Insights from NeurIPS 2024
Reflections on the future of AI from NeurIPS 2024, covering inference-time compute, unconditional generation, hardware efficiency breakthroughs, and the evolving relationship between AI and creativity.
Jeffrey Urban
I enjoy working on problems where standard solutions don't yet exist.
My career has followed a pattern: identify hard problems early, design and build solutions, deploy to production, scale up, then move on to the next challenge. I think of this as:
None → 0 → 1 → n
I've built production systems for fleet-scale applications in automotive telematics, developer tools, edge AI, clean energy, and nuclear safety. I work across deep learning, large-scale data processing, cloud infrastructure, and embedded systems. My cross-domain expertise provides the most advantage in software for physical systems.
I was inspired by a conversation with Sepp Hochreiter following his talk at NeurIPS 2024 about the future of deep learning, that we will return to diversity of models, tailored to specific applications. I run Banche Labs, my personal studio exploring applications of deep learning and generative AI.
I care deeply about developer tools. I build practical utilities that enable my work, from terminal UI applications to remote system management tools. I'm drawn to interdisciplinary thinking, finding insights at the intersections between fields of study and application domains.
I'm among the first graduates of UT Austin's Master's in AI, where I served on the teaching staff for courses in AI ethics and generative modeling. When I'm not working, I enjoy driving sports cars, flaneuring in foreign cities, and attending academic and industry conferences.
I'm interested in opportunities in Boston, New York City, and between, or remote.
Mapping the Landscape of Deep Learning: Topic Discovery from Book Embeddings
I created a dataset of deep learning concept embeddings, extracted hierarchical topic structure, and created text and concept indexes across 158 books about deep learning comprising 50k book-pages.
An Intuitive Explanation of SGD vs Gradient Descent
An intuitive exploration of why Stochastic Gradient Descent often outperforms traditional gradient descent in machine learning optimization, from data efficiency, a focus on progress, and leveraging randomness.