At this point, Tesla's FSD is almost certainly more "baked" than the vast majority of software you've ever used. The amount of engineering and compute time that have gone into it are colossal.
That said, something being excessively baked does not mean it is good.
Thats an irrelevant argument (and unless you work there directly on this just empty baseless words).
The point is - it didnt deliver, and still doesnt. Its a securities fraud out in the open, but clearly from a guy who is above the threshold of applicable law
And it has been limited by cost cutting decisions resulting in not using lidar or radar, meaning it has far less data to work with than other self driving efforts.
I won't trust Tesla FSD for that reason alone. You've got a way to have far more reliable distance sensing and you choose not to use it? That's crazy.
Roads were not built for LIDAR. Building a system that only requires the same sensing that humans use is a system that will work anywhere (if it works).
As soon as you buy into LIDAR (or any other non-human sensing mechanism) you introduce the possibility that your solution could have strict upper bounds on scale.
Which is indeed what we have seen so far with Waymo – it may have fewer issues per mile than Tesla FSD, but every one of those miles is expensively pre-mapped and requires $30k of extra hardware on every vehicle.
And for the use-case of "driverless taxi service in a well-established metro area", the Waymo solution does currently seem to be the better one. But that is also clearly not the ambition Tesla is pursuing.
"Roads were not built for LIDAR" is a strange take. LIDAR detects range between the sensor and some object. Do that in a scanning approach and you have an idea of whether there are objects in front of you, the distance to them, and their rough shape. None of that functionality depends on how a road is manufactured, it's physics.
The $30k figure you're giving for Waymo's LIDAR cost is quite old. As often happens, a new technology got cheaper over time.
"But LIDARS that are $5,000 today are forecast to be under $200 when bought by the millions, meaning the bill-of-materials for the extra hardware should drop below $2,000, and even $1,000, in time."
Fair, I mentioned how roads are built, but I really meant "the entirety of human wheeled transport is built around human sensing capability". But roads matter too. For example have you ever driven through a small Italian town where there are walls on every side? If I was a LIDAR-equipped human I think I would've been overwhelmed and unable to drive at all. I also would've "sensed" various turns a little too late, as LIDAR happens in a straight line, which can occlude facts about the world that are (more) obvious with vision.
In perfect conditions LIDAR is helpful, but what about the million situations where conditions are not perfect? e.g. what about rain/snow/fog? What about telling the difference between a cardboard box and a metal one? If you want a vehicle that can operate in all conditions and situations on the road today, your vision component needs to be incredibly robust to the point of (possibly) obviating the need for LIDAR.
That $30k figure is actually the lower bound on the current Waymo fleet as it exists today. The $5k figure from the article is the estimated cost per LIDAR sensor (there are 4 on a Waymo vehicle) if you bought them today, which is still $20k total. Additionally the vehicle cited in the article isn't on the road yet, and therefore seems a bit premature to cite as a reference for real-world cost. Imagined future costs are just that, imagined. Those are uninteresting in the same way that Tesla's promises of vision-only capability are uninteresting. What matters is what is actually achieved.
If Waymo achieves the range of driving conditions you can currently operate Tesla FSD, I will be impressed. Likewise if Tesla achieves the safety and consistency per mile that Waymo has, I will be impressed. The question is: which hurdle is higher?
I already explained this. There is no such thing as a free lunch, and by focusing on LIDAR (the cost of LIDAR itself, spending valuable onboard compute on processing and fusing LIDAR data, etc) you are by necessity leaving fewer resources for your vision system. This matters a lot if there are situations where you effectively can't rely on the LIDAR at all, which actually seems to happen a lot in a driving context, assuming you want your vision to work on all roads at all times (that a human could drive).
So yes, if you have infinite resources you should obviously use LIDAR. But we don't so its not nearly as clear.
Yes, I have been making that point for the entirety of this conversation.
1. Tesla needs to figure out how to achieve performance of their system without LIDAR. However since humans do so with two eyes, there is strong precedent that this is possible.
2. Waymo needs to figure out how to self-drive in situations when LIDAR cannot be relied upon. If they cannot do this, their system will never achieve Level 5 (but that's OK! Level 4 taxi service seems like a good business). If they can do this, then it is not clear they ever needed LIDAR in the first place.
> It has been designed by brilliant engineers whose only goal is that it works for its intended purpose.
Is this statement actually true? From what I heard, it was designed by highly overworked stressed engineers working in pretty bad workplace conditions. They work there, because there are not many other places to work at and doing similar work.
And their primary goal was to produce as fast as possible.
One thing I wonder is, if BYD was allowed to compete in the western markets, would the scale of deployment yield better data and thus a better FSD experience?
That said, something being excessively baked does not mean it is good.