FSD is improving really quickly. I love it, but drivers here seem very aggressive. I turn FSD on and a giant pickup drives on my bumper. FSD tries to pass and the car being passed speeds up. And it's hard to trust 100% to me. LOVE Park Assist. Wish I could buy that on its own. Great video!
When he said "It did the right thing, crossing the double yellow line and going to the turning lane" Nooooooooo! It absolutely didn't do the right thing, just because you don't want to wait 4 more seconds or at worst an other red light, it doesn't mean that FSD should do what humans do because they are impatient. If we really want to save lives we have to stop saying that FSD should drive like a human. There are cases where FSD should violate traffic laws because there is no other alternative but for convenience purposes FSD should be following all traffic laws including speed limits.
@@pranshukrishna5105 FSD is going to solve a huge problem car accident deaths. When FSD is complete it's going to lower car insurance to $50 a month or maybe less. That's what is going to drive adoption not how much people like how it drives.
Tesla could be running a small robotaxi fleet like Waymo if they were willing to have human teleoperators. But it looks like they're waiting for FSD to be fully independent, perhaps also for the dedicated vehicle to be in production.
Europe has voted a law that authorize types of FSD like this to come in September 2024 ! Probably need to submit to the authorities some kind of testing to show before, but won't take long ! 👌
HW3 will not be noticeably limited in smoothness/comfort. As to safety, we'll be talking about a difference in the ballpark between 1 accident per 100 million miles vs 1 accident per billion miles. HW3 has compute equivalent to small mammals (e.g. squirrel) who's brains control complex musculo-skeletal systems that navigate complex environments, finding food, evading predators, pursuing prey, etc. Driving safely, efficiently and comfortably from point A to B is going to be well within that capacity. Where HW3 will be limited is in situations that need a deeper human understanding to make a decision. As robotaxi, such cars will just avoid routes which take them into such scenarios. If such a route is unavoidable, the robotaxi dispatch network will send a HW4+ car to pick up that fare. But HW3 will be fine for >>99% of routes, and its performance on those routes will be most likely be indistinguishable from HW4 by human passengers. _Eventually_ as robotaxi become ubiquitous and are permitted to drive "superhumanly" (i.e. faster, closer following distance, split-second timing of 'gaps'), then we'll notice the old HW3 models are a bit less superhuman than newer models (assuming we notice the driving performance at all and don't have our noses in books, watching movies, reading X, or sleeping...) What turns out to be much more difficult than Elon thought is getting the most out of that hardware. It's not just number of parameters, but how they are organized into a model and how that model is trained. With HW4 they have some breathing room -- a less optimized model can run in HW4 and be as good as the most optimal possible model in HW3. It's going to take quite a bit more work to create the optimal model that "fits" in HW3. This is a big reason they need to expand their super-computing training clusters as fast as possible. They need to train many different variations of model architecture & training data set to figure out what is most optimal within the constraints of HW3, so they need training itself to go as quickly as possible. Here is a speculation I give high chance of being correct: internally, Tesla has models which run on more powerful hardware than HW4, maybe more powerful than HW5, which they know already drive far, far better and safer than human (note: they don't need special hardware for this -- the same hardware they use for training can run these models. Or they can use HW3 or HW4 to test such models but run them in a "slow motion" simulation, where, for example, 10 or 100 seconds of real time is use for 1 second of simulated driving time -- that gets you a test of the abilities of an AI with 10 or 100 times as much compute).
Very good analysis, 100% agree, having no wheel at all will happen, hard to say when but not long than 5 years for sure. To say that let's just have a look in the past at the beginning of FSD beta (2020) where it couldn't almost stay in the lane... Then imagine that much improvement in the future in that not even 4 years, because yes technology never stops improving. Only one thing can stop this : money, but don't think Tesla is in a very shape... 🤞
Forget about HW3/4 - let’s hear what has happened to the thousands of dollars put into Tesla’s coffers over the years by purchasers who were told the vehicles they bought were FSD hardware capable?
Tesla just announced the release of FSD for Tesla MCU1 cars, their very first cars, before HW2.5, before HW2 there was MCU1. Those cars now have a version of FSD. All HW3 cars are FSD capable up to the ADAS level Elon promised. It might not be complete yet, there will be configurations that lag behind the others, but nobody has been left at a dead end.
Hi TechGeek Tesla, I love your content! I'm interested in discussing a potential collaboration. Could you let me know the best way to reach you? Thanks!😊
Do I interpret it correctly that in the USA it is not allowed to not stop at a stop sign even if there is no other traffic in sight at all, but it IS allowed to cross a DOUBLE yellow line? I just looked up what the potentially financial consequences would be in Europe: in Germany that would cost €30.- if nobody got endangered, €35.- if somebody was endangered, and here at home in Switzerland up to 350.- bucks. There might be more than just one reason why FSD won't be a thing in Europe within the next few years at the very least...
@charangohabsburg1 You can cross the double yellow in a case of emergencies such as if that is the only way to prevent a accident. That would make sense. For example you are in the process of crossing a intersection on green traffic signal, but someone takes the last available spot in the lane ahead and the car behind you blocks your retreat option. Your car is now sitting in the middle of the intersection and the light ahead turns red. Your car is now in the path of a perpendicular traffic. Assuming that there are no available turns to the left or right lane and your only option to get out of that situation is to cross the double yellow line, for a short distance, in order to get to a safe spot and not block the traffic, and you can and did it safely, then that action shouldn't be punished, imo. even though it's a violation of the road rules. Ie. It's better to break a minor rule to be safe, than to create a road hazard by being stuck in the path of oncoming traffic. It's a bit of a conundrum, similar to a trolley problem. Edge case for sure that will have to be solved by the fsd supervisors and lawyers. etc.
Thanks for the videos! While I appreciate your enthusiasm it is important to note that Tesla FSD vision-only can never be anything more than Level 2 - full supervision required. There are zero safety authorities that would ever approve a system that has zero alternative back-up sensor suites like every other actual working, approved self-driving system uses like Mercedes and Waymo. Tesla, itself, fully admits that vision-only will fail in common, everyday conditions in their FSD limitations section: "Visibility is critical for Full Self-Driving (Supervised) to operate. Low visibility, such as low light or poor weather conditions (rain, snow, direct sun, fog, etc.) can significantly degrade performance." Obviously "direct sun" occurs multiple times almost every day and still causes FSD to fail. Musk's own expert engineers warned him that vision-only could never work as a robotaxi, years ago and they have been proven right. At this moment, based on crowdsourced data from hundreds of FSD drivers, the very latest version of Tesla FSD 12.5.* is still over 3,500 times less reliable than an actual working, government-safety-approved robotaxi like Waymo. Musk's engineers were not being negative but simply factual.
I think all Teslas with HW3/4 will reach Level 4 autonomy. HW2.5 might make it to Level 4. I doubt HW2.5 or HW3 will ever be certified Level 5. However, knowing that the oldest platform, MCU1, is getting FSD, anything is still possible.
I think the time has come now to split hw3 and hw4 versions, ther are still so many gains on HW4 to be had and they should grab them now and let them trickle to HW3. maybe HW3 will get there but maybe it will need an upgrade. oh, and if you think the government will allow you to turn your car into a auto taxi and let you make 10s of thousands of bucks you have another thing coming. At best the car will do little more than pay for itself, after taxes and service charge.
![30 ชั่วโมงคนเดียว บนรถไฟชั้น 3 โซเวียต [Around The World Ep. 2]](http://i.ytimg.com/vi/MlE9boAVqDA/mqdefault.jpg)
TechGeek Tesla, You're the best! I just had to subscribe!
Great vid, you look very natural on cam.
Thank you for mentioning Europe!
FSD is improving really quickly. I love it, but drivers here seem very aggressive. I turn FSD on and a giant pickup drives on my bumper. FSD tries to pass and the car being passed speeds up. And it's hard to trust 100% to me.
LOVE Park Assist. Wish I could buy that on its own.
Great video!
Excellent video and discussion!.
I am really excited to see your next downtown Chicago video! Will it handle (or avoid) dwarves, goblins and trolls this time?
Great video! Keep ‘em coming!
When he said "It did the right thing, crossing the double yellow line and going to the turning lane"
Nooooooooo!
It absolutely didn't do the right thing, just because you don't want to wait 4 more seconds or at worst an other red light, it doesn't mean that FSD should do what humans do because they are impatient. If we really want to save lives we have to stop saying that FSD should drive like a human. There are cases where FSD should violate traffic laws because there is no other alternative but for convenience purposes FSD should be following all traffic laws including speed limits.
FSD has to be adjusted to human driving otherwise it would be unpopular
@@pranshukrishna5105 FSD is going to solve a huge problem car accident deaths. When FSD is complete it's going to lower car insurance to $50 a month or maybe less. That's what is going to drive adoption not how much people like how it drives.
Thanks!
@@rickmiles3701 Thank you! You don’t know how much that means to me. I really appreciate it!
Finally some unique opinions on something, thank you!
Tesla could be running a small robotaxi fleet like Waymo if they were willing to have human teleoperators. But it looks like they're waiting for FSD to be fully independent, perhaps also for the dedicated vehicle to be in production.
I'm to the point where I don't like riding in an ICE vehicle. Over three years of Tesla driving spoiled me.
yeah but Starsky and Hutch was out there with ya
Hah I thought I was the only one that noticed that.
Europe has voted a law that authorize types of FSD like this to come in September 2024 ! Probably need to submit to the authorities some kind of testing to show before, but won't take long ! 👌
We need FSD in EUROPE !!!
HW3 will not be noticeably limited in smoothness/comfort. As to safety, we'll be talking about a difference in the ballpark between 1 accident per 100 million miles vs 1 accident per billion miles. HW3 has compute equivalent to small mammals (e.g. squirrel) who's brains control complex musculo-skeletal systems that navigate complex environments, finding food, evading predators, pursuing prey, etc. Driving safely, efficiently and comfortably from point A to B is going to be well within that capacity. Where HW3 will be limited is in situations that need a deeper human understanding to make a decision. As robotaxi, such cars will just avoid routes which take them into such scenarios. If such a route is unavoidable, the robotaxi dispatch network will send a HW4+ car to pick up that fare. But HW3 will be fine for >>99% of routes, and its performance on those routes will be most likely be indistinguishable from HW4 by human passengers. _Eventually_ as robotaxi become ubiquitous and are permitted to drive "superhumanly" (i.e. faster, closer following distance, split-second timing of 'gaps'), then we'll notice the old HW3 models are a bit less superhuman than newer models (assuming we notice the driving performance at all and don't have our noses in books, watching movies, reading X, or sleeping...)
What turns out to be much more difficult than Elon thought is getting the most out of that hardware. It's not just number of parameters, but how they are organized into a model and how that model is trained. With HW4 they have some breathing room -- a less optimized model can run in HW4 and be as good as the most optimal possible model in HW3. It's going to take quite a bit more work to create the optimal model that "fits" in HW3. This is a big reason they need to expand their super-computing training clusters as fast as possible. They need to train many different variations of model architecture & training data set to figure out what is most optimal within the constraints of HW3, so they need training itself to go as quickly as possible.
Here is a speculation I give high chance of being correct: internally, Tesla has models which run on more powerful hardware than HW4, maybe more powerful than HW5, which they know already drive far, far better and safer than human (note: they don't need special hardware for this -- the same hardware they use for training can run these models. Or they can use HW3 or HW4 to test such models but run them in a "slow motion" simulation, where, for example, 10 or 100 seconds of real time is use for 1 second of simulated driving time -- that gets you a test of the abilities of an AI with 10 or 100 times as much compute).
So when the cat went around the traffic " Truck" it could it tell no car was in front of the truck or did it just follow the traffic?
i dunno, i think robotaxi will require additional hardware layers / sensors. i love using fsd but it’s fun until it almost kills ur fam :P
I'm on my 3rd family, still love FSD! 😊
Very good analysis, 100% agree, having no wheel at all will happen, hard to say when but not long than 5 years for sure. To say that let's just have a look in the past at the beginning of FSD beta (2020) where it couldn't almost stay in the lane... Then imagine that much improvement in the future in that not even 4 years, because yes technology never stops improving. Only one thing can stop this : money, but don't think Tesla is in a very shape... 🤞
13:01 Starsky & Hutch alert...
Forget about HW3/4 - let’s hear what has happened to the thousands of dollars put into Tesla’s coffers over the years by purchasers who were told the vehicles they bought were FSD hardware capable?
Tesla just announced the release of FSD for Tesla MCU1 cars, their very first cars, before HW2.5, before HW2 there was MCU1. Those cars now have a version of FSD. All HW3 cars are FSD capable up to the ADAS level Elon promised. It might not be complete yet, there will be configurations that lag behind the others, but nobody has been left at a dead end.
What is this talk about the hardware? Didn't Elon said 2 years ago "hardware features are complete for level 5"?
Hi TechGeek Tesla, I love your content! I'm interested in discussing a potential collaboration. Could you let me know the best way to reach you? Thanks!😊
Do I interpret it correctly that in the USA it is not allowed to not stop at a stop sign even if there is no other traffic in sight at all, but it IS allowed to cross a DOUBLE yellow line?
I just looked up what the potentially financial consequences would be in Europe: in Germany that would cost €30.- if nobody got endangered, €35.- if somebody was endangered, and here at home in Switzerland up to 350.- bucks. There might be more than just one reason why FSD won't be a thing in Europe within the next few years at the very least...
@charangohabsburg1
You can cross the double yellow in a case of emergencies such as if that is the only way to prevent a accident. That would make sense. For example you are in the process of crossing a intersection on green traffic signal, but someone takes the last available spot in the lane ahead and the car behind you blocks your retreat option. Your car is now sitting in the middle of the intersection and the light ahead turns red. Your car is now in the path of a perpendicular traffic. Assuming that there are no available turns to the left or right lane and your only option to get out of that situation is to cross the double yellow line, for a short distance, in order to get to a safe spot and not block the traffic, and you can and did it safely, then that action shouldn't be punished, imo. even though it's a violation of the road rules. Ie. It's better to break a minor rule to be safe, than to create a road hazard by being stuck in the path of oncoming traffic.
It's a bit of a conundrum, similar to a trolley problem. Edge case for sure that will have to be solved by the fsd supervisors and lawyers. etc.
Thanks for the videos! While I appreciate your enthusiasm it is important to note that Tesla FSD vision-only can never be anything more than Level 2 - full supervision required. There are zero safety authorities that would ever approve a system that has zero alternative back-up sensor suites like every other actual working, approved self-driving system uses like Mercedes and Waymo. Tesla, itself, fully admits that vision-only will fail in common, everyday conditions in their FSD limitations section: "Visibility is critical for Full Self-Driving (Supervised) to operate. Low visibility, such as low light or poor weather conditions (rain, snow, direct sun, fog, etc.) can significantly degrade performance." Obviously "direct sun" occurs multiple times almost every day and still causes FSD to fail. Musk's own expert engineers warned him that vision-only could never work as a robotaxi, years ago and they have been proven right. At this moment, based on crowdsourced data from hundreds of FSD drivers, the very latest version of Tesla FSD 12.5.* is still over 3,500 times less reliable than an actual working, government-safety-approved robotaxi like Waymo. Musk's engineers were not being negative but simply factual.
HW3 will have a better fsd in 2025.
HW3 fsd just came out🎉
I think all Teslas with HW3/4 will reach Level 4 autonomy. HW2.5 might make it to Level 4. I doubt HW2.5 or HW3 will ever be certified Level 5. However, knowing that the oldest platform, MCU1, is getting FSD, anything is still possible.
I love my Tesla model 3 and Y, but I hate to say but you will never see Tesla robotaxi. I kinda doubt HW 3 will be upgradable to what HW4 will have.
I think the time has come now to split hw3 and hw4 versions, ther are still so many gains on HW4 to be had and they should grab them now and let them trickle to HW3. maybe HW3 will get there but maybe it will need an upgrade.
oh, and if you think the government will allow you to turn your car into a auto taxi and let you make 10s of thousands of bucks you have another thing coming. At best the car will do little more than pay for itself, after taxes and service charge.