Matt Deitke

Wiki Powered byIconIQ
Matt Deitke

We've just announced IQ AI.

Check it out

Matt Deitke

Matt Deitke is an American artificial intelligence (AI) researcher at (MSL). He is known for his work in multimodal AI, embodied AI, and large-scale 3D datasets, having led the development of notable projects such as Molmo, Objaverse, and ProcTHOR. [1] [2]

Education

Deitke attended the University of Washington, where he completed his undergraduate studies at the Paul G. Allen School of Computer Science & Engineering. He subsequently enrolled in the university's Ph.D. program, continuing his research in collaboration with the Allen Institute for AI (AI2). In December 2024, Deitke announced that he was leaving the Ph.D. program before its completion to co-found a new company. [1] [2] [4]

Career

Deitke began his research career at the Allen Institute for AI (AI2) in Seattle when he was 18, working as a research scientist while concurrently pursuing his university education. His work at AI2 focused on creating robust, multimodal AI systems. During his tenure, he led several significant research projects that resulted in widely used open datasets and models.

In November 2024, Deitke co-founded Vercept, a startup dedicated to building autonomous designed to operate internet-based software. The company successfully raised $16.5 million in funding from investors, including former Google CEO Eric Schmidt.

In July 2025, Deitke announced he had joined (MSL), a division focused on advancing AI capabilities. His recruitment was part of a broader, high-profile talent acquisition effort by Meta to assemble a leading team of AI researchers.

Meta Recruitment and Compensation

Deitke's transition to Meta in 2025 garnered significant media attention, primarily due to the reported value of his compensation package. According to reports, Meta initially offered Deitke a "low-ball" package valued at approximately $125 million over four years, which he declined. After a personal meeting with Meta CEO Mark Zuckerberg, the offer was reportedly doubled to roughly $250 million, with potentially $100 million paid in the first year. This was cited as one of the largest employment packages in recent corporate history and reflective of the intense competition for elite AI talent.

The scale of the offer prompted commentary on the valuation of top AI researchers. MIT economist David Autor remarked, "When computer scientists are paid like professional athletes, we have reached the climax of the ‘Revenge of the Nerds!'" On an earnings call with investors, Zuckerberg explained the company's strategy: "We’re building an elite, talent-dense team... it really does make sense to compete super hard and do whatever it takes to get that, you know, 50 or 70 or whatever it is, top researchers to build your team." He added, "There’s just an absolute premium for the best and most talented people." [3] [6] [7]

Deitke has led or been a key contributor to several influential projects in the field of artificial intelligence, particularly in the areas of vision-language models, 3D data, and embodied AI.

  • Molmo and PixMo: A family of open-source vision-language models developed from scratch. The project also released PixMo, a novel dataset with detailed image captions, question-and-answer pairs, and 2D pointing data. The 72-billion parameter Molmo model demonstrated performance comparable to or exceeding that of several proprietary models on academic benchmarks.
  • Objaverse: A large-scale, open dataset of 3D objects. The initial release contained over 800,000 annotated 3D models. It was designed to advance research in generative models, 2D instance segmentation, and open-vocabulary object navigation.
  • Objaverse-XL: An expanded version of the Objaverse dataset, containing over 10 million 3D objects. This dataset was used to train Zero123-XL, a foundation model for 3D generation tasks such as image-to-3D and text-to-3D.
  • ProcTHOR: A platform that uses procedural generation to create large-scale, diverse, and interactive simulated 3D environments. By scaling up the diversity of training data, ProcTHOR was shown to significantly improve the generalization and performance of embodied AI agents on various tasks.
  • Phone2Proc: A system that generates semantically matched, simulated training environments from a short video scan of a real-world space taken with a smartphone. Training an object navigation agent in these custom-generated scenes was shown to improve sim-to-real transfer performance from 35% to 71%.
  • RoboTHOR: An open simulation-to-real embodied AI platform. The project involved creating computationally modeled digital twins of real-world apartment layouts, allowing researchers to study how well agents trained purely in simulation can transfer their skills to physical robots.
  • AI2-THOR: Deitke was a contributor to this interactive 3D simulation framework for visual AI research. The platform provides near-photorealistic indoor scenes where agents can navigate and interact with objects, enabling research in areas like visual navigation and task completion.

These projects have been published at major AI conferences and are widely recognized for their contributions to open science and the advancement of AI research.

Deitke has received several awards for his research contributions.

  • Best Paper Honorable Mention, CVPR 2025: Awarded for the paper "Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models."
  • Outstanding Paper Award, NeurIPS 2022: Awarded for the paper "ProcTHOR: Large-Scale Embodied AI Using Procedural Generation." This is a significant honor in the AI research community, given to a small fraction of the more than 10,000 submissions to the conference.
  • Outstanding Reviewer Award, CVPR 2023: Recognized for his contributions to the peer-review process for the Conference on Computer Vision and Pattern Recognition.

In addition to these awards, Deitke was a contributor to the second edition of Richard Szeliski's textbook, Computer Vision: Algorithms and Applications, published in 2022. [1] [2] [4] [5]

REFERENCES

HomeCategoriesRankEventsGlossary