Tesla’s push for unsupervised self-driving reveals that consumer vehicles with Hardware 4 (HW4/AI4) computers run scaled-down, distilled versions of AI models designed for the Cybercab, which features more powerful FSD hardware. According to a source, Tesla develops FSD for the Cybercab first, then distills it for HW4 systems. This distillation process trains massive “teacher” models in data centers, which are then optimized into smaller “student” models for consumer cars. The Cybercab serves as the ideal hardware setup, while consumer cars receive lighter versions. The FSD v14 versions in vehicles are also distilled models of what runs natively on the Cybercab. HW4 likely requires less optimization than Hardware 3 (HW3/AI3), which has about 15% of the memory bandwidth of its successor. The Cybercab’s next-gen processor features more RAM than HW4, and Tesla announced AI4+ chips during its Q1 earnings call earlier this year, expected to come with 64GB of memory or potentially even more.