Nick Frosst
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v1 · 2,167 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v1 · 2,167 words
Add missing citations, update stale details, or suggest a clearer explanation.
Nick Frosst is a Canadian artificial intelligence researcher, entrepreneur, and musician. He co-founded the enterprise AI company Cohere in 2019 with Aidan Gomez and Ivan Zhang, and he leads research at the firm. [1][2] Before that he was the first employee hired into Google Brain's Toronto office, where he worked with Geoffrey Hinton on capsule networks, on methods for distilling neural networks into interpretable models, and on adversarial examples. [3][4] Frosst is also the lead singer and songwriter of the Canadian indie band Good Kid, which he formed with fellow University of Toronto students and which has grown into a successful act in its own right. [5][6]
Frosst was born on January 5, 1993. [6] He is based in Toronto, the city where Cohere and Good Kid both started. [3][5] He has become a recognizable voice in Canadian technology, both for his research record under Hinton and for arguing in public that large language models are useful tools rather than systems on the verge of human level general intelligence. [7][8]
Frosst studied at the University of Toronto, where he was enrolled at Woodsworth College and earned a bachelor of science in computer science and cognitive science, graduating in 2015. [4][6] His path into research was unusual. While working at a board game cafe in the Koreatown neighbourhood of Toronto during his undergraduate years, he fell into a conversation with a customer about the computability of a game, and that exchange led to an introduction to the York University vision researcher John Tsotsos and a position as an undergraduate research assistant. [4]
He later connected with the Google Brain team while doing software work for Google in Waterloo and spending part of each week in Toronto, and he deliberately introduced himself to Geoffrey Hinton, who was building a research group in the city. [4] Frosst has said the experience taught him a great deal about how to design an experiment and how to talk to others about the work he was doing. [4]
Frosst was among Hinton's earliest hires at the Google Brain laboratory in Toronto, and several profiles describe him as the office's first employee. [3][7][9] He worked there as a machine learning researcher from roughly 2016 to 2020, on a small team led by Hinton that focused on fundamental questions about neural networks rather than near term products. [4][7] His research concentrated on three connected areas: capsule networks, interpretability, and adversarial examples. [9]
Capsule networks were one of Hinton's long running ideas, an attempt to build vision models whose internal units, called capsules, encode the properties of an object or object part and reach agreement about what is present in an image rather than only detecting the presence of features. Frosst was the middle author, between Sara Sabour and Hinton, of the 2017 paper "Dynamic Routing Between Capsules," presented at the Neural Information Processing Systems conference, which set out a routing by agreement procedure and reported strong results on the MNIST digit benchmark and on recognizing overlapping digits. [10] He also contributed to follow up work on capsule routing. [11]
On interpretability, Frosst was the first author with Hinton of the 2017 paper "Distilling a Neural Network Into a Soft Decision Tree," which proposed taking the knowledge captured by a trained neural network and transferring it into a soft decision tree. [12] The aim was to make individual decisions easier to inspect by relying on a hierarchy of choices rather than distributed representations, a form of knowledge distillation aimed at explanation rather than only at compression. [12] He returned to interpretable models in the 2021 paper "Neural Additive Models," written with Rishabh Agarwal, Rich Caruana, Hinton, and others, which fit a separate neural network to each input feature so that the contribution of each feature can be read off directly. [13]
Frosst's work on adversarial examples, the small and often imperceptible input changes that can fool a classifier, also drew on the capsule idea. With Yao Qin, Sara Sabour, Colin Raffel, Garrison Cottrell, and Hinton he published "Detecting and Diagnosing Adversarial Images with Class-Conditional Capsule Reconstructions" at the 2020 International Conference on Learning Representations, and a related paper, "Deflecting Adversarial Attacks," appeared the same year. [14][15] In a separate line of research he was the first author, with Nicolas Papernot and Hinton, of the 2019 paper "Analyzing and Improving Representations with the Soft Nearest Neighbor Loss," which introduced a measure of how entangled the representations of different classes are inside a network and showed how adjusting that entanglement can improve generalization and uncertainty estimates. [16]
Frosst left Google to start Cohere in 2019 with Aidan Gomez, whom he had met on the Toronto team, and Ivan Zhang. [1][3] Gomez had been one of the eight authors of the 2017 transformer paper "Attention Is All You Need," the architecture that underlies modern large language models. [2][17] The founders have said they left because the research breakthroughs they were close to were not yet reaching the people who might build products with them. [1]
Cohere develops large language models and sells them to businesses rather than to consumers. [1][2] The company has positioned itself around data security and deployment inside a customer's own cloud or on its own servers, an approach pitched at regulated industries such as finance and healthcare that are reluctant to send sensitive data to outside systems. [7][18] Its model families include the Command series of text generation models and the Embed and Rerank models used for search and retrieval. [2] In 2025 the company released North, an agent platform that bundles its models with search and customizable AI agents and can run in private or air gapped environments. [19]
Within Cohere, Frosst leads research. [2][3] Reporting in 2025 placed the company among the best funded artificial intelligence developers in Canada. In August 2025 Cohere said it had raised 500 million dollars at a valuation of about 6.8 billion dollars, with investors including Radical Ventures, Inovia Capital, AMD Ventures, Nvidia, the pension manager PSP Investments, and Salesforce Ventures, and a further close the next month lifted the valuation to about 7 billion dollars. [19][20] Geoffrey Hinton was among the company's early backers. [1]
Frosst's research sits at the meeting point of deep learning and interpretability. [9][12] His recurring concern is making neural networks easier to understand and to trust, whether by distilling them into decision trees, by fitting additive models whose parts can be read individually, or by studying how class representations are arranged inside a network. [12][13][16] Capsule networks and adversarial robustness were central to his early work, and they connect to the same theme, since both ask what a model has actually learned to represent. [10][14] Since joining Cohere his published work has shifted toward large language models, including studies of training data and benchmarks and contributions to the multilingual Aya family of models released by Cohere's research lab. [21][22] He has also written on the limits and risks of language models, including a critique of a widely used language benchmark and work on filtering training data to reduce harmful outputs. [21][23]
Frosst is the lead singer of Good Kid, an indie band he formed in Toronto in 2015 while he and his bandmates were studying at the University of Toronto, most of them in computer science. [5][6][24] The group began as a hobby, and its other members are Jonathon Kereliuk on drums, Michael Kozakov on bass, and David Wood and Jacob Tsafatinos on guitar. [6][24] Their first single, "Nomu," came out in October 2015, and their first extended play, also titled Good Kid, followed in June 2018. [6] The band's music is usually described as indie rock or power pop, and Frosst has cited Two Door Cinema Club and Bloc Party as influences. [5][6]
Good Kid gained a wider following after several of its songs were used in streams of the video game Fortnite around 2021. [6][24] The band has since played the Lollapalooza festival in Chicago in 2024, toured internationally, and was nominated for the Juno Award for breakthrough group of the year in 2024. [5][6] By 2026 it had reached several million monthly listeners on streaming services, and it released its debut studio album, "Can We Hang Out Sometime?," on April 3, 2026. [5][6] Frosst has described Cohere as his life's work and music as something he does to unwind, and he has said the two are additive, with the band practicing twice a week and members often putting in a full day of remote work before evening shows on tour. [5]
Frosst is publicly skeptical of the idea that artificial general intelligence is near, and he has argued that AI will streamline routine tasks rather than replace people. [6][8] He has taken part in onstage discussions with Hinton in which the two describe themselves as occupying different ends of the spectrum on artificial intelligence risk. [7] When Hinton argued that large language models are approaching a form of consciousness, Frosst responded that such models are more conscious than a rock but less conscious than a tree, a line he used to push back on strong claims about machine awareness. [7] He has also called the scenario of a language model creatively breaking security systems unlikely. [7] On the prospects for the field in Canada, Frosst has been more optimistic than Hinton, saying that the country invented the technology and has every right to be a leader in it. [7]
Frosst and his Cohere co-founders have appeared on Canadian lists recognizing leading figures in artificial intelligence, including Maclean's AI Trailblazers list and The Logic's ranking of innovation leaders. [6] His research papers with Hinton, especially the work on capsule networks and on distilling networks into soft decision trees, are widely cited in the deep learning literature. [10][12]
| Item | Detail |
|---|---|
| Full name | Nicholas Frosst |
| Born | January 5, 1993 |
| Nationality | Canadian |
| Education | University of Toronto (Woodsworth College), BSc in computer science and cognitive science, 2015 |
| Known for | Co-founding Cohere; capsule networks and interpretability research with Geoffrey Hinton; lead singer of Good Kid |
| Google Brain | First employee, Toronto office, about 2016 to 2020 |
| Company co-founded | Cohere (2019, with Aidan Gomez and Ivan Zhang) |
| Role at Cohere | Co-founder; leads research |
| Notable papers | "Distilling a Neural Network Into a Soft Decision Tree" (2017); "Dynamic Routing Between Capsules" (2017); "Analyzing and Improving Representations with the Soft Nearest Neighbor Loss" (2019); "Neural Additive Models" (2021) |
| Music | Lead singer and songwriter of Good Kid (formed 2015, Toronto) |
| Recognition | Maclean's AI Trailblazers; The Logic Innovation Leaders |