Brandon Duderstadt

Brandon Duderstadt is a cofounder of Nomic and a coauthor of research on GPT4All, Nomic Embed, and Nomic Embed Vision. These projects address access to language models, reproducibility of text embeddings, and a shared representation space for images and text. His later work focuses on computational semiotics: studying how meaning spreads and changes through communication among people and AI systems.[2][5][6][7][9]

Duderstadt coauthored GPT4All: An Ecosystem of Open Source Compressed Language Models, submitted in November 2023, with Yuvanesh Anand, Zach Nussbaum, Adam Treat, Aaron Miller, Richard Guo, Ben Schmidt, the GPT4All Community, and Andriy Mulyar. The paper describes GPT4All's evolution from a single model into an open-source ecosystem intended to democratize access to LLMs. The authors frame accessibility as a problem distinct from model performance, citing costly infrastructure, restricted web interfaces, and unavailable code and technical reports.[5]

With Zach Nussbaum, John X. Morris, and Andriy Mulyar, he coauthored Nomic Embed: Training a Reproducible Long Context Text Embedder, first submitted in February 2024 and revised in February 2025. The report describes nomic-embed-text-v1, an English text embedding model with an 8,192-token context. Its reproducibility approach includes releasing curated training data alongside training code and model weights; the code and weights use an Apache 2.0 license. The authors report that the model outperforms OpenAI Ada-002 and text-embedding-3-small on the short-context MTEB and long-context LoCo benchmarks.[6]

Duderstadt, Nussbaum, and Mulyar extended that work in Nomic Embed Vision: Expanding the Latent Space, submitted in June 2024. The image embedding model has open code and weights and shares the same latent space as nomic-embed-text, bringing image and text representations into a common space for multimodal tasks.[7]

In Towards Computational Semiotics, Duderstadt proposes studying how signs acquire and change meaning by measuring their use. His research agenda treats humans, language models, and combined human–model systems as “sign producing machines.” Concrete proposals include a language model conditioned on geographic location and publication date, analysis of how training-data composition relates to generated sequences, and simulations of how messages change as they pass through populations of language models.[9]

An August 22, 2025 update to that essay announced the establishment of Calcifer Computing to construct a nooscope. Duderstadt describes this as a theoretical instrument for measuring the collective semiotic state of a population—an intended means of studying the spread and distortion of ideas. On his personal website, he describes founding Calcifer to focus on AI's long-term effects on the “epistemic environment.” He also lists teaching Introduction to Computational Semiotics and advising independent studies as a Whiting Fellow at Johns Hopkins, alongside early-stage investing with Factorial Capital.[9][4]

Before this computational semiotics agenda, Duderstadt's work included medical-imaging models and the Atlas prototypes around which he decided to build a company.[4][8]

Duderstadt says that before cofounding Nomic, he trained the first CT and MRI models at Rad AI. Historically, Latent Space introduced him as Nomic.ai's CEO in its December 2023 NeurIPS startup episode, where his segment covered GPT4All and Atlas.[4][1]

In Movement Composition: Starting Nomic, Duderstadt recalls spending months building Atlas prototypes in his spare time and saving money while apprehensive about leaving his job. After completing a prepared rooftop jump twice, he applied the same reasoning about moving from preparation to action to starting a company. His conclusion was: “If right now is the inevitable moment, and I'm serious about "jumping" to Atlas full time, I should stop jumping across rooftops.” He reports becoming self-employed three months later.[8]