Tactic: Shorten the CEO-to-Engineer Feedback Loop
To make better decisions, companies need to eliminate filters between executives and engineers, argues Kieran Gilmurray. He points to the tight loop between Musk and Karpathy at Tesla on GPU scaling as a model for making reality-based decisions quickly, a key method for engineers to build influence with the C-suite.
Andrej Karpathy served as Tesla's Director of AI from 2017 to 2022, leading the Autopilot vision team and reporting directly to Elon Musk. This direct line was critical as the team tackled the immense challenge of making Full Self-Driving a reality. The core technical problem was training neural networks using massive volumes of video data collected from Tesla's fleet. To handle this, Karpathy's team utilized a supercomputing cluster with 5,760 NVIDIA A100 GPUs, which at the time was considered one of the top five most powerful supercomputers in the world. This intense demand for processing power led to the development of "Project Dojo," Tesla's custom-designed supercomputer. Dojo was engineered specifically for AI training, aiming to enable larger, more complex neural network models more efficiently and cost-effectively than off-the-shelf hardware. However, the strategy evolved. By August 2025, reports indicated the Dojo project had been shut down, with the team reassigned to other data center projects. The former head of the Dojo project, Ganesh Venkataramanan, subsequently left along with other team members to found a new AI chip startup. After leaving Tesla in July 2022, Karpathy briefly returned to OpenAI in 2023 before leaving again in February 2024. He has since started his own company, Eureka Labs, focused on AI education. The principle of a short feedback loop is most critical in capital-intensive R&D. When a company invests heavily in a novel technology with high uncertainty, direct, unfiltered communication between the CEO and lead engineers allows for rapid risk assessment and directional pivots without bureaucratic delay. For engineering leaders, establishing this culture involves creating psychological safety for unfiltered feedback and implementing structured, frequent check-ins. The goal is to make candid, data-driven conversations a routine part of the operational workflow, not a rare event.
