I feel like this question is like when the Supreme Court had to decide what makes something pron vs art. A lot of Tesla fanboys keep saying this is pron. but for the rest of us, when comparing it to Waymo it’s obvious that it’s not there.

@JDTurk It's obvious enough Tesla is not at Waymo's level but this market doesn't require that. There's nothing obvious about how to resolve this market. Hopefully the FAQ has things pinned down enough that with more time it will become obvious.

(For those just tuning in, I actually want to taboo words like "obvious". I'd like us all to demonstrate epistemic humility and a spirit of cooperation in figuring out what's most true and most fair.)

Tesla now has a California-like (e.g. human in the driver's seat) pilot in Germany (Eifelkreis Bitburg-Prüm, a city of 100K people).
https://www.linkedin.com/posts/mwvlw-rlp_frohe-botschaft-zum-jahresende-aus-der-ugcPost-7409248393060958209-y_NN/

Musk claims he did a rider-only drive himself, and Ashok has posted another video of such a test. No word on extending to actual customers yet for this.

An updated homepage for Robotaxi https://www.tesla.com/robotaxi
We’re bringing autonomous rides to you today—starting with Model Y. Autonomous Robotaxi rides are currently being offered in Austin, Texas. To get started, download the Robotaxi app.

Unsupervised FSD testing now carrying person in the backseat (but a Tesla employee, for now) https://x.com/philduan/status/2002843697273262468

@dreev does this count as “removing safety monitors” yet?

@MarkosGiannopoulos I'd say we should wait for customer trips with only the customer in the car to check off that milestone. I'm also forgetting now what we decided about the significance of that milestone. I'm still thinking that Elon Musk's seeming admission in November that Tesla had not cracked unsupervised self-driving as of then is pretty strong evidence for NO for this market, whether or not Tesla does crack it this month or early in 2026.

@dreev In regard to the "admission" thing, see past discussion https://manifold.markets/dreev/will-tesla-count-as-a-waymo-competi#d5huzbcese6

In any case, Musk's comments are immaterial. The hard facts are
- Robotaxi in Austin in June was running an unreleased version of v14
- Robotaxi in June had paying customers in cars that drove themselves without a person in the driver's seat
- v14 is now public and shown to be a significant improvement over v13
- The June launch was not a one-off event; Tesla has expanded the service area and the number of cars (now 30 vs 200 of Waymo - a company several years ahead), and also has a California operation (130 cars)
- Tesla is about to remove the safety monitor (already doing rider-only tests with employees in the backseat). This means achieving in 6 months what took Waymo 2.5 years.

@MarkosGiannopoulos

- Robotaxi in Austin in June was running an unreleased version of v14 but tesla owners have had several releases 14.1,14.1.1 through 14.1.7 and 14.2, 14.2.1 and 14.2.1.25. Given the times over which these have been released it makes little sense to assume all these were available to Tesla at end of August. The driverless tests could easily be another version like the v15 order of magnitude larger version or v14.3
-Safety monitor in front passenger seat has remained there for nearly 4 months now. It is possible to put safety monitors in a level 4 system and you wouldn't expect many interventions and for it to be easy to quickly remove them. 4 month delay suggests there are issues and excessive interventions they are still ironing out. This suggests system at end of August was not quite at level 4 ability. Even if you believe they had ability at end of August, Tesla wasn't feeling confident enough to remove safety monitors so they haven't launched it by end of August.
- Yes v14 is better than v13 but this doesn't mean v14.0.x or whatever that was available at end of August was good enough. The long delay before safety monitors are removed and the several later versions in this period tends to suggest what was available at end of August was not good enough and Tesla still has not launched without safety monitors yet

- June launch was not a one-off event, fair enough there has been expansion to SF bay area. However I would question: when does the exponential growth start if there are still zero passenger rides without safety monitor? Is the launch when the exponential growth pattern of rides & taxis without safety monitors starts?

-Maybe they are about to remove or maybe v15 testing continues for a month or two before safety monitors are removed. We surely now want to wait to at least see what version is first used without safety monitors? Achieving in 6 months what Waymo took 2.5 years, well this is suggesting lots of fast progress and if still not launched then it seems more reasonable to tend to deduce they weren't ready for level 4 at end of August rather than they were ready but are being very cautiously slow with the roll-out while not needing to make further refinements.

Length of delay before removing safety monitors, number of different versions and reports of erratic driving sufficient to start regulatory investigations and a few accidents with very few cars operating. AFAICS this is all pointing to not quite level 4 capable at end of August 2025 as well as not launched in time at the level required.

@ChristopherRandles "when does the exponential growth start if there are still zero passenger rides without safety monitor? Is the launch when the exponential growth pattern of rides & taxis without safety monitors starts?"

The launch was in June when you had the first paying customer in a car that drove itself. It's as simple as that. This market is not about "exponential growth". This market is about the launch of the service.

The safety monitor has now been removed for Tesla employees. My estimate is they will have non-Tesla customers doing rider-only this week to keep Musk's promise.

@MarkosGiannopoulos It had to be "level 4" robotaxis which don't need near constant supervision. They weren't at that level at June. Maybe they had improved a bit by end of August but several different versions since and monitors still not removed is not a good impression that level 4 service is reached yet. In June, they launched a level 2 service.

>The safety monitor has now been removed for Tesla employees.
You mean there is more testing that has recently started (not offering rides). Maybe this indicated a major new version that they want tested before allowing safety monitors to be removed. Maybe the start date of this testing might indicate when they actually have a system they trust to be level 4 capable? The launch of the service surely has to come after the capability.

@ChristopherRandles "It had to be "level 4" robotaxis which don't need near constant supervision. They weren't at that level at June." ... "The launch of the service surely has to come after the capability."
If your expectation was that there would be zero safety personnel in the car, with customers on board on day 1, it's entirely not based on any past experience of such programmes (Waymo, etc). The service was Level 4 in June, and they also needed to have some kind of safety monitor, as this was an initial launch with paying customers.

@MarkosGiannopoulos
"My expectation", what does my expectation matter?
Elon Musk repeatedly stated, prior to June 2025, that Tesla would launch a robotaxi service in Austin with "no one in the car" as a fully unsupervised, paid service. This claim was made during various public appearances and earnings calls in early 2025.
https://www.fool.com/earnings/call-transcripts/2025/01/29/tesla-tsla-q4-2024-earnings-call-transcript/
"So, we're going to be launching unsupervised full self-driving as a paid service in Austin in June. So, I talked to the team. We feel confident in being able to do an initial launch of unsupervised, no one in the car, full self-driving in Austin in June." 

This is the background to the question.

They didn't do that, they weren't ready to do it, so they launched a level 2 service while they built up not only a bit more confidence in the system under operational conditions but also improvements to the system through a series of software updates.

Back then I hoped they would succeed without too much delay, but they clearly have delayed doing it.

Guys guys guys, please demonstrate epistemic humility. A car with no one in the driver's seat is probably more than level 2. But the lack of scale and the possibility of safety monitors with real-time kill switches and unknown amounts of tele-operation means we can't act like we know it's level 4 either.

The research y'all are doing is insanely helpful and I'm very grateful and you've shifted my own probability in both directions at various times. Let's just maintain a fully collaborative, truth-seeking tone. Tendentiousness backfires. When readers can tell what you want to be true, they take the evidence you're offering with lots of salt.

Some of the things, off the top of my head, that we're waiting on in order to feel more confident about the fairest resolution:

1. Verdict on physical kill switches

2. Degree of tele-operation

3. Safety record, with confidence intervals (e.g., zero deaths in a million miles is weak evidence of superhuman safety)

4. What FSD version the robotaxis had on August 31

5. Transparency about disengagements (part of driver-out permitting in California?)

6. What Musk meant by calling unsupervised autonomy "close" in November

I continue to expect this to get less murky with the benefit of more hindsight.

Finally, I started out asking for more epistemic humility but everyone here is 1000% better than what I've seen in other markets, where people just jawbone each other about how "obviously" passenger seat safety monitors do / don't count as supervision.

In my opinion, the relevant definition of "supervised" is not just about a human literally watching but rather about being ready to intervene in real time in order to be safe enough. That's what we're waiting to be sure about.

@dreev "the relevant definition of "supervised" is not just about a human literally watching but rather about being ready to intervene in real time in order to be safe enough."

Well, the established player, Waymo, just reminded us that they rely heavily on remote monitoring

Navigating an event of this magnitude presented a unique challenge for autonomous technology. While the Waymo Driver is designed to handle dark traffic signals as four-way stops, it may occasionally request a confirmation check to ensure it makes the safest choice. While we successfully traversed more than 7,000 dark signals on Saturday, the outage created a concentrated spike in these requests. This created a backlog that, in some cases, led to response delays contributing to congestion on already-overwhelmed streets. We established these confirmation protocols out of an abundance of caution during our early deployment, and we are now refining them to match our current scale. While this strategy was effective during smaller outages, we are now implementing fleet-wide updates that provide the Driver with specific power outage context, allowing it to navigate more decisively.
https://waymo.com/blog/2025/12/autonomously-navigating-the-real-world

Autonomously navigating the real world: lessons from the PG&E outage

At Waymo, our mission is to be the world’s most trusted driver. We know trust is built through consistent behavior over time—earned through every mile we drive and every interaction we have with the community. This past Saturday, as a widespread PG&E outage cut power to nearly one-third of San Francisco, our service was put to the test. With power now restored, we want to share an account of our operations during the outage and how we are evolving to better serve the city.

So, at what point do we accept that some remote monitoring is ok and has little effect on whether the car does in fact drive itself? :)

@MarkosGiannopoulos Is that a genuine question? The excerpt you quoted answers it beautifully. Namely, it illustrates exactly where we've set the bar for counting as level 4. If the car can get confused by unusual situations, stop, and phone a human to ask what it should do, and, importantly, the human merely tells it the answer and does not take control -- that's still level 4.

I wrote the following post to answer all the Waymo whataboutism, based on previous discussion in this market: https://agifriday.substack.com/p/waymo

I'd summarize it as no real-time control at driving speeds or monitoring with the ability to disengage in real time.

PS: I found some evidence that Tesla robotaxis in San Francisco at least sometimes needed the human backup driver to take over through dead intersections. A nice irony at 0m49s in that video: a Waymo autonomously navigating a dead intersection while the Tesla's backup driver is manually navigating it. And something confusing in the video: the display that normally shows the car's path and other traffic and pedestrians shows only the Tesla. Did the power outage somehow spoil its ability to detect other cars? Only at one point (right around when the Waymo appears) does it register another car as an obstacle on the display.

@dreev Waymo cars also sometimes end up needing a human to take over. Example https://x.com/Cyber_Trailer/status/2004671911847231904

@MarkosGiannopoulos I can't tell if we're missing each other on this question. If the car is confused or has a problem and it stops autonomously and waits for human assistance, that's allowed for level 4 autonomy. Level 5 is the one where the car has to be able to handle everything a human can handle. If a human is ever need in real time while the car is driving, that's at most level 3.

Review of things we know:

1. Waymo is level 4

2. Tesla is way more than level 2 (in spirit even if not technically)

3. Tesla has not yet driven a customer without human "supervision"

Review of things we don't know for sure:

1. Whether Tesla is level 3 (let alone level 4)

2. Whether Tesla's human supervision is necessary

3. The role of tele-operation for Tesla robotaxis

4. Whether the safety monitors have their finger on a kill switch

5. Whether Tesla can scale up to hundreds of unsupervised cars

6. How far we are from vision-only level 4 autonomy for the masses

>"4 Whether the safety monitors have their finger on a kill switch"

If they don't have finger on a kill switch but can intervene via main screen (as seen) and are required to be eyes on road / paying attention while the car is driving is this the same as having figure on a kill switch or at least supervision that restricts what is going on to level 2?

These seems extremely likely to me by rules requesting passengers do not speak to monitor when ride is underway and the likelihood this is required under the level 2 permitting used.

Note also FAQ 3: If there's a backup driver we're going to presume that it's because interventions are still needed, even if rarely.

So even if there is no kill switch, I still think it is supervised not unsupervised. If monitors were removed within a couple of months using same version software as was in use at end of Aug then I could see an argument that the software was ready and launched even if not done in a level 4 way by the deadline so it was level 4 ability operating in level 2 way, out of caution before a larger scale roll out of a level 4 service. With 11? software versions coming out and monitors still not removed nearly 4 months after end of Aug then it is still supervised.

So on your don't know for sure list
1. Tesla is operating and regulated as level 2, they may be close or even at level 4 capable now but not at end of Aug 2025.
2 Supervision is presumed necessary for this question by presence of monitors per FAQ 3. While this might well not be necessary now under latest software versions there were enough interventions, erratic driving reports and official investigations to say supervision was necessary back in June 2025 and fixes aren't likely immediate so almost certainly still necessary in July and August.
3456 we don't know for sure but I don't see the need to know in order to resolve the question no.

@ChristopherRandles Let me start by reemphasizing that I personally don't see this market getting to a YES at this point. But it's not that easy to declare a NO at this point either. For starters, FAQ3 is specifically referring to backup drivers in the driver's seat, right? As in someone who can take over the driving in real time if the car messes up. (If you also read FAQ2, it's made more explicit.) The passenger seat safety monitors are a highly confusing and ambiguous middle ground between level 2 and level 4. Being able to intervene via a touchscreen is quite slow and limited. I don't think we have any footage of that happening at normal driving speed, do we?

Supervision, in the context of self-driving autonomy levels, means a human watching in real time with the ability to intervene. I think there are unanswered questions about kill switches and tele-operation before we can give a definitive verdict on that.

(Quick review of the levels: 2 = human monitoring, ready to intervene in real time if car messes up; 3 = human not monitoring but ready to take over in real time if the car beeps; 4 = human never needed in real time; 5 = human never needed ever.)

To also reemphasize, I know people have very strong intuitions in both directions about the safety monitors. To some it feels painfully obvious that that's supervision and not fundamentally different from a backup driver in the driver's seat. To others it feels night-and-day different. If you can't take over the actual driving at a moment's notice then, the Tesla proponents argue, the AI is meaningfully the one in control. In that framing, the human's ability to give the AI a stop order is only supervision in the most literal way.

So just remember, whatever arguments you make, show respect and sympathy for those with opposite intuitions.

(Not to say I'm accusing you of failing to do that, @ChristopherRandles, and I think some of your arguments -- like how many new FSD versions have come out between August 31 and whenever the safety monitors are finally removed -- are persuasive.)

@dreev On your open questions

1. Whether Tesla is level 3 (let alone level 4)
Level 3 requires a fallback driver ready to take over. This was never the case in Austin because the safety monitor was on the passenger seat. I guess we can call Level 3 what Tesla is doing in California. In Austin, Tesla is now further demoing Level 4 (but they need to go beyond a couple of demo videos on Twitter) https://x.com/philduan/status/2002843697273262468

2. Whether Tesla's human supervision is necessary

Since June, Tesla has been confident enough not to have someone in the driver's seat. They will have customers doing rider-only rides pretty soon, within 6-7 months since the June launch. What further criteria do you have for this?

3. The role of tele-operation for Tesla robotaxis

This is unlikely to prove one way or the other, especially if you want to focus on the summer period. No special hardware has been noticed in the cars that would indicate they are teleoperated.

4. Whether the safety monitors have their finger on a kill switch

Will that still be a point of discussion if, in a couple of weeks, rider-only is launched for customers?

5. Whether Tesla can scale up to hundreds of unsupervised cars
They went from 10 to 160 if you count California as well. This is within the "launch" description of the market, I think.

6. How far we are from vision-only level 4 autonomy for the masses
Why is this part of resolving the market? The question is whether Tesla launched a robotaxi service.

@dreev

(/Attempting to understand other sides view)

I think we need to break this down:
>"To also reemphasize, I know people have very strong intuitions in both directions about the safety monitors. To some it feels painfully obvious that that's supervision and not fundamentally different from a backup driver in the driver's seat. To others it feels night-and-day different. If you can't take over the actual driving at a moment's notice then, the Tesla proponents argue, the AI is meaningfully the one in control. In that framing, the human's ability to give the AI a stop order is only supervision in the most literal way."

I am not sure about the "Tesla proponents" label. I think I am a Tesla supporter but in this case accept they haven't done it in time. What I see you arguing here for 'yes resolution arguers' is that what matters is "driving control" of the vehicle and if the interventions are more like changing the destination than taking over the driving then that should be considered good enough for a yes resolution. I do not agree with that point of view. The criteria level in the question title is level 4 and you said "4 = human never needed in real time" so humans needing to watch practically every second if not 100% of time is not at level 4.

We can perhaps try to break this down into various standards of 'humans needed' and 'vehicle control' to see what we think is the more important viewpoint.

1. No humans in car or remotely watching unless car reports it is unsure. I think Waymo are at this level and it is level 4 even if there are very occasional accidents.
2. Human in car is supposed to watch all the time but never intervenes.
3. Human in car but entirely an air hostess role than a pilot role. Human doesn't need eyes on road so this qualifies as level 4
4. Human in car but mainly an air hostess role than a pilot role. Human doesn't need eyes on road most of time but might be called to give advice after car stops if car is uncertain of what to do.
5. Human in car but mainly an air hostess role than a pilot role. Human doesn't need eyes on road most of time but might be called to give advice either while approaching situation or after car stops if car is uncertain or suspects it might become uncertain of what to do. (If it doesn't get advice in time then it tries to take safer option like stopping and waiting for advice.)
6. Human in car but mainly an air hostess role than a pilot role. Human doesn't need eyes on road most of time but might be called to take over driving with x seconds notice.


My view on these is that
6 would be level 3 not level 4 so should resolve no
4. seems to be level 4 like waymo so the question could be resolved yes.
5. is arguable a better system than 4 so while the human might occasionally take driving decisions while vehicle still moving I could see a yes resolution being more appropriate than no
3. qualifies as level 4 and should resolve yes
2. is a level 4 capable software being used in a level 2 manner. I don't believe the software was at this capability level in June and almost certainly not in July or August.
1 seems a clear level 4 resolve yes situation.

My suggested resolutions for these situations don't completely align with humans needing eye on road pretty much whole time, but it is fairly close. (Supposed to keep eyes on road but never intervene can be treated as not 'needing' to but 5 is a somewhat awkward case.)

As the descriptor in the title is level 4 and not about "driving control", I don't see any reason to morph the question into being about "driving control" rather than general understanding of what level 4 means which you have said is "human never needed in real time".

Coming back to the question "Are interventions more like changing the destination than taking over the driving?" I would argue that if human needs eyes on the road to know when to intervene then this is supervision not merely an instruction to change the destination. If they never need to intervene then I would concede that the supervision isn't necessary and it is level 4 being used in level 2 way. This 'never need to intervene' brings us back to only needing one accident or intervention or one critical disengagement for a no resolution (like indicated in FAQ9) and we know we have had some of these since August 2025.

The Robotaxi Tracker site now includes data on crashes from NTHSA https://www.teslarobotaxitracker.com/nhtsa?provider=tesla

NHTSA Incident Reports | Robotaxi Tracker

ADS crash reports from NHTSA Standing General Order 2021-01.

We have yet to see any catastrophic headlines like this for Tesla

“Waymo said Saturday that it was stopping service across San Francisco after numerous online videos showed its autonomous vehicles snarling traffic during the citywide blackout.”

“Many of the videos and images showed Waymos stuck one behind the other, with human drivers passing them by.”

https://missionlocal.org/2025/12/sf-waymo-halts-service-blackout/

Waymo halts service during massive S.F. blackout after causing traffic jams

Numerous autonomous vehicles caused traffic jams across San Francisco after a PG&E outage hit 1/3 of the city.

FSD having no issues

https://x.com/tesla_ai/status/2002663584263016589?s=46

@MarkosGiannopoulos Catastrophic? It makes sense to me that supervised self-driving would muddle through, trusting the human supervision, while unsupervised self-driving would be ultra-conservative and stop driving when it sees an environment very different from training and with traffic lights gone dark. "Better safe than sorry" is basically Waymo's motto.

I was bracing for something very different when you said "catastrophic headlines"! (And even so, Waymo will hit a quarter billion rider-only miles in the new year so at some point we really should be ready to learn of a Waymo literally killing someone and still correctly view it as life-saving technology.)

@dreev well, multiple cars freezing in the middle of the road one behind the other and blocking traffic in multiple cases is not something that has happened with Tesla.

@MarkosGiannopoulos True, but that might be because the supervising human takes over in such situations, right? There are so many ways it's so tricky trying to compare Teslas and Waymos. The orders of magnitude difference in scale, for starters. (In which direction? Depends if we're talking about unsupervised or not! Way more FSD miles in private Teslas; way more Waymo miles with empty driver's seats. The scale is vastly skewed and we can't even agree in which direction.) And the cherry-picking media coverage is just an epistemic hellscape. Everyone has an agenda.

I've been bearish on Tesla and predicted they wouldn't launch robotaxis at all this past summer. I was wrong. But the bulls expected a scale-up to millions of cars by now, so they were wrong too. Prediction is hard. Especially about Tesla.

But I think it's going to happen. We've gone from a year away in Musk-reality to a month away. Maybe a Musk-month is about a year? I'm not betting NO in the 2026 robotaxi markets. (Or maybe I am, depending on the odds, who knows.)

I did originally, back in April, predict it wouldn't happen for Tesla till Musk relented on lidar. I think in retrospect I ought to have said "either lidar or another year-ish of AI progress".