People's opinions of FSD come in many forms. Many people without experience in actual self-driving technology are quite impressed by it. That might include you. But you may be puzzled that pretty much everybody with actual experience thinks it's a very poor effort at present, with a great deal more distance to go, unlikely to be achieved prior to other projects (from tech companies, not car companies) going commercial, and possibly not for a great deal of time, if ever, on the current sensor configuration.
Why this disconnect? People of course are not satisfied at appeal to authority and expertise, that is not enough, though it certainly is a hint when the inexperienced like it and the experienced find it wanting. That is reason to try to examine your own view. In trying to understand this difference, I have come upon several factors:
- The gamble: While many agree with the Tesla team that driving with just cameras might be possible some day, even Tesla admits it requires breakthroughs which do not yet exist, and for which the path is not known. A breakthrough could come next week, or take years. As such, one can have optimism about FSD because you have optimism about Tesla's ability to be the one to make this breakthrough. On the other hand, approaches with more than cameras no longer need any more fundamental breakthroughs, though they do need more work. One can bet either way.
- The "Geofence" illusion: Many think that Tesla's attempt to drive without maps is a feature, which lets it someday handle more territory. They deride the pilot projects which work in only a few cities, imagining them inferior to a system which doesn't work but doesn't work in many more cities. They think mapping is inherently too difficult and expensive, or that handling extremely rare surprise areas where maps are out of date is somehow harer than handling never having maps in the first place. (They also think that surprises for out of date maps must be common rather than extremely rare.) In fact, making the vehicle work at all anywhere is the hard problem. Expanding the map region is the easy problem. (You can watch my Youtube video and article on why Tesla would get working faster, not slower, if they moved to requiring maps, if you want more detail. )
- The false confidence: Many are impressed that they drove a trip without needing interventions. Though I have not managed my to get my Model 3 to do that yet, I have seen reports and videos of those who have. People say "wow" but don't understand how spectacularly unimpressive that is. To be ready for production, you need a car to be able to drive for a whole human lifetime without an intervention to prevent a significant accident. An hour, a day, a week? They mean nothing. Really. Waymo's cars have done 10 human lifetimes without ever being at fault for an accident in reality or in sim, and they're still only in pilot projects, though very close to release. And you're impressed by an hour? Instead, the fact that it often can't go 5 minutes is the real datum you should pay attention to.
- The neural network leap of faith: There are many people who feel that the AI capacity of deep neural networks is unbounded. That if you just throw enough training and a big enough processor at it, there is nothing, even driving, that it can't solve. There are experts who believe this, and there are experts who don't believe it and they have arguments. But none would say they are sure. This is part of the gamble. However, it is important to know that when it comes to neural network expertise, a wide consensus is that Google (which spawned Waymo and provides tech to it) has by far the most advanced neural network expertise, as well as the highest performance neuromorphic processors. So if you believe in this approach, Tesla is not the likely winner. Tesla is not without its advantages, but they are fewer than many imagine.
- The bold claims: Tesla definitely makes the boldest claims about when they will make it work. Unfortunately the track record on those claims is now so poor that they must be disregarded. I am at a loss to understand why Elon Musk keeps making those claims. He is somebody of great achievement which is what has led people to have faith and accept these bold claims, but when it comes to predicting the progress of FSD, he is no longer somebody of great achievement in that particular area. I am happy to trust his predictions on space and EVs and several other areas, but no longer on this.
- The generation of antagonism: While being brash is often the path to success, you can overdo it, and antagonize people at all levels, from the California DMV to the President of the USA, to the point that it slows you down rather than gives you freedom. I fear Tesla has tripped over that line.
- The "Beta" name: Tesla FSD is not even remotely close to a beta. In software development, a project moves to beta testing when it is in a near-release state, when internal alpha testing is not finding bugs fast enough to keep the team occupied. A product is not a beta if it is regularly getting major rewrites of components and experts more of those, along with many new features. That's a prototype, not a beta. Some companies have in recent history certainly stretched the meaning of beta quite a bit, but none nearly so much as calling FSD a beta. Don't be fooled by that name.
arstechnica.com
