Ivan Zhang is a Cohere co-founder with a background in research infrastructure and software engineering. His published work includes unsupervised cipher learning, neural-network pruning and machine translation.[1][3][4][5][7]
Start with Unsupervised Cipher Cracking Using Discrete GANs →
GitHub 
X Background
Zhang’s résumé records computer science study at the University of Toronto from 2014 to 2016 and explicitly marks the degree incomplete. His early engineering work included Ranomics, where he built software for customers to access genomics datasets and worked on deployment infrastructure. He later developed backend services at Vision Critical / Pressly, including integrations with existing services and tools for fellow engineers.[3]
Career
Zhang co-founded the independent research group FOR.ai. His résumé describes designing cloud infrastructure for experiments on GPUs and TPUs and improving researchers’ workflows. At Cohere he is listed as co-founder, alongside Aidan Gomez and Nick Frosst. He and Gomez also co-authored the announcement of Cohere For AI, a research community intended to expand participation in machine learning.[1][3][6]
Contributions
Learning a cipher without paired examples
Zhang co-authored CipherGAN, a system for recovering a mapping between encrypted and ordinary text. The method learns from separate collections of ciphertext and plaintext, rather than receiving a matching translation for every example. The paper adapts a generative adversarial method to discrete symbols, addressing a training problem that arises when the inputs are words or characters rather than continuous images. Its experiments use shift and Vigenere ciphers. The contribution concerns learning structured mappings; it does not demonstrate breaking modern cryptography.[4]
Training a network to survive pruning
Targeted Dropout tackles the cost of redundant neural-network parameters. During training, the method selects weights or units that might later be removed and randomly drops members of that set. The network therefore learns while already experiencing the kind of damage that compression will cause. Zhang is a named co-author with a FOR.ai affiliation. The paper’s focus is a simple training strategy that makes later pruning less disruptive, rather than a new general-purpose model architecture.[5]
From research prototypes to translation
Zhang is a co-author of North Small Translate. The report’s author list identifies the team with Cohere. His credit here is paper co-authorship; it does not establish sole responsibility for the model.[7]
Tools, writing & talks
Writing, research & talks
Start here
Read this for a concrete example of learning from two unpaired collections. The cipher task makes the goal easy to grasp before the paper introduces adversarial training.
The cipher problem gives a concrete entry into Zhang’s research without requiring knowledge of enterprise AI.
Co-author[4]North Mini Code is Cohere's first open-source coding model. It is deliberately small and efficient, built for agentic performance and released for community iteration. Early responses contest how it benchmarks against rival small coding models.
Catalog workOriginal source- Targeted Dropoutopenreview.net
A short research paper on teaching networks to tolerate removal of weights or units. It is most useful after learning the basics of dropout and model compression.
Co-author[5] - Introducing Cohere For AIcohere.com
Zhang and Aidan Gomez announce a research community and explain its intended role in broadening participation in machine learning.
Co-author[6] - North Small Translatearxiv.org
A Cohere translation report listing Zhang as a co-author; read it for his participation in a larger research team.
Co-author[7]