CMU professor, former Apple AI chief
Russ Salakhutdinov
Profile
Ruslan “Russ” Salakhutdinov is a UPMC Professor of Computer Science at Carnegie Mellon University and one of the quiet heavyweights of the deep learning revival. He did his PhD at the University of Toronto under Geoffrey Hinton in the mid-2000s, and the two co-authored Reducing the Dimensionality of Data with Neural Networks in Science (2006) — the deep autoencoder paper that, along with Hinton’s work on deep belief nets, convinced a skeptical field that training deep models was actually possible. If you’ve ever wondered which papers made “deep learning” a real term, that’s one of them.
He spent the rest of the 2000s and 2010s pushing generative models forward — Deep Boltzmann Machines, Semantic Hashing, and a long run of probabilistic model work that influenced how a generation of researchers thought about representation learning. In 2016 Apple pulled him in as its first Director of AI Research, a move that was widely read as Apple finally taking AI seriously. He stayed until 2020, built out the research org, and shaped the on-device ML direction that ended up running quietly inside hundreds of millions of iPhones.
After Apple, he returned full-time to CMU and pivoted hard toward multimodal agents — models that don’t just read text but click, scroll, and use tools on real websites. With his students he shipped VisualWebArena in 2024, one of the first serious benchmarks for vision-language agents on actual web tasks, and followed it with work on adversarial robustness of those same agents. In 2024 he joined Meta’s Superintelligence Lab as a VP of Research working on computer-use agents; as of late 2025 he wrapped up that stint and is back at CMU.
For developers learning AI, Salakhutdinov is worth following for two reasons. First, his CMU lecture courses — 10-707 Deep Learning and 10-703 Deep Reinforcement Learning — are some of the cleanest free material on the fundamentals, with slides available on his homepage. Second, his current research line on web-navigating agents is exactly where the “LLM + tools” story is going next, and his papers are readable without a PhD.
Key Articles & Papers
Reducing the Dimensionality of Data with Neural Networks Deep Boltzmann Machines Semantic Hashing An Efficient Learning Procedure for Deep Boltzmann Machines Dropout: A Simple Way to Prevent Neural Networks from Overfitting VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks Dissecting Adversarial Robustness of Multimodal LM AgentsSpotify Podcasts
