Nick Frosst is a Cohere co-founder whose research spans capsule networks, interpretable prediction and language models. Publishing as Nicholas Frosst, he helped develop ways to represent visual objects and expose how neural networks make predictions.[1][4][5]
Start with Neural Additive Models →
X Background
Career
Frosst co-founded Cohere with Aidan Gomez and Ivan Zhang. His personal biography describes it as an enterprise-focused foundation-model company, meaning a company that develops general-purpose models for business applications. Cohere’s leadership page lists him as co-founder. His X biography identifies him as the singer of Good Kid.[1][2][7]
Contributions
Representing parts and whole objects
In Dynamic Routing Between Capsules, Frosst, Sara Sabour and Geoffrey Hinton proposed grouping neurons into capsules that represent an object or one of its parts. A capsule encodes both the likelihood that something exists and properties of that thing. Lower-level capsules send predictions upward; agreement between predictions strengthens the corresponding higher-level representation. The 2017 paper tested this method on handwritten digits, including overlapping digits. It offered an alternative way to preserve relationships between parts of an image, rather than treating recognition as a collection of independent feature detections.[4]
Predictions people can inspect
Frosst co-authored Neural Additive Models, which separates a prediction into contributions from individual input features. Each feature gets a neural network, and the model combines their outputs. This lets a user inspect how changing one input changes the prediction while retaining nonlinear relationships that a straight-line model would miss. The paper reports competitive results on classification and regression datasets. Its value is a deliberately constrained model structure: an explanation follows from the calculation itself.[5]
Translation as a model task
The 2026 North Small Translate report lists Frosst among its co-authors. The team describes an open-weight translation model built on the Command A+ foundation, trained for translation, editing existing translations and estimating translation quality. It combines supervised training, preference optimization and reinforcement learning. The authors emphasize throughput and report results across 50 languages. This is collaborative research; the author list establishes Frosst’s participation without assigning him sole responsibility for the model.[6]
Tools, writing & talks
Writing, research & talks
Start here
Neural Additive Modelsarxiv.orgA useful introduction to interpretable prediction. The paper shows how separate neural networks for individual features can produce a combined prediction whose components remain visible.
Start with the additive-model paper to see a concrete problem, an understandable design choice, and its tradeoffs.
Co-author[5]- Dynamic Routing Between Capsulesarxiv.org
The original capsule-routing paper explains how agreement between representations of object parts can activate a representation of a whole object. Familiarity with neural networks helps with the equations.
Co-author[4] - North Small Translatearxiv.org
A technical report on a translation model trained for document translation and related editing tasks. It explains the training stages and evaluates multilingual performance.
Co-author[6]