Rivian May Be Much Closer to FSD Than You Think

[hudesigns]

Look at tesla robotaxi service. Even to scale to 20 cities WITH safety monitors they have to build depots in every city they operate. Even the path to delivering 1000 cars on the robotaxi network is hard.

Elon mentioned in one earnings call that tesla will raise money to fund robotaxi if they need to. That is a company with (37? billion dollars) and they might need to fundraise money to scale self driving

The idea that these cars will overnight turn on and be self driving will never happen. And it's not a sensor issue.

Elon said he was expecting Tesla robotaxis to be available to "probably half of the population of the U.S." by the end of 2025 according to electrek July 23 2025 article. Is this just rumor or Tesla has a chance to achieve it? In October Elon stated that Tasla aimed to have 500 robotaxis in Austin by year end but this month he also said Tasla will "roughly double" its robotaxis fleet in Austin to 60 by year end. There are so much information out there for us to sort out. Now you have AI "news" in the mix, don't know what to believe any more. I do also hear FSD 14.2 is really really good and hope its true.

Sponsored

 

[sunydrm]

Elon said he was expecting Tesla robotaxis to be available to "probably half of the population of the U.S." by the end of 2025 according to electrek July 23 2025 article. Is this just rumor or Tesla has a chance to achieve it? In October Elon stated that Tasla aimed to have 500 robotaxis in Austin by year end but this month he also said Tasla will "roughly double" its robotaxis fleet in Austin to 60 by year end. There are so much information out there for us to sort out. Now you have AI "news" in the mix, don't know what to believe any more. I do also hear FSD 14.2 is really really good and hope its true.

There is no chance tesla has large scale robotaxi anytime soon.

There is a small chance that tesla removes the safety monitor within the next 30-60 days

if they remove the safety monitor and drive on the highway then the safety is there to scale. Maybe tesla next year has a couple thousand. Most likely they'd shoot for more than waymo numbers, but that may not happen

Realistically tesla can be in 20 cities with ~300 cars per city. 6000 cars would likely surpass waymo

14.2 is very safe. With bumper camera integration and the "perfect" mapping that robotaxi geofence areas have, it is really good

Rivian will go through hell just like tesla did for many years. They will have incorrect maps with bad lane information, wrong speed limits, etc. and try to generalize driving based on this data. Then for unsupervised they'll have to undo a lot of this work to build a system that drives with correct information.

 

[hudesigns]

There Rivian will go through hell just like tesla did for many years. They will have incorrect maps with bad lane information, wrong speed limits, etc. and try to generalize driving based on this data. Then for unsupervised they'll have to undo a lot of this work to build a system that drives with correct information.

But I thought the whole point of vision+neuronet approach which led to recent breakthrough is to not rely on rule based HD maps and let monitor lens on vehicle see environment as human in realtime and neuronet processes in realtime. HD map is not singularly crucial any ways in current technology. Otherwise how FSD can drive every little road out there mapped or not.

 

[sunydrm]

But I thought the whole point of vision+neuronet approach which led to recent breakthrough is to not rely on rule based HD maps and let monitor lens on vehicle see environment as human in realtime and neuronet processes in realtime. HD map is not singularly crucial any ways in current technology. Otherwise how FSD can drive every little road out there mapped or not.

FSD drives every road on earth. Not with 100% correctness. Right now the miles per safety critical intervention are probably 50K. But depending on where you live, the miles per intervention could be 50-100 miles. The issue is mostly lane selection where the car incorrectly chooses the lane (turning lane, straight lane, etc.) Depending where you live parking may be perfect, or in a big city it chooses the wrong parking spots or doesn't look far enough around the parking lot. It mostly circles near the front.

Freeway of course the car goes for 10s of 1000s of miles with the only intervention being toll booths. Or sometimes that minimum speed 40 bug in GA depending on the version

Imagine driving for the first time in a new city but the person giving you directions is feeding you wrong information at every turn. Tesla solved a very hard problem

The other issue is reading street signs in real-time is hard. Tesla has speed limits stored in the maps (sometimes incorrect) but it also reads them. In GA it reads a minimum speed 40 sign as 40mph. So on the interstate it will sometimes drop from 70 down to 40.

Tesla rectified this problem by having the car drive the natural speed of traffic. The issue is now you can have a scenario where you are in a construction zone 50-55mph and going 75 because the car is taught to ignore speed limits

FSD v14 now more literally follow the maps and speed limits. Which as long as the maps are improved can fix a lot of disengagement issues.

 

[Dave Cundiff]

Sounds like you're repeating the verdict from that $239 million dollar lawsuit or whatever it was. In that case it was autopilot, the driver was on the floor and he pressed the accelerator pedal where it disabled auto emergency braking. Tesla's logic is that human input while using ADAS will override auto emergency braking

My understanding is that Tesla had full information about the Florida crash, but chose not to share it with local police (and actively misled the police officer who asked for all relevant electronic records and logs, telling him no such record existed). If Tesla had possessed evidence that the operator pressed the accelerator pedal prior to that crash, they appear to have missed the chance to share it with police and with the court.

Juries hate it when people lie to police. They especially hate it when wealthy corporations lie to police in order to protect their image and profits.

What happens on appeal is anyone's guess, because "money" also "talks" -- loudly -- in the American legal and political systems. But citizen juries generally try their best to be fair, using all the facts presented to them.

 

[sunydrm]

My understanding is that Tesla had full information about the Florida crash, but chose not to share it with local police (and actively misled the police officer who asked for all relevant electronic records and logs, telling him no such record existed). If Tesla had possessed evidence that the operator pressed the accelerator pedal prior to that crash, they appear to have missed the chance to share it with police and with the court.

Juries hate it when people lie to police. They especially hate it when wealthy corporations lie to police in order to protect their image and profits.

What happens on appeal is anyone's guess, because "money" also "talks" -- loudly -- in the American legal and political systems. But citizen juries generally try their best to be fair, using all the facts presented to them.

This is not what happened. Tesla collects crash data from all accidents. This is submitted to the NHTSA and also to anyone that asks for it. What tesla said they did not have was the crash video.

Tesla collects the videos from cars to train FSD including all accidents. What they do is anonymize the video so it is not linked to any car's vin number, person's identity, vehicle, etc.

Technically Tesla does have the video and based on the time of the accident maybe they can go through that day's videos and find it. That would just be against the purpose of collecting anonymous videos.

A hacker found the video on the autopilot computer and if you have the video you can match it to a video on your own servers.

There was no defect of autopilot in that the user was on the floor, did not see the warning that shows that auto emergency braking is disabled (if you step on the accelerator pedal while autopilot is enabled). He admitted to pressing the accelerator pedal with his hand and the crash data shows this. Tesla was 0% at fault in this regard.

The reason they found Tesla at fault (which was BS) is that they claim tesla's marketing of FSD (which was not out yet) led to the driver believing he could do this. The other issue is the driver likely does not have much money and so Tesla is an easy target.

Because of this settlement, Tesla chose to settle many other lawsuits, which Tesla have even less credibility. Tesla has won many of these already.

Keep in mind the "lying" to the police (about not having the video) had nothing to do with the outcome of this trial. Tesla was not even found to have lied to the police.

Tesla has since changed their policy and for accidents they do collect the crash video remotely if possible. This is now linked to your vin. Rivian does not remotely collect crash videos.

 

[Dave Cundiff]

Thanks for the prompt response, @sunydrm.

My source for the information about Tesla's alleged misleading of the police was Electrek articles posted by Fred Lambert, with the most specific information in Fred's 2025-10-31 update: https://electrek.co/2025/10/31/judg...ul-deliberate-violations-fatal-crash-lawsuit/ .

Fred describes himself as an early Tesla booster who has become very disillusioned with Musk's and Tesla's behavior. I'm not aware of serious factual challenges to Fred's journalism.

***

NBC's local report on the trial says this: "At the beginning of the trial, Tesla claimed it could not locate the car’s crash data, but after the attorneys representing the families had success in retrieving the data, the automaker said it was able to locate the same data on its own servers. 'It was intentional,' the [plaintiff's] attorney said. 'Why do people hide things? Because they don’t like what they see.' However, Judge Beth Bloom, who presided over the case, did not find sufficient evidence that Tesla was intentionally hiding the evidence by not initially providing it. The judge did require the company to reimburse the families for the costs paid in getting the data themselves." (https://www.nbcmiami.com/news/local...9-fatal-autopilot-crash-in-key-largo/3685939/)

The gap between Tesla's data-retrieval abilities and its data-retrieval claims is also addressed in this law-faculty blog: https://law.marquette.edu/facultybl...s-242-million-in-first-verdict-against-tesla/ .

***

These don't directly address Fred's point about Tesla's alleged misleading of the police. But another document does:

Both the accelerator claims, and the withholding-of-evidence claims, are described in this summary of the case: https://acrobat.adobe.com/id/urn:aa...7-1f993d336566?viewer!megaVerb=group-discover . I ran out of time to read this in detail, but the jury's verdict presumably took into account all the evidence, and the jury found "clear and convincing evidence" that McGee was 67% liable for plaintiffs' injuries while Tesla was 33% liable.

***

Sorry I don't have time for more detail. I hope this helps provide a factual basis for evaluating all the evidence presented in the Benavides case, and the jury's thinking.

Very best wishes!

 

[sunydrm]

se, did not find sufficient evidence that Tesla was intentionally hiding the evidence by not initially providing it. The judge did require the company to reimburse the families for the costs paid in getting the data themselves

Fred Lambert makes a lot of dubious claims. He makes a living writing anti-tesla articles and therefore is incentivized to exaggerate a little bit. It would be the same as the realtesla subreddit where you simply post something negative and you get free karma. While there are good posts on there, I would bet 50% of the information is wrong.

But the actual answer was exactly as I described. Tesla did not collect crash video at the time of the accident that is linked to a car. They would have had to go through their video library and use forensic analysis to find said video. That would be a huge security violation to do that.

Tesla's video library has videos inside people's garages or outside their homes.

 

[Dark-Fx]

Tesla spent more than 10 billion on training compute and storage.

I love all these arguments like spending money is a plus, when Tesla's DOJO failure exists. All that effort and money down uselessly down the drain. Tesla is constantly throwing away their old shit they worked so hard on in this space.

 

[sunydrm]

I love all these arguments like spending money is a plus, when Tesla's DOJO failure exists. All that effort and money down uselessly down the drain. Tesla is constantly throwing away their old shit they worked so hard on in this space.

DOJO was 1 billion dollars but most of that was towards actually building the training compute. They use it

DOJO was not dropping multiple billions of dollars to develop training compute. Most of that 1 billion was on the compute itself. it was not a failure you think it was.

The bigger "failures" were cybertruck and the 4680 battery packs. Those have not delivered. The 4680 packs are being used in the cybercab if that does scale. The cybertruck is dead unless it sells due to self driving in the future.

When I was referring to spending money, I am referring to total spend. Tesla employs only Americans to work on self driving and they have over 1000 data labelers. They have probably 800 working on AI/FSD as well. They are easily spending 300 million a year

By comparison I think these chinese companies have 5000 a year working on this problem. Tesla is basically a tier 1 AI company at this point and what is the talent like in Serbia?

 

[Donald Stanfield]

FSD drives every road on earth. Not with 100% correctness. Right now the miles per safety critical intervention are probably 50K. But depending on where you live, the miles per intervention could be 50-100 miles. The issue is mostly lane selection where the car incorrectly chooses the lane (turning lane, straight lane, etc.) Depending where you live parking may be perfect, or in a big city it chooses the wrong parking spots or doesn't look far enough around the parking lot. It mostly circles near the front.

Freeway of course the car goes for 10s of 1000s of miles with the only intervention being toll booths. Or sometimes that minimum speed 40 bug in GA depending on the version

Imagine driving for the first time in a new city but the person giving you directions is feeding you wrong information at every turn. Tesla solved a very hard problem

The other issue is reading street signs in real-time is hard. Tesla has speed limits stored in the maps (sometimes incorrect) but it also reads them. In GA it reads a minimum speed 40 sign as 40mph. So on the interstate it will sometimes drop from 70 down to 40.

Tesla rectified this problem by having the car drive the natural speed of traffic. The issue is now you can have a scenario where you are in a construction zone 50-55mph and going 75 because the car is taught to ignore speed limits

FSD v14 now more literally follow the maps and speed limits. Which as long as the maps are improved can fix a lot of disengagement issues.

You make a lot of claims but have yet to link a single source to them, while others here have referenced articles and lawsuits. Could you start backing up the things you say with source material? Frankly I find it hard to take anyone’s trust me bro sourcing on the internet.

 

[Time2Roll]

Another delusional Tesla apologist spotted. Denying that FSD has killed anyone is pure delusion.

I am no apologist however the primary cause of death has been reckless driving and misuse of the FSD.
Still need eyes on the road, hands near the wheel, and stay in the speed limit already.

 

[BigSkies]

I don't care enough to read through this whole chain or engage in internet arguments about Tesla. But I know a few people in the industry, and have at least a passing knowledge based on second or third hand comments from people that know what they're talking about.

The Economist also had a good article about how autonomous driving in last week's edition. I recommend it to anyone who wants to know the market better.

A few points in no particular order:

1. Full Self Driving is marketing gobbledlygook and doesn't really mean anything specific. Anyone referring to FSD as anything beyond Tesla's specifically not Full Self Driving software probably doesn't know what they're talking about.

2. Incrementally improved ADAS and no-driver taxi's share underlying technology, but are fundamentally different businesses.

Anyone conflating the two probably doesn't know what they're talking about. While consumers think of apps like Uber as a "finished product", don't forget that Uber is investing about $3B a year (cash flow from Investments) , and has about $15B in Opex every year. That's actually ~50% more Opex than Tesla currently has ($10B in 2024).

3. As it stands today, operating a self-driving taxi service is still significantly more expensive than having a driver. The Economist quoted a BCG study that self-driving companies have costs of about $7-9/mile to operate with traditional ride hailing at $2-3 per mile, and personally owned vehicles at about $1/mile. A McKinsey study estimates it will take about a decade of development to get to $2/mile for self-driving. This is at best a slow-burn business to grow. It will more likely be a capital inferno until costs come way down.

4. Neither improved ADAS or self-driving will be businesses where a company just flicks a switch and the service is magically available. This is a capability that will slowly improve each year until it reaches maturity decades from now. Equities should be priced accordingly.

5. Barriers to entry are much lower than assumed when Tesla first started selling their not-full-self-driving package. Various companies, partners, and licensing opportunities exist for different parts of the tech stack. Rivian (and Ford, GM, etc) can choose which parts to license and which parts to build in house.

6. I believe the hope consumers will pay thousands of dollars for an improved ADAS package is hopelessly naive. Even among the demographic most likely to pay for improved ADAS (Tesla drivers), uptake is estimated to be in the ~15% range. If only 15% of Tesla drivers will pay for this, what percentage of GM drivers would?

 

[Donald Stanfield]

I am no apologist however the primary cause of death has been reckless driving and misuse of the FSD.
Still need eyes on the road, hands near the wheel, and stay in the speed limit already.

Agreed, but when the marketing and press point to a system that is better than it is in reality Tesla is responsible for the abuse of the system.

 

[sunydrm]

You make a lot of claims but have yet to link a single source to them, while others here have referenced articles and lawsuits. Could you start backing up the things you say with source material? Frankly I find it hard to take anyone’s trust me bro sourcing on the internet.

https://teslafsdtracker.com/

There is a site called the FSD community tracker. It is not the most accurate source of data because the sample size is pretty small, but it does a good job of classifying ratio of disengagements. The safety critical miles are also not accurate because they automatically decide safety critical or not based on intervention type.

80% of disengagements of FSD are mapping issues. Probably closer to 95% of disengagements are issues that can be fixed by having information in the maps which are not directly a mapping issue, i.e. you can have information about a nasty speed bump or pothole in the maps. Direct mapping issues are navigation and lane selection.

Tesla sandbags their performance (to where FSD doesn't appear to improve) because they do not independently fix mapping issues to give individual users a better experience. Which means the 80% of disengagements mostly does not change unless the underlying mapless driving has huge leaps in performance. You can of course report mapping issues to tomtom and they do get fixed.

If you fix mapping issues then the miles per intervention (putting other cars, vulnerable road users at risk, etc.) is extremely low and probably would be more or less the same as the milers per safety critical intervention. That's not to say that there would be zero interventions for the car getting stuck in a parking lot (that is blocked off) or something similar. This is not an issue for unsupervised.

The new version of FSD 14.2 is completely full-featured driving. It pulls over for emergency vehicles, detects accidents and close roads ahead and reroutes by turning around, etc. The new versions of the software will have reasoning and be able to read complex parking signs. Even current FSD goes through drive thru and stops at the various speakers and windows.

If we look past all the incorrectly labeled safety critical disengagements (the site auto assigns them and you can not specify), I would bet the miles per safety critical disengagement is in the 50K range.

FSD 13 was about 1000? but it had a bug where it would sometimes run a red light after being stopped and the coast is clear. That was fixed in robotaxi and I would guess that version was closer to 10,000 (which was definitely not ready for unsupervised). You can also show a big improvement beyond that by geofencing any areas that would have an issue crop up if you drove that intersection 1000s of times.

Keep in mind for unsupervised you'd need millions of miles per true safety critical disengagement (within the operating domain which can exclude say heavy snowfall) on the highways. Ideally if you see any type of issue with any ADAS on the highway it is not good.

Sponsored