MobileEye/nVidia have all their auto hardware on cheap, low power, and low bandwidth processors.

The plain fact is that no one uses off the shelf hardware that can readily address the issues you claim to exist (high definition cameras, processing power). I guess none of these billion dollar companies figured out how to string together multiple off the shelf GPUs.

What does that mean?

It means nothing. I can only hope he doesn't think that each individual car is coming up with its own algorithm for imaging.

ETA: Or that he thinks robotic vision is single threaded. Both are equally absurd.

Going right over your heads.

Here's a challenge. Try and explain:

a) What each car is going to be "learning"/what data it is going to be storing
b) What data will be uploaded over the air from customer cars to Tesla
c) How that helps Tesla refine their autonomous driving algorithms from level 2/3 (where it is now) to level 4, compared to say a more traditional route the majors/Waymo are taking. I'm specifically interested in what advantages they will gain from this process over having a fleet of say 20 of their own test cars where they can download the full data sets.

Everyone was excitedly claiming "Tesla will be collecting tons of data, that's a big advantage", but no one has ever fleshed out even the slightest bit of what they would collect, and how it provides them an advantage in creating level 4. The advantages I can see:

- Camera feed from disengagement moments if there's sufficient bandwidth to upload
- Catching sensor errors
- Decision data (useless without the camera feed)
- Mapping common driving routes/positions so that the location-specific algorithms can be more reliable.

But Waymo and GM have already solved all of the problems here. Waymo has 1 disengagement in thousands of miles, drving far more complex routes (city, full traffic, roadworks, markings missing, signs, traffic lights, all kinds of variables). GM, as you can see, drives the city without incident very easily.

There's nothing in the data above that helps you get to level 4 in way that driving 20 test cars doesn't. Further, there's no way to put it into a learning algorithm without slow human intervention. It's all ugly hand tinkering. There's no way to parallelize what you're getting back.

If actually think it through, you'll realize the claimed advantage is nothing but nonsense. But I doubt any of you have the brain capacity to think it through. You couldn't even think through the logic of whether Tesla was ahead in autonomous driving or not - a simple thing with mountains of obvious evidence that they weren't purely from the way in which their algorithms failed, which tells you all need to know about the algorithms - I laid that out for you and you STILL couldn't see it - so how could you possibly think through something more difficult?

The bolded has to be a joke. Why would the camera feed even matter? You basically claim that a person learns driving by playing Need For Speed in third person view.

I have no idea how this whole concept in detail works but why camera footage alone would be of specific use makes no sense to me.

On holiday and on my phone so will be short, but Tesla are gathering video(multiple images) from cars. Machine learning basically is 3steps:
1. Gather data = parallelizable
2. Train model = parellelizable
3. Inference = parallelizable

No bottlenecks for this anymore, it is mostly the problem of designing the model that is hard to parallelize, since humans don't scale well...

If Tesla stores sensor data and data about the driver's actions especially when autopilot is off, they can work on training models that try to emulate the drivers' actions.

Right, but how do you make sure that you're not training the model using a bad driver here?

But in machine learning the bottleneck is always in 1. Actually 1b, which is, in this problem, how do you differentiate from good driving you want the model to emulate and bad driving you want it to avoid. That's the bottleneck.

Take a good driver and get him drunk. EZ game

It doesn't matter. You can simply have a hyper aware average driver that will outperform almost every human.

What in the world does that mean? We were talking about using the decisions of Tesla drivers in the wild to train the model, which I pointed out was problematic as there is no way to easily classify good and bad driving.

No one is talking about using their decisions.

Beyond that, even if you were to train a decision model based on aggregated drivers, it doesn't matter. As I clearly stated above. You don't need to classify to come up with a better driver entirely by mimicking the decisions of the entire population.

Perhaps you and TS can discuss this. You seem to be on similar wavelengths with tech :/

Uh, I was directly responding to someone who was suggesting that

But you don't need to classify the data to train the model? How does that work?

While it isn't very easy to classify individual good and bad drivers, there are enough metrics that allow you to prune the extremes. Hard brakes, fast accelerations, unstable steering, etc.

I'm certainly no genius, but this stock sure seems to be that rare opportunity where market value is so divorced from any rationality that as an investor you have to pounce. This company is a huge dog to survive as a going concern, let alone justify anything approaching this valuation.

That's not how it works. That can never be how it works. A lot of Tesla bulls without tech knowledge (not you, just in general) seem to think it is, though. Musk's PR (aka lies and spin) has convinced a good portion of the population that Tesla are leading on autonomous driving and have an unassailable advantage. It's completely insane. Just another example of fake news in action and how a fraud/charlatan can manipulate people.

Seems like a reasonable approach would be to identify, ex ante, certain outcomes that are obviously negative (accidents, near accidents that have to be avoided by sudden lane changes or hard stops) and then statistically model the factors that predict those negative outcomes in observed driving data.

That may end up creating an autonomous vehicle that's overly cautious, rather than a vehicle that's just good, but doesn't that approach seem plausible?

Sure, but those events are sparse, and not always the driver's fault. Not saying it can't be done but it's far from a simple problem.

What are you talking about? No one says that. Of course this is not how it works.

Your knowledge isn't convincing though. I asked you a question and you haven't replied to it. It made no sense so I am excited to hear your explanation for it. Also, heltok who works in the field said the opposite.

You get caught up in your hate but you make no sense.

I make a lot of sense, but it's not worth the time explaining.

As for "hating" Musk, he's just a fraud, not a figure of hate. It's a shame to see otherwise intelligent people get caught up in his religion though. Yet more objective evidence of his blatant fraud toward shareholders:

tesla went up after these tweets

tooth has been calling musk a dishonest promoter (and worse). after the sheep bought musks dishonest tweets the market digested this along with his history of missed deadlines and hype.....tank

i hate to be greedy but is 265 by september too much to ask?

if sentiment is really changing in this (which i think this looks like) this can go bad pretty quick. they're gonna need new cash within a year. i see a lot of bad news that were swept under the rug in the last year are being brought up again (like their safety issues).

/but then again, musk is pretty good at spinning and there are a lot of people who eat it up

If he ramps Model 3 well then that'll keep it simmering for a while longer.

If Model 3 gets delayed or messes up, and Model S/X is really declining as we have with the news today (safety issues, downgraded by a safety testing group; people reportedly fire in SoCal sales teams), then it's $150 very fast.

Market matters too. The guys here didn't believe it even when I screamed it at them, but the run to $380 was just a high beta with a high short percent in a massive big cap tech bull run. Had little to do with Tesla. Now that's deflating a little, chuck a bit of bad news on top, and your profit gets wiped out incredibly quickly

It's why you don't drink the KoolAid and believe your thesis is being validated when it pops to huge new highs. It means nothing.

It is difficult. But it seems like gathering data from half a million cars instead of a few dozen would provide a fairly large advantage? Seems like something that is solved by Tesla's information advantage.

In earnest, did you really expect a well thought out, reasonable viewpoint?