PeopleofSuperintelligence

Researcher

Nimit Sohoni

Nimit Sohoni is an AI researcher at Cartesia whose work spans efficient training, robust classification and compact recurrent language models. He did his PhD at Stanford in Christopher Ré's Hazy Research lab, where he wrote a 2019 report on low-memory neural network training and, in 2020, introduced the GEORGE method for hidden subclasses. In 2022, he coauthored Monarch, a class of structured matrices that speeds up ViT and GPT-2 training by 2x, and the paper was named an Outstanding Paper Runner Up at ICML 2022. After two years as a quant at Citadel, he joined Cartesia and coauthored the Llamba models, distilled from Llama-3.x into the Mamba architecture, in 2025.

Now at
Cartesia

Timeline

  1. 2019

    Low-memory neural network training

    As a PhD student at Stanford in Christopher Ré's Hazy Research lab, he is first author of a 2019 technical report on how much memory neural network training needs. With Christopher Aberger, Megan Leszczynski, Jian Zhang and Ré, he profiles WideResNet and DynamicConv Transformer training and tests sparsity, low precision, microbatching and gradient checkpointing.

  2. 2019

    Near-optimal quasar-convex optimization

    In 2019, he coauthors a paper with Oliver Hinder and Aaron Sidford that gives near-optimal accelerated first-order methods for smooth quasar-convex functions, a class of nonconvex functions that includes the convex and star-convex ones. The paper also derives a lower bound on the gradient evaluations that any deterministic first-order method needs.

  3. 2020

    Finding hidden subclasses with GEORGE

    In November 2020, he is first author of "No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problems", with Jared Dunnmon, Geoffrey Angus, Albert Gu and Ré. It proposes GEORGE, which estimates unlabeled subclasses by clustering and uses them in a distributionally robust optimization objective to measure and reduce hidden stratification.

  4. 2022

    Monarch structured matrices

    In April 2022, he coauthors "Monarch: Expressive Structured Matrices for Efficient and Accurate Training" with Tri Dao, Beidi Chen, Christopher Ré and others. Monarch matrices are parameterized as products of two block-diagonal matrices, and they speed up ViT and GPT-2 training by 2x with comparable model quality.

  5. 2022

    An ICML outstanding paper runner-up

    The Monarch paper receives an Outstanding Paper Runner Up award at ICML 2022.

  6. By 2024

    From Citadel to Cartesia

    He works as a quant at Citadel for two years. By December 2024, he has joined Cartesia as an early member of its research staff.

  7. 2025

    Llamba distilled language models

    In February 2025, he coauthors Llamba with Aviv Bick, Tobias Katsch, Arjun Desai and Albert Gu: recurrent language models distilled from Llama-3.x into the Mamba architecture, using less than 0.1% of the training data typically used for models of similar size. In March 2025, Cartesia releases the weights for Llamba-1B, Llamba-3B and Llamba-8B.

  8. As of October 2026

    AI researcher at Cartesia

    He is an AI researcher at Cartesia. In January 2026, he appears on The Peterman Pod to discuss the differences between AI research and quant careers.

Career

  1. Since 2024CartesiaAI researcher
  2. Until 2024CitadelQuant

Education

  1. Stanford UniversityPhD

From the biography.

  1. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Stanford University. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.
  2. Google had the most people shaping AI every year from 2014 to 2025Includes Stanford University. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  3. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes Stanford University. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.
  4. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Stanford University. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Nimit Sohoni yet. Send us a free-licensed photo.

  1. [1904.10631] Low-Memory Neural Network Training: A Technical Report arxiv.org. Checked 22 October 2026.
  2. Hazy Alumni Life Updates (2024) · Hazy Research hazyresearch.stanford.edu. Checked 22 October 2026.
  3. Ex-Citadel Quant and AI Researcher On Breaking In, Tech vs Finance Careers – The Peterman Pod – Apple Podcasts podcasts.apple.com. Checked 22 October 2026.
  4. [1906.11985] Near-Optimal Methods for Minimizing Star-Convex Functions and Beyond arxiv.org. Checked 22 October 2026.
  5. [2011.12945] No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problems arxiv.org. Checked 22 October 2026.
  6. [2204.00595] Monarch: Expressive Structured Matrices for Efficient and Accurate Training arxiv.org. Checked 22 October 2026.
  7. ICML 2022 Awards icml.cc. Checked 22 October 2026.
  8. [2502.14458] Llamba: Scaling Distilled Recurrent Models for Efficient Language Processing arxiv.org. Checked 22 October 2026.
  9. Llamba: scaling distilled recurrent models for efficient language processing | Cartesia cartesia.ai. Checked 22 October 2026.