Pushmeet Kohli
Pushmeet Kohli
Biographical profile of Pushmeet Kohli, VP of Research at Google DeepMind. Documents his role leading the AI Safety and Alignment team, his prior career at Microsoft Research where he worked on computer vision and learning systems, and his position as one of the senior research executives at DeepMind with responsibility for the organization's published safety work and the Frontier Safety Framework.
Quick Assessment
| Dimension | Assessment |
|---|---|
| Primary Role | VP of Research, Head of AI Safety and Alignment, Google DeepMind |
| Education | PhD in Computer Vision, Oxford Brookes / Oxford University |
| Prior Role | Principal Researcher, Microsoft Research Cambridge (computer vision, learning, optimization) |
| DeepMind Responsibilities | Oversight of DeepMind's safety, alignment, and responsible-AI research; senior author or co-author on the Frontier Safety Framework and related documents |
| Research Background | Computer vision, structured prediction, learning to optimize, scientific applications of ML |
Overview
Pushmeet Kohli is a senior research executive at Google DeepMind, where he holds the title of VP of Research and leads the AI Safety and Alignment team. He is one of DeepMind's most senior figures with explicit responsibility for safety, and his name appears as a senior author on DeepMind's Frontier Safety Framework documentation, AlphaFold-team publications, and a range of scientific-ML publications.
Kohli's background is in computer vision and structured prediction rather than the AI-safety-as-existential-risk tradition that motivated some of DeepMind's earlier safety hires. His role is thus partly a translation function — providing senior management coverage for safety work while bringing a mainstream-ML research-management background to it.
Background
Education and Early Career
Kohli completed his PhD on computer vision at Oxford Brookes University in collaboration with Oxford University, supervised by Philip Torr. His doctoral work focused on energy minimization methods for image segmentation and the related field of structured prediction — assigning labels to outputs with rich combinatorial structure (e.g., pixels in an image, words in a sentence) where the labels are interdependent.
Microsoft Research (2007–2017)
After his PhD, Kohli joined Microsoft Research Cambridge as a postdoc and was subsequently promoted to Principal Researcher. His decade at Microsoft Research produced a substantial body of work across:
- Computer vision: Object recognition, semantic segmentation, 3D reconstruction
- Structured prediction: Energy-minimization methods, graph cuts, dual decomposition
- Learning to optimize: Methods for training models that learn to solve optimization problems
- Probabilistic programming: Contributions to the broader probabilistic-programming research community
Kohli was a productive publication-record researcher during this period, with hundreds of papers and a high citation count in the computer-vision community.
DeepMind (2017–Present)
Kohli joined DeepMind in 2017 in a senior research role and was subsequently promoted to VP of Research. His DeepMind portfolio has expanded over time and now spans:
Scientific Applications
Kohli has been a co-author on multiple DeepMind scientific-ML papers, including AlphaFold, AlphaTensor (matrix-multiplication algorithm discovery), and FunSearch (mathematical-conjecture solving via LLM-guided search). His role on these papers is typically as a senior research-management author rather than a hands-on technical contributor, reflecting his oversight responsibility for the research-applications portfolio.
AI Safety and Alignment
In 2023, Kohli was given additional responsibility for the AI Safety and Alignment research function. This expanded role consolidates several previously distinct research groups including DeepMind's:
- Frontier safety / dangerous capability evaluations
- Alignment and scalable oversight research (including the team formerly led by Jan Leike before his 2021 departure)
- Mechanistic interpretability (under Neel Nanda)
- Responsible-AI and policy-adjacent research
He is a senior author on the May 2024 Frontier Safety Framework document, DeepMind's policy commitment to evaluating frontier models against critical capability thresholds and implementing safeguards accordingly.
Public Profile
Kohli is moderately publicly active — he gives conference talks and academic-venue presentations, but is less media-engaged than peers like Demis Hassabis or Allan Dafoe. His public communications focus on technical research content and DeepMind's published positions on safety.
Critics inside the AI safety community have at times raised concerns about whether a researcher whose background is in computer vision can credibly lead a function focused on x-risk-relevant safety research. Defenders argue that senior management of the safety function does not require deep prior immersion in the existential-risk literature — what it requires is institutional weight, technical credibility, and a willingness to push back against capability-side pressure.
See Also
- Google DeepMind
- Neel Nanda — Mechanistic interpretability team lead reporting through Kohli's organization
- Victoria Krakovna
- Rohin Shah