Denis Yarats identifies himself as Perplexity’s co-founder and CTO in an indexed public LinkedIn profile. His earlier published work studies how reinforcement-learning systems can learn useful behavior directly from images, with public implementations including DrQ and DrQ-v2.[1][3][4]
Start with The perplexities of information retrieval →
GitHub 
X Background
Yarats’s GitHub profile describes doctoral study in AI at New York University. An indexed LinkedIn snapshot lists NYU from 2018 to 2022 and research in reinforcement learning and natural-language processing. Degree completion remains unverified. The LinkedIn introduction identifies him as Perplexity’s co-founder and CTO; its crawl was reported as September 2026 when checked on October 26. Direct LinkedIn access returned HTTP 999.[2][1]
Career
At Perplexity, Yarats works on an AI-native search product. In his Practical AI interview, he discusses how the company approaches information retrieval and the accuracy and validation of answers. The episode offers a different view of his work from the reinforcement-learning repositories: it concerns selecting and using information for a search experience. These sources document his research and his leadership role separately; they do not establish that the visual-control algorithms are components of Perplexity’s search system.[1][5]
Contributions
Learning from pixels without a separate world model
DrQ is a research implementation credited to Yarats, Ilya Kostrikov and Rob Fergus. Yarats and Kostrikov are marked as equal contributors. The project studies reinforcement learning directly from image observations: the agent sees pixels and must learn which actions earn rewards. DrQ adds image augmentation to the learning process, using transformed views to help regularize the model. The repository includes training instructions, evaluation logs and comparisons on image-based DeepMind Control Suite tasks. It provides inspectable code for the method, rather than only a reported score.[3]
Improving training efficiency with DrQ-v2
DrQ-v2 extends that research with changes to the learning algorithm, reward estimation, exploration schedule and implementation. Its authors are Yarats, Rob Fergus, Alessandro Lazaric and Lerrel Pinto. The project reports improved sample efficiency and training time on the control tasks it evaluates, including learning humanoid locomotion from pixel observations. Those are task-specific research findings, not evidence that an agent can handle every real-world visual problem. The published code makes the experiment’s dependencies and procedure available to other researchers.[4]
Ideas
The Practical AI conversation is the accessible starting point for understanding the search problem he is working on. Developers interested in the underlying research record can move to DrQ and then DrQ-v2, comparing what changes between the implementations. Running the research code requires the listed machine-learning and simulation dependencies, and the repositories preserve historical environments. Reading the method descriptions and training setup is useful even without reproducing the experiments.[5][3][4]
Tools, writing & talks
Writing, research & talks
Start here
The perplexities of information retrievalpracticalai.showA Practical AI interview about Perplexity’s search and answer-validation approach. It introduces the retrieval problem without requiring a reinforcement-learning background.
Start with the interview for a nontechnical introduction to his current search work, then use the code repositories for the earlier research.
Interview guest[5]