ImageNet started as a Princeton project
Fei-Fei Li began building ImageNet in 2006 while an assistant professor at Princeton, working from the premise that progress in visual recognition was bottlenecked by data rather than by algorithms. In 2007 she worked with Princeton professor Christiane Fellbaum, a creator of WordNet, whose roughly 22,000 noun categories became the hierarchy ImageNet's images were sorted into. The dataset was first presented in 2009 at the Conference on Computer Vision and Pattern Recognition. [2]
Labeling ran from July 2008 to April 2010, using roughly 49,000 workers across 167 countries on Amazon Mechanical Turk to sort and label more than 160 million candidate images down to about 14 million hand-annotated ones across more than 20,000 categories, with bounding boxes on around a million. Li later moved to Stanford, where she still directs the Stanford Institute for Human-Centered AI, but the dataset's foundational work, choosing the WordNet structure and building the labeling pipeline, was done at Princeton. [2]
Atlas interpretation: ImageNet mattered less for its existence than for the annual competition built on top of it. When AlexNet won that competition in 2012 by a wide margin using a convolutional neural network, it was ImageNet's scale and consistent labeling that made the win legible as a result rather than an anecdote: a big enough, clean enough benchmark for a new method to beat the old ones by a margin nobody could dismiss. [2]
A university-wide initiative for large models
Princeton launched the Princeton Language and Intelligence initiative, PLI, on September 26, 2023, to build research capacity around large language models specifically, rather than AI in general. Sanjeev Arora, the Charles C. Fitzmorris Professor in Computer Science, directs it, drawing on faculty from more than a dozen departments through a seven-person executive committee and a ten-person research advisory committee. PLI's stated aims include deepening the technical understanding of how large models work, applying them across other academic disciplines, and studying their safety, policy and ethical implications, explicitly framed as keeping some of that expertise in universities rather than leaving it entirely to private labs. [3]
The university backed the initiative with a computational cluster it described as the largest in academia at the time, funded through the endowment alongside postdoctoral fellows and research scientist positions. Princeton expanded that hardware in March 2024 with an additional 300-GPU cluster dedicated to academic AI research, aimed partly at making large-scale model experiments possible for researchers outside the small number of institutions and companies that could otherwise afford them. [3][4]
PLI's published research areas include alignment work aimed at making language models more helpful, honest and harmless, applications of large models to genomics and precision health, and computational analysis of historical and classical texts, alongside broader work on AI safety and policy. That spread mirrors the initiative's cross-departmental structure more than a single technical bet, in contrast to a company like Poolside or a lab like OpenAI, each organized around one method or one product line. [3]
Sources
- History | Princeton University
Princeton University
- ImageNet
Wikipedia · Sep 9, 2026
- Beyond ChatGPT: Princeton Language and Intelligence initiative pushes boundaries of large AI models
Princeton University School of Engineering and Applied Science · Oct 6, 2023
- Princeton invests in new 300-GPU cluster for academic AI research
Princeton University · Mar 15, 2024