Greg Kamradt

Greg Kamradt is president and a board member of ARC Prize Foundation. His AI work spans long-context retrieval evaluation through Needle in a Haystack, helping run ARC Prize, and explaining interactive evaluations of agents. He also runs Leverage with a team of independent AI experts, helping businesses build with AI.[4][7][8][10][3][12]

Kamradt previewed ARC-AGI-3 onstage at the AI Engineer World’s Fair in 2025 and authored its March 25, 2026 release announcement. The first fully interactive benchmark in the ARC-AGI series uses turn-based environments handcrafted by a team of human game designers. Agents receive no instructions, rules, or stated goals: they must explore, discover how the environment works and what winning means, then carry their learning across increasingly difficult levels.[2][5]

The benchmark evaluates learning through action, rather than only a final answer. Its score definition makes efficiency consequential: 100% means an agent can beat every game as efficiently as humans. In his announcement, Kamradt framed the challenge as the gap between AI that follows instructions and AI that can “genuinely explore, learn, and adapt in unfamiliar situations.”[5][6]

The ARC Prize 2025 technical report, submitted January 15, 2026, credits François Chollet, Mike Knoop, Gregory Kamradt, and Bryan Landers as authors. They identify refinement loops as a defining development of 2025: repeatedly optimizing a program for a particular task using feedback, through approaches such as evolutionary program synthesis and application-layer refinement of commercial AI systems. They also argue that frontier reasoning remains constrained by knowledge coverage, creating new forms of benchmark contamination.[9]

Kamradt’s involvement began before those interactive evaluations. In his October 2024 account, he recalled contacting Zapier cofounder Mike Knoop for a recorded conversation after hearing him speak about internal AI adoption. They stayed in touch, and Knoop later recruited him to help run a $1 million ARC-AGI competition with François Chollet. Kamradt accepted the opportunity to work on a small, autonomous team and partnered with Bryan Landers to help run the competition.[10]

Kamradt described the team’s work as extending beyond awarding prizes: preparing technical guidance, maintaining data, testing new models, validating public leaderboard submissions, managing the competitor community, and visiting university AI programs. He explicitly credited Landers with the ARC Prize website’s design and frontend work.[10]

For Kamradt, benchmarks also shape what researchers pursue: “Benchmarks also serve as a compass for AI research.” He distinguished mastery of a familiar skill from acquiring new skills outside training data, using Deep Blue’s strength at chess as an example of the difference. He nevertheless qualified the ambition: solving ARC-AGI would not itself constitute AGI.[10]

He described ARC Prize as a change of pace from work between the AI API layer and customer value toward more fundamental AI concepts. Nonprofit work also offered a goal besides revenue, which he found refreshing. That shift accompanied continued application work through Leverage.[10][3][12]

Before joining ARC Prize, Kamradt’s evaluation work included Needle in a Haystack, whose license names him as its 2023 copyright holder. The repository preserves historical runs for GPT-4-128K on November 8, 2023, and Claude 2.1 on November 21, 2023. The test varies both context length and the position of information within that context, making retrieval performance at different locations visible.[7][8]

The current v2 toolkit sweeps combinations of context length and needle depth, scores responses, and records one JSONL result per combination. It supports single-fact lookup, multi-fact recall, UUID retrieval, and chains of UUID links requiring multiple retrieval steps. These are specific retrieval and reasoning tasks. The project also documents a v2 correction to reported depths in older multi-needle tests, where earlier insertions had inflated the token count.[7]

His earlier career supplied experience in analytics and product operations. In Value Through AI, Kamradt recalled starting in corporate financial planning and analysis at NetApp, discovering that he preferred analytics to finance, and attending the Zipfian/Galvanize Data Science Bootcamp in 2015. He then led growth teams for Salesforce’s Sales and Service Clouds. After four years, he left and later joined Digits as its first business/operations hire, working across monetization, product analytics, finance, and a six-month interim product-management role.[11]

Kamradt positioned his AI interests “between The Model and Customer Value,” emphasizing concrete B2B applications that create value for users. Teaching was part of that work: he presented Introduction To Building AI Applications: LangChain 101 at Devoxx Belgium in October 2023.[11][13]