FSD Brain

FSD Brain: Inside Tesla's FSD AI

How FSD's AI actually works, plus every FSD version, rollout and Tesla AI update we track. Every item links to its source.

Explainer: Tesla's approach to autonomy

First published 24 Oct 2025. This is Tesla AI lead Ashok Elluswamy's own write-up of his talk at the International Conference on Computer Vision (ICCV). It's background on how FSD works, not new approval news for Asia.

• One network, end to end. FSD is a single neural network: camera pixels, speed, audio, maps and navigation go in, steering and acceleration come out.
• Learned, not hand-coded. "Codifying human values is incredibly difficult. It is easier to learn them from data." Tesla calls everyday trade-offs like the one below "mini-trolley problems".

• Huge scale. About 2 billion input tokens (7 cameras, 30 s of history, 48 kHz audio) get reduced to 2 outputs. The fleet produces the equivalent of 500 years of driving every day.
• Thinking ahead. In drizzle, the AI braked early because it predicted the car ahead could spin, hit the barrier and bounce back into its path.

• Reasoning on board. A small natural-language reasoning model already runs in FSD v14.x. Tesla's generative Gaussian Splatting rebuilds the 3D scene in about 220 ms.
• A neural world simulator. It generates all 8 camera views in real time (someone drove inside it for 6+ minutes). Tesla uses it to test new models, create tricky scenarios and train with reinforcement learning. The same tech carries over to Optimus.

Source: Ashok Elluswamy on X, "Tesla's approach to Autonomy" (24 Oct 2025) · tesla.com/AI

FSD and Tesla AI updates

FSD versions, rollouts and Tesla AI news, newest first. Updates automatically from our news feed.