Hacker Newsnew | past | comments | ask | show | jobs | submitlogin



"Against" is a pretty clear winner at this point. The chances of a level 5 self driving car being available for purchase in the next six years are ~0.


It doesn't say "available for purchase." It says "will be commercially available for passenger use in major cities."

Waymo is already there, in at least 2 cities but not all weather conditions. Expanding to 10 cities and all weather conditions in the next 6 years sounds more than plausible.

I had sided with Atwood (skeptic) but now it seems like Carmack "has already won".


Is "all weather conditions" a global benchmark, or is it "all weather conditions occurring in the cities of operation?"

At least in SF, I haven't seen a day where Waymo refused availability for the weather. Rode it all over town in the rainstorms this weekend, and found it a little jarring to see the windshield covered in rainwater cause the computer doesn't GAF.

If it doesn't need to handle snow to win this bet, it seems it could handle Seattle-tier rain already.


They're allowed to have limits like "local driverless taxis don't operate outside SF city limits or below 35 degrees with precip in the forecast" etc. at level 4, but to meet level 5 (per the bet) it has to be able to "drive everywhere and in all conditions," [0] which adds a lot of really difficult edge cases.

Situations that come immediately to mind:

- Driving in the hurricane lane on the shoulder during an evacuation

- Reversible lanes and streets

- Sizing up an icy hill and figuring out whether it's safe to keep going

- Ferries

- Knowing a baseball entering the road from behind a parked car will probably be followed by a child

- Understanding traffic police, sign turners, "follow me" trucks, etc.

[0] https://www.sae.org/binaries/content/assets/cm/content/blog/...


I think each of these is already handled, or at least most. They say 99.4% of uptime in record inclement weather, which seems like it should satisfy "all". https://waymo.com/blog/2023/08/the-waymo-drivers-rapid-learn... I don't think they really mean "all" (like it shouldn't need to handle a lava flood). Just "all a human might do". This feels superhuman already.

I actually see the main thing right now that would mean this bet is "not currently won by Carmack" is that they are not officially offering freeway access in its commercial product: https://waymo.com/blog/2024/01/from-surface-streets-to-freew... But this seems minor, and I can't imagine it taking more than 2 years to allow freeway driving in multiple metros.

I can't fathom what would need to happen to derail this particular bet from being satisfied in Jan 2026 let alone Jan 2030.

(Note: if it wasn't for Waymo, I think this timeline would be much less clear. Tesla/Cruise feel much less predictable.)


Level 5 means there is no instance when Waymo intervenes from home base to tell the car what to do to get around an obstacle, and no instance where an emergency responder drives the car to move it out of the way. If the car is told to take the passengers to Arrowhead stadium, and it is directed verbally by stadium staff to go to parking lot 3, it goes to parking lot 3. Waymo will probably roll out level 4 robotaxis to a major city soon, but it's very hard to see them getting to level 5.


Emergencies are noted exceptions to Level 5, so the "A first responder insists on driving the car" case is still level 5.

Waymo has level 4 taxis in two US cities and is running tests elsewhere too. The usual reaction from the "This can't be done because somehow driving a motor vehicle is a uniquely human ability" people in those cities seems to be head-in-the-sand refusal to believe. The good news is that if they choose to believe a Waymo doesn't exist and step in front of it, it'll probably brake politely to a halt like they're any other asshole.

Level 5 mostly requires more range. I don't think this bet is a sure thing, but it's certainly possible that Waymo's reason you can't go from say New York City to Birmingham Alabama in 2030 is something that sounds like a taxi company reason rather than an "Our AI can't do that" reason.


>Understanding traffic police, sign turners, "follow me" trucks, etc.

I can confirm that in my experience, Waymo handles these kinds of situations fine. Better than many humans in SF seem to, ha! Compared to my past experiences with Cruise, where the car would become instantly paralyzed at the sight of flashing yellow lights, so it could phone home for human intervention.

Icy conditions seem like a big open question to me. As a human driver, there's a big difference between driving on a road slick with rain and a road slick with snow and ice..but maybe there is not much of a difference to a self driving car? Certainly, other humans on the road behave differently in snowy/icy conditions than in rainy conditions, and the self driving vehicle needs to share the road with them.


The bet specifically exempts natural disasters and emergencies.


I wonder when humans will meet level 5 by those standards?


> If it doesn't need to handle snow to win this bet, it seems it could handle Seattle-tier rain already.

Seattle-tier rain probably isn't what you think. But if it operates in Seattle, it does need to handle snow --- at least enough to recognize it's snowing and find somewhere flat to chill out until the roads become drivable again.


Waymo is already there, in at least 2 cities but not all weather conditions.

A few weeks ago I saw a video of someone making his first Waymo ride, there was some construction site and the vehicle stopped and waited for some human to take over control. They are not there in any weather and given that the last 10 % of a problem have a tendency to consume 90 % of the effort, they might have a long road ahead even if it already mostly works.


> Waymo is already there, in at least 2 cities

I find it odd that a self-driving system only works in 2 cities. Is the model overfitted to only know some specific locations? How does that work?

Like, I can understand it struggling in snow or heavy rain. I can understand it struggling if the lines on the road are faded. But given a well-marked intersection in one city, and another well-marked intersection in another, why would it fail in one?


> Is the model overfitted to only know some specific locations?

Yes. Every inch of drivable area is LIDAR-mapped. The opposite of the antifragile approach, which is harder to execute but much more durable.


> Yes. Every inch of drivable area is LIDAR-mapped.

If this is true, then I find it incredibly absurd that people would try to act like Waymo is farther than Tesla in self-driving tech.

Tesla is trying to make self-driving work everywhere without pre-mapped data beyond what you'd get from Google Maps.


That's my position on it. Sure, it functions better, but you can drop a Tesla into an area it's never seen before and it will drive relatively competently. HD maps are a crutch, and ignoring that produces fragile solutions that just require massive overhead to function.


I'll agree once Waymo can operate in any weather. Including snowy, meaning the road surface is completely covered by snow and ice plus heavy snowing.


I live in Seattle, and I can tell you with some certainty that most Seattle drivers are incapable of driving in all Seattle weather conditions...


That is a bit of an exaggeration. Yes, they do have problems in the rain, snow, or when it is really sunny out. But we are pretty good at driving in overcast but dry weather.


Waymo is level 3 at best. It is active in 1 flat city right now (and partially in another), and needs frequent interruption from remote drivers to get it out of hairy situations. You really think it will be fully rolled out in the 10 most populous cities in America in the next 6 years? You really think it'll work without any human interaction (local or remote)? This thing won't last 5 minutes in NYC or Philadelphia .


It says "any" of the 10 most populous cities, not "all."


SF is a flat city?


The parts of Phoenix where Waymo operates is... SF is "(partially in another)"


Waymo operates in all of SF


It’s actually in LA now, which to me is a testament to its ability


"Passenger use" is the bet -- not purchase.


> "Against" is a pretty clear winner at this point.

You seem very confident that "against" will win after only two years have elapsed since the bet was made, when there are still EIGHT YEARS of time remaining on the bet.


2024 to 2030 is six years, not eight.


The hell are you talking about. Waymo is already there.


"For" is a pretty clear winner at this point.

Waymo is already at level 4 at least on surface streets, and they claim 99.4% uptime in rain, high winds, and thunderstorms during last winter. [1] They're testing highway service in Phoenix now.

There might be a debate about the word "all" for the distinction between levels 4 and 5, but Waymo has six years to erase that doubt.

[1]: https://waymo.com/blog/2023/08/the-waymo-drivers-rapid-learn....


Disagree. Waymo is currently at level 3. From your link, emergency responders have the ability to put the car in a manual mode to get it out of the way if it is stuck. Waymo also has employees make decisions on behalf of the car remotely when it finds itself in a tricky situation.

Level 5 fully autonomous would mean the car is able to drive in icy and snowy conditions (adjusting how it drives based on traction), it would be able to respond to verbal commands from construction workers, and it would be able to detect when a hazard in the road prevents it from continuing, plan a new route, and detour.

There isn't much incentive for Waymo or its competitors to create level 5 autonomous cars when there is a plenty large market in places where there is good weather and good signal for an employee to tell the car what to do if necessary.


Waymo cars don't have a driver, and they don't let a passenger take over driving. That's level 4: https://www.sae.org/binaries/content/assets/cm/content/blog/...


Perhaps it straddles the line between 3 and 4. I don't see much of a distinction between the passenger taking over, and a first responder taking over. Would Waymo put cars on the road that don't have pedals and a steering wheel if it were cheaper to do so? I don't think they would. I think they want a backup in case the car has to be driven by a human.


No, it's very clearly level 4.

Here's how you can tell so that you don't get confused as easily in future. Level 3 cars can require the occupant to drive the vehicle, "When the feature requests: You must drive the car". Waymo even offers service to people who cannot drive whether that's legally (they aren't licensed) or physically (blind people can use Waymo but can't drive a car). They are simply not required to take over driving.


They have remote operators that the car can request help from. That's level 3.


It's been reported that the remote operators can't drive the car, they can only guide the decision-making process.


> guide the decision-making process

What does it mean exactly? Because this can mean everything really. I mean you could guide decision making process by "now press break, now reverse and...", is it still driving?


I'll take the "for" side of the bet then. You game?


Waymo is not going to be working everywhere in the US in all conditions by 2030 (and it doesn't need to, either).

AFAIK, that's what L5 means.


The bet is to be available is one of the top 10 US cities.


It will be in major cities and that’s what the bet is.


The bet requires the cars to be SAE level 5 autonomous. You wouldn't notice the difference as a passenger between level 4 used in appropriate conditions, and level 5 car. But the bet does require the car to be level 5 nonetheless.


There's a big leap to go from geofenced Phoenix to NYC


It does not have to go to NYC at all. If Waymo can show profitability is some city it is good enough. If it gets love in one city is is good enough for the rest of the world to put a red carpet before them.


The bet doesn't seem to require service in NYC. It says:

  By "major cities" we mean any of the top 10 most populous cities in the United States of America.


It has literally been in San Fran for over a year where some parts are arguably harder than NYC.


I don't think NYC itself(terrain) is bad. It's dead simple. It's the driver and everyone else allowed anywhere near the street that is the problem.


That happens in San Fran too. Random parking on the street for deliveries, bicycles, construction, pedestrians milling about on the road etc.


Is anyone at Level 5, though? At all? In even a prototype?


What is level 5? Just widespread level 4? Because Waymo is definitely at level 4.


They're level 3. There are remote operators that the car can pull in if it needs help.


It's really interesting to see Carmack take such risky positions, such as betting FSD Level 5 by 2030 along with his all-in bet on VR. The VR thing didn't go nearly as well as he had hoped... and I suspect, like VR, the FSD thing will also not come to fruition due to it's grossly underestimated difficulty.


Carmack is not a poor man, so a $10k bet or two (for charity, mind you) is more of a friendly wager than a make/break gamble.

And he's clearly been optimistic most his life, and it's paid off a number of times.

And self-driving has progressed further and faster than I have thought it would, though I don't think we'll get Level 5 by 2030.


> Carmack is not a poor man, so a $10k bet or two (for charity, mind you) is more of a friendly wager than a make/break gamble.

The point was he seems to consistently grossly underestimate the difficulty of problems outside of his domain expertise.


He actually wrote the Netflix VR app for Oculus, and wrote a really detailed post on it: https://netflixtechblog.com/john-carmack-on-developing-the-n...

VR is not _at all_ outside his domain expertise.


The issue with VR is that it almost entirely doesn't require his type of expertise for it to succeed. It's almost entirely a hardware problem, in terms of cost, wearability/usability, quality, along with the other issues like nausea etc. None of which Carmack's vast expertise lends itself towards. Writing an app for Netflix isn't exactly using his skills.

We see this often with engineering types, many of us included. Vastly underestimating difficulty of challenging problems and naively believe they're simple to solve.


I would argue its mostly a hardware problem because lots of engineers, including Carmack, have already explored and solved a ton of the hard software problems e.g. https://web.archive.org/web/20140719085135/http://altdev.co/...


No one who correctly estimates the difficulty of writhing a game engine would ever start. Some things require optimism.


The two examples were self driving and VR.

> The point was he seems to consistently grossly underestimate the difficulty of problems outside of his domain expertise.

Carmack made 3d games and game engines. I don't think VR was outside his domain of expertise.

And for the self driving one, that bet is still on


Personally I think VR is outside the domain of expertise of 3D games and game engines. To think that VR is within that domain is exactly the error -- VR as an overall product requires different consumer behavior than buying a game for a PC you already have. It requires buying a new kind of device or buying into a whole new form of interaction. It requires remembering to use that device and think about what games you have for it. It requires engineers to ship high FPS and low lag to prevent nausea to a degree unheard of for PC gaming.

Facebook's VR plan seems to hinge on being "even more ubiquitous than mobile" <https://officechai.com/learn/mark-zuckerbergs-email-explaini...>, which is a level of constant use that 3D games don't really have.

I am not a VR or 3D game expert, but even I can list other key differences. I bet you can too if you try. I think those are reasons to think VR is not really the same domain. I am curious what you think.


While I mostly agree that the hardware and UX challenges are mostly out of his domain,

> It requires engineers to ship high FPS and low lag to prevent nausea to a degree unheard of for PC gaming.

This is very much Carmack's speciality, and why his focus was on making mobile VR happen [0]. There are very few people who can outperform him at that, and most of them worked for Oculus :-)

[0]: https://www.gamedeveloper.com/design/q-a-carmack-reveals-the... (this predates the Quest, but it should give you an idea of the problems he was tackling)


Fair enough. I think the UX challenges are tougher than FPS, and secondarily the hardware challenges are also tougher than FPS, so I probably should have omitted the FPS mention in my comment to keep things focused. :D


With the amount of time he's put into VR over the last decade I think it's fair to say he's as much an expert as anyone else at this point.


He became expert in the domain enough to decide to walk away...

He's now working on a startup in the AGI field, which will also probably go nowhere for him.

He gets to work on things that excite him - what a place to be in life. We can all envy that - but he's not very good at gauging problems and consistently underestimates their difficulty/time-to-market.


He also tried (is trying?) rockets.


Tried, (more or less) failed, and gave up (because it was costing too much, basically). Search for "Armadillo Aerospace".


I am a self driving skeptic but seems like a good bet for Carmack. I would expect Waymo to open up to consumers before 2030 in SF, even if more expensive and take way longer than a normal car. It could be little more than a tech demo and he would still win the bet.


Waymo's already open in SF and Phoenix, with LA and San Antonio to come, and it's only 2024.

There's a wait-list you have to join, but anyone can download the app and get in line.


Why is Waymo so slow to roll out?

Pick an optimal path for each metro, only operate on good weather days, roll out nationally.


The major companies that have tried to go faster than Waymo have been destroyed by safety incidents (Uber, Cruise). Waymo's caution may be well-warrented.


Destroyed seems a bit harsh. Tesla is moving forward.

FSD12 appears to be a major advancement. https://twitter.com/WholeMarsBlog/status/1750448446790922634


Because they're careful? Especially after the recent Cruise debacle - SF is relatively friendly to their efforts, and once they can point to a few years of problem-free operation in SF, other cities might allow them in, but trying to expand the service too early could prove a very costly mistake if a major issue is found and they get thrown out again...


>Why is Waymo so slow to roll out?

Ignoring the safety path, this other issue just scale. Manufacturing the cars would be a big problem. Getting a ton of them on the road is another. I wonder if all those sensors that the car uses could ramp up to a national level in a year.

I've been in the Waymo beta for a couple months now. At peak times, it's more expensive and slower than an Uber simply because there aren't enough cars on the road.


I doubt that fees are quite covering expenses yet. I expect they'll roll out quite a bit quicker once expansion decreases their losses rather than increasing it.


you've just eliminated a lot of useful use cases. Who needs a car/service that only drives a limited set of "optimal routes" in good weather?

Also, "optimal routes" tend to quickly become suboptimal due to traffic, road closures etc.


I was like wtf are they talking about I’ve been using self driving cars for half a year


John carmack's won the bet, 6 years early.


It's rather more than a tech demo already.

https://youtu.be/P6sw4EKegp4?si=Yafm5a-ufCGizumy


What does "everywhere in all conditions" mean for level 5? Across a flooded road? Through 8" of snow? Up "The Chute" in Sand Hollow?


> except in the case of natural disasters or emergencies

Whatever a human driver with a non-commerical license does on a day to day basis is a good measuring stick.


This is an oddly imprecise position to me. It's unclear whether an 8" snowfall is a natural disaster. In places where this is a regular event, it is not treated as a natural disaster. Some human drivers drive in it just fine, though most prefer not to.

What about a 3" snowfall? That is a more common event, and a much larger number of human drivers stay on the roads despite the snow.


The ability to safely and successfully get where you want to go would not be diminished by the fact it was an autonomous vehicle. 8" of snow in LA would be a disaster because they don't have the road maintenance infrastructure to handle that. It's not that people in LA are dramatic compared to people in Michigan.


> oddly imprecise position to me

Yes. It is intended to be so. Because being pedantic is now how you define these levels.

The ultimate goal is to run a large fleet of robo taxies wherever possible and Level 5 tech can do that everywhere on globe. One can get pedantic and ask if this would work in Antartica or on Dead Horse Bay or whether driving in sand dunes of Saudi Arabia is possible. But the folks who are trying to pedantic wont be happy with any definition either here.


The text of the official standard explicitly mentions floods, defining Level 5 to be when the software "can operate the vehicle on-road anywhere within its region of the world and under all road conditions in which a conventional vehicle can be reasonably operated by a typically skilled human driver"

It adds: "However, there may be conditions not manageable by a driver in which the ADS would also be unable to complete a given trip (e.g., white-out snow storm, flooded roads, glare ice, etc.) until or unless the adverse conditions clear. At the onset of such unmanageable conditions the ADS would perform the DDT fallback to achieve a minimal risk condition (e.g., by pulling over to the side of the road and waiting for the conditions to change)."


It means that it needs to be able to get the vehicle into a safe situation regardless of conditions. That can mean needing human help once it is there in some cases.

This doesn't mean that it needs to magically cross destroyed bridges, as the popular culture definition often implies.


Standard roads are probably the intent. Being pedantic is just moving goalposts.


The problem is that freezing rain caused Portland to come to a standstill, but in Canada, that’s just an annoying day with elevated rates of accidents.

In Texas, a light breeze takes out the entire state, whereas in Canada, most of our workforce is expected to at try to continue working after a 20 inch snow dump the night before.

I hope you can see how there’s reason to being pedantic.

If level 5 is all conditions human drivers do regularly for day to day activities, 2030 is an idiotic bet and anyone not living in a bubble could see that.


A lot of that doesn't have anything to do with cars or drivers, but is rather about city infrastructure.

Does the city own a fleet of snowplows? Does it keep massive stores of road salt available? Do schools and offices even have heating systems at all?

On the other hand, driving carefully on snow and ice is a skill in itself that has to be learned. And in snowy places I do know people who simply don't drive in some conditions because they're well aware they don't have those skills, even growing up there -- while others enjoy the challenge. So that part is a fair question of which humans we're talking about when we talk about human-level driving skill.


Regardless of city infrastructure, I've seen schools in North Carolina close after a light snow (like a few millimiters). In Stockholm this winter I regularly feel my car drifting when entering roundabouts, and a lot of roads have transformed from 2-lane streets into 1.5-lane streets.

> So that part is a fair question of which humans we're talking about when we talk about human-level driving skill.

These California-based companies assume everywhere is California, and still want to release those cars internationally.


It doesn't even have to be snow. Heavy thunderstorms can close schools too.


Did you typo "light freeze?" Texas does terribly with ice but it is regularly windy as hell here, among the windiest of all US states. Even when a tornado tore up three miles of North Dallas a few years ago, it didn't take out even the local traffic, let alone the whole state.


There's an additional aspect however: on days with e.g. freezing rain, human drivers might risk it, but companies like Waymo will probably decide to play it safe and not let their taxis run, because they know how much bad publicity an accident would be...


In that case, SAE Level 5 driving may have a place for “sub-division” by weather, the same way that CAT-III ILS offers a range of operation options based on visibility.


Do we know where officially Waymo is on the SAE J3016 scale ?

I used Waymo in San Francisco, and to my not-expert eye, it did perform very well already.


Waymo is level 4. SAE calls out “local driverless taxi” as an example of level 4.

https://www.sae.org/blog/sae-j3016-update


"No" seems easy to say right now, but technology is accelerating at a breakneck pace. If AGI can happen within a decade, I reckon 2030 self driving is not a crazy bet at all.


John Carmack pandering to buzzword communities to try to stay relevant?

Yeah. That sounds right.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: