Matan Grinberg

Matan Grinberg co-founded Factory with Eno Reyes to bring autonomy to software engineering. His approach centers on building useful products with existing foundation models, automating work developers want to delegate, and judging results across an engineering organization. The two appeared together on Latent Space’s The AI Coding Factory on May 29, 2025; ELC’s speaker biography identifies Grinberg as Factory’s CEO and co-founder.[1][2][3]

In his Training Data interview, Grinberg explained the decision to build on models from established labs rather than compete in foundation-model training. Factory could benefit as those models became better, cheaper, and faster while concentrating on products engineers actually used. He treated fine-tuning and training as capabilities to earn through product success: “If you’re unable to build a product that people are actually using, with these incredible models, then chances are fine tuning and training will not save you.”[3]

Grinberg described Factory’s focus on code review, documentation, testing, debugging, and refactoring as deliberate. In his view, automating only the enjoyable work of creating software would leave engineers with the maintenance tasks they least wanted to do. His lesson from building autonomous tools was succinct: “Engineers love ownership.” He argued that Factory should remain aligned with developers as their roles evolved, rather than make them defensive about automation.[3]

For Grinberg, strong model reasoning alone does not make an agent productive inside an enterprise. He illustrated this with a hypothetical Jeff Dean entering an unfamiliar legacy codebase: even an exceptionally capable engineer needs to understand dead code, past decisions, engineering practices, communication, and organizational priorities. The requirement is a deep understanding of the engineering system, extending beyond the code itself.[3]

Factory divided that work among task-specific autonomous systems called droids. Reyes explained the technical approach as inference-time reasoning and decision search, with different cognitive architectures for different tasks. He described balancing rigid entry points and consistent error recovery with flexibility during problem solving. This was Reyes’s explanation of the company’s architecture, alongside Grinberg’s account of its product choices.[3]

Grinberg offered Factory’s own GitLab integration as a concrete example of delegation. He recalled that the team had postponed the integration because it wanted to work on AI instead. Its code droid specified the implementation steps and completed the resulting subtickets while the team monitored it for breakage. He presented this as work they had considered assigning to an intern, demonstrating planning and implementation with human oversight.[3]

Grinberg favored organization-wide outcomes over intermediate measures such as autocomplete acceptance rates. Reyes described how a faster individual task could still produce churn or block other engineers, pointing to measures such as open-to-merge time and code churn. Grinberg distilled the goals into saving engineering time, increasing speed, and improving code quality, with engineering velocity as the business-facing objective. These measures, he argued, gave engineering leaders a clearer way to explain value to executives.[3]

Asked about competition among applications using the same models, Grinberg emphasized execution speed and sustained commitment to the mission: “The only way you can really win is by executing faster and being more obsessed.” He also emphasized collaboration in AI rather than individual genius, praising Mistral’s open-source mission and the speed of its model development.[3]

Before Factory, Grinberg worked in theoretical physics at Princeton and UC Berkeley, studying string theory. Recalling his Princeton undergraduate years, he said he emailed Juan Maldacena at the Institute for Advanced Study seeking an adviser. After a chalkboard meeting, he worked overnight on a problem Maldacena had introduced; according to Grinberg, solving it led to the advisory relationship and a joint paper.[2][3]

Grinberg and Maldacena co-authored Proper time to the black hole singularity from thermal one-point functions, first submitted on November 2, 2020. The paper argues that the proper time from a black-hole horizon to its singularity can be extracted from thermal expectation values of certain operators outside the horizon. Its proposed method uses fields coupled to higher-curvature terms and requires varying the field’s mass.[4]

Grinberg recalled realizing about a year into his Berkeley PhD that he was pursuing theoretical physics because it was difficult rather than because he loved it. His research shifted from physics toward machine learning in physics, then increasingly toward AI. He sought advice from investor and former string theorist Shaun Maguire, whose career offered an example of moving beyond academia. In parallel, Grinberg reconnected with Reyes at a LangChain hackathon. He recalled subsequently getting coffee with Reyes and beginning sustained collaboration on code generation and what became Factory.[3]