Early days but I'm confident that your prediction of it'll never happen is wrong. I use them all the time, not perfect but extremely impressive already.They would say that, wouldn’t they?
Don’t do yourself a disservice, it’s really not as clever as you think it is. It’s pretty easy to trip up something like ChatGPT, whether it’s asking it to write some code and then calling it out when the sample given is doesn’t work or something as simple as asking it to summarise what happened in a random football match from 15 years ago.
Great. When it can do what the human brain can do, let me know. Crunching a load of data doesn’t impress me, nor does the amount of compute power and electricity it uses.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
UK FSD Discussion
- Thread starter Buckminster
- Start date
Re roads, I'd agree, but actually, I'd say that in some respects US signs are harder for autonomous vehicles. Fewer standardised symbols and more textual signs, for one.
Sadly FSD doesn't (yet) understand text. This is a big deal in America. The biggie is turn on red. In the US, you can sometimes turn right on a red light as long as you treat it as a stop sign and then give way to other traffic and any pedestrians crossing. Whether you are allowed to do this by default depends on the state, but either way, the default can be overridden by a textual plate under the traffic lights. Which the car doesn't currently pay any attention to.
Drew57
Active ember
Would someone from 2000 years ago believe what we have today was possible? I don't think so.
Imagine being confronted by a smart phone, global communications, an aircraft, space travel, a drone, limb transplants, microsurgery etc etc. Most of what we take for granted would literally blow their minds so I'm prepared to believe that fully autonomous vehicles will definitely happen.
Just because it doesn't fit with what we currently know doesn't mean it can't.
Imagine being confronted by a smart phone, global communications, an aircraft, space travel, a drone, limb transplants, microsurgery etc etc. Most of what we take for granted would literally blow their minds so I'm prepared to believe that fully autonomous vehicles will definitely happen.
Just because it doesn't fit with what we currently know doesn't mean it can't.
I imagine if they train it on enough data of those junctions it’ll learn that no one does a protected right and it won’t do it either. I suspect also if the signs are similar for that it will learn that sign means no protected right, even if it cannot read the text directly simply by the fact that anytime that sign is visible, no one does a protected turn.
What’s hard to know of course is if the car does this at the moment or if humans are telling it explicitly in the training data where it’s not allowed to do this.
Exactly and not only that but while we don’t get it in this country, you can watch videos of it in the US and see the progress that is being made. Ask me a few years ago before FSD Beta landed and I’d have told you it was probably 20 - 30 years out but they shocked me with just how capable even the early FSD Beta’s were. They might crack this close to fully in the US maybe in 5 - 10 years but it’s already extremely impressive.
I’d love to have it here and see how it progresses over the years in front of my eyes.
What concerns me, both for the UK and Europe…is that this neural network is based on millions of miles of film…this I assume was gathered over many years by a large cadre of FSD beta drivers in North America….where is our data going to come from ?….we are told that Tesla staff have been testing things…but this is going to be extremely limited…probably only fully covering Tesla service centres.
Can someone tell me I am wrong ?
Can someone tell me I am wrong ?
Tesla is actually using a combination of captured clips from the fleet (not just FSD) and simulation, as they can effectively ask all cars in a country to capture video clips of scenarios they are interested in.
They also use shadow mode where the neural net can capture clips when it predicts something different to what the driver is doing, even though the driver is in control and might not even have FSD.
So the more Tesla cars on UK roads the more data they can collect, although the chances of FSD Beta in the UK soon probably aren't that great.
Perhaps the new autonomous rules on motorways might help, which is the only place I use autopilot anyway and I do have FSD.
Based on the YouTube videos I've watched, I'm pretty sure that the car was just ignoring "no turn on red" signs. It's down to the human driver to intervene in those cases.
Maybe sometimes the restrictions are mapped, but pretty sure that even quite recently there was no recognition of the sign.
Yes this is one of their smart moves by putting all the hardware into every car sold they probably by now have a massive amount of data from the UK and other countries. Just don’t think they’ll have done any or little training on it yet is all.
Would all your years of driving be completely useless if you drove in the US ? Thought not.
I think this is perhaps an oversimplification of how we would theorise how FSD works. It doesn't really 'learn' how to drive, it's given specific instructions as to how to behave at traffic lights, road junctions etc. This will clearly change between different countries traffic regulations and how drivers behave. The core of how it recognises where it is in the world and what other cars are doing is what it learns and is largely the same between different countries, and only needs a small amount of relearning.
Anyway, they have all our cars collecting training clips for them, whether we are using FSD or not. I'm fully aware I will have been tagged as a negative example of what not to do.
JupiterMan
Active Member
there have been many FSDBeta videos this weekend showing no right on red junctions being treated correctly, the fleet is learning this already.
I recall Elon saying at some point that 90% of the learning can simply apply in Europe, and that they have only 10% specific training to do on European roads and signage.
Sorry if this sounds like trolling but I don't see the point of paying any attention to what is going on in the States regards FSD Beta. It has never beared any relevance to anything that happens in the UK/Europe. Our autonomous features regress while the States moves forwards.
I just don't get the excitement over it, from the perspective of a UK driver. I struggle to even rustle up any real excitement for changes or proposed changes to the UNECE, since unless it changes in a way that basically permits "FSD Beta" to be deployed pretty much "as is", I don't see it actually leading to anything actually happening with our cars.
I've had "FSD" as an option on my car coming up to 4 years now. "Automatic driving on city streets" was promised by the end of 2020. I have no confidence it will be here as it exists in the States in the next 4 years.
I just don't get the excitement over it, from the perspective of a UK driver. I struggle to even rustle up any real excitement for changes or proposed changes to the UNECE, since unless it changes in a way that basically permits "FSD Beta" to be deployed pretty much "as is", I don't see it actually leading to anything actually happening with our cars.
I've had "FSD" as an option on my car coming up to 4 years now. "Automatic driving on city streets" was promised by the end of 2020. I have no confidence it will be here as it exists in the States in the next 4 years.
Think that was how they did it before FSD Beta 12, it was a massive bunch of nested if and then statements probably
I think they said 300k lines of code. I'm sure a lot of that might still be around but they are working to remove a lot of it and now train it on neural networks of driving clips. It's interesting watching some of the FSD Beta 12 videos as one of the things it's learnt how to do is break the speed limit a bit on some roads where most humans do the same It also goes slower than the speed limit when most humans also seem to. It's learning to drive like we do.I suspect it will need to be trained on UK data because of how much more busy our roads are, it'll need to be more assertive and prepared to jump into smaller gaps than it does in the US. But yes for the most part a lot of that training might carry over. Avoiding humans and animals, going around parked cars and so on is kind of a universal thing the world over.
I think this is two different things, one is talking about the progress of the technology which is moving forwards pretty quickly and the other is when this will come to the UK. I think it's improving pretty quickly in the US and agree with your assessment that we won't see it for years in this country.
The autopilot that we get in the UK is dead, they don't really work on it and haven't done for years. I don't expect it to improve until they bring the changes they've made in the US over here. I think they could do this piecemeal as in just for motorways and dual carriageway A roads for instance as a starting point. I don't think there's probably any laws stopping them doing this even now but they still don't bother. I suspect they feel that autopilot as it is in this and other countries is good enough vs the competition that they just aren't interested in spending time on it at the moment.
Having said that you can see them maybe making progress to this. As I understand it FSD Beta 12 in the US is using FSD 11 on their interstates so they can still split these out and run different code depending on the roads. Also the camera tracking in the cabin that's now enabled to check you are paying attention, it's annoying but first steps towards them maybe improving this and offering something like Blue Cruise where it's hands off on motorways for instance. Depends how good they can get the eye tracking as most use those infrared cameras which do this easier but Tesla with their vision approach don't have those extra sensors but potentially outside of night time might be possible with their camera.
This is the crux of my point though.
This thread is titled "UK FSD discussion" so I posted my thoughts regards whatever scraps from the table we might get. I take your point that the discussion has broadened to talk about FSD Beta as it exists in the States, etc though.
In the past however many years Tesla have made it pretty clear that they aren't prepared to really provide customised software to Europe, on any front. It has taken 3 years past when they started fitting the hardware for them to deliver adaptive headlights software, in no small part I'll bet because the NHTSA in the States has taken steps to approve it.
My issue really is not that I think this technology isn't progressing, rather that realistically the landscape we find ourselves on with the UNECE and Tesla's unwillingness to customise software for us (assuming this is even practical for FSD Beta) means that all we can really do is look at what they get in the States and marvel at it, and that's basically it. People talking about how we'll benefit from "the single stack" etc are wilfully ignoring the fact that we haven't benefitted from any of it up until now. As said, our AP has gone backwards in functionality in recent years. Nothing has improved in any meaningful way.
From my layman's perspective it makes sense to focus on a single market and get the system to a certain level of competence before adjusting it for other jurisdictions. The USA makes as much sense as any to be that starting point. How far down the list will the UK be when it starts getting tailored for other countries is anyone's guess but I am not optimistic that we are high on that list.
Pink Duck
Active Member
It feels like we're in regulator watch territory at present, with the wider FSD Beta 12.3 roll-out to US happening. Depending if that goes well other global regulators might take note - particularly if safety rates improve because of it. However, with Tesla not publishing their annual safety report there's no real metric to judge as yet. Those with FSD 12.3 also note that the confidence of 12.2 has been dialled down, and I guess that makes sense since avoiding the PR hit should there be injury/fatality under FSD is crucial.
Boiling the ocean with FSD tyring to tackle all countries was never going to be a smart move, and one could argue that even within North American trying to be all things to all people in all situations is still doing to much when maybe nailing self driving on freeways to the point that they could have automation was a smarter move, but they are where they are.
NA seems to take a different more relaxed approach to regulations, they seem to step in when it goes wrong more than require every box ticked before they sanction it as ok, its a culture which has enabled them to grow as a country albeit at the expense of some lives. Boeing is living the consequences of that culture right before our eyes and the backlash that comes, seemingly the public disatisfaction having mopre of an impact that the regulatory framework. While Musk may be head strong, he may also be getting some strong advice as to the hard that would result if they for L3 and get it wrong.
Europe and ROW seem different, and there will be many boxes to tick before release. The question for me is whether Tesla decide pushing for L3+ in the US and having what you might consider true FSD is the bigger prize than Level 2 in Europe, however good that looks. That won't be solely in their gift to give, and just looking at the Waymo and the like and how they seem to plateau just short of the line for public purchase may encourage a european push.
On adaptive headlights, I think this is actually one of the tangible benefits of the work in the US. I think Tesla have what they think is an acceptable standard of understanding the cars surroundings. They are confident enough to know where to blank, and it should take simple maths to actually blank out that area with the headlights. Same with parking and summon. Personally, I don't use AP on the motorways as much as others so I very much doubt I'd use FSD in cities and town centres. If they can make AP more reliable on motorways, and the FSD engine should be able to provide some of that, then I may well use it more, so I'd take a general improvement in what we have and fixing the gripes.
I guess I'm the type that wants 99.9999% good performance on a subset of things in preference to 95% performance over a wider set. When its my life on the line, I just want whats provided to work, otherwise don't bother providing it.
NA seems to take a different more relaxed approach to regulations, they seem to step in when it goes wrong more than require every box ticked before they sanction it as ok, its a culture which has enabled them to grow as a country albeit at the expense of some lives. Boeing is living the consequences of that culture right before our eyes and the backlash that comes, seemingly the public disatisfaction having mopre of an impact that the regulatory framework. While Musk may be head strong, he may also be getting some strong advice as to the hard that would result if they for L3 and get it wrong.
Europe and ROW seem different, and there will be many boxes to tick before release. The question for me is whether Tesla decide pushing for L3+ in the US and having what you might consider true FSD is the bigger prize than Level 2 in Europe, however good that looks. That won't be solely in their gift to give, and just looking at the Waymo and the like and how they seem to plateau just short of the line for public purchase may encourage a european push.
On adaptive headlights, I think this is actually one of the tangible benefits of the work in the US. I think Tesla have what they think is an acceptable standard of understanding the cars surroundings. They are confident enough to know where to blank, and it should take simple maths to actually blank out that area with the headlights. Same with parking and summon. Personally, I don't use AP on the motorways as much as others so I very much doubt I'd use FSD in cities and town centres. If they can make AP more reliable on motorways, and the FSD engine should be able to provide some of that, then I may well use it more, so I'd take a general improvement in what we have and fixing the gripes.
I guess I'm the type that wants 99.9999% good performance on a subset of things in preference to 95% performance over a wider set. When its my life on the line, I just want whats provided to work, otherwise don't bother providing it.
