Hacker News Reader

Comment by cal5k | Hacker News Reader

This whole thread smacks of this type of cynicism - apparently a lot of tech people now really, really hate Elon Musk.

Andrej Karpathy addressed this point in his recent Lex Fridman interview - more sensors aren't necessarily better, and for each sensor you introduce a whole host of supply chain issues and physical constraints. By sticking to one sensor type that contains all of the information theoretically necessary to make decisions they've simplified the problem.

They bet pretty big on this approach... including by making their own silicon. I don't know how anyone could really believe that they're not earnestly pursuing eventual "Full Self-Driving" capability, even though they may still be far from that goal.

What they've achieved thus far is really impressive. That took a LOT of work.

View on HN

Andrej Karpathy designed a system that is literally 1000x worse in miles between disengagements than the systems deployed by Waymo and Cruise. It reached that level more slowly than just about every other company in the space and is improving more slowly than other companies that are already tens to hundreds of times further ahead. For crying out loud, they have not even figured out “Do Not Enter” signs, school bus stop signs, and not running into children and strollers yet after literal years of development. His opinion on the subject relative to the real experts, who do use more thorough sensor suites, is worthless.

That Andrej Karpathy admits that they hamstrung themselves by focusing on BoM, development costs, and production logistics when they have received payment for and are currently shipping a defective safety-critical product demonstrates either their total reckless disregard for all established safety process or total ignorance. Neither of which is acceptable when doing regular engineering, let alone safety-critical engineering.

That interview was a load of bullshit to justify the decision to stick with cameras.

The only reason why they picked cameras is because they are cheap, and they can put them in vehicles now, and promise they will be self-driving in the future.

Tesla saw that there is a huge demand for self-driving vehicles, so they decided to sell them, even though they don't have them. I wouldn't call that "earnest".

> Andrej Karpathy addressed this point in his recent Lex Fridman interview - more sensors aren't necessarily better, and for each sensor you introduce a whole host of supply chain issues and physical constraints.

That's a bullshit justification. Going to camera-only and trying to overcome that limitation with "AI" was hubris. All of the competition working on FSD realized cars need more and more varied sensors, not fewer. Humans use a variety of sensors between primary perception and proprioception to drive vehicles.

The biggest problem with camera-only is the system needs to recognize objects to understand them as road hazards. Humans don't need to recognize a thing in the road to think "don't hit that thing". It doesn't matter is an object in the road is a crumpled newspaper, a cinder block, or a cinder block covered in a newspaper, you avoid it all the same. Vision augmented with radar and LiDAR can help just say "that's a thing in the way" more reliably than vision alone. Glints from the setting Sun or windows aren't going to cause radar to think some solid object is no longer a solid object. A bad angle or odd shadow won't cause LiDAR not be able to pull an object from the background.

> This whole thread smacks of this type of cynicism

Both Musk stans and Musk haters go wrong when they conflate his companies with his persona. Tesla is obviously a highly innovative company, and will likely be Muskless at some point, and probably better for it. Then they can make more reasonable claims about FSD and not be under multiple investigations, including a criminal probe by the DOJ for misleading customers.

Reply on news.ycombinator.com