a more pragmatic approach. She has always believed that AI learns from real-world examples, not from algorithms. She emphasizes that AI should reflect the physical world we live in and address real-world issues like fairness, employment, and public investment. This approach is more in line with the true potential of AI than the extreme positions.
---
In summary, Li Feifei’s insights highlight the critical role she played in the development of AI and the current challenges faced by the industry, as well as the hypocrisy of Silicon Valley’s rhetoric about the future of AI.