Andy Konwinski

Andy Konwinski is a computer scientist and co-founder of Laude Institute, Databricks and Perplexity. His AI work spans product leadership, coding-agent evaluation and research funding. He committed $100 million to launch Laude Institute and organized the Konwinski Prize to encourage open-source AI systems that solve real software issues, with a test set to be collected after the submission deadline.[2][4][6]

Konwinski launched Laude Institute on June 23, 2025, with fellow co-founders Chris Rytting, K. Tighe, Justin Fiedler and Lindsey Gregory. His $100 million pledge anchored its funding. Laude combines grantmaking with hands-on support and operates through open publication, open source and open discourse.[4][3]

At launch, Konwinski described two forms of support: Slingshots, providing fast grants and hands-on help for early-stage research, and Moonshots, backing long-horizon labs in areas such as scientific discovery, healthcare and workforce reskilling. He linked Moonshots to Shaping AI, a project he co-authored with Dave Patterson, Jeff Dean, John Hennessy and others. His stated aim was to help researchers move discoveries into use: “breakthrough ideas matter, but only if you ship them.”[4]

At Laude, Konwinski worked within the collaboration that co-created Terminal Bench and Harbor with Ludwig Schmidt’s group at Stanford. He also co-founded Laude Ventures, where Laude’s directory lists him as a general partner, alongside his institute co-founder and board roles.[2][3]

Before Laude’s launch, Konwinski announced the Konwinski Prize at NeurIPS in December 2024. He offered $1 million to an open-source AI that “breaks 90%” on a new version of SWE-bench, which tests systems on real issues from GitHub repositories. He restricted the competition to open-source code and open-weight models, explaining that he wanted participants to build on one another’s work.[6]

Konwinski’s concern was that SWE-bench publishes its test set, allowing submissions to train on it; the underlying public repositories could also appear in model training data. He assembled a team to adapt SWE-bench’s collection pipeline and implement evaluation on Kaggle’s infrastructure. The planned safeguard was to collect a new test set after the submission deadline, reducing exposure to the competition’s test examples rather than relying on an already public set.[6]

He credited Chris Rytting, Justin Fiedler, Alex Shaw, K. Tighe and Lindsey Gregory as the core K Prize team, Kaggle as a partner, and SWE-bench creators John Yang and Carlos E. Jimenez for advice and help with the collection pipeline.[6]

In a July 2, 2026 essay, Konwinski argued that concentration of AI power should itself be treated as a risk. He advocated a research commons spanning academia, industry and the public interest, with frontier-scale compute, access to advanced models, research talent and operational support. His safety argument emphasized researchers’ ability to challenge assumptions and identify mistakes: “Safety is not secrecy. Safety is error correction.”[5]

Konwinski acknowledged that more capable AI could enable cyberattacks, biological-weapons development and other severe harms, and explicitly rejected “openness at all costs.” He also conceded that even a well-funded commons would not match the largest private labs’ budgets. He argued that this relative scarcity could encourage work on architectures, alignment techniques and evaluations those labs might overlook.[5]

Konwinski’s earlier work provides context for his later emphasis on deploying research. He earned a computer-science PhD at UC Berkeley. In his Laude launch account, he recalled arriving in 2007 and working with Matei Zaharia during his first year to improve Hadoop’s compute engine and contribute the work back to open source. He described that experience as changing his understanding of what it meant to be a computer scientist.[2][4]

He co-created Mesos and co-authored its 2011 NSDI paper with Benjamin Hindman, Matei Zaharia, Ali Ghodsi, Anthony D. Joseph, Randy Katz, Scott Shenker and Ion Stoica. The related May 26, 2010 Berkeley report lists him as Andrew Konwinski; the USENIX conference record uses Andy Konwinski.[2][7][8]

Mesos addressed sharing a cluster among different computing frameworks, such as Hadoop and MPI. The team’s design used two-level scheduling through resource offers: Mesos decided how many resources to offer each framework, while each framework chose which resources to accept and which computations to run. This retained framework-specific scheduling decisions while allowing frameworks to share the same machines.[7]

During graduate school, Konwinski also worked at Google on the Omega cluster scheduler and was part of the team that created Spark. He subsequently co-founded Databricks, served as VP Product AI/ML, created the Data+AI Summit and co-authored Learning Spark. At Perplexity, he was a co-founder, one of the first investors and the founding president.[2]