The Researcher's Field Guide: Papers, Labs, People, and How to Get Hired

Born in Nizhny Novgorod, Russia; family emigrated to Israel then Toronto. Sutskever went to the University of Toronto as an undergraduate knowing little about AI — he met Geoffrey Hinton there, who gave him a research project improving the Stochastic Neighbor Embedding algorithm. He completed his bachelor's, master's, and PhD in computer science at Toronto under Hinton's supervision.

In 2012, Sutskever built AlexNet with Hinton and Alex Krizhevsky — the ImageNet-winning CNN that established deep learning as computer vision's dominant paradigm and, more broadly, the practical starting point of the modern deep-learning era. After a short postdoc with Andrew Ng at Stanford, he returned to Toronto and joined Hinton's spinoff, DNNResearch — which Google acquired four months later, hiring Sutskever as a research scientist at Google Brain. He went on to co-found OpenAI, become its Chief Scientist, and later found SSI.

As detailed in this site's full SSI investigation, Sutskever's post-OpenAI research agenda centers explicitly on generalization: he has publicly argued that current LLMs generalize "dramatically worse than people," and has proposed continual learning via an internal "value function" as the mechanism to close that gap — a direct, career-spanning line from AlexNet's supervised pattern-recognition breakthrough to a research bet that current pattern-recognition-based learning itself needs a successor.

Pure maths undergrad at Cambridge (graduated 2020), with quant-finance internships at Jane Street and Jump Trading along the way. After graduating, he weighed quant finance against grad school — but having spent years reading about AI catastrophic-risk arguments, he wanted to work on reducing them directly. He took an offer to work on mechanistic interpretability with Chris Olah at Anthropic, worked there as a language-model interpretability researcher, then did a period of independent interpretability research before joining Google DeepMind's mechanistic interpretability team in 2022.

He joined DeepMind expecting to be an individual contributor — when the team lead stepped down, he stepped up with no prior management experience and ended up running the team at 26. A specific, concrete practice he credits for his early impact: he committed to writing one blog post per day for a month specifically to break through perfectionist paralysis — those posts helped seed the field of mechanistic interpretability as it exists today.

EE diploma from ESIEE Paris (1983), PhD in computer science from Sorbonne Université (1987), followed by a postdoc at the University of Toronto — putting him, like Sutskever, through the Toronto deep-learning lineage early in his career. He joined AT&T Bell Labs in 1988, becoming head of its Image Processing Research Department by 1996, where his work on convolutional networks became foundational to modern computer vision.

He joined NYU in 2003, founding its Center for Data Science (2011-2014), then became founding director of Facebook AI Research in 2013 and Meta's Chief AI Scientist from 2018. Over that period he became an increasingly vocal advocate for self-supervised learning, world models, and energy-based frameworks as an alternative to purely generative, language-model-centric AI. He left Meta in 2025 to found AMI Labs (Advanced Machine Intelligence) in Paris, built explicitly around his world-models research agenda — AI that understands the physical world, not just text.

Undergraduate degree from the University of Freiburg, Germany; PhD in reinforcement learning theory from the Australian National University. He worked as an alignment researcher at DeepMind, where he helped prototype RLHF, before joining OpenAI in 2021 — where he rose to co-lead the Superalignment team, a group explicitly tasked with solving superintelligence alignment within four years, and contributed to InstructGPT, ChatGPT, and GPT-4's alignment work.

He left OpenAI and joined Anthropic in May 2024, the same month, to lead its Alignment Science team. He's best known as a co-author of the 2017 "Deep Reinforcement Learning from Human Preferences" paper — the foundational work behind RLHF, meaning he helped invent the technique he then spent years applying and refining at two different frontier labs.

BS in Physics and Mathematics from Vanderbilt University, an MPhil in theoretical condensed matter physics from Cambridge, then a PhD in Chemistry from the University of Chicago — applying machine-learning techniques specifically to protein dynamics. In between, he spent three years at D.E. Shaw Research building molecular-dynamics simulations, before returning to Chicago in 2011 to apply ML directly to the physics of protein folding, completing his PhD there in 2017.

He joined Google DeepMind in 2017 as a research scientist and led development of AlphaFold2 and AlphaFold3, building new ML architectures specifically to let AI systems learn from biological and chemical data at a level of accuracy that transformed structural biology. He shared the 2024 Nobel Prize in Chemistry for this work, and — notably, as of 2026 — has since joined Anthropic after his time at DeepMind.