omgwtfbyobbq
Active Member
Me after noticing they went from -$520 million to +$516 million in non-GAAP net income in a single quarter...
You can install our site as a web app on your iOS device by utilizing the Add to Home Screen feature in Safari. Please see this thread for more details on this.
Note: This feature may not be available in some browsers.
Exactly - they’re just dog-fooding instead of actually doing their own research. Fools. It’s disappointing the laziness off media. And not all those people on the panel have an ‘agenda’.Gene Munster: it’s not a rabbit out of a hat, they’ve said they were going to do this. Love it.
Gene Munster: If Tesla CEO Elon Musk didn't 'mess it up,' the stock would be over $400
Actually, it almost came to that!1? or 2?... 1? or 2?... (slap) 1? or 2?
Correct.
Ah... but it doesn't unless you actually *do* that, and they're not doing that.
You have to hire perfect drivers and then drive these perfect drivers through every possible situation. If you're just using random incompetent schmoes... you're not going to be giving the network accurate feedback as to whether it's wrong.
This is what you, and practically everyone else, has missed. I do NOT want a "self-driving" car which is as crappily incompetent as the average American driver!!!! And neither does anyone else. There are stupid, incompetent, dangerous things which *the majority* of American drivers do. How are we going to train the car to not do that?
Tesla is doing great at the tech and missing the human factors. Which is actually totally normal for them. I hope they eventually wake up about the human factors.
When they finally get that right -- when they really start looking to experts to figure out what the car *should* do in situations where most American drivers get it wrong -- then I'll expect self-driving to be done in about 2 years. They haven't started yet.
Now do you get where I'm coming from?
Half of that seems easy to me and I have ideas already; if I do, it won't take long for engineers to think of it once they're assigned to look at it. Others have already started tackling the other half. There's lots of potential.concrete
Yeah I heard about it. It's really 'big'. Except that neural nets can be infinitely big but have asymptotic performance typically at fairly small sizes. Going 'big' is usually the realm of academics trying to beat other academics by .0001% on a digit recognition task.
Reinforcement learning uses a good/bad reward stimulus to propagate penalties back to decisions made earlier. If the NN is not controlling the car there is nothing to penalize. Further, the dimensionality of the problem is absurd. An NN might not know that a purple car is dangerous if it never saw one before because it doesn't generalize like we do and our intuitions are broken when we think about it. Heck even the visualizations that you see generally prove this isn't end to end training because those visualizations are of conceptualizations that are handed off from the perception layer and then used by the procedurally coded planning layer.
Just give me some actual evidence for end-to-end training.
I'm saying if volume production by 2020 has to be GF1. Any other guesses?
I interpreted it as a way to say Tesla is in a good position competitively, and those opportunities still exist for the other companies in the future.With the final speech about how they did everything they could to help the other car companies, and how they will still help them (anyone can join the Supercharger network!), but how the other car companies simply wouldn't take them up on it, and how this has given Tesla their massive unbeatable advantage?
Yep. I wouldn't call it cold-blooded, exactly -- more "rubbing it in".
One percent profit on parts/service? Not a chance. Assume you have not seen a parts bill for body shop repairs?
This is where Tesla can start making serious money but hopefully balanced with what owners will find "reasonable" or they could kill resale value on out of warranty cars. That could bode very badly for Model 3 owners in 4 years/120k miles.
This isn’t really true. They aren’t likely going to be just grabbing 100% of all user data and labeling that as perfect driving. Rather, they can pull useful information from that data of both safe driving and examples of unsafe driving(what not to do). Better is that, with the right metrics, they get some level of automatic labeling of good vs bad driving by treating it as a reinforcement learning task, with detected crash or near miss being punished.
I don't think so: most human drivers are near perfect
My limited understanding of AI and Python is that you need to tell the computer what is a pass or fail which is called Supervised Learning. I think you're describing Unsupervised Learning which would result in the average Joe schmoe out there. I believe you just need to throw out the bad behaviors, and maybe there's an algorithm for that even... when the car wrecks
One percent profit on parts/service? Not a chance. Assume you have not seen a parts bill for body shop repairs?
This is where Tesla can start making serious money but hopefully balanced with what owners will find "reasonable" or they could kill resale value on out of warranty cars. That could bode very badly for Model 3 owners in 4 years/120k miles.
I agree. But they are going to have to stop making some model 3 to make the model Y in Fremont.
I doubt they want to make that call yet. If they saw model 3 demand still very strong mid next year then it would be hard to change what is working. That could push the model Y to 2021ish, perhaps in Sparks.
Given Tesla's production constraints the model Y is very much a safety play. Tesla could not be sure that buyers would not look at the model 3 and say "nice, but it's a sedan". But fortunately for Tesla many of us were not willing to wait for the SUV.
If the model 3 is supply constrained in 2019 and 2020 Tesla maximizes revenue by not making the model Y until they add a U.S. factory.
I used to think as you do about that, but today and recently I realized that it can compare like decisions even further down in the chain, because it has examples of an infinite number of human decisions to look at, many of which the learning can realize is similar to its own decisions and compare.(and shadow mode is insufficient because it's not actually controlling the car to measure its decisions).
I haven't had enough experience to see exactly how long that is to do. With many projects, it starts to get complete after some time; I don't think there's anything unique about neural learning that makes it not true with that topic.the long-tail of exceptions which takes you straight into the same problem AI always had before NN came about which are the exceptions.
I disagree with your point here.Autonomy in 99.99% of situations is like near-singularity type AI.
Since I don't know, I'll leave this part of your message quoted. But I'm more optimistic. For full disclosure, I've been more optimistic about this for a long time and been wrong, and less optimistic than Elon initially, matched him later, and more optimistic than him after that having been unsatisfied with their speed of progress. What I'm saying here is that there's a lot of hand waving going on about new tech, and claiming either that it has been solved or won't be able to be solved for a very long time seem like extreme claims.To get there before that is a process of iterating on problem specification, figuring out hacks and compromises and 'just good enough' and so on. I think Neroden is right to be skeptical. My guess is it takes 5 years before we begin to see a reasonable and non-hand wavy path to it.
I’m still saying flying customers to California is cheaper then trucking the cars to customers.It takes 30 days to get a car from the factory to the customer? Are they kidding? Most of the customers were in California.
Ok, Elon will improve to 10 days, which is really good.
I agree. But they are going to have to stop making some model 3 to make the model Y in Fremont.
I doubt they want to make that call yet. If they saw model 3 demand still very strong mid next year then it would be hard to change what is working. That could push the model Y to 2021ish, perhaps in Sparks.
Given Tesla's production constraints the model Y is very much a safety play. Tesla could not be sure that buyers would not look at the model 3 and say "nice, but it's a sedan". But fortunately for Tesla many of us were not willing to wait for the SUV.
If the model 3 is supply constrained in 2019 and 2020 Tesla maximizes revenue by not making the model Y until they add a U.S. factory.
I agree they could push it out. But to your point on people going "ahh I'll wait for the Model Y instead of getting a Model 3" I think that's why they might start production sooner than we think once they unveil it. And no I'm saying they would make the Y entirely in GF1 not Freemont.