Knightshade
Well-Known Member
It can, and does (as certain hacker folks with direct view of what the car "sees" have shown in videos for years).
But does not bother spending the resources to display them without FSD.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Why does the UI miss so many oncoming vehicles?
- Thread starter MikeyC
- Start date
It can, and does (as certain hacker folks with direct view of what the car "sees" have shown in videos for years).
But does not bother spending the resources to display them without FSD.
OxBrew
Active Member
OK, then why not just fake it a bit, add some simple smoothing filters, keep the cars/trucks from looking like they're jumping around, especially if they really know where they are? Just seems kinda bush league, not a good look, does not give me the warm fuzzies about any of it.
A lot of people confuse the displayed information on the screen with what a device knows or does. It's not a direct relation, as some are pointing out here. The computer can choose to display only some cars, it could choose not to display any cars at all. That doesn't mean it's not aware of them. It doesn't mean FSD wouldn't drive around them. It just means they're not displayed.
Without hijacking the thread, I've seen other cases of people inferring what a computer or device does from what it displays. Displayed information is useful for the human, not the computer, and it's not always 100% related.
Without hijacking the thread, I've seen other cases of people inferring what a computer or device does from what it displays. Displayed information is useful for the human, not the computer, and it's not always 100% related.
Exactly. If they want us to buy FSD it's doing a horrendous marketing job.
Take a look at the FSD beta UI then. That is showing A LOT of information. That is to help the human figure out if the beta is doing things right or not. It's there so the human can QA/QC what the fsd computer does. Once the beta is finished I bet they'll tone it down again.
The fact is that there's so much information they could show that our brain would overload, if you include calculated information, inferred properties, confidence rates etc. They have to cut stuff.
The display on screen for the normal FSD doesn't have a goal of attracting buyers (marketing). Once you see it, you have already bought into it. It's probably there to give some confidence to the driver without attracting attention too much.
The fact is that there's so much information they could show that our brain would overload, if you include calculated information, inferred properties, confidence rates etc. They have to cut stuff.
The display on screen for the normal FSD doesn't have a goal of attracting buyers (marketing). Once you see it, you have already bought into it. It's probably there to give some confidence to the driver without attracting attention too much.
OxBrew
Active Member
Wait, what? Buggy, laggy, incomplete "FSD-lite view" is supposed to give me some confidence about something? It's doing the opposite.
If it's not supposed to be marketing for "FSD", then it's doing that really well, making me run the other direction.
So where exactly am I supposed to look that would give me positive FSD vibes? Nothing Tesla's feeding me is doing that. Online actual test videos are not doing that. The whole thing is an epic cluster.
smatthew
Active Member
In 2018/2019 the cars really did dance. It wasn't fixed with smoothing filters - but by improving the neural net that was coming up with the bounding boxes. I think "faking it" on a driver display would be a really really bad idea.
roblab
Active Member
The longer Tesla keeps trying, even though they might miss, the closer they get to a functional FSD. You watch. Tesla will have FSD. And it will work. I don't have to be a fanboi to say that, either, because Tesla has done all sorts of things that Ford and GM thought was too hard for them to do -- until Tesla did it.I think I said most functional.
Customizable is a plus.
Let's say Ford beats Tesla to city navigation with radar and vision how many people will stick around waiting for a non-existent cybertruck assuming Ford could crank out Mustang and F150 EV's? I'm not saying anyone is even close yet, but the longer Tesla keeps shipping broken software the more opportunity there is.
OxBrew
Active Member
you missed my point. Everyone says the system actually knows where the cars really are, and the stutter in the FSD preview is just an artifact peculiar to the free AP. If true, clean up the stutter, show us what you really know. What I have now is embarrassingly bad.
stopcrazypp
Well-Known Member
That's not what people are saying. What you see in free AP and production FSD has the stutter because that is what the perception system sees. It only looks at individual frames and also does not merge what is seen by different cameras (so when a car straddles two cameras, it tends to flicker).
FSD Beta is different and is using something far more advanced, as I described in this post (I suggest looking at the video to see the changes):
Why does the UI miss so many oncoming vehicles?
OxBrew
Active Member
Please no FSD people. It's not ready. Do it as a toy and be ready to take over at any given moment. Just yesterday afternoon the FSD UI shows a traffic light green even though it is actually red when direct sun light was shining on the light. Granted there were multiple traffic lights and others were showing red correctly. What does FSD do in such case? It's just not worth the risk.
There's no intelligent if a machine has to be trained constantly. It's just a computer executing human instructions given in the past.
FSD will be feasible only after there's a standard for computer to talk to computers (car to car to infrastructure). Computer to human interaction is unsafe. We can't react fast enough comparing to a machine. Phantom braking induced accident is considered a human error or a FSD error? Predicability is hard to quantify yet extremely important with regard to safety.
There's no intelligent if a machine has to be trained constantly. It's just a computer executing human instructions given in the past.
FSD will be feasible only after there's a standard for computer to talk to computers (car to car to infrastructure). Computer to human interaction is unsafe. We can't react fast enough comparing to a machine. Phantom braking induced accident is considered a human error or a FSD error? Predicability is hard to quantify yet extremely important with regard to safety.
Knightshade
Well-Known Member
This belies a deep lack of understanding of what machine learning is and how it works.
It is most assuredly not "just a computer executing human instructions given in the past"
The "has to be trained constantly" bit is odd phrasing too. Humans constantly have to train to get better at many tasks as well- including driving.
Someone who has 10 years experience driving, all else being equal, will drive far better than someone who has never driven before.
Training in AI's world == adding new scenarios/considerations so human can code new "if then else" statements. Your word "training" in human world == Full Self Learning. Even going to school or taking a class is "self imposed". There's no "training" (forced upon by outside force) taking place.
I'm a programmer. I deal with "AI" (claimed to be) everyday. It's false advertising in a sense like company redefining what "organic" means so they can label a product as organic.
Do you consider Google's Alpha Go intelligent? Or just faster computation than human? Can Alpha Go learn to draw the same way a human could learn a new hobby? Even though we could use the same word "training" in both situations, but the meaning behind them are very different. Google needs to come up with a new "machine learning" drawing algorithm different from the original chess machine learning algorithm. It's not like there's an universal "machine learning" algorithm for learning everything.
Intelligent has to equal taking responsibility for the decision it makes. A car will not be responsible for anything in the forceable future; therefore, it is not intelligent.
Just my opinion of course.
You're a programmer, but are you an expert in AI/ML? There's a huge difference, as I've personally realized being in software development and working with AI/ML specialists at my job. Step 1 is thinking you understand AI, step 2 is realizing it's complicated and limited, and there are steps after to figure out the proper application, proper learning. I haven't reached the end but I defer to the specialists. You seem to have stopped at 2.
There was a time a few years ago when it didn't see ANY oncoming cars, so in a way it's an improvement! Basically, you will have to wait for the "trickle-down" from the FSD beta stack that will happen as FSD beta gets closer to completion, since the new NN stack for the beta is far better at detecting cars (as it has to be!).
This is incorrect. NN training is a mechanical process that generates an NN that, within certain constraints, can identify data patterns in an input dataset that has a certain "resemblance" to the training set. There is no traditional if/then/else involved here; there is no C/C++ etc source code in the car that runs to determine if the car can (for example) see a pedestrian (though there is code to decide what to do once the pedestrian is seen). There is, of course, no "intelligence" involved, since the NN, once trained, is purely mechanistic (and yes, "AI" is a bad name, I prefer "Artificial Mimicry", which is closer to what NNs achieve).
This isnt quite correct. Yes, the cameras can "see" the same thing with and without the FSD beta,vin the sense that the video pixel feeds are the same, but the beta stack is FAR better at actually discriminating that there is a vehicle at a certain location. The old stack could not discriminate cars etc anywhere near as well as the FSD beta stack (and, in fact, this is really the primary development that makes FSD beta possible).
Similar threads
