I dunno how many of you work in software but "total re-write on new architecture" is akin to dropping a nuke.

They are literally admitting that their proposed solutions to FSD don't work.

This guy is constantly glazing Elon and Tesla on his channel. Is he a shill?

I assume that is V11.

But yeah, FSD turned out to be a bit harder than we thought. Many people were wrong, Elon included, me included. Back in 2012 before the neural net revolution we thought we were almost there:


11 years later deep learning has made a lot of progress but self driving cars are not here yet. But a lot of money is getting thrown at it by various actors and a lot of progress on the problems are being seen. I am still bullish that FSD will be solved by someone in the next few years and think Tesla is a good candidate to do it. I could be wrong, time will tell. Good for you to never have been wrong on predicting what Tesla would do.

What do you me "we" Kemosabe?

Does this statement apply to Elon? He might have more info as to the actual work involved in FSD.

I don't work in software but I found that Facebook did a complete rewrite in 2020. They said "A complete rewrite is extremely rare, but in this case, since so much has changed on the web over the course of the past decade, we knew it was the only way we’d be able to achieve our goals for performance and sustainable future growth"

Is it possible that dropping a nuke and using a decade of experience would be much more likely to achieve Tesla's end goal rather than continue to build out a failing architecture?

https://engineering.fb.com/2020/05/0...book-redesign/

Sure it is certainly possible.

The main difference is Facebook was able to leverage improvements made and hardened by others. The webpage / webapp tech stack has advanced greatly the last 20 years and they had to rearchitect to move into the modern age.

Tesla doesn't have the benefit of leveraging improvements made and hardened by others. It screams hard pivot to try and resolve a glaring deficiency. And hey, you gotta solve the problems, but it shouldn't lend confidence.

They'll probably partner with Waymo lol

There are plenty of improvements by the other in the machine learning space that Tesla are using. Nvidia, Meta, Openai, Deepmind etc.

And it's not a total nuke. They still have their dataset which is the main part of making neural networks. What they mainly changed was increasing the scope of their neural network from just doing perception and letting c++ code do the control to letting the neural network do the control also. This allowed them to remove 99% of their c++ code(from 300k lines to 3k lines) in the FSD stack. See this video:
https://youtu.be/y57wwucbXR8?t=358

They just grew the scope of the software 2.0 part a lot getting close to 100% being software 2.0.

Do you really believe any of that nonsense you just typed on your keyboard?

Leveraging big data and cutting-edge AI, they have implemented a disruptive neural-symbolic framework that drives a frictionless transition from rigid legacy systems to scalable cognitive architectures. This next-gen deep learning engine autotunes network topology and loss functions, enabling 10x gains in prediction accuracy and business value for them.

They have aligned core competencies to incentivize out-of-the-box synergies, moving the needle from siloed niches to an integrated omnichannel offering. Partnership is in their DNA - they're fusing human talent with machine learning to holistically transform the paradigm from process-centric to customer-centric.

Their roadmap catalyzes a future-proof ecosystem, connecting all stakeholders via a distributed ledger of tokenized incentives. This will organically enable them to deliver turnkey Blockchain-as-a-Service solutions. They eat their own dogfood, leveraging Agile sprints to drive exponential organic growth.

Through harnessing big data and pioneering AI advancements, they are on the verge of implementing a disruptive neural-symbolic framework, facilitating a seamless transition from antiquated legacy systems to infinitely scalable cognitive architectures. The precision and business value of their next-generation deep learning engine will be enhanced via automatic optimization of network loss functions.

By incentivizing unorthodox synergies, their realignment of core competencies will shift the paradigm from isolated niches to a unified omnichannel offering. Human ingenuity will be fused with machine learning mastery through their partnership-driven DNA, wholly transforming mindsets from process-oriented to customer-centric.

I'm not super into TSLA, but isn't their problem hardware? They insist on using video rather than other sensors which makes safe navigation impossible, because something as simple as bright sunlight can obstruct the sensors. You can rewrite the software anyway you want, but the hardware don't work. All the self driving taxis I've seen use more sensors than video.

yes, but if they add more sensors now they either have to retrofit old cars or they might get sued by all the people who paid thousands of $ for their old autopilot stuff.

i still remember when they removed radar sensors because allegedly they didn't need them anymore and then just reintroduced them a couple of months later. in reality it was just the height of the chip crisis and they needed to get those cars out the door to make the quarter.

Which problem are you referring to? That it didn't stop for one red light during a 40min drive? What hardware would solve that? A better camera? The camera could see it, it was just the software just failed to assign it to the right lane. Waymo solves that problem with software aka HD-map and localization(using lidar, but camera could do localization reasonably well also).

I mentioned one example: sunlight. I don't think it's possible to implement self driving with cameras only.

Of course it's not.

But Elon has decreed that LIDAR is not a solution.

So you're just gonna have to deal with the fact that HAL is going to be driving you off a cliff.

It should be noted that in RGB images there is information that humans struggle to see but computers see fine. A human struggles to see the difference between 255 255 254 and 255 254 254 but to the computer there is a clear edge. A good example illustrating this:
https://youtu.be/bcZFQ3f26pA?t=47

Tesla takes this a step further by not even converting the sensor data into RGB but instead staying at what Elon calls "photon count" which is technically not true, but I guess it correlates with that so close enough. This likely improves low/high light sensitivity.

As for it being impossible to drive with cameras only. To disprove a strong statement such as something being impossible you just need one example. And we do have one example: humans.

Humans have sun visor and movable head piece. Usually sunglasses and hats. If there was enough redundancy, maybe. Why invest in fancy or many cameras when you could use a more reliable sensor?

I don't know much about Tesla tech, but their FSD is a marketing gimmick. Which is fine. I'm sure it's profitable for them. No taxi service with autopilot or startup attempting it uses cameras exclusively to my knowledge.

That video is processing low light images. I was under the impression that looking directly into the sun was not a low light situation. Can the computer distinguish between 1,1,1 and 1,2,2?

Humans have eyelids. How well do cameras see through heavy rain? Fog? Mud? The blood of innocents?

The CyberTruck will never be able to drive through a mud puddle without completely shutting down... which kind of defeats the purpose of having a tough and mighty truck.

When will Elon think of little wipers to put over each lens?

Humans also struggle with driving in fog and heavy rain. But in general car cameras see a lot better through rainy windows than humans do.

Don't think they expect robotaxis to drive through heavy mud, but at slower speeds that will probably not be a problem.

Directly into the sun is a high light situation. And yes a computer can tell the difference between 1,1,1 and 1,2,2.

The current vehicle setup has 5 cameras looking forward so there is some amount of overlap. If you find this to be an important issue for your bear/bull thesis, why don't you rent a car and try to drive into the sun on autopilot and see if it fails. Here is an example of someone trying it:

No thanks. I'd prefer testing be done on closed tracks with professional drivers.

Based on my experience with Subaru Eyesight, it's cameras are incapacitated by heavy precipitation, fog on the glass shielding the cameras, and bright sunlight. I don't think these can be fixed with software and more cameras, but they're sort of edge case scenarios. Tesla probably does a pretty good job and considering the expense of lidar, it's not practical for consumer models. So they don't really have an option.