Aravind Srinivas

Research

Srinivas's 2021 Berkeley dissertation, "Representation Learning for Perception and Control", is a thesis by articles on two themes: self-supervised representation learning and neural network architectures built on self-attention. It collects the papers on data-efficient image recognition with contrastive predictive coding (CPCv2), CURL, reinforcement learning with augmented data (RAD) and Bottleneck Transformers.³

CURL uses contrastive learning to make reinforcement learning from raw pixels more sample efficient, while BoTNet studies self-attention in place of convolutions in vision backbones. Srinivas is also a co-author of the 2021 Decision Transformer paper, which frames reinforcement learning as sequence modeling with a Transformer.⁵⁶⁷

Perplexity

According to an AWS Startups article written with Perplexity's team, Srinivas worked as an AI researcher at OpenAI before co-founding Perplexity in 2022 with Denis Yarats, Johnny Ho and Andy Konwinski. Perplexity builds an answer engine that responds to questions with cited answers.²

In March 2026 NVIDIA announced the Nemotron Coalition of AI labs working on open frontier models, and quoted Srinivas, as cofounder and CEO, on Perplexity contributing its model development work to the coalition.¹

Work and contributions

• CURL Published research

  First author, with Michael Laskin and Pieter Abbeel, of "CURL: Contrastive Unsupervised Representations for Reinforcement Learning" (2020); CURL is one of the articles presented in Srinivas's Berkeley dissertation.⁵³

• Bottleneck Transformers (BoTNet) Published research

  First author of "Bottleneck Transformers for Visual Recognition" (2021), which replaces the spatial convolutions in the final three bottleneck blocks of a ResNet with global self-attention; the paper is one of the articles presented in Srinivas's Berkeley dissertation.⁶³

Timeline

1. May 2019Co-authored "Data-Efficient Image Recognition with Contrastive Predictive Coding" (CPCv2).⁸
2. Apr 2020Published CURL, on contrastive representations for reinforcement learning, as first author.⁵
3. Jan 2021Published "Bottleneck Transformers for Visual Recognition" as first author.⁶
4. Jun 2021Co-authored "Decision Transformer - Reinforcement Learning via Sequence Modeling".⁷
5. Aug 13, 2021Completed the Berkeley EECS dissertation "Representation Learning for Perception and Control".³
6. 2022Co-founded Perplexity with Denis Yarats, Johnny Ho and Andy Konwinski.²
7. Mar 2026Represented Perplexity in NVIDIA's announcement of the Nemotron Coalition for open frontier models.¹

Other documented roles

Roles the sources document that are not confirmed as current.

• AI researcher, OpenAI²⁴
• PhD student in computer science (EECS), advised by Pieter Abbeel, University of California, Berkeley³⁴

Where they work

• Perplexity 3 people in the wiki

Related people

• Denis YaratsAlso at Perplexity
• Andy KonwinskiAlso at Perplexity
• Johannes HagemannReinforcement learning · Language models
• Niki ParmarLanguage models · Computer vision
• Bob McGrewReinforcement learning · Language models
• Noam BrownReinforcement learning · Language models
• Lilian WengLanguage models · Reinforcement learning
• Wojciech ZarembaReinforcement learning · Language models
• Weizhu ChenLanguage models · Reinforcement learning
• Nick FrosstLanguage models · Computer vision