I just experienced Tesla’s newest version of FSD (version 12.3) with my friend Will from ALL Electric! Tesla nerds already know that v12.3 is a big deal. Everyone else, I assure you that this is a big breakthrough. Watch the video and then read the rest of the post to hear why I think 12.3 may or may not be The One that finally starts to solve FSD.
FSD 12.3
Somewhere around 2015, I read an article about how Google was testing autonomous cars on the San Francisco roads. I’m a nerd and embraced robot cars immediately. My thought was this:
I love this. My kids may never have to own a car and maybe won’t even have to learn how to drive!
That didn’t play out well. My 8 year old is now 17 and has a car, a driver’s license, and a minor fender bender (not her fault) under her belt. Oh well.
I’ve grown increasingly skeptical of autonomy, especially Tesla’s approach. Elon Musk has promised autonomy many, many times. And then the results failed to impress. But then, I read about Tesla’s new end-to-end network approach with version 12. I won’t go deep into the technical details, but the technology is amazing. Here is my two sentence explanation:
Tesla teaches the car how to drive by showing the training computer lots and lots and lots of driving videos. The computer learns how to drive from watching these videos.
Again I was skeptical, but continued to pay attention to Tesla’s efforts. The first version of 12 was released to the public in February and testers started talking about it differently:
It drives more like a human.
It has fixed problems that I’ve had with FSD for years.
And then v12.3 came out and the accolades continued. Perhaps the most important praise came from Rob Maurer of the now defunct Tesla Daily podcast. I’ve listened to Rob for years and know him to be fair in his evaluation of all things Tesla. So when he posted the following on X, I got excited:
The next thing to do was ping everyone I know with FSD software to see if they had v12.3 yet. Tesla deploys software in a phased releases and there is no way to predict who gets it first.
On Tuesday, my friend Will let me know that his car had received the update. I dropped whatever nerd stuff I was doing and got my ass to Will’s house. I parked my Y in front of Will’s and we hit the road. I have much more to say, but first watch the video:
Thoughts on V12.3
My car came with a 3 month trial of FSD 11. Much of the time, the software wasn’t great. It would:
- Drift out of its lane into exit lanes when it was supposed to go straight.
- Activate a turn signal when there was just a gentle curve in the road.
- Stop in merge lanes.
And every once in a while, it would do something terrifying:
- Steer towards a pedestrian in a crosswalk. (BAD CAR! YOU CAN’T KILL HUMANS! YOU HAVEN”T EVEN ATTAINTED CONSCIOUSNESS YET! WTF!!!)
- Steer towards a road that was closed and blocked with a sign.
- Phantom brake when it mistook the shadow of an overpass for an obstruction in the road.
And when it did OK, the driving was jerky and not human-like.
The drives were anxiety inducing and I eventually gave up. My 3 month FSD trial expired and I saw no need to pay for it.
And Then Came 12.3
I glanced down at my Fitbit as Will and I began our FSD ride to Costco. My heart was racing with anxiety. My resting heart rate is normally somewhere in the 60s, but was now hovering around 100. Oof.
But then the car started to go and it drove like a human:
- It stayed in the part of the lane I would stay in.
- It stopped like I would expect a good human driver to stop.
- It navigated perfectly through a construction zone where the road was down to one lane. (I have more to say about this below)
- It navigated a simple traffic circle perfectly.
- In an act of extreme bravery, Will let the Y navigate a Costco parking lot in mid-afternoon. The lot was a sea of other cars and pedestrians. The car did amazing. It gently waited for its turn to move, weaving through the chaos perfectly.
My Two Big Takeaways
The car did it? What?!??: There were multiple points in the drive where the car performed so well, I just assumed Will had taken control. Sometimes it’s difficult to tell if the human is driving because the driver still has to touch the wheel and pay attention. Will acknowledged that he only corrected the car one time; pressing the accelerator to encourage the car to move.
I calmed down: As Will’s Y started the drive home from Costco, I realized that I was no longer freaking out. My heart rate was down. It just felt like a human was driving. I wasn’t paying super close attention like I had with every other FSD ride I have ever taken. The car drove that well.
What I Still Need To See To Be Convinced
It’s all about the edge cases. Allow me to explain.
Most of driving isn’t that difficult. It’s easy to teach a computer (or a 15 year old) what to do at a stop sign, traffic signal, or even a roundabout (note that a lot humans still really suck at this!).
However, there are loads of situations that can only be resolved with higher-level thinking and reasoning. The car must figure out how to interact with construction zones, humans directing traffic, accidents, crappy lane markings, and the infinite number of other difficult situations that humans can easily reason their way out of.
The only way Tesla can solve these difficult situations is if FSD can perform some level of artificial general intelligence. Elon Musk has hinted that Tesla is working on it.
So what would I need to see to truly believe Tesla will solve autonomy? Here you go:
I mentioned that the car successfully navigated through a difficult construction zone. However, there were other cars in front of it. It’s known that at least in previous FSD releases, FSD was trained to follow other cars.
I’d like to see the car navigate through complex situations without a lead car. When I see the Tesla navigating through construction zones, road closures, dealing with emergency vehicles, and driving in severe weather, I’ll really believe it.
For all I know, v12.3 may already be capable of doing some of these things. I just haven’t had enough time to spend with it.
For now, I’m cautiously optimistic. But also, I’m more optimistic than I’ve ever been about Tesla’s autonomous future. This year could be a very, very, interesting year for Tesla.
I’m going to sign up for FSD so I can test more myself. Stay tuned.
In the meantime, thanks Will for the drive! Be sure to check out his excellent ALL Electric channel for more on Tesla, ebikes, scooters, and everything else electric.
More 1500 Days!!!
You can also find me (and the dinosaurs) at:
Mile High FI podcast:
MindyOnMoney podcast!
Also here:
- Facebook: Facebook group and page
- YouTube: My channel is mostly devoted to home improvement, but I have some other material coming up soon too.
- Instagram: Pretty pictures of dinosaurs, sunsets, and nail guns!
- Twitter: Spontaneous, often insane, ramblings
- Coworking space: On the surface, MMM HQ is a coworking space. Look a little deeper and you’ll see that we’re really building community. The members of MMM HQ are some of the finest people I know
- Buying a Tesla? Use my referral code to get some perks!
My only question after reading is: what is FSD? Perhaps I have not followed Tesla technology enough to know what the acronym is.
Self-drive technology certainly looks exciting, but I know I will be a very very late adopter – I’ll let others be the guinea pigs for now.
I love that every article I read about self driving cars always has the phrase “but the driver still has to do essential driving action x”. Then it’s not autonomous driving. Lol.
Two thoughts:
1) This technology is still in it’s infancy and it’s a very difficult problem to solve. There are just so many edge cases that you need to solve at least a basic version of artificial general intelligence. I think we’ll see it this decade, but we’re still a couple of years off.
2) Even though there is no widely deployed autonomous car fleet, this software in its rudimentary form will do a lot of good. A car with a suite of sensors will prevent many accidents that an inattentive human would otherwise cause. So even if autonomy isn’t solved, I’ll bet we start to see traffic deaths decrease as cars with these sensors start to become the norm.
Hey Carl,
Brought home a new Model Y LR on Weds after being a stock holder for over a decade. Ownership has done nothing but renew my faith in the company and the stock. As for FSD, I know full autonomy is the goal, but I don’t view solving it as an all or nothing proposition. I would buy a monthly subscription for an FSD service that handles all normal driving conditions and situations better than a human. Sounds like Tesla is very close to this benchmark. I’m okay with taking control through the occasional construction zone or if its snowing sideways, Solving all edge cases is needed for robo taxi, but there’s still tremendous value in FSD that covers 90-95% of all driving situations and conditions, as long as the limitations are known and defined, allowing the driver to avoid them or take control for short periods. As long as the system is 100% reliable outside its known/defined gaps, it offers tremendous value. I would buy a month of FSD today if I was driving cross country. Having FSD would allow me to drive more hours per day and arrive less stressed and burnt out at my destination.
Agree Tesla is on the right path with neural net learning and vision. Impressed with Tesla’s ability to shift approach and ignore sunk costs, as they did with LIDAR and code writing approaches.
Justin!
I’m so glad you like the car!
Robotaxi: I wonder if Tesla will roll out similar to Waymo? By that, I mean first deploy to warm places where the weather is calm and streets are better defined (Phoenix is easier than Boston!).
The other thing you didn’t mention is the tremendous value of having a sophisticated set of sensors on the car. Even if one doesn’t have FSD, the cameras are going to prevent a lot of accidents.
I know I’m three weeks late on this, but here goes:
You can’t compare Tesla to Waymo. Waymo has LIDAR and cameras, so isn’t fooled by, say overpass shadows. Also, Waymo was rolled out carefully and thoughtfully with extensive support and monitoring staff and limited service areas and users. Tesla sold a Autopilot and FSD add ons to anyone who believed the hype and had some extra $ then rolled out their beta test on an unwilling and unwitting public.
You talk safety, but NHTSA has some serious concerns about Tesla driver aids (as that is all they are). We’ve seen lack of training and risk compensation (i.e. people drive faster and less attentively when wearing a seatbelt, have brake assist, etc and especially when they think the computer can drive the car) going back to the early Autopilot deaths. These features have been around a while and we haven’t seen a decline in deaths, In fact they are rising and insurance data showed Tesla had the highest crash rate of all manufacturers.
Driving is all edge cases. SAE Level 2 or 3 systems like this make computers do what they are terrible at (decipher complicated and new conditions) AND require humans do to what they are terrible at (pay close attention and be ready to respond with no warning over long periods). You weren’t driving, but you described your own relaxation on the return trip…how attentive are you going to be on your 100th trip to Costco on FSD when something unusual happens? The trouble is edge cases can’t be predicted and come at you at 50 mph.
Transportation safety is my profession and I get worked up by technology is our savior handwaving when it is contrary to reality. We may eventually get there, but we are decades away from robotaxis and the in between ground with partial AV integration in the fleet and partial AV functionality is the uncanny valley of safety. You mention that general AI may be needed to solve this and who knows when that will happen.
I think we’re in agreement that Tesla has a long way to go. What FSD does now is amazing, but there is no way it’s ready for a Robotaxi fleet. My car likes to stop in merge lanes and jump into right turn lanes when it’s supposed to go straight. I think Tesla can do it with only camera sensors and some level of AGI, but I have no idea when.
One thing that I’d ask you to elaborate on is this: “We may eventually get there, but we are decades away from robotaxis…” Although very limited, Waymo seems to be showing promise. Did you mean a Tesla based Robotaxi?
I was primarily referring robotaxis in the way the Elon has been promising them as an income stream to customers for a decade. (I wonder if it the owners weren’t such fanboys there would have been a class action suit by now). I agree that would seem to require significant advances at a much faster pace than we’ve seen.
Waymo is truly impressive with what they have done. I’ve interacted with them professionally and appreciate their commitment to safety. However, given then anticipated cost of their vehicles with extensive sensors, plus monitoring, maintenance, user control, etc. I expect they are a very long way from being financially viable compared to just putting a low-skilled person in the driver seat. However, their eventual goal may be autonomous transit as well which is another use case and financial dynamic.
To go further, robotaxis bring in some new problems. We’ve seen increased vehicle miles traveled from Uber/Lyft, largely due to so much unproductive driving to reach the pickup point. This issue would get much worse and risk girdlock with robotaxis plus rich 6 years have autonomous cars and empty cars are sent to collect the drycleaning and the groceries. And especially when people can reclaim their commute times for productive purposes while lounging in the back of a car, so live twice as far from work requiring twice the road capacity. Cities are already geometrically limited and more cars just won’t fit, whether EV or AV. The capacity improvements to AVs are less than you would think.
You make a lot of valid points. I can see people sending a robotaxi to a store to grab a gallon of millk.
I’ve also wondered about the class action thing with Tesla. Upon a lot of consideration, FSD is doing pretty well now. Most of my interventions are me hitting accelerator because it’s so hesitant. Tesla is still a while away from a driverless solution though.
Do you follow the EVOTL space?
The question is how far away is “FSD is doing pretty well” from “FSD is effectively perfect” and how much money and time does it take to get there? And will it be profitable in the end (see estimated cost of Waymo vehicles plus R&D).
EVOTLs! You really know how to push my buttons! 😉 I saw another booster presentation last week. I could see a use case in rural/inaccessible areas for UAVs. Zipline in Rwanda delivering blood and medicine to rural clinics is very very cool and has been successful for years and should be expanded to anywhere it makes sense. Electrified general aviation could be cool and reduce emissions for a small slice of the pie. Would require new, lighter battery technologies that are just around the corner, right?
However, the idea that cities will be better by a swarm small people or freight carrying EVOTLs buzzing between buildings to avoid traffic is near delusional. First, get past the new thing bias and replace them with helicopters in your head and see what problems arise. EVOTLs are very loud, very small, and have very limited payload and range. We would need to retrofit roofs and build new buildings to create not just landing zones, but also charging, loading, maintenance, passenger waiting, etc areas. Flight times are very short, so what happens when you are in a rush hour queue to land and have minutes of power left? You would need thousands in use to make a difference. Lastly and most importantly, while some aspects of flight are easier for autopilot than on roads, there are no fender benders in the sky and weather is a challenge. Any crash is killing all passengers plus people on the ground and/or in buildings.
The solution these people are looking for is not EVTOLs and is certainly not The Boring Company. It is a bit more urban density with a lot more mixed use (see 15-minute city). It is a mix of walking, biking, buses and trains, some of which might be automated and should electrified. Cars are the problem for urban congestion, crash risk, air quality, taking up a lot of space, and more. Flying cars (and EVs and AVs for that matter) doesn’t really solve any of them, just maybe perhaps replaces air pollution with noise pollution if you ignore the additional power generation needed. But none of that is exciting to tech bros and venture capital, so it isn’t funding or publicized. Whew!
Glad to hear about the improvements on the auto driving technology front.
Does this change your mind about downsizing your Tesla holdings percentage?
Great question: I have such a hard time selling anything. This definitely makes it harder. From a stockholder perspective, if Tesla solves FSD, the stock would easily double. From a humanity perspective, car accidents go way down, folks who can’t drive have freedom, and car ownership starts to go away. It would be sooooo good for society.
I’m still not convinced that Tesla has solved it yet though. There are an infinite number of crazy scenarios that humans can reason themselves out of. While I’ve seen bits of FSD driving that have amazed me, I want to know that Tesla has AGI figured out which is what it will need to solve the complex driving problems.
That’s really cool! As I’m years away from buying a new EV, I purposefully don’t follow emerging tech much. You’re literally my only source of this stuff, and I really enjoy these updates.
Questions:
1-when you were down in AZ, did you try the waymo taxis? Would love to hear how that compares.
2-We’re moving to a more bike friendly ‘hood, and I’ll finally be able to bike the kids (4&7) to school. I’m thinking of getting one of those Radwagons for kid hauling. Do you or any of your fellow Longmontians have experience with those or similar? Any recommendations? I recall that you retrofitted bike back in the day, but it seems like the new options have come a long way.
Thanks Carl!
That looks great, hope it becomes normal for most vehicles.
Once Tesla FSD or any autonomous vehicle is able to deal with driving like this – SA Taxi drivers:
https://www.youtube.com/watch?v=OxfhbpVFk6w&ab_channel=YouTubeViewers
Or is able to deal with roads that have no markings, no lighting and drivers that disregard most traffic requirements (lanes, incoming traffic, side-walks and ignoring traffic lights) like in South Africa, Pakistan or Vietnam, then I’ll be convinced FSD is possible haha