Russell Kaplan

Russell Kaplan, identified by Hg as Cognition’s founding president, described his work on AI software engineering through Devin in an interview with Jason Richards. Previously, he led Scale AI’s machine-learning organization and worked on Tesla Autopilot. His decision to work on Cognition rested partly on a technical argument: coding offers an executable feedback loop, allowing models to generate code, run it and test whether it works.[3]

In his Hg interview with Jason Richards, Kaplan explained that leading machine learning at Scale gave him visibility into foundation-model data pipelines. He became particularly optimistic about coding because generated code and tests can supply automated feedback, making data acquisition easier to scale than in many other domains. He identified that reasoning, alongside his longstanding interest in programming, as a reason to work on Cognition.[3]

Kaplan described Devin as asynchronous task delegation: an engineer assigns work, the agent writes and tests code, iterates on errors and returns a pull request. Human review remains part of that workflow. He emphasized traces that identify the person who requested the work and record the agent’s actions, arguing that this separation makes autonomous agents auditable.[3]

His stated research goal was to “get Devin promoted to senior engineer.” Drawing on engineering leveling guides used at Scale, he distinguished completing defined tasks from taking a defined outcome and breaking it into tasks—and, beyond that, defining outcomes independently. At the time of the Hg conversation, he characterized Devin as closer to the first stage. Moving from tasks to outcomes was the research direction, rather than an assertion that Devin had already reached senior-engineer capability.[3]

Kaplan summarized Cognition’s approach as “for us, speed is the strategy.” He also offered a concrete retrospective qualification: starting with large enterprises helped the team address existing codebases and environments, but he wished it had opened Devin to self-service earlier. Broader availability, he said, accelerated customer feedback and product improvement.[3]

Kaplan participated in the discussions that brought Cognition and Windsurf together in July 2025. In their July 14, 2026 retrospective, Scott Wu and Jeff Wang placed him in the room alongside Scott, Jeff and Graham during the Friday-to-Monday process. The authors explained the combination through complementary capabilities: Cognition sought a go-to-market organization, Windsurf sought engineering, and their products—a cloud agent and an IDE—served different forms of developer work.[7]

Kaplan’s public argument extends beyond generating code. He told Richards that engineers would increasingly review and coach AI agents, while more people would help shape software creation. An April 23, 2026 Latent Space recap attributed another enterprise argument to him: buyers want model flexibility and infrastructure covering the full software development lifecycle, rather than attachment to one model lab.[3][2]

Before Cognition, Kaplan’s work centered on computer vision, training systems and the data underlying machine learning. His Stanford, Tesla and Scale roles provide the earlier context for his technical responsibilities and emphasis on data quality.[4][3]

Kaplan earned bachelor’s and master’s degrees at Stanford, where he researched computer vision in the Stanford Vision Lab under Fei-Fei Li and helped establish TreeHacks, the university’s hackathon. He recalled testing graduate study alongside applied machine learning, finding Tesla Autopilot work more enjoyable than writing research papers and choosing Tesla full time. As a senior machine-learning scientist there, he worked on the vision neural network, including multitask learning and large-scale distributed training infrastructure.[4][5][3]

Kaplan recalled sitting beside his manager, Andrej Karpathy, and repeatedly seeing him label data despite having a labeling team. In Kaplan’s account, Karpathy explained that inaccurate labels constrain the entire system and produce downstream errors that are difficult to catch. Kaplan linked that lesson to his own focus on data quality and to joining Scale through the sale of his startup.[3]

After Tesla, Kaplan founded and served as CEO of Helia AI, a startup for real-time video understanding that Scale acquired in 2020. By October 2022, he was Scale’s director of engineering and led Nucleus, its data-management platform for machine-learning teams. He subsequently led Scale’s machine-learning organization.[4][3]

In Latent Space’s July 19, 2023 Llama 2 discussion, Kaplan described Scale’s open-source LM engine library for fine-tuning and serving language models on Scale or users’ own infrastructure. He discussed planned SQL fine-tuning and possible retrieval and tool-use datasets. He also said that reinforcement-learning-from-human-feedback labeling increasingly required domain experts, including programmers and licensed professionals.[1]

Kaplan predicted that fine-tuning a model specifically for one tool could help agents overcome errors that compound across long action chains. He also grounded demand for self-hosted models in his experience with enterprise customers: deployment inside a virtual private cloud and compliance requirements could determine whether a system was usable. These were his July 2023 expectations and customer observations, rather than claims of completed tool-specific releases.[1]