31:13 This is an extremely important issue in the field of AI in general that you're describing, and I'm really glad you brought that up. It's called the black box problem. Whether the AI does the right thing or not, it's impossible to trace how it reached the conclusion it did. There's been a huge push to make AI explainable and more transparent (as in outputting not only acceleration and steering, but the reasoning that led it to make that action) since right now, we have literally no way of knowing what the algorithm is really doing. Transitioning to this completely different kind of AI architecture would require a full overhaul, and that's especially difficult since AI is used all over the place already.
What do you want it to print? A bunch of weights and biases? But I guess you can make some kind of memory storage for path planning confidence level and look at those. Neural network is currently best method to crack FSD now. Procedural programming won't work.
@@qui11 Exactly. Conventional programming won't work as well as a neural net for FSD. I'm not saying we shouldn't use AI, I'm saying we need to figure out how to make AI safer so that we are in control of it, not the other way around. For example, back when cars were a new invention, we didn't decide to stop using cars when we found out they were dangerous because we knew the potential they had to revolutionize transportation. Instead, we invented safety measures like seatbelts, crumple zones, and airbags to make them more safe and driver-focused. This is what needs to happen with AI.
@@qui11 You want it to explain itself in text. So "Waiting for person to cross" or "Breaking for red light" or "Lane change because x". And it should even be able to explain why it chose a certain path. The occupancy network was cool.... yet they scrapped that. Now I'm not sure if FSD can handle completely novel problems. Like tools sticking out of a truck. Or a car painted on a bus.
Remember when FSD is on, the right thumbwheel adjusts the max speed. if you're going to fast , or slow you can either spin it down to slow 1mph at a click, or a quick flip changes it 5mph, or close to.
The camera suctioned in the middle of the view of the road is distracting. It's hard to see what the driver would see, in order to perceive how it's doing at driving.
Great video and some deep questions. I have been in cars with other people driving, some you are relaxed with and some make you nervous!. Also driving a slow works van ,getting in a fast car and seeing how close you get to a stop line before the driver brakes! It’s a surprise but their car stops much quicker than my van. I think fsd needs to get to a point where it drives most safely but also makes the most efficient progress. Then everyone will just have to get used to how it drives. For example it can change lanes in places where, a human could not be sure it was safe ,because it can see in 8 directions at once , no blind spots. It can also go from accelerating to braking in one millionth of a second ,lightning fast reactions. This could lead to driving closer to the car in front , because it knows it can out react it! And lastly if you were looking at your phone you would not even notice the driving unless something quite serious happened. Thanks for an interesting video.
This is one of the best (if not the best) FSD driving videos I have seen. I watch all the other standard ones and cannot make it to the end without throwing my shoes at the TV. The first time I hear the words "... I would have bla bla bla" I throw a shoe and stop the video. These videos should be about evaluating Self Driving not about your driving habits. Your friend was super cool.
One thing to note about why FSD doesn't currently do auto-park on it's own while in FSD, Elon has said recently that currently the auto-park code is completely separate from the FSD inference model. I wouldn't doubt it if auto-park, in it's current form, isn't performed by an inference model at all. So combining that behavior with FSD might be a lot of effort. It obviously needs to be done, but it's not trivial.
@@markfrantz4612 well they might use NN’s while in the Autopark mode but those NNs are likely completely separate and trained separately. And maybe Autopark isn’t as “end-to-end” as FSD.
Another great video. Love to hear the thoughts of a Tesla "wait-and-seer." 😅 Colin made a good point in the video about how predictable the car's actions become once you've ridden with FSD for a while, and that is a good thing from a passenger perspective. Massive driving style changes from iteration to iteration could be jarring. Along those lines, at one point you mention it doesnt swerve to avoid a pothole that you would have avoided. This is a bit of a conundrum. On one hand, obviously to save our tires and suspension we would love for FSD to do some fancy maneuvering and miss them. But as a passenger, that evasive maneuver could be startling if you dont also see the same pothole and anticipate the cars quick steering input. Tough decisions to be made by the Tesla team on what level of unpredictable behavior we can stomach.
Four FSD Teslas arrive at a four-way stop at practically the same time ... Do they? a) sit there and play Toybox games with each other until other cars arrive b) just sit there c) perform for each other with their matrix headlights d) proceed normally
C - because unlike the people driving cars, the FSD does the full stop so the Tesla can "see" it and follow the "program". When FSD is hesitant is when two or more cars are at the stop at the same time and one of them is not doing a full stop.
I like your drives with FSD. It gives me a better understanding of improvement from the previous version. I try to keep track of different scenarios, most interesting things are not been tested by early testers yet. My opinion of FSD 12.5 shows readiness, on a scale from 1 to 10: 1 - Not working, 3 - Working, not useful, 5 - Good, 9 - Like human, 10 - Better than human. 7 - Highway: Performs well, but not perfect(FSD11). 5 - Exit from Highway: Adequate, but requires more refinement, miss exist or change lane too late(FSD11). 7 - Change Lane: Generally effective, but occasional issues. 8 - Intersection: Handles most scenarios smoothly. 8 - Roundabouts: Navigate effectively, though not flawlessly. 4 - Parking Lots: Functional but can struggle with complexity. 3 - Parking on the Street: Needs better spot finding to avoid blocking traffic. 8 - Reverse Parking: Decent, but not consistently reliable. 2 - Gate: Needs significant improvement for reliability. 7 - Signal to Other Cars: Generally good, but improvements are needed. 7 - Left Turn: Handles most left turns well. 4 - Unprotected Turns: Often cautious or hesitant. 5 - Speed Bumps: Recognizes and slows down appropriately. 3 - Constructions on Road: Needs better detection and navigation. 4 - Road Blocks: Struggles to handle unexpected obstacles. 8 - Rain: Performs adequately, but not in heavy rain. 3 - Snow: Limited capability in snowy conditions. 5 - Fog: Manages light fog, and struggles with heavy fog. 4 - Other Cars Breaking Rules: Handles some scenarios, but not all. 8 - Pedestrians: Good at detecting and reacting. 5 - Other Road Vehicles: Adequate, but can be improved. 7 - Traffic Lights: Generally accurate, occasional misreads. 7 - Driving Smoothness: Mostly smooth, with minor jerkiness at times. 2 - Potholes or debris on the road: I don't see any lower speed or avoid 6 - Speed Limits: Adheres well, but sometimes slow to adjust. 3 - Emergency Vehicles: Not recognized can be improved. 5 - Supervision: Nice that no hands on the wheel are required. 1 - Front Parking: Major improvements are needed. 1 - Reverse in Street: Significant challenges remain. 1 - Actual Smart Summon(ASS): 1 - Self-park(Banish): 1 - Ferry Crossings: 4 - Train Crossings: 1 - Charging Stations: My assessment is open for discussion, but this is my opinion based on the videos that have been published so far. If there haven't been any new updates, I base this on earlier versions. I assume my opinion may change in the future, but right now, I don't see any area where FSD is as good as or better than a human driver, even though it can sometimes surpass humans in certain scenarios.
Your assessment is rational unlike that. I would point to the fact that FSD has a lower accident rate than humans. That would seem to indicate it's better than humans in some scenarios.
@@macmcleod1188 Most humans rarely have accidents, with only a small percentage experiencing frequent incidents. However, I often see drivers needing to take control in almost all situations, even with version 12.5.1 of FSD technology. While FSD is generally better than most human drivers, it still makes mistakes even in the best scenarios. For FSD to be considered better than a human driver, it would need to go at least a year without any mistakes or accidents that a human driver couldn't avoid. I hope we reach that level of reliability soon.
@@annawilson4571 humans have an accident every 8 years on average. Humans also have an accident roughly every 150,000 miles of driving. Self-driving Tesla have a record of only having an accident every 450,000 mi. I do agree with your point that as long as a human needs to take control, FSD should top out an eight.
@25:21 U R spot on. Munich Germany (in the 80s) No Left turns, all right turns because they found it safer and actually made less traffic jams. Was hard to get used to, but I didnt even notice the unfamiliarity feeling after a few trips into the city.
Using the camera for rain sensing also means that it needs light... When driving in badly/non-lit areas at night, the number of wipes drops dramatically. Also, the software needs a "water repellant applied" setting.
My biggest issues. Starts off way too damn fast, even on chill mode. I'm trying to keep my tires more than 5,000 miles. The NHTSA stop. Dude, that thing is going to get me rear-ended. It makes people behind me think I'm brake checking them. Literally no human driver on earth stops that far back, so no one expects it. Other than that it's nearly perfect, but those two things are just holding me back.
The change in behavior where FSD stops at the Stop Sign and then immediately starts to creep has changed the dynamics considerably for people behind me. I often would use the accelerator pedal when I had someone right behind me so they would be annoyed and/or honk. Now I never do with v12.5.1 and not a single driver behind me has honked or appeared agitated. A nice change. I definitely agree the NHTSA insistence on stopping is stupid but Tesla has done a reasonable job with v12.5.1 to address that.
“No human driver on Earth”? Just because the USA has a silly way to define when to yield or not, doesn’t mean the rest of the world makes the same mistake.
@@z4zuse People generally expect you to stop just shy of where the road meets the end of the other road. Not 20 feet back. Therefore it confuses people behind you
I agree that starts and stops need to be dialed down. The main road in my town has multiple stop lights not very far apart. When a light turns green, if there are no cars directly in front of me, FSD takes off trying to reach the 40mph speed limit very quickly, only having to slow down or stop abruptly within another few seconds. Some people may find this acceptable, but it is a bit too aggressive for me. At least give me an option in Chill mode to smooth this out a bit. Regarding stops, I have no issue with the NHTSA stop requirement the way it is now implemented in 12.5 and 12.5.1 (12.3.6 was not good). But yes, the completed stop is too far back from the crossroad in most cases. I think that it should be closer to the actual crossroad itself.
I actually had my Model Y stop for a yellow light on 12.4 (I think) then went on reverse to get off the cross walk. Maybe it goes on reverse if its a safety issue.
You are correct about the preference for right turns. When I visit a friend, I back into his driveway. When I leave, FSD wants to turn right and loop around, while I always go left and save about 3 blocks. Both routes come together at the same, unavoidable left turn. Same thing leaving my home, I'm one lot from an intersection and it always wants to turn right from my driveway and take me almost 3 blocks out of my way.
I think the no rain sensor approach is also because they are thinking so far into the future where it doesn’t matter if the human in the car can see. Only the car needs to see. However, I think they learn a lot using only the camera to determine how much water is on the windshield, which the human may not be able to see anymore. It's all about picking the right trainingdata. 27:48 ⛈
Waymo is level 3, as there is a remote operator ready to take the wheel on demand. Autonomous levels are a very outdated notion anyway. It doesn't take into account features like geofencing, choosing a destination, the energy consumption of the system, road conditions, parking or the price/need of extra equipment from the factory car and such.
Waymo has a much more advanced vehicle than Tesla, and Waymo is in business today, giving rides to 1000's of passengers every month. Meanwhile, Tesla robotaxi is *years* *behind* *schedule* .
@@DerekDavis213 Excuse me to say it, but this is nonsense. Tesla is in business today, selling and renting FSD all over the US and Canada, making millions of rides, and not just in a little square of sunny towns here and there. The "much more advanced vehicles", IIRC are Jaguar I-Paces, which design dates from 2018 and didn't change since then, which where already rather bad back then, plus hundreds of thousand dollars of very ugly, bulky and power hungry equipment (which is probably why the chose an electric vehicle with a big battery). The only advantages of Waymo today are an hailing app and the ability to have one person supervise several cars at once.
Excellent video!! Please do it again soon! I am questioning whether or not the occasional re-routing that FSD does when it disregards map data, and proceeds in doing its own thing, is simply the inference computer making the “wrong” choice. When I first got v12.5, I put in a destination that I’ve gone to many times. But this time my car went into the left turn lane at a stop light and turned on the left turn signal. That was at odds with map routing; the shortest and fastest route was straight ahead; there wasn’t much traffic at all. As I waited at the intersection, I was ready to take over because I didn’t know if it might just try following the mapped route from the wrong lane (as sometimes would happen with 12.3.6). However, my car did indeed turn left. Afterward the map updated with a new route, which was a longer and slower. So idk. V12.3.6 would often lane dance as it approached this (and other) intersections. I have not experienced any lane uncertainty at all with 12.5 or 12.5.1. So when my car moved into the left turn lane quickly and confidently I wondered what was going on. One possibility that comes to mind is that FSD is sometimes still making a wrong choice (different than map route), but now it commits to it (no dancing, proper signal). It has since always chosen the straight ahead route, even in heavier traffic. I may be way off base, and I would like to believe that AI is truly at work here, but if FSD is purposely choosing a different route, some direct evidence would be nice.
The red light situation: It is not learning between goes. Small things that might seem insignificant to a human can influence the confidence of an AI system. During the first go FSD simply wasn't sure enough that the object was a red light so it didnt stop. They can simulate the scenarios exactly as they take the recording data and can feed it through the exact same model and it will ignore the red light again. When you drive there are small differences that change the confidence intervals so it can be hard to replicate the scenario where it doesnt stop. Tesla also might not really care what caused fsd to drive off the cliff as much as tesla cares to never let it happen again. They store a lot of critical situations that caused problems in the past and validate the new models on these situations before releasing them. You need to be wise about what you validate and train on because if you have overabundance of boring scenarios FSD will not treat critical situations seriously enough. If you just feed it critical situations it will be too timid or jerky in regular driving scenarios. (this actually happened with one of their versions) You do not have infinite storage space nor time to validate so there is a balance there as well.
22:35 Look into the progression of 'Alpha Go' the AI that plays the game GO. The short of it; early training was done using human game data, later versions created their own game data and trained off of playing itself. Tesla is also able to create its own input data now (all sensors) and video using AI. Once the new training system is online, things are going to get wild.
It's a bit unfair to be hypercritical of FSD when no other cars comes close. For instance, Mercedes Level 3 speed is capped and is only legal in CA and AZ. And FSD is video only and overall is a very hard problem. I think Tesla engineers are heros.
6:47 That’s the reason I bought FSD in 2021 (not knowing when- or if it would be implemented in my country ‘Belgium EU’), in a medical emergency it could be faster to navigate your car to a hospital ER than calling- and waiting for a ambulance to get you there the fastest way…
The ambient lighting in the '24 Model 3 is OK, but if it is on while driving, I can see it on the windshield like a heads-up display, right where I don't want anything. Not Tesla's best design .
I don’t dislike it, he’s coming from an Ioniq 6 which has really nice designs on the door panels Adjusting the brightness is coming in a software update
@@MetzTech Fair enough. I enjoy your videos. Just tiring of reading the word “hate” on social media these days. To be fair, rarely if ever, do I see it on your feeds. 👍
I hate sitting at stop signs , I’m from the uk and we have roundabouts most of the time so lots less stoping. I would also take the path of least resistance!
Tesla cameras can not only see visible light range but also extend into the near-infrared (NIR) spectrum. Which is why FSD performs better than humans at night.
I have this theory that Tesla will launch a robotaxi pilot program with FSD (unsupervised) in parallel with FSD (supervised). When the performance of the unsupervised version surpasses that of the driver supervised system, they will go to the regulator and ask for regulatory approval for the whole fleet. The measure will be miles driven between accidents. This makes sense from both an ethical standpoint as a technical standpoint.
But you and i are not perfect neither. We are mistakenly we are perfect and killed 43,000 of us on roard last year. It is funny how ridiculous we are when we make critical judgement 😂
They do take gyroscope and accelerometer data. There is a video on how the planner calculates multiple paths and picks the smoothest one from the proposed ones. In order to train the system to pick smooth paths you need all that data. In a way this kind of approach of generating multiple possibilities and picking the best one seems to be used a lot in AI. Next version of ChatGPT will most likely use a search algorithm that generates multiple answers during training.
12:45 hardware V4 has IR emitters mostly for short-range illumination. The AI takes in the raw camera feed, so I imagine it has a wider sensitivity than a human eye. Wonder if future camera versions will have an iris to adjust light input like humans, that would be cool.
Can you test if this version know how to read the meaning of solid yellow line and broken yellow lines (passing and no passing zone)? A wall street analyst testing pre - FSD 12.5 claims he has to disengage because it tried to pass on a no passing zone. This test case requires two cars, the lead car will intentionaly drive slow so that the tesla following it will be forced to overtake it. Thanks. A standard broken yellow line tells you that passing is allowed (a passing zone). When there is one broken line and one solid line, traffic traveling adjacent to the broken line can pass (passing zone), and traffic traveling adjacent to the solid line cannot pass (no-passing zone). Two regular solid yellow lines mark where crossing the center for passing is prohibited (a no-passing zone). It applies to traffic traveling in either direction.
I am totally new to FSD so a couple of off the wall newbie questions. Can you make your own route for it to follow? What about off road, can you make off the road route? Can a legless invalid operate it now?
Sure you could do a custom route. You could probably make it go off-road but it’s not going to know what to do. I’m sure there are ways to customize the vehicle so that someone with those types of impairments could operate using their hands
Cool review. Disagree on rolled stops, regulators need to get real, 2 or 3 mph vs 0 *where safe?* Zero may be required at any time based on conditions, but it's not magically safer than a continuous creep or walking sp roll. Full stops cause other drivers anxiety, anger when they aren't needed, wasteful of energy, less comfortable. (Immense driver judgement required lk driving in general)
My FSD is going way too fast in the city. It tries to go from 0 to speed limit in the city street in a hurry and breaks hard. Why? I’m on chill mode. Way too aggressive. Seems to be worse after upgrading to 12.5.1
3:55 Mainly for regions where FSD is under review for admission by the government (Elon tweeted that Europe is planned to rollout in Q1 2025 if EU rules allow it) 🤔
I find it really odd that tesla would allow the car to do that much over the speed limit. I could understand like +4 mph over, but realistically, it shouldn't speed, should it? Do they allow that because so many people speed normally?
AI is about data and train models and designing and evolving new models so that the loop closes or coverges on a behavious or ability to drive in the same weather and road conditions as a human. To be able to navigate and plan and to be as safe as doing all of this depend trafic conditions and the type of other road users. All these variables we as haman have also been trained. We only become compitent at driving after 5-8yrs, after our passing the test.
So freaking close. You can tell the model wants to put you in a spot because that’s what people Do at the end of a drive. It just can’t do it right yet
Definitely! We do the Metz Talk podcast together if you want to check those out. It’s been a minute since we recorded an episode but we have a nice backlog of episodes and will have more coming in the near future! @metztalk
I've driven FSD for 2 years now and it sucks. It's 99% accurate and that 1% will kill you Is that good enough for elon? How about 99.9% accurate - that good enough elon? The truth is elon has no intention of telling us "how good" FSD will have to be to become "ready" and this is on purpose FSD was NEVER designed to operate in a Tesla with a steering wheel and 2 pedals it was designed to run a RoboTaxi Want proof? Engage the new FSD and fiddle with your iPhone above the steering wheel FSD will be happy and the cop who just pulled you over will cite you for reckless driving and the judge will suspend your license for a year and you told both of them "elon said FSD was ready" and both laughed When FSD is "ready" that $10,000 I paid for it means nothing since I still have to drive the car...
I don't get it. You're complaining it's too assertive, but don't take it off the assertive mode? Your brain clearly works completely differently to mine!
Neural Network outcomes are inherently based on probability. When the neural network probability of success is far above human success rates, Tesla would probably take full responsibility. Great video. Thank you!
Tesla FSD will probaly max out @lv3 it is a good system if u are behind it but nothing more. Mobileye/Waymo are just far better because they have the extra hardware needed for lv4/lv5
Yep. I've had a comma ai on my Toyota rav 4 since 2021. Works really well on highway. Just doesn't start off fast enough and have enough steering wheel torque on my rav 4 for it to be good in city. It can do city pretty well in some cars though.
@@qui11 I don't have many miles on the highway with Tesla. I have 35,000 with Comma so take this with a grain of salt. Open pilot is insane on the highway. I trust it more than any other system. It's never done anything crazy, period. It doesn't do auto lane change, so it's not as advanced, but it's really good at lane keep. The e2e model in the city wasn't that great, but that's because my Rav4 just doesn't allow enough steering wheel torque. It can do stop signs and stop lights, though. FSD will just take more time for me to "trust" it fully because my first drive on highway it tried to lane change into a semi truck. I know that was the old v11 system though. But again open pilot doesn't have auto lane change. Obviously, in the city FSD wins hands down.
11:55 😂 silly humans that's the wrong lever. If their turn signal was on, FSD probably would have gone when they didn't advance. Maybe their signal is out 🤷♂️
The main focus for markets now is Nvidia, which has powered a large chunk of the S&P 500’s recent earnings. Nvidia’s stock, up more than 90% this year, rose 2.5% in New York on Monday, sending the Nasdaq 100 index to another record high. I'm still looking for companies to make additions to my $350K portfolio, to boost performance. Here for ideas....
@@MetzTech clips sent in by the fleet are reviewed to find out if they show anything interesting, from a driving standpoint. When the system asks me why I disengaged I usually say something like "Destination, thanks for the ride" or "Destination, you can't back into my garage". Of course, disengagements for cause are reported as much as possible.
3:40: Super wrong there! Having people honk at and almost rear-end you at stop signs does NOT put anyone at ease. 7:00: Technically that is correct but it's also an example of why the "levels" are pretty much useless. Like, one of the other companies got regulatory approval to be officially autonomous, but only on the highway, behind lead car, in good weather, on certain stretches, at particular speeds, etc. So, basically, Tesla's software could easily be Level 4, if you put enough constraints on the operational area, but it would be a pointless gimmick. If it can be L4 in something the size of a city, in essentially all conditions, THAT is where it's meaningful, but that's not part of the level definitions, hence them being crap. 29:30: It's obviously still beta. I don't know why people act like the name change means anything. Until it legally does not need a driver, it's beta (really alpha, since not feature-complete). Colin's a smart guy!
![CAMPปลิ้น | EP.83[2/2] ประสบการณ์ขนหัวลุกจากตำนานวงการผี!](http://i.ytimg.com/vi/T_kFecQw7Ds/mqdefault.jpg)
100% he is not a hater, he knows more about FSD that a LOT of people.
He said himself, he is a “wait-and-seer” so you are not wrong
He’s a smart dude. I think 12.5 is converting a LOT of people
is he a programmer ?, he probably understands the technology but not convinced
Cool seeing that knowledge being experienced
Yes he is
Your friend is great, so clever and eloquent in describing the experience. Your conversation really added a lot to the drive.
Glad you enjoyed!
“I don’t like the ambient lighting”
Homie it’s day time
😂
@@MrGiggity890 That had me ROTFLMAO 😂
I love the lighting.
31:13
This is an extremely important issue in the field of AI in general that you're describing, and I'm really glad you brought that up. It's called the black box problem. Whether the AI does the right thing or not, it's impossible to trace how it reached the conclusion it did. There's been a huge push to make AI explainable and more transparent (as in outputting not only acceleration and steering, but the reasoning that led it to make that action) since right now, we have literally no way of knowing what the algorithm is really doing. Transitioning to this completely different kind of AI architecture would require a full overhaul, and that's especially difficult since AI is used all over the place already.
Great point
What do you want it to print? A bunch of weights and biases? But I guess you can make some kind of memory storage for path planning confidence level and look at those.
Neural network is currently best method to crack FSD now. Procedural programming won't work.
@@qui11 Exactly. Conventional programming won't work as well as a neural net for FSD. I'm not saying we shouldn't use AI, I'm saying we need to figure out how to make AI safer so that we are in control of it, not the other way around. For example, back when cars were a new invention, we didn't decide to stop using cars when we found out they were dangerous because we knew the potential they had to revolutionize transportation. Instead, we invented safety measures like seatbelts, crumple zones, and airbags to make them more safe and driver-focused. This is what needs to happen with AI.
@@qui11 You want it to explain itself in text. So "Waiting for person to cross" or "Breaking for red light" or "Lane change because x". And it should even be able to explain why it chose a certain path.
The occupancy network was cool.... yet they scrapped that. Now I'm not sure if FSD can handle completely novel problems. Like tools sticking out of a truck. Or a car painted on a bus.
Remember when FSD is on, the right thumbwheel adjusts the max speed. if you're going to fast , or slow you can either spin it down to slow 1mph at a click, or a quick flip changes it 5mph, or close to.
Doesn’t work on Automatic speed offset
The camera suctioned in the middle of the view of the road is distracting. It's hard to see what the driver would see, in order to perceive how it's doing at driving.
Clipped to the top of the screen could be much better
Not my normal position. Tough with two people
@@MetzTechjust had to move it a bit to the right.. and maybe just maybe a bit higher.
Great video and some deep questions.
I have been in cars with other people driving, some you are relaxed with and some make you nervous!.
Also driving a slow works van ,getting in a fast car and seeing how close you get to a stop line before the driver brakes! It’s a surprise but their car stops much quicker than my van.
I think fsd needs to get to a point where it drives most safely but also makes the most efficient progress.
Then everyone will just have to get used to how it drives.
For example it can change lanes in places where, a human could not be sure it was safe ,because it can see in 8 directions at once , no blind spots.
It can also go from accelerating to braking in one millionth of a second ,lightning fast reactions.
This could lead to driving closer to the car in front , because it knows it can out react it!
And lastly if you were looking at your phone you would not even notice the driving unless something quite serious happened.
Thanks for an interesting video.
This is one of the best (if not the best) FSD driving videos I have seen. I watch all the other standard ones and cannot make it to the end without throwing my shoes at the TV. The first time I hear the words "... I would have bla bla bla" I throw a shoe and stop the video. These videos should be about evaluating Self Driving not about your driving habits. Your friend was super cool.
Thanks so much!
One thing to note about why FSD doesn't currently do auto-park on it's own while in FSD, Elon has said recently that currently the auto-park code is completely separate from the FSD inference model. I wouldn't doubt it if auto-park, in it's current form, isn't performed by an inference model at all. So combining that behavior with FSD might be a lot of effort. It obviously needs to be done, but it's not trivial.
Methinks they will use neural net for that too, they’re just focused on driving rn.
@@markfrantz4612 well they might use NN’s while in the Autopark mode but those NNs are likely completely separate and trained separately. And maybe Autopark isn’t as “end-to-end” as FSD.
Interesting! Hope it isn’t too complicated to combine
Another great video. Love to hear the thoughts of a Tesla "wait-and-seer." 😅 Colin made a good point in the video about how predictable the car's actions become once you've ridden with FSD for a while, and that is a good thing from a passenger perspective. Massive driving style changes from iteration to iteration could be jarring.
Along those lines, at one point you mention it doesnt swerve to avoid a pothole that you would have avoided. This is a bit of a conundrum. On one hand, obviously to save our tires and suspension we would love for FSD to do some fancy maneuvering and miss them. But as a passenger, that evasive maneuver could be startling if you dont also see the same pothole and anticipate the cars quick steering input. Tough decisions to be made by the Tesla team on what level of unpredictable behavior we can stomach.
Very fair point. I think in that specific instance it could’ve just slightly weaved to avoid. But yea overall it’s complicated
Four FSD Teslas arrive at a four-way stop at practically the same time ...
Do they?
a) sit there and play Toybox games with each other until other cars arrive
b) just sit there
c) perform for each other with their matrix headlights
d) proceed normally
Hopefully C
C - because unlike the people driving cars, the FSD does the full stop so the Tesla can "see" it and follow the "program". When FSD is hesitant is when two or more cars are at the stop at the same time and one of them is not doing a full stop.
It all depends on what settings they have.
10:42 - Snake's tongue animation should be the official name of that animation. Kudos friend.
I like your drives with FSD. It gives me a better understanding of improvement from the previous version. I try to keep track of different scenarios, most interesting things are not been tested by early testers yet.
My opinion of FSD 12.5 shows readiness, on a scale from 1 to 10: 1 - Not working, 3 - Working, not useful, 5 - Good, 9 - Like human, 10 - Better than human.
7 - Highway: Performs well, but not perfect(FSD11).
5 - Exit from Highway: Adequate, but requires more refinement, miss exist or change lane too late(FSD11).
7 - Change Lane: Generally effective, but occasional issues.
8 - Intersection: Handles most scenarios smoothly.
8 - Roundabouts: Navigate effectively, though not flawlessly.
4 - Parking Lots: Functional but can struggle with complexity.
3 - Parking on the Street: Needs better spot finding to avoid blocking traffic.
8 - Reverse Parking: Decent, but not consistently reliable.
2 - Gate: Needs significant improvement for reliability.
7 - Signal to Other Cars: Generally good, but improvements are needed.
7 - Left Turn: Handles most left turns well.
4 - Unprotected Turns: Often cautious or hesitant.
5 - Speed Bumps: Recognizes and slows down appropriately.
3 - Constructions on Road: Needs better detection and navigation.
4 - Road Blocks: Struggles to handle unexpected obstacles.
8 - Rain: Performs adequately, but not in heavy rain.
3 - Snow: Limited capability in snowy conditions.
5 - Fog: Manages light fog, and struggles with heavy fog.
4 - Other Cars Breaking Rules: Handles some scenarios, but not all.
8 - Pedestrians: Good at detecting and reacting.
5 - Other Road Vehicles: Adequate, but can be improved.
7 - Traffic Lights: Generally accurate, occasional misreads.
7 - Driving Smoothness: Mostly smooth, with minor jerkiness at times.
2 - Potholes or debris on the road: I don't see any lower speed or avoid
6 - Speed Limits: Adheres well, but sometimes slow to adjust.
3 - Emergency Vehicles: Not recognized can be improved.
5 - Supervision: Nice that no hands on the wheel are required.
1 - Front Parking: Major improvements are needed.
1 - Reverse in Street: Significant challenges remain.
1 - Actual Smart Summon(ASS):
1 - Self-park(Banish):
1 - Ferry Crossings:
4 - Train Crossings:
1 - Charging Stations:
My assessment is open for discussion, but this is my opinion based on the videos that have been published so far. If there haven't been any new updates, I base this on earlier versions. I assume my opinion may change in the future, but right now, I don't see any area where FSD is as good as or better than a human driver, even though it can sometimes surpass humans in certain scenarios.
Your assessment is rational unlike that. I would point to the fact that FSD has a lower accident rate than humans.
That would seem to indicate it's better than humans in some scenarios.
@@macmcleod1188 Most humans rarely have accidents, with only a small percentage experiencing frequent incidents. However, I often see drivers needing to take control in almost all situations, even with version 12.5.1 of FSD technology. While FSD is generally better than most human drivers, it still makes mistakes even in the best scenarios. For FSD to be considered better than a human driver, it would need to go at least a year without any mistakes or accidents that a human driver couldn't avoid. I hope we reach that level of reliability soon.
@@annawilson4571 humans have an accident every 8 years on average.
Humans also have an accident roughly every 150,000 miles of driving.
Self-driving Tesla have a record of only having an accident every 450,000 mi.
I do agree with your point that as long as a human needs to take control, FSD should top out an eight.
@25:21 U R spot on. Munich Germany (in the 80s) No Left turns, all right turns because they found it safer and actually made less traffic jams. Was hard to get used to, but I didnt even notice the unfamiliarity feeling after a few trips into the city.
To me there’s a big difference between 12.5 and 12.5.1.
There is. Video coming soon
@@MetzTech Still does some bad stuff though. Too slow in general.
It’s been very good so far for me.
Blown away by that short cut through the gas station. Crazy.
I’m so impressed with how well it drives at this point. Speed is the biggest issue this version, but that seems an easy change.
Using the camera for rain sensing also means that it needs light... When driving in badly/non-lit areas at night, the number of wipes drops dramatically. Also, the software needs a "water repellant applied" setting.
Good video. Always good to get inputs from a non-Tesla guy. Keep it up!
Thank you 🙏🏼
It won't break the speed limit if you limit the speed. Like you said it is all on you self supervising.
My biggest issues. Starts off way too damn fast, even on chill mode. I'm trying to keep my tires more than 5,000 miles. The NHTSA stop. Dude, that thing is going to get me rear-ended. It makes people behind me think I'm brake checking them. Literally no human driver on earth stops that far back, so no one expects it. Other than that it's nearly perfect, but those two things are just holding me back.
NHTSA really overstepped their bounds with that decision. Setting a minimum of 1-3mph would have been much better, if they really feel the need
The change in behavior where FSD stops at the Stop Sign and then immediately starts to creep has changed the dynamics considerably for people behind me. I often would use the accelerator pedal when I had someone right behind me so they would be annoyed and/or honk. Now I never do with v12.5.1 and not a single driver behind me has honked or appeared agitated. A nice change. I definitely agree the NHTSA insistence on stopping is stupid but Tesla has done a reasonable job with v12.5.1 to address that.
“No human driver on Earth”? Just because the USA has a silly way to define when to yield or not, doesn’t mean the rest of the world makes the same mistake.
@@z4zuse People generally expect you to stop just shy of where the road meets the end of the other road. Not 20 feet back. Therefore it confuses people behind you
I agree that starts and stops need to be dialed down. The main road in my town has multiple stop lights not very far apart. When a light turns green, if there are no cars directly in front of me, FSD takes off trying to reach the 40mph speed limit very quickly, only having to slow down or stop abruptly within another few seconds. Some people may find this acceptable, but it is a bit too aggressive for me. At least give me an option in Chill mode to smooth this out a bit.
Regarding stops, I have no issue with the NHTSA stop requirement the way it is now implemented in 12.5 and 12.5.1 (12.3.6 was not good). But yes, the completed stop is too far back from the crossroad in most cases. I think that it should be closer to the actual crossroad itself.
You sold your mate down the river just to get clicks. That's not cool. He is not a hater. He said he is a "wait and see er".
Just giving him a hard time, relax 🤣
I like the ambient lighting not being ambient lol
Total fine for me too
I actually had my Model Y stop for a yellow light on 12.4 (I think) then went on reverse to get off the cross walk. Maybe it goes on reverse if its a safety issue.
You are correct about the preference for right turns. When I visit a friend, I back into his driveway. When I leave, FSD wants to turn right and loop around, while I always go left and save about 3 blocks. Both routes come together at the same, unavoidable left turn. Same thing leaving my home, I'm one lot from an intersection and it always wants to turn right from my driveway and take me almost 3 blocks out of my way.
Happens to me almost every time I use FSD. It’s super interesting
Is FSD being trained by UPS drivers? 🤔 (Look up "UPS right turn policy" if you haven't before. It's quite interesting)
I think the no rain sensor approach is also because they are thinking so far into the future where it doesn’t matter if the human in the car can see. Only the car needs to see. However, I think they learn a lot using only the camera to determine how much water is on the windshield, which the human may not be able to see anymore. It's all about picking the right trainingdata.
27:48 ⛈
⚡️⚡️⚡️
This guy was great, but definitely not a hater, he even said it himself, that he is a "wait and see'r" :)
I was just messing with him
Waymo is level 3, as there is a remote operator ready to take the wheel on demand.
Autonomous levels are a very outdated notion anyway. It doesn't take into account features like geofencing, choosing a destination, the energy consumption of the system, road conditions, parking or the price/need of extra equipment from the factory car and such.
Waymo has a much more advanced vehicle than Tesla, and Waymo is in business today, giving rides to 1000's of passengers every month. Meanwhile, Tesla robotaxi is *years* *behind* *schedule* .
@@DerekDavis213 Excuse me to say it, but this is nonsense.
Tesla is in business today, selling and renting FSD all over the US and Canada, making millions of rides, and not just in a little square of sunny towns here and there.
The "much more advanced vehicles", IIRC are Jaguar I-Paces, which design dates from 2018 and didn't change since then, which where already rather bad back then, plus hundreds of thousand dollars of very ugly, bulky and power hungry equipment (which is probably why the chose an electric vehicle with a big battery).
The only advantages of Waymo today are an hailing app and the ability to have one person supervise several cars at once.
@DerekDavis213 Waymo is geo fenced.
Excellent video!! Please do it again soon!
I am questioning whether or not the occasional re-routing that FSD does when it disregards map data, and proceeds in doing its own thing, is simply the inference computer making the “wrong” choice. When I first got v12.5, I put in a destination that I’ve gone to many times. But this time my car went into the left turn lane at a stop light and turned on the left turn signal. That was at odds with map routing; the shortest and fastest route was straight ahead; there wasn’t much traffic at all. As I waited at the intersection, I was ready to take over because I didn’t know if it might just try following the mapped route from the wrong lane (as sometimes would happen with 12.3.6). However, my car did indeed turn left. Afterward the map updated with a new route, which was a longer and slower. So idk. V12.3.6 would often lane dance as it approached this (and other) intersections. I have not experienced any lane uncertainty at all with 12.5 or 12.5.1. So when my car moved into the left turn lane quickly and confidently I wondered what was going on. One possibility that comes to mind is that FSD is sometimes still making a wrong choice (different than map route), but now it commits to it (no dancing, proper signal). It has since always chosen the straight ahead route, even in heavier traffic. I may be way off base, and I would like to believe that AI is truly at work here, but if FSD is purposely choosing a different route, some direct evidence would be nice.
The red light situation:
It is not learning between goes. Small things that might seem insignificant to a human can influence the confidence of an AI system.
During the first go FSD simply wasn't sure enough that the object was a red light so it didnt stop.
They can simulate the scenarios exactly as they take the recording data and can feed it through the exact same model and it will ignore the red light again.
When you drive there are small differences that change the confidence intervals so it can be hard to replicate the scenario where it doesnt stop.
Tesla also might not really care what caused fsd to drive off the cliff as much as tesla cares to never let it happen again.
They store a lot of critical situations that caused problems in the past and validate the new models on these situations before releasing them.
You need to be wise about what you validate and train on because if you have overabundance of boring scenarios FSD will not treat critical situations seriously enough. If you just feed it critical situations it will be too timid or jerky in regular driving scenarios. (this actually happened with one of their versions)
You do not have infinite storage space nor time to validate so there is a balance there as well.
You think the recent Crowdstrike bug was catastrophic? Wait until a bug sent out in an FSD update causes a 100-car pileup on the freeway.
I don't think I've seen a fsd vid without the gopro overheating 😂
I haven’t made one yet…. It’s so frustrating 😂
22:35 Look into the progression of 'Alpha Go' the AI that plays the game GO.
The short of it; early training was done using human game data, later versions created their own game data and trained off of playing itself.
Tesla is also able to create its own input data now (all sensors) and video using AI. Once the new training system is online, things are going to get wild.
Slowing down BEFORE the curve is the way to make it extra smooth imo
It's a bit unfair to be hypercritical of FSD when no other cars comes close. For instance, Mercedes Level 3 speed is capped and is only legal in CA and AZ. And FSD is video only and overall is a very hard problem. I think Tesla engineers are heros.
Definitely agree
Your buddy reminds me of James Stephenson so much!! Great video, great guest! Kudos!
Thank you 🙏🏼
6:47 That’s the reason I bought FSD in 2021 (not knowing when- or if it would be implemented in my country ‘Belgium EU’), in a medical emergency it could be faster to navigate your car to a hospital ER than calling- and waiting for a ambulance to get you there the fastest way…
I dunno I kind of like the ambient lighting. its understated
The ambient lighting in the '24 Model 3 is OK, but if it is on while driving, I can see it on the windshield like a heads-up display, right where I don't want anything. Not Tesla's best design .
It should allow us to adjust brightness.
I don’t dislike it, he’s coming from an Ioniq 6 which has really nice designs on the door panels
Adjusting the brightness is coming in a software update
@@rafaelricarte6227Coming on new Update 2024 26.1
@@MetzTech It would be nice if you could adjust different segments by different amounts too.
I couldn’t do it, Louis. I would be too nervous. Love the video, though.
Truly not a hater. Even said so at the beginning. You don’t need to make that up do you?
Just giving him a hard time 😉
@@MetzTech Fair enough. I enjoy your videos. Just tiring of reading the word “hate” on social media these days. To be fair, rarely if ever, do I see it on your feeds. 👍
I hear you! Thanks for watching
I hate sitting at stop signs , I’m from the uk and we have roundabouts most of the time so lots less stoping.
I would also take the path of least resistance!
You don't have you keep your hands on wheel any longer? Is there a camera directed @ your eyes to keep them on the road?
Correct
Tesla cameras can not only see visible light range but also extend into the near-infrared (NIR) spectrum. Which is why FSD performs better than humans at night.
I have this theory that Tesla will launch a robotaxi pilot program with FSD (unsupervised) in parallel with FSD (supervised). When the performance of the unsupervised version surpasses that of the driver supervised system, they will go to the regulator and ask for regulatory approval for the whole fleet. The measure will be miles driven between accidents. This makes sense from both an ethical standpoint as a technical standpoint.
But you and i are not perfect neither. We are mistakenly we are perfect and killed 43,000 of us on roard last year. It is funny how ridiculous we are when we make critical judgement 😂
Very true. I always try to remind people of that.
They do take gyroscope and accelerometer data. There is a video on how the planner calculates multiple paths and picks the smoothest one from the proposed ones. In order to train the system to pick smooth paths you need all that data.
In a way this kind of approach of generating multiple possibilities and picking the best one seems to be used a lot in AI. Next version of ChatGPT will most likely use a search algorithm that generates multiple answers during training.
Very cool! Thanks
Good video. Wipers are MUCH better than they used to be.
True
Maybe it blew the stop light because it felt it was the safest thing to do?! Hopefully safety is the most important thing.
12:45 hardware V4 has IR emitters mostly for short-range illumination. The AI takes in the raw camera feed, so I imagine it has a wider sensitivity than a human eye. Wonder if future camera versions will have an iris to adjust light input like humans, that would be cool.
This guy is smart, and for someone with very little FSD experiences he knows his stuff. You should have him ride shotgun more often.
Definitely!
I don’t understand haters who haven’t experienced the thing they hate on. It’s like they can’t admit they don’t know something.
He’s not actually a hater I was just giving him a hard time
@@MetzTech
Yes I can see that. 😁
Can you test if this version know how to read the meaning of solid yellow line and broken yellow lines (passing and no passing zone)? A wall street analyst testing pre - FSD 12.5 claims he has to disengage because it tried to pass on a no passing zone. This test case requires two cars, the lead car will intentionaly drive slow so that the tesla following it will be forced to overtake it. Thanks.
A standard broken yellow line tells you that passing is allowed (a passing zone).
When there is one broken line and one solid line, traffic traveling adjacent to the broken line can pass (passing zone), and traffic traveling adjacent to the solid line cannot pass (no-passing zone).
Two regular solid yellow lines mark where crossing the center for passing is prohibited (a no-passing zone). It applies to traffic traveling in either direction.
I am totally new to FSD so a couple of off the wall newbie questions.
Can you make your own route for it to follow?
What about off road, can you make off the road route?
Can a legless invalid operate it now?
Sure you could do a custom route.
You could probably make it go off-road but it’s not going to know what to do.
I’m sure there are ways to customize the vehicle so that someone with those types of impairments could operate using their hands
Very good video and good discussion. I don't think your friend is a hater, but I understand the need to bait some views as a small channel.
Just giving him a hard time, he’s definitely not a hater!
statement of the year " i don't feel like a humans driving " well no shit buddy
Cool review.
Disagree on rolled stops, regulators need to get real, 2 or 3 mph vs 0 *where safe?* Zero may be required at any time based on conditions, but it's not magically safer than a continuous creep or walking sp roll.
Full stops cause other drivers anxiety, anger when they aren't needed, wasteful of energy, less comfortable. (Immense driver judgement required lk driving in general)
That gopro is mounted to cover the most important part of the video
How is Waymo going to take me from downtown to my home out in the country? Only fsd has that capability.
True that
It still won't take the proper off ramps for me on 12.5.1... And it won't cross bycicle lines... So it's good but not perfect... YET
ballsy you let the car go way over with a police 😂😂
May have been sweating a bit lol 😂 would’ve made a great thumbnail tho
My FSD is going way too fast in the city. It tries to go from 0 to speed limit in the city street in a hurry and breaks hard. Why? I’m on chill mode. Way too aggressive. Seems to be worse after upgrading to 12.5.1
Zillions of humans “drive off a cliff” every year , I’m sure it will be very rare for fsd to do that! So overall fsd will save countless lives
"My Dinner with Andre" in a Tesla!
You are quite smart for your looks. But your friend is even smarter.
Wait what do you mean for my looks?!
it drives as well as any newborn would, since it looks like you're in a 2024 M3
😂
Great videos guys!
Thank you! 🙏🏼
3:55 Mainly for regions where FSD is under review for admission by the government (Elon tweeted that Europe is planned to rollout in Q1 2025 if EU rules allow it) 🤔
I find it really odd that tesla would allow the car to do that much over the speed limit. I could understand like +4 mph over, but realistically, it shouldn't speed, should it? Do they allow that because so many people speed normally?
I think it’s based on the surrounding vehicles
Don't like how fast it travels over speed limit and accelerates, then select absolute speed and maybe select chill.
10:10 they need to tie in the auto park and let it park. Ironically, it’s the only way it will park.
Well done!
Thank you 🙏🏼
Hi there...how about for your next FSD 12.5.1 do a nightime with rain..that should stress test it even more
I already filmed my FSD 12.5.1 video but I can definitely do a nighttime drive video
AI is about data and train models and designing and evolving new models so that the loop closes or coverges on a behavious or ability to drive in the same weather and road conditions as a human. To be able to navigate and plan and to be as safe as doing all of this depend trafic conditions and the type of other road users.
All these variables we as haman have also been trained. We only become compitent at driving after 5-8yrs, after our passing the test.
Actually interesting conversation! Your friend has good insights. Also I can't help but compare this to WholeMars's friend Hannah. Night and day
Thank you 🙏🏼
Yeah….i saw that one… lol
He wanted to say he loves it
I think he does now…he’s been converted
epic video!
Thank you 🙏🏼
Seems funny that they haven’t yet combined the functions of FSD and AutoPark into one smooth fluid motion. So close.
So freaking close. You can tell the model wants to put you in a spot because that’s what people
Do at the end of a drive. It just can’t do it right yet
"Are you an FSD hater?"
"No."
Title: "FSD HATER FIRST DRIVE!"
😂
I like to mess with him 😂
Your friend seems like a very intelligent guy. Is he a software engineer, by chance?
He is!
@@MetzTech Awesome. I was nodding along with basically everything he said. Would love to hear from him again in future videos!
Definitely! We do the Metz Talk podcast together if you want to check those out. It’s been a minute since we recorded an episode but we have a nice backlog of episodes and will have more coming in the near future!
@metztalk
@@MetzTech Yeah I'll check it out, thanks!
In defense of FSD.. you never brake into a turn.. you accelerate out of a turn so you don't lock up your brakes in case of an emergency.
I know where ur at louisville Ky everyone who lives there knows that tower at 2:07 thats ruths chris steak house. nice video
Bingo!
WOW WOW WOW
I've driven FSD for 2 years now and it sucks.
It's 99% accurate
and that 1% will kill you
Is that good enough for elon?
How about 99.9% accurate - that good enough elon?
The truth is elon has no intention of telling us "how good" FSD will have to be to become "ready"
and this is on purpose
FSD was NEVER designed to operate in a Tesla with a steering wheel and 2 pedals
it was designed to run a RoboTaxi
Want proof?
Engage the new FSD and fiddle with your iPhone above the steering wheel
FSD will be happy
and the cop who just pulled you over will cite you for reckless driving
and the judge will suspend your license for a year
and you told both of them "elon said FSD was ready"
and both laughed
When FSD is "ready" that $10,000 I paid for it means nothing since I still have to drive the car...
Excellent, thanks
Thanks for watching!
I don't get it. You're complaining it's too assertive, but don't take it off the assertive mode? Your brain clearly works completely differently to mine!
Definitely not complaining. In fact it’s not assertive enough
Neural Network outcomes are inherently based on probability. When the neural network probability of success is far above human success rates, Tesla would probably take full responsibility.
Great video. Thank you!
Thank you!
Biggest issue was the auto wipers
They honestly aren’t that bad. At least not now.
Tesla FSD will probaly max out @lv3 it is a good system if u are behind it but nothing more. Mobileye/Waymo are just far better because they have the extra hardware needed for lv4/lv5
7:17 comma AI is also e2e
Yep. I've had a comma ai on my Toyota rav 4 since 2021. Works really well on highway. Just doesn't start off fast enough and have enough steering wheel torque on my rav 4 for it to be good in city. It can do city pretty well in some cars though.
@@drewhjava how is current comma 3x compared to FSD 12.5? Better, smoother?
@@qui11 I don't have many miles on the highway with Tesla. I have 35,000 with Comma so take this with a grain of salt.
Open pilot is insane on the highway. I trust it more than any other system. It's never done anything crazy, period. It doesn't do auto lane change, so it's not as advanced, but it's really good at lane keep. The e2e model in the city wasn't that great, but that's because my Rav4 just doesn't allow enough steering wheel torque. It can do stop signs and stop lights, though.
FSD will just take more time for me to "trust" it fully because my first drive on highway it tried to lane change into a semi truck. I know that was the old v11 system though. But again open pilot doesn't have auto lane change. Obviously, in the city FSD wins hands down.
TH-camrs really need to learn to adopt the AutoTopNL way of recording
You like POV? I have a few of those!
11:55 😂 silly humans that's the wrong lever. If their turn signal was on, FSD probably would have gone when they didn't advance. Maybe their signal is out 🤷♂️
What’s happens when there’s too much glare?
Nothing really
The main focus for markets now is Nvidia, which has powered a large chunk of the S&P 500’s recent earnings. Nvidia’s stock, up more than 90% this year, rose 2.5% in New York on Monday, sending the Nasdaq 100 index to another record high. I'm still looking for companies to make additions to my $350K portfolio, to boost performance. Here for ideas....
a Tesla will not autopark forward in a stall. the fact it went to the stalls is more luck than thought.
Drive it through Manhattan
All video clips used by Tesla for training are carefully curated. Actual humans look at them and approve them for use.
Good to know!
@@MetzTech clips sent in by the fleet are reviewed to find out if they show anything interesting, from a driving standpoint. When the system asks me why I disengaged I usually say something like "Destination, thanks for the ride" or "Destination, you can't back into my garage". Of course, disengagements for cause are reported as much as possible.
3:40: Super wrong there! Having people honk at and almost rear-end you at stop signs does NOT put anyone at ease.
7:00: Technically that is correct but it's also an example of why the "levels" are pretty much useless. Like, one of the other companies got regulatory approval to be officially autonomous, but only on the highway, behind lead car, in good weather, on certain stretches, at particular speeds, etc. So, basically, Tesla's software could easily be Level 4, if you put enough constraints on the operational area, but it would be a pointless gimmick. If it can be L4 in something the size of a city, in essentially all conditions, THAT is where it's meaningful, but that's not part of the level definitions, hence them being crap.
29:30: It's obviously still beta. I don't know why people act like the name change means anything. Until it legally does not need a driver, it's beta (really alpha, since not feature-complete).
Colin's a smart guy!
So you are pro rolling stop signs?
Dot releases!
34:00 Definitely NOT😅
Fsd hater would be Fred from electrek
That whole brand is a hater