Tesla discontinues lane centering from Autopilot: TechCrunch

[Bdr100]

I still can't find a single lawsuit, much less an article citing any lawsuits. I even entered your AI text from above and didn't get a single suitable link. Just some on the removal of radar and USS, not on actual software features removed. I'm not saying you're wrong - just that the data isn't available online that I could find.

This video is a pretty good summarr/speculation as to why Tesla made this change:

His reasoning can be summarized as:
• Maintaining two codes bases is harder; AP was never ported to CyberTruck, and Tesla thinks the future is FSD.
• The CA DMV lawsuit on the naming/branding that led to CA regulators issuing new rules (change name/marketing or assume L3 responsibility).
• He discounts Elon's pay package terms as a reason, although he allows for it at the end.

He's mostly a Rivian channel, and perhaps unsurprisingly, touts the Rivian solution method as superior, avoiding the question of performance IRL.

But, I do think he's spot on with the two code base reasoning, and I'll add that Elon lives in the future, but almost always gets the timing wrong. He forced the removal of radar and USS before the software was ready, and it turned out to be around a year before it was ready. He's also killed the development of new Tesla vehicles, claiming that UFSD will make those obsolete before they're ready. But, as usual, he'll be wrong on the timing, probably measured in years, not months. Tesla will continue to suffer for a while as a result, but it's got a strong enough balance sheet to survive Elon being way too optimistic.

BTW, if you do take a trial of FSD(Supervised), make sure you look at the settings and choose the one that makes sense for you. Upthread someone reported that it was too slow to get going, and that could be because it was in "sloth" or "chill" instead of "standard" or "hurry." There's even a "Mad Max" if you're really impatient and are paying close attention. Also note that regulators previously ruled that Tesla wasn't doing "full stops" at stop signs (they wanted more than a millisecond of stopping, even though the law doesn't say that), so FSD does full stops behind the white line, then creeps forward to see, and then finally goes. It's like it's taking a DMV driving test, but that's what regulators have decided it needs to do, not Tesla's fault.

Kyle on OOS goes in to this in his recent torching of Tesla for doing this. He brings up the concept of Tesla maybe not wanting to maintain two different code bases. His simple solution is that Tesla can simply have two versions of FSD. A free version that comes with all cars and is geofenced to highways, does not include point to point navigation, that would make it very similar to current Auto-pilot. Then if you want all the added features of full FSD, you pay for the upgrade. This is blatantly an attempt to force higher adoption of full FSD.

I am not an expert on Tesla and never owned one but unless there is a contract exception that all Tesla owners signed that stated that Auto-pilot was a limited feature and that it could be significantly limited in the future at Tesla’s discretion, then Tesla is going to fact a massive class action lawsuit for selling cars with Auto-pilot being a major selling point and then taking it away later.

When I was looking at various cars a year ago as I started my SUV search, I looked at BMW and was intrigued by their “Highway Assist Plus” which is very similar to GM Supercruise. Unlike Supercruise, which has a subscription cost, Highway Assist Plus does not have subscription cost but did have a footnote on their site stating that the functionality would be limited to 10 years.

 

[Smorgasbord]

His simple solution is that Tesla can simply have two versions of FSD.

Unfortunately, I suspect that solution isn't simple. FSD is an E2E AI codebase where a differentiating factor is the "weights," not code. There just aren't places to insert "if-the-else" logic - they'd have to restructure the transformers and then re-train, and keep both those moving forward. And while the idea of geo-fencing to highways is interesting, that would be adding features not in AutoPilot today. Like I said, Elon's living in the future where the only reason to buy a Tesla is to get FSD, but that everyone will want FSD. That future is further away than Elon thinks.

At some point pretty soon (but again not as soon as Elon thinks), FSD will be good enough that most people will not miss nor want AP. I suspect he's happy to let other automakers have their highway-focused products that do TACC, lane-keeping, and change lanes for you when you put on the turn signal. I agree with Travis Ketchum that's what people want today. But, also, when you drive with FSD today you start to realize how it's actually safer than driving manually, and FSD is still getting better. The robotaxi efforts highlight what it's capable of, and also what the trajectory is.

unless there is a contract exception that all Tesla owners signed that stated that Auto-pilot was a limited feature and that it could be significantly limited in the future at Tesla’s discretion, then Tesla is going to fact a massive class action lawsuit for selling cars with Auto-pilot being a major selling point and then taking it away later.

At this time it's unclear whether Tesla is actually taking away AP from cars that have it today, or is (more likely, IMO) just not offering it for new cars being sold. Much like Mercedes no longer offers their "Drive Pilot" Level 3 product anymore.

 

[PCMc]

My opinion, and this is the same I posted on the comparable Tesla Owners Forum, is that removal of autopilot has nothing to do with trying to push people to FSD, but I believe is driven by the recent California DMV ruling that found Tesla's use of the name Autopilot to be misleading and violates state law. The DMV gave Tesla 60 days to remedy that or California would pull their dealer's license, prohibiting Tesla from selling vehicles in California.

You can read the CA DMV news release here: https://www.dmv.ca.gov/portal/news-...mv-finds-tesla-violated-california-state-law/

The case also took aim at Tesla marketing of FSD, which is what prompted Tesla previously to now start referring to FSD (Supervised). I just received another email from Tesla today regarding my ability to transfer my fully paid FSD from my current Model S to a new vehicle. Everywhere in that email it consistently refers to FSD (Supervised) and also includes the footnote "*Features require active supervision and do not make your vehicle autonomous.". That language is quite a departure from what was used previously.

While at times I now fall in the camp of being someone that owns a Tesla Model S somewhat in spite of the company, the latest versions of FSD (FSD 14.2.2.4) is pretty amazing at what it can do. I've made trips to a couple of different stores/coffee shops locally today. Each time I started FSD (Supervised) from within my garage, only to have it gracefully back out of my garage, drive me across town, pull into the parking lot of my intended destination. Upon arriving back home it pulls into my driveway, lines up with the open garage door and proceeds to pull into the garage. Over the 2100 miles I put on FSD 14.2.x versions since late December I'm at about 99.5% having been driven by FSD (Supervised). It's quite amazing the evolution from the capability of my first Model S purchased June 2016 with the MobileEye AP1 system. It's to the point where I honestly don't care if it ever gets to a true L3/4/5 system. I'm totally OK with the minimal supervision I need to give it in spite of it's remaining quirks.

 

[Tâm]

The $999 Comma.ai might be a bargain to gain Autosteer without paying Tesla FSD.

Unfortunately, Lucid is not listed in its compatibility list.

Maybe if there's enough demand, Lucid would get in the compatibility list.

 

[Tym]

If I were a lucid, I’d add it to basic dream drive. Just to show them who’s boss.

 

[Tâm]

If I were a lucid, I’d add it to basic dream drive. Just to show them who’s boss.

Agreed. It's excellent opportunity for Lucid to be competitive and add lane centering as standard while Tesla just took it away.

 

[Miami_Lucid_Man]

Agreed. It's excellent opportunity for Lucid to be competitive and add lane centering as standard while Tesla just took it away.

One could hope. This has always confused me because of changes they’ve made over the years but would for example Pures on basically dreamdrive even have the tech built in for lane keeping?

 

[gaskiller]

Ex-14-year Tesla owner here:

FSD is now pretty great.

I hate it, too slow to react, hesitant and you still have to watch the road. I would rather just drive until they have fully functional robotaxis. Not to mention it can kill you....though not very common. The fact that Tesla doesnt have redundancy and puts cost over safety is a big warning sign for me.

 

[Smorgasbord]

The fact that Tesla doesnt have redundancy and puts cost over safety is a big warning sign for me.

Those aren't facts, but whatever.

Anyone wanting to know how Tesla is approaching autonomy should spend a half-hour (or less if you play 1.25X) on:

I have hope that the Nvidia-based system Lucid and Mercedes (and I assume others) are leveraging will work out, and that those systems will approach Tesla's current L2++ capabilities for us Gravity owners without expensive hardware retrofits. Mercedes has already acknowledged that their early CLAs will have the older Orin processor, with Thor coming at the end of this year. Gravity, as we know, uses a dual-Orin setup, which may not be enough to do all that L2++ requires. Everyone from Mobileye to Tesla has underestimated the processing power they needed in vehicles, so I'm not expecting greatness. Maybe Earth, with Thor, will do well, though.

 

[Tym]

One could hope. This has always confused me because of changes they’ve made over the years but would for example Pures on basically dreamdrive even have the tech built in for lane keeping?

They currently don’t, but all Lucid has ability to alert when you drift off lane, so it is possible with the right software engineers

 

[TahoeMichael]

Those aren't facts, but whatever.

Anyone wanting to know how Tesla is approaching autonomy should spend a half-hour (or less if you play 1.25X) on:

I have hope that the Nvidia-based system Lucid and Mercedes (and I assume others) are leveraging will work out, and that those systems will approach Tesla's current L2++ capabilities for us Gravity owners without expensive hardware retrofits. Mercedes has already acknowledged that their early CLAs will have the older Orin processor, with Thor coming at the end of this year. Gravity, as we know, uses a dual-Orin setup, which may not be enough to do all that L2++ requires. Everyone from Mobileye to Tesla has underestimated the processing power they needed in vehicles, so I'm not expecting greatness. Maybe Earth, with Thor, will do well, though.

I've seen this video. I think Ashok glosses over huge parts of the safety strategy, and how they're approaching solving the well known weaknesses of modern machine learning systems. For example, it's a deep research problem to understand why large neural networks "hallucinate", and how one can fix problems of hallucination. I can list plenty of other challenges in AI (I work in this area for another large tech company) that Tesla doesn't appear to be investing in solving. Given Tesla's long term trend of not investing in fundamental AI R&D in an attempt to solve that (and other) problems, I personally am not eager to depend on their AI. I think it's actually pretty concerning that they're saying they need an extra 3 billion hours of driving data for their training. More training data is great, but their insatiable thirst for data in an attempt to solve reliability problems is questionable. In contrast, Waymo is showing a better safety record with a significantly smaller data set for AI training. There's multiple reasons for Waymo's success (e.g. they have a more sophisticated simulation strategy than Tesla does), but they're clearly benefitting from being part of Alphabet's overall AI investment, and the deep commitment Alphabet has made in fundamental research on AI. One of the really cool things about Waymo's approach is that they've invested deeply in their AI understanding when it's in over it's head, and how to safely come to a stop. FSD has gotten better in that regard, but they're suffering from too many years of leaning on the crutch of human intervention to backstop against AI failures.

Tesla is devoting substantial resources to FSD, and so I do think they'll eventually get there. Particularly since most AI research teams are still publicly sharing their results, so Tesla is likely to be able to benefit from the fundamental R&D happening elsewhere. But I'd prefer other people to be guinea pigs for their engineering. Particularly since you can't really trust what they say about their technology, as they're obviously engaging in stunts and hype in an effort to persuade people to keep the stock price high.

 

[Smorgasbord]

I've seen this video. I think Ashok glosses over huge parts of the safety strategy,

Well, you can't expect a huge amount of details in one presentation, especially since many of those details would be considered company confidential.

I think it's actually pretty concerning that they're saying they need an extra 3 billion hours of driving data for their training.

Elon said that 10 Billion miles (not hours) were needed, and he's apparently talking about seeing enough real-world scenarios to not get tripped up on edge cases. For instance, while Waymo overall has a good safety record, we've seen recent crashes, including a case where two Waymo vehicles crashed into each other.

Your comments about Waymo's "more sophisticated simulation strategy" actually only reflect the complexity of Waymo's sensor suite. Same is true for Nvidia. Both have to not only create realistic looking 360º camera views, they also need to create perfectly matching LiDAR and radar outputs. Tesla, since it only uses cameras, only needs the visuals.

they're suffering from too many years of leaning on the crutch of human intervention to backstop against AI failures.

That's an interesting view, but Waymo was the same for years as well. Waymo started with a smaller problem - 35mph and under, heavy pre-mapping, tons of sensors all over the vehicle, so both companies have their crutches. As for "too many years," Tesla essentially started over in early 2023 when they started down the fully end-to-end path - they ended up throwing their whole hand-coded driving policy code away, and I suspect the NN recognition code has been relegated to what's displayed visually on the screen. That said, Waymo has also had many revisions in its approach; we just didn't get to see it as we weren't behind the wheel for the changes.

Here's a recent video of a few Waymos getting confused around each other:

That all said, I do understand a hesitancy to trust Tesla as they've been fast and loose with what they've rolled out for years. I find it fascinating that Waymo's timing of rolling out new cities suddenly took a turn to the upside (4 new cities just announced, whereas the first 6 cities took 6 years) right after Tesla started doing some driverless taxis in Austin and projected to have 9 by the end of Q2 (which I don't think they'll make, btw). I suspect one of Waymo's issues is simply whether it can modify enough Zeekrs or Ioniq5s quickly and cost-effectively enough to supply all these cities with enough cars to make service via the Waymo app viable in terms of wait times for customers. That's purportedly why they went with Uber in Atlanta, so it's interesting they didn't go that route with these new 4 cities.

Back to Lucid, I think we're all waiting to see when the modified Gravitys with Nuro software build on top of Nvidia's Alpomayo hit the streets. And then to see what, if anything, from Nvidia makes it back into our personally owned Gravitys, since we don't have the same sensor suite.

 

[tmichael]

WOW, you guys are a knowledgeable group of people!!! I find it fascinating to read posts from people who have experience in the area of self-driving.

As exciting as it would be to go downtown to a baseball game from my home in the back seat with no driver and dropping me off right at the entrance to the ballpark, I am hopeful that simply getting me from the expressway entrance to the off ramp without slamming on the brakes due to the exit speed limit sign or following a car so far back that another one slides in front (pushing me back again, etc., etc.) becomes possible. I don't need it to handle pedestrians, bicycles, cars coming out or going in parking places - I just want it to drive down a divided expressway, with most cars going the same speed and in the same direction. I want self-driving to first learn to crawl, then walk, then run. It has worked since the dawn of man, maybe it will work this time...

I DO, however, NEEDd it to get me to my doctor's office visits in 10-15 years... Just sayin'...