Full Self Driving in Australia

[arnyswart]

So, now that Tesla is moving to hardware 4 it seems like our cars will not be upgradable.

I note that full self upgrading can no longer be bought in my car but can be bought through the app.

What are others thoughts?

 

[paulp]

So, now that Tesla is moving to hardware 4 it seems like our cars will not be upgradable.

I note that full self upgrading can no longer be bought in my car but can be bought through the app.

What are others thoughts?

My thought is that anyone who bought FSD thinking it would give them autonomous driving or a robotaxi wasted their money, and it wont be any different for hardware 4. But if you just wanted the tesla marketting definition of FSD then version 3 will give you that. Tesla Aust are probably now avoiding version 3 upgrades whilst they assess the various representations that have been made, and their liability to various owners who purchaced based on those representations.

 

[moa999]

On a compute basis HW4 looks to be about 60-70% more powerful, yet it's dealing with processing cameras with 400+% more pixels.

It's a gradual improvement, not a step change.

What HW3 can do is still improving. Remember Tesla's first Dojo Exapod (which is intended to be the main training computer for FSD) still hasn't opened yet.

 

[Chuq]

I think it's the same as before - assuming FSD arrives, HW3 will still be capable. If you have already purchased it and you have an older HW vehicle, Tesla will have to "make good" at no cost (if they can't upgrade to HW4, this might mean HW3, but it's up to them).

 

[Hairyman]

I had read that Tesla wouldn’t be upgrading the hardware to HW4 unless HW3 and the available software would not outperform a human.

What counts as outperforming humans wasn’t described. At the moment it seems the software not the hardware is the primary issue. Tesla also seems focussed on better safety outcomes with the driver assistance packages, rather than an ability to navigate the car without intervention or supervision.

I suspect that actually fully describing how a human drives and navigates a dynamic environment is much harder than might be imagined. I do note that other manufacturers are also having phantom braking issues, which is one of my main irritations at the moment.

The software is still quite a way from the expectations that have been raised. I have seen quite an improvement in performance of FSD as time has gone past.

It does ebb and flow in the level of skill it displays. I think this is an issue of moving from one optimisation to the next better level. That’s typical for automated evolutionary processes. You are at one peak and all around is lower performance, but at some distance is another higher peak.

Moving from one to the next means that in the interim things are worse.

After that performance gets constrained by the wider strategy. Windscreen wiper rain detection would be an example. Tesla seems to want to go for a particular in house system that might be bringing other efficiencies and future outperformance, but other systems are better at present.

Then there is the difference in semantics for driving that means the car will drive like its machine learning model rather than just like a human. When I‘m using it I pay attention to how it behaves to learn how it drives and responds. It lets me anticipate situations where it is going to underperform and I take over in advance. I experiment with leaving it alone when conditions are safe to find out how things are progressing.

I bought FSD because I'm a nerd interested in epistemology and algorithmic learning, as well as all the cool stuff.

What I have seen lately is a big improvement in the automated high beam headlight control, smoothness of lane following through curves and lane following across intersections.

Across intersections I‘m talking about where the middle of the intersection is right on top of a crest or in the middle of a bend, both on multi-lane roads where there can be a lot of ambiguity about which is the correct lane to choose on the opposite side.

We don’t have access to FSD for urban environments, but I think it will be interesting when we get it.

Level 2 autonomy is probably the most awkward for the driver. In a situation where there is confusion it is a big cognitive load to suddenly have to decide is the car the problem or do I have an external threat. In those conflicts I just take over completely and then reflect on what was the actual issue. It saves the time and risk of deciding the car is the problem when the problem is actually external and you have delayed your corrective action.

I do find driving less fatiguing on long highway trips using FSD, but I’m more in a position of supervising a novice rather than letting an experienced driver just get on with the job.

So far I have had the active avoidance features save me from a likely collision when another vehicle moved into my lane. I haven’t had any circumstances when I didn’t see an error coming and so was put at immediate risk.

Others have widely varying experiences and I think that says a lot about how well humans can translate driving principles across very different conditions. I understand the frustrations that come from disappointed expectations.

I do think some of those expectation problems come from having someone in control of the company whose approach is fairly aggressively erratic. I’m not especially wedded to Tesla, but at the moment they still seem to be delivering the best value overall. I have an eye on the alternatives and there are some standouts in certain areas, but nothing that manages as well as a whole.

 

[EcoCloudIT]

I had read that Tesla wouldn’t be upgrading the hardware to HW4 unless HW3 and the available software would not outperform a human.

What counts as outperforming humans wasn’t described. At the moment it seems the software not the hardware is the primary issue. Tesla also seems focussed on better safety outcomes with the driver assistance packages, rather than an ability to navigate the car without intervention or supervision.

I suspect that actually fully describing how a human drives and navigates a dynamic environment is much harder than might be imagined. I do note that other manufacturers are also having phantom braking issues, which is one of my main irritations at the moment.

The software is still quite a way from the expectations that have been raised. I have seen quite an improvement in performance of FSD as time has gone past.

It does ebb and flow in the level of skill it displays. I think this is an issue of moving from one optimisation to the next better level. That’s typical for automated evolutionary processes. You are at one peak and all around is lower performance, but at some distance is another higher peak.

Moving from one to the next means that in the interim things are worse.

After that performance gets constrained by the wider strategy. Windscreen wiper rain detection would be an example. Tesla seems to want to go for a particular in house system that might be bringing other efficiencies and future outperformance, but other systems are better at present.

Then there is the difference in semantics for driving that means the car will drive like its machine learning model rather than just like a human. When I‘m using it I pay attention to how it behaves to learn how it drives and responds. It lets me anticipate situations where it is going to underperform and I take over in advance. I experiment with leaving it alone when conditions are safe to find out how things are progressing.

I bought FSD because I'm a nerd interested in epistemology and algorithmic learning, as well as all the cool stuff.

What I have seen lately is a big improvement in the automated high beam headlight control, smoothness of lane following through curves and lane following across intersections.

Across intersections I‘m talking about where the middle of the intersection is right on top of a crest or in the middle of a bend, both on multi-lane roads where there can be a lot of ambiguity about which is the correct lane to choose on the opposite side.

We don’t have access to FSD for urban environments, but I think it will be interesting when we get it.

Level 2 autonomy is probably the most awkward for the driver. In a situation where there is confusion it is a big cognitive load to suddenly have to decide is the car the problem or do I have an external threat. In those conflicts I just take over completely and then reflect on what was the actual issue. It saves the time and risk of deciding the car is the problem when the problem is actually external and you have delayed your corrective action.

I do find driving less fatiguing on long highway trips using FSD, but I’m more in a position of supervising a novice rather than letting an experienced driver just get on with the job.

So far I have had the active avoidance features save me from a likely collision when another vehicle moved into my lane. I haven’t had any circumstances when I didn’t see an error coming and so was put at immediate risk.

Others have widely varying experiences and I think that says a lot about how well humans can translate driving principles across very different conditions. I understand the frustrations that come from disappointed expectations.

I do think some of those expectation problems come from having someone in control of the company whose approach is fairly aggressively erratic. I’m not especially wedded to Tesla, but at the moment they still seem to be delivering the best value overall. I have an eye on the alternatives and there are some standouts in certain areas, but nothing that manages as well as a whole.

I have FSD in our X (latest HW before 4.0 etc.) and just normal AP in a 3 year old Model 3...with the exception of the auto lane changing etc. there is zero difference in how they perform.....I do not think EAP is any different than AP in Australia with the exception of enabled features (lane change, summon etc.)...no actual software change per se....so all the waffle about different code base in baseless in the case of Australia and EAP/AP.

So, in Australia you are not testing FSD, more so you are testing any improvements that Tesla filter down, if any, from FSD in the US to general release in EAP/AP.

 

[paulp]

I had read that Tesla wouldn’t be upgrading the hardware to HW4 unless HW3 and the available software would not outperform a human.

What counts as outperforming humans wasn’t described. At the moment it seems the software not the hardware is the primary issue. Tesla also seems focussed on better safety outcomes with the driver assistance packages, rather than an ability to navigate the car without intervention or supervision.

I suspect that actually fully describing how a human drives and navigates a dynamic environment is much harder than might be imagined. I do note that other manufacturers are also having phantom braking issues, which is one of my main irritations at the moment.

The software is still quite a way from the expectations that have been raised. I have seen quite an improvement in performance of FSD as time has gone past.

It does ebb and flow in the level of skill it displays. I think this is an issue of moving from one optimisation to the next better level. That’s typical for automated evolutionary processes. You are at one peak and all around is lower performance, but at some distance is another higher peak.

Moving from one to the next means that in the interim things are worse.

After that performance gets constrained by the wider strategy. Windscreen wiper rain detection would be an example. Tesla seems to want to go for a particular in house system that might be bringing other efficiencies and future outperformance, but other systems are better at present.

Then there is the difference in semantics for driving that means the car will drive like its machine learning model rather than just like a human. When I‘m using it I pay attention to how it behaves to learn how it drives and responds. It lets me anticipate situations where it is going to underperform and I take over in advance. I experiment with leaving it alone when conditions are safe to find out how things are progressing.

I bought FSD because I'm a nerd interested in epistemology and algorithmic learning, as well as all the cool stuff.

What I have seen lately is a big improvement in the automated high beam headlight control, smoothness of lane following through curves and lane following across intersections.

Across intersections I‘m talking about where the middle of the intersection is right on top of a crest or in the middle of a bend, both on multi-lane roads where there can be a lot of ambiguity about which is the correct lane to choose on the opposite side.

We don’t have access to FSD for urban environments, but I think it will be interesting when we get it.

Level 2 autonomy is probably the most awkward for the driver. In a situation where there is confusion it is a big cognitive load to suddenly have to decide is the car the problem or do I have an external threat. In those conflicts I just take over completely and then reflect on what was the actual issue. It saves the time and risk of deciding the car is the problem when the problem is actually external and you have delayed your corrective action.

I do find driving less fatiguing on long highway trips using FSD, but I’m more in a position of supervising a novice rather than letting an experienced driver just get on with the job.

So far I have had the active avoidance features save me from a likely collision when another vehicle moved into my lane. I haven’t had any circumstances when I didn’t see an error coming and so was put at immediate risk.

Others have widely varying experiences and I think that says a lot about how well humans can translate driving principles across very different conditions. I understand the frustrations that come from disappointed expectations.

I do think some of those expectation problems come from having someone in control of the company whose approach is fairly aggressively erratic. I’m not especially wedded to Tesla, but at the moment they still seem to be delivering the best value overall. I have an eye on the alternatives and there are some standouts in certain areas, but nothing that manages as well as a whole.

The alternative cars depend on your personal criteria, personal perception of value, and assuming EV is an essential for everyone. For my criteria tesla are very wide of the mark. Rampant acceleration and automation are so far down my list they do not even register, but thats just me.

 

[sulrich70]

Tesla Aust are probably now avoiding version 3 upgrades whilst they assess the various representations that have been made, and their liability to various owners who purchaced based on those representations.

THIS. Expect a class action...

 

[ForeverFree]

I have FSD in our X (latest HW before 4.0 etc.) and just normal AP in a 3 year old Model 3...with the exception of the auto lane changing etc. there is zero difference in how they perform.....I do not think EAP is any different than AP in Australia with the exception of enabled features (lane change, summon etc.)...no actual software change per se....so all the waffle about different code base in baseless in the case of Australia and EAP/AP.

So, in Australia you are not testing FSD, more so you are testing any improvements that Tesla filter down, if any, from FSD in the US to general release in EAP/AP.

Hi, US owner here. Was lucky enough to drive Model 3 with FSD for three weeks in Sydney over the recent summer holidays. (evee is great!)

Noticed that FSD is very different there than here. For instance, car tried to stop for every green traffic signal on the road up to the Blue Mountains.

Here, it actually tries to fully self-drive … but just isn’t very good at it.

By the way, Sydney and surrounding NSW are amazing!!

 

[donmac]

Noticed that FSD is very different there than here. For instance, car tried to stop for every green traffic signal on the road up to the Blue Mountains.

We don't actually have FSD. Currently, it is restricted to North America.
We do have some of the subfeatures, such as "Traffic Light and Stop Sign Control" as you observed.
Over the past year it has become better at automatically proceeding through green lights when there are vehicles ahead in the same or adjacent lanes.
It considers roundabouts as being similar to traffic lights and stop signs, but could do with more improvement there.

 

[Hairyman]

Hi, US owner here. Was lucky enough to drive Model 3 with FSD for three weeks in Sydney over the recent summer holidays. (evee is great!)

Noticed that FSD is very different there than here. For instance, car tried to stop for every green traffic signal on the road up to the Blue Mountains.

Here, it actually tries to fully self-drive … but just isn’t very good at it.

By the way, Sydney and surrounding NSW are amazing!!

Glad you had a great time. Thanks for the insight.

 

[blaqmamba]

When I ordered my first Tesla in 2016, they said FSD was a few years away. I was pretty sure back then (7 years ago), that actual FSD was decades away, not years. And here we are in 2023 and I am still convinced actual FSD is decades away (unfortunately). Can't blame Elon for aiming for the stars to achieve great things, but these ambitions should be kept separate from their consumer marketing strategies.

 

[Nobody Famous]

THIS. Expect a class action...

Will it need that when we have Australian Consumer Law?

 

[Maximillan]

When I ordered my first Tesla in 2016, they said FSD was a few years away. I was pretty sure back then (7 years ago), that actual FSD was decades away, not years. And here we are in 2023 and I am still convinced actual FSD is decades away (unfortunately). Can't blame Elon for aiming for the stars to achieve great things, but these ambitions should be kept separate from their consumer marketing strategies.

 

[Jane4]

I had read that Tesla wouldn’t be upgrading the hardware to HW4 unless HW3 and the available software would not outperform a human.

What counts as outperforming humans wasn’t described. At the moment it seems the software not the hardware is the primary issue. Tesla also seems focussed on better safety outcomes with the driver assistance packages, rather than an ability to navigate the car without intervention or supervision.

I suspect that actually fully describing how a human drives and navigates a dynamic environment is much harder than might be imagined. I do note that other manufacturers are also having phantom braking issues, which is one of my main irritations at the moment.

The software is still quite a way from the expectations that have been raised. I have seen quite an improvement in performance of FSD as time has gone past.

It does ebb and flow in the level of skill it displays. I think this is an issue of moving from one optimisation to the next better level. That’s typical for automated evolutionary processes. You are at one peak and all around is lower performance, but at some distance is another higher peak.

Moving from one to the next means that in the interim things are worse.

After that performance gets constrained by the wider strategy. Windscreen wiper rain detection would be an example. Tesla seems to want to go for a particular in house system that might be bringing other efficiencies and future outperformance, but other systems are better at present.

Then there is the difference in semantics for driving that means the car will drive like its machine learning model rather than just like a human. When I‘m using it I pay attention to how it behaves to learn how it drives and responds. It lets me anticipate situations where it is going to underperform and I take over in advance. I experiment with leaving it alone when conditions are safe to find out how things are progressing.

I bought FSD because I'm a nerd interested in epistemology and algorithmic learning, as well as all the cool stuff.

What I have seen lately is a big improvement in the automated high beam headlight control, smoothness of lane following through curves and lane following across intersections.

Across intersections I‘m talking about where the middle of the intersection is right on top of a crest or in the middle of a bend, both on multi-lane roads where there can be a lot of ambiguity about which is the correct lane to choose on the opposite side.

We don’t have access to FSD for urban environments, but I think it will be interesting when we get it.

Level 2 autonomy is probably the most awkward for the driver. In a situation where there is confusion it is a big cognitive load to suddenly have to decide is the car the problem or do I have an external threat. In those conflicts I just take over completely and then reflect on what was the actual issue. It saves the time and risk of deciding the car is the problem when the problem is actually external and you have delayed your corrective action.

I do find driving less fatiguing on long highway trips using FSD, but I’m more in a position of supervising a novice rather than letting an experienced driver just get on with the job.

So far I have had the active avoidance features save me from a likely collision when another vehicle moved into my lane. I haven’t had any circumstances when I didn’t see an error coming and so was put at immediate risk.

Others have widely varying experiences and I think that says a lot about how well humans can translate driving principles across very different conditions. I understand the frustrations that come from disappointed expectations.

I do think some of those expectation problems come from having someone in control of the company whose approach is fairly aggressively erratic. I’m not especially wedded to Tesla, but at the moment they still seem to be delivering the best value overall. I have an eye on the alternatives and there are some standouts in certain areas, but nothing that manages as well as a whole.

We paid for FSD back in 2020 and regret doing so, since during Covid, we didn't travel much and are still catching up with travel.I like the Cruise Control and use it exclusively and consistently and know that the car has kept me from several collisions. We have to cross near some interstates coming to and from town and people often think they can dart out in of front of me. My favorite thing is adjusting the speed using the "rolling balls" on the steering wheel (I am sorry I don't know the official name.) And have driven for hours on state roads here in NM which can be very curvy and laden with big rigs and felt very safe. I love the one peddle driving and not having to step on the brake or the gas. Feels less stressful.

 

[hooty]

“eLoN dIdNt UsE tHe WoRd ‘PrOmIsE’”
- any common garden variety fanboi

 

[Quickst]

Now that M3 highland has been available for press review in Australia, any news about when HW4 will be available in the Australian M3 Highland?

 

[Zinzilla]

Interesting thing, after all the talk of Dojo, is that Tesla have bought a big Nvidia cluster (10 000 H100s). Dojo and the H100 are quite different architectures, perhaps they're finding one's better for training one type of NN and the other for another (and they already had an A100 cluster, which must be well overdue for retirement).

Nvidia is powering a mega Tesla supercomputer powered by 10,000 H100 GPUs

Elon Musk hopes to crack fully autonomous vehicles with an incredibly powerful supercomputer

www.techradar.com

 

[Jules22]

Now that M3 highland has been available for press review in Australia, any news about when HW4 will be available in the Australian M3 Highland?

The new M3 comes with HW4