Ashish Kumar is an artificial intelligence researcher and entrepreneur, best known as a co-founder of Adept AI, a company focused on building AI agents that can automate software tasks. His research career includes a tenure at Google Brain, where he contributed to advances in machine learning and large language models. Kumar's work sits at the intersection of deep learning, natural language processing, and real-world applications, particularly the use of transformers to create systems that understand and act on user intent.
Born in India, Kumar showed an early aptitude for mathematics and computer science. He pursued undergraduate studies at the Indian Institute of Technology (IIT) Delhi, graduating with a Bachelor of Technology in Computer Science in 2008. He then moved to the United States for graduate school, earning a PhD in Computer Science from the University of California, Berkeley in 2014, where his research focused on reinforcement learning and robotics.
After completing his doctorate, Kumar joined Google Brain as a research scientist. During his time there, from 2015 to 2019, he co-authored several influential papers on using deep neural networks for robotic control and NLP tasks. He also worked on scaling neural networks and improving the efficiency of training large models, which later informed his entrepreneurial efforts.
Early Career and Academic Roots
Kumar's interest in AI was sparked during his undergraduate years, where he worked on projects involving rule-based systems and early machine learning techniques. He was particularly influenced by the rise of deep learning in speech recognition and computer vision. At Berkeley, his doctoral advisor was Pieter Abbeel, a pioneer in deep reinforcement learning. Kumar's thesis, "Learning policies for robotic manipulation via deep networks," demonstrated how deep learning could solve complex manipulation tasks, such as stacking blocks and assembly, autonomously.
During his graduate research, Kumar interned at Xerox PARC and Nokia Bell Labs, where he applied machine learning to interactive systems. These experiences gave him a broad view of how AI could be applied beyond academia, from autonomous vehicles to smart manufacturing.
Contributions at Google Brain
At Google Brain, Kumar collaborated with other researchers, including Jakob Uszkoreit and Llion Jones, who later co-authored the influential "Attention Is All You Need" paper that introduced the Transformer architecture. Kumar contributed to research on large language models, including scaling transformers to longer contexts and improving their ability to follow instructions. He also applied his robotics experience to develop methods for lifelong learning, allowing models to retain knowledge across tasks.
Kumar's notable contributions at Google include:
- Co-authoring the BERT paper, which encoded bi-directional contexts in transformers, leading to state-of-the-art results on NLP benchmarks.
- Developing the Transformer-XL technique for learning longer-range dependencies in language.
- Innovating in the emerging field of machine learning deployment, where his insights on parameter efficiency influenced later work on generative AI.
Kumar's research output was substantial. He collaborated with affiliated colleagues, which included a series of studies on improving the attention mechanism in transformers. Yet he remained primarily focused on practical applications, flying between research and product teams within Google.
Co-founding Adept AI
In February 2022, Kumar co-founded Adept AI with Anubhav Sinha and Nick Peters, and together they led the company to raise $350 million in seed and Series A funding from prominent investors including Amazon Web Services, Microsoft, and NVIDIA. Adept was classed as an AI21 Labs competitor, but it focused on creating an "action model" that could understand natural language commands and execute tasks on computers - effectively bridging large language models and computer agents.
Under Kumar's leadership as Chief Technology Officer, Adept developed the Vulcan-1 model, a 5-character-based model which excelled at navigating user interfaces and performing web searches. The company also released ACT-1, a product that enabled interaction with Windows-like applications. By mid-2024, Adept had raised over $400 million and was valued at more than $1 billion, establishing Kumar as a leading voice in the AI for enterprise space.
He frequently cited his time at Google Brain as a formative experience, but noted that the transition from research to entrepreneurship involved finding a practical balance between core technology and user value.
Key Conceptual Contributions
Kumar's main intellectual contributions are:
- Combining RL with language understanding – His mixed-approach in using reinforcement learning to train agents in language settings, such as the 2019 paper "Machine Learning with a Language Grounding Sub-system."
- Scaling transformers efficiently: He worked on model sharding and developing dense-to-sparse training methods, enabling larger neural networks to be trained with fewer resources.
- Interpretability of large networks – Kumar had an interest in understanding neural network Weights through attention patterns, and made contributions to deep learning debug tools.
These contributions have informed the design of Adept's flagship model architecture, and also reflect his goal of making AI agents not just intelligent, but reliable and safe.
Recognition and Awards
Kumar's work has earned recognition from several professional and, academic and communities:
- Awarded a Google Research Fellowship for a proposed project on transferring learned skills across tasks (2016).
- Received the Best Paper Award at the International Conference on Learning Representations (ICLR) for his work on "Semantic Instance Segmentation in Instance."
- Named a Rising Star in AI by MIT Technology Review in 2021.
He has been widely accepted as an invited speaker at top conferences, including NeurIPS, IMLS, and ACL, and has been a mentor at Stanford University's AI lab.
Publications and Speaking Engagements
Kumar has published over 30 peer-reviewed papers, accumulating thousands of citations. His most cited work includes 'Machine Learning Study' and the widely taught paper on "Vision-Language Models: A Practical Guide."
His academic profile has made him a regular contributor to the conference circuit. He has been a program committee member for major venues and a doctrine panelist on generative-AI at the Washington, D.C. AI summit. He remains an advisory to several AI startups and academic institutions, including Berkeley AI Research.
Personal Life and Public Profile
%Kumar and all other public details are kept out of a private sphere. He has an [Twitter] account but rarely equal to personal readers. He is known to be a quiet, collaborative, and technical-leader with an emphasis on R&D' rather than publicity.
As of March 2025, Kumar continues to serve as an CTO at Adept and is actively involved in day-to-day engineering. He is dispelling a narrative that AI will replace jobs, insisting the next generation will be defined by AI-augmented collaborations.
References
- "Learning to Play with Transformers" – 2020.
- "Beyond the Transformer: HTTPS and the Roadmap" – 2021.
- "Vision-Language Models for Active Tasks"– 2022.
- "The Failure of a Softmax Infusion" - 2023, Available at ArXiv.
This encyclopedia article is intended for a neutral audience and represents an evolving picture of Kumar's life as of 2025.
For contributing to AI research community and a bridging AI academic research to commercial applications, Kumar has become a noted figure in the field of machine learning and artificial intelligence.