Daniel Bear

Daniel Bear is an AI researcher whose work includes self-supervised, multimodal models for domains where labeled data are difficult or impossible to collect. As of September 18, 2026, Noetik listed him as Executive Vice President and Chief AI Officer. His earlier research includes Counterfactual World Modeling, a collaborative framework for using one unsupervised visual model across multiple perception tasks.[2][4]

Bear endorses combining interdisciplinary teams with large, multimodal datasets of real human biology. In his undated Noetik team statement, he argues that AI can reveal “new, therapeutically crucial biology that would otherwise go unseen.” His emphasis is on preserving biological complexity in the data used to build models.[2]

He appeared alongside Noetik CEO Ron Alfa in an April 20, 2026 Latent Space episode about AI for cancer treatment development. The discussion centered on Noetik’s use of multimodal human-tumor data and models intended to help match treatments to patients.[1][2]

Before Noetik, Bear worked in experimental neuroscience and computational vision, and was founder and CEO of an unnamed stealth SSL startup. His Harvard and Stanford research provides context for his later work with biological data and self-supervised models.[2][3]

Bear earned an A.B. in Molecular and Cellular Biology and a Ph.D. in Neuroscience at Harvard. In Michael Greenberg’s lab, he studied how electrical signals induce lasting neuronal gene expression. For his doctorate in Bob Datta’s lab, he investigated how odor molecules become olfactory perceptions. Bear and his collaborators discovered a new kind of chemoreceptor protein and a sensory pathway whose activation elicits instinctual behavior.[2][3]

As a postdoctoral researcher in Dan Yamins’s lab at Stanford, Bear investigated whether neural networks given biologically realistic architectures and sensory experience could learn representations resembling those in animal brains. He built self-supervised, multimodal models for settings where collecting labels was difficult or impossible.[3][2]

In the June 2, 2023 preprint Unifying (Machine) Vision via Counterfactual World Modeling, Bear, Kevin Feigelis, Honglin Chen, Wanhee Lee, Rahul Venkatesh, Klemen Kotar, Alex Durango, and Daniel L.K. Yamins introduced a unified, unsupervised network that could be prompted to perform different visual computations.[4]

The framework combines two mechanisms. Structured masking encourages the prediction model to capture low-dimensional structure in visual data, exposing physical components of a scene through small sets of visual tokens. Counterfactual prompting compares predictions for real inputs with predictions for slightly modified inputs to derive visual representations without task-specific training. The authors reported readouts for keypoints, optical flow, occlusions, object segments, and relative depth on real-world images and videos.[4]