Photo headshot of Department Researcher Ivan Bercovich.
Researcher
ibercovich@ucsb.edu
Personal Website
ScOp Venture Capital
CS Capstone
Education
B.S. Electrical Engineering and Mathematics, University of Massachusetts Amherst
Ph.D. Statistics and Applied Probability, University of California, Santa Barbara (Unfinished)
Campus Affiliations
Co-Head of Capstone, Project Design & Development (Refer to website shared above)
Bio

Ivan currently makes investments through his fund, ScOp Venture Capital. Prior to investing full time, he spent ten years at Graphiq, a startup eventually sold to Amazon, where he remained for three years until 2020. At Graphiq, his team built one of the largest general knowledge graphs and developed natural language understanding and generation technology, which became Alexa's pre-LLM Q&A engine. At Amazon he managed a team of 140, which gave him a lot of time to think about and experiment with management and organizational approaches to optimize product and software development and foster innovation. He holds a B.S. in Electrical Engineering and Mathematics from the University of Massachusetts Amherst and was briefly a Ph.D. student in Financial Mathematics at UC Santa Barbara.

Research

Since 2024, Ivan has been primarily interested in the big implications of AI and how it might impact society in ways we haven't prepared for. He has gone back into research mode and collaborated on many academic publications, focused on the problem of reward hacking in agentic environments: its implications, detection, and mitigations. He leads Terminal Wrench, a dataset of reward-hackable environments and exploit trajectories. He has also concluded that benchmarking is essential both to advance AI and to do so safely. The speed of advancement and the voracious appetite for RL environments (environments and benchmarks are essentially the same thing) have made it increasingly difficult to maintain a high bar. As a core author of Terminal Bench, focused specifically on task quality, he has looked at hundreds of tasks and developed intuition for the limits of current approaches. He believes verifier quality is one of the most important problems in AI in 2026.