Steve

On the dog bite side topic: Sometimes dogs miss. We ask them to bite a small toy from our hand during play. So, we should be forgiving when they make a mistake. The more you play the more accurate they will get with both aim and appropriate bite force for the activity. Some breeds instinctively bite soft as default and learn when to bite hard (Labradors). Some instinctively bite hard as default and learn when to bite soft (Staffordshire aka Pit). It can take extensive training to overcome.

Vision-only FSD wasn't a religious stance, it was a "burn the boats" strategy. It focused effort on truly generalized superhuman driving intelligence, and it worked. Vision is both sufficient and necessary (with humans as existence proof). Radar and lidar are neither. They are a crutch that wastes engineering effort, as well as precious local compute and battery. Early solutions used these crutches to solve early DARPA local driving challenges. It was the only way at the time. But it seeded academia and a generation of autonomy students, born inside this paradigm box. It's true that one can contrive edge cases where radar or lidar can perceive something before vision can. But this is a mental trap, a local minima concealing a better way. Unpopular Opinion: Once robust, superhuman, generalized driving is on its Nth 9 of safety, and the vision only paradigm is well established, and once there is spare compute and battery, I can see Tesla adding other sensors types to chase down the infinite safety tail. Think FSD v17+.

With that killer attach rate, annualized FSD revenue is exciting. But a forward projection is even more exciting. It gives a hint at a FSD becoming a material earner. Lets assume a similar ongoing attach rate, a low churn, and then compound revenue. Obviously many people who are paying this year will continue paying next year, on top of the new subscribers. Even if you assume zero growth in deliveries next year, 2027 revenue from FSD will likely be closer to $1.7 billion from FSD. (This would be an ARR of about $2.2 billion at the end of 2027.) Also, this is almost entirely profit. Analyst estimates are way too low. It is literally software and the marginal incremental cost is next to nothing. (You could say those costs are almost entirely baked into the base car.) So, it's well over 90% profit.

Tesla has already 1. provided video evidence that this was in fact one of their Roadster gas thruster and fan test facilities 2. stated that they are building a much larger test and demo track and facility, and it will be publicly accessible to Roadster owners A more reasonable assumption would be that Roadster testing needs outgrew this tower and it became obsolete.

Think about the portion of the economy that is or can be digital only: sales, finance, CAD, office productivity, marketing, customer service, games, entertainment media, communications, art, many kinds of research, etc. All of these can improve as fast as AI improves, and can be substantially further automated by moving only bits. Change there can happen insanely fast. Then also add the portion of the economy that involves making physical goods, but where scaled production requires moving atoms mostly just at the very end of the supply chain, and where that is already outsourced to highly efficient fabs. This is everything from microchips to medicine. Many of those manufacturing endpoints run under capacity most of the time, and even the ones already fully loaded can have capacity displaced by higher earning / paying developers with AI automation. 2030 might be a very different world.