CEO of Waymo John Krafcik is leaving
blog.waymo.com
blog.waymo.com
Google demands a lot of its "other bets" companies. Some might say it asks too much. Mostly it seems to me that they want a 'moonshot' company that has the original profitability of search advertising, full stop. And while it would be great for them if they found it, there is a lot to be said for having a bunch of businesses that just make anywhere from a few million to a billion dollars a year in revenue.
I don't know if the attitude has changed since I was there but people creating $20M/year business revenue streams were not considered "successful" back in the day. I found that somewhat self defeating.
[1] https://gist.github.com/ChuckM/ff5fc8c800c7fe9160483b68ec45a...
It’s too bad CEOs don’t write their PR like you have done. Like when they drive companies into the ground, lay everyone off, and then they exit with millions.
You accurately describe the Other Bets climate when I left X in 2015, and I hear it's even worse these days from friends who recently left when Loon shut down.
In the case that the successful business is Ads/Search... it's tough to compete. I'd imagine even within Ads/Search there is a resource curse between AdX/AdWords/Youtube and every other ad business.
The kind of people drawn to "chief" positions are not interested in shrinking their empires.
And yes, splitting is part of a solution. I think discipline is another part.
I thought the entire purpose of "other bets" was to pursue ideas that have the potential to become $XXB/year revenue streams. So of course they want 'moonshot' companies.
For example, I doubt the AirBNB founders early-on and/or 5-10 years ago even imagined they'd have multiple billions of revenue.
That's the thing though, it isn't a waste of time.
One can hear very similar logic from people with investments. They will say "The stock market is returning X% / year and is way better than those bonds with only 3%/year. Investing in bonds is a waste of time."
They say that because they have yet to internalize the value of having a diversified their investments. Not everything goes up all the time.
It only makes sense for Google if the Search Ads business were to never ever lose its profitability. And yet, it is losing its profitability. As a result, Google has to aggressively cut back on expenses, remove projects, end of life products, etc as that cash cow slowly deflates.
Consider then the alternative where there are 10, 20, even 30 business lines within Google generating $10 - $30M of profit each. 30 businesses at $30M is only $900M, less than 5% of their revenue, but those businesses are SOLID and provide a supply of management talent, consistency, and some bucks to keep the lights on elsewhere.
That is diversification of execution risk. It works the same way investment diversification works, it adds other, less high margin, businesses to the portfolio that are all revenue positive.
A company like Google can use those businesses to experiment with alternate user support models, management schemes, policies, and communications. All of that helps the "main" company to mature in its thinking about how to be a business. Sadly, executives who have never had any experience other than one wildly successful business tend to think exactly like you do, "Why would I waste time on piddling little products when I've got more cash than I know what to do with being pumped out by my main business?"
Short answer: "Things change."
If something happened to their core business, it's unlikely that those tiny projects (<1% of revenues combined) would save them. The more likely thing is that many of those small projects fail over time and it becomes death by a thousand cuts.
What you said is mostly correct and is exactly what they are doing. The only problem is that at the scale of a trillion dollar company, they need 10, 20, even 30 business lines generating $XB - $XXB of revenue each.
Their reasoning was that the resource usage to net profit numbers wasn't "good enough." The comparison was always search advertising.
Thanks!
Even with the weight of the Google brand, creating new businesses is HARD.
It'd be roughly 1,000X easier to squeeze an extra $900M in revenue out of search than it would be to incubate 30 new mid-size companies.
Instead of going on an insane boondoggle where your brand image is trashed by creating literally thousands of failed companies (the only way you're going to end up with 30 successful ones over the $30M hurdle rate)...why wouldn't Google just buy those 30 companies? They have enough cash on hand to buy 99.9% of Silicon Valley startups outright.
And even then, would the 30 companies they buy grow faster than their core business...or even the S&P 500 at 9% a year? Because otherwise they might as well just dump that money in existing products or return it to shareholders.
If Google buys a bunch of businesses that grow slower, then their valuation and stock price drops dramatically. If investors wanted to own a random sampling of 100 mid-size companies, they'd buy the appropriate index fund! They buy Google because they want a concentrated bet, not an index fund.
This is nowhere near as easy or simple as you think it is.
For all of Google's PR efforts around moonshots and only hiring "the best talent," a vast majority of their revenue still comes from only one product they incubated on their own, the google search engine. The next biggest bucket comes from external acquisitions (DoubleClick, YouTube, Android).
I think the fact that Google hasn't incubated any big success in the last decade is a good thing! It leaves more room for others to take their place. Why would we want one company to dominate everything forever?
First, you have all of the infrastructure for a business already in place. Even when I was there it was straight forward to get resources allocated to a project.
Next, you have billions of "seats" in that users the world over are already using Google branded services every day, have a reasonably good impression on the brand, and typically a low barrier to "trying out something new."
Finally, you have a tremendous amount of smart, experienced, people you can call on for advice for free! I know a lot of people have left but when I was there it wasn't uncommon to have the argument about a thing settled by the person who invented the thing weighing in on the argument. While it became clear to me that Google and I were not compatible long term, it was still intoxicating to walk around bounce ideas of some really really sharp people who could trim months off idea research and development.
As a result of that, starting businesses within Google was akin to scoring in baseball where you got to start at third base (or maybe second base if it was a longer stretch). That is significantly easier than starting from scratch.
They could not even succeed with one of their messengers, there are so many of them out there. What makes you think they will succeed with 30 business lines?
There is as much reason to shut down a profitable company as to throw away gold. That's literally what you are doing, and no one is too big to throw away money.
But these companies were founded inside Alphabet, and grew within Alphabet's infrastructure. This is how they gained a decent chunk of their initial bootstrapping advantage, as described upthread by Chuck. To decouple them out into something suitable for sale is extra time and cost, and beancounters never like time and cost.
So, ironically, part of what makes these attempts possible at all is part of what ends up killing them.
The challenge is that market cap isn't durable. Specifically businesses that just provide net revenue, even if it is only a "small" amount relative to the overall company, are helpful in mitigating shocks to the market.
So it may be intoxicating that the stock going up 10% adds a few billion dollars to your net worth, but it can also go down 10 or 20% and that wealth goes poof![1]
[1] The reality is that paper wealth is not worth anything, only after it has been diversified can you really count it as part of your net worth. Again, this is my lived experience :-).
Over that long time, I'm sure there are benefits for the executives' personal stock(don't CEO's diversify every few years?) , and for Google in general(for ex. Better talent).
In fact, all my emails and communications are written in this style all the time with everyone - what shall I do now wise one? (serious question)
You wanna start a band or whatever. Maybe we can start like an Elks club thing?
I've job hopped looking for a place where people don't engage in the Kabuki but apparently all the worlds a stage. Maybe I'll make one of those lil ships in a bottle this year.
Given my experience and my interest in systems I have become something of a 'journeyman' engineer in the traditional sense. I like helping folks and I like figuring things out. I charge money for that because it weeds out the freeloaders but I don't charge a lot of money. My terms are that if its not working out, we'll part friends, but we will part. As a result I've managed to meet some manager types whose only tool in their manager toolbox is "I'll fire you if you don't do that." (which is a pretty weak toolbox).
It is an exceptional thing to be where you are, at the age you are. How you choose to spend the years can be very satisfying and impactful, or a source of anger and regret.
It's the "I stumbled onto this giant pile of cash, and if you don't stumble onto one too, you obviously aren't stumbling hard enough" attitude of management.
Well, when Larry page wrote “we also like that it means alpha-bet”[1], a lot of readers in tech probably mistook him at his word. When it comes to investing, “alpha” means winner. As in somebody else loses. As in what gets optimized is _the investor_, not necessarily the community. The analogy to investment return above benchmark is a trademark Googley delusion—- it’s rationalizing a position using math or science even when the position (applauding investing culture) is exclusionary or even toxic to a substantial segment of the population (e.g. the 99% who either don’t invest or who form the ETF benchmark). Larry isn’t talking about excellence, he’s talking about dominating, and slipping that through with a cute abc.xyz domain name.
While it’s very interesting to hear these stories about Google, perhaps the lesson learned is just how blind people are to Google’s greed when technical adventures are dangled in front of them.
Waymo has had a ton of notable departures:
* Urmson made a good amount of money and left for Aurora.
* The founders of Nuro cashed in $40m each and left for their own thing.
* Levandowski made nearly a quarter billion and took off.
* Drago worked on Streetview and more Google-centric things, then went to Zoox to make about $100k (lol), then returned to Waymo.
* Now the era of Krafcik is coming to an end.
The perhaps unique thing about the self-driving problem is that all of the above individuals made tons of money without having delivered equitable value to end-users. At least not today. When Google was bleeding headcount to Facebook, both companies were making bank. It's surreal to see people minted with money for life and yet deliver so little value. That sort of arbitrage usually only happens on Wall Street.
I think it's worth reflecting on the era of Krafcik as a general success-- he was brought in to do hard work and he generally did a good job. But by no means did he (nor any of his predecessors) solve the "team" problem. Krafcik himself couldn't stick to the team he helped shape, nor the userbase he helped grow, at least nt long enough to actually deliver widespread value on the scale of his own compensation.
It also seems like you're not really also applying in the fact self driving cars are approved in how many places? And it takes approximately 5 years to get a standard car to market, getting a self driving car that has been tested in just 10 states, it's unreasonable to expect the value to end users you're expecting to see here. Getting the car to market seems to be the actual problem and that seems more of a legal issue of no one wanting to say yes on the major scale. This is a very slow moving industry, electric cars are basically a solved problem, yet we'll be seeing people buying combustion engine cars for decades.
How long has this project been going on for? 12-years? Seeing management change over that amount of time seems normal. It would be notable if we were seeing 1 or 2 people a month leave for 6 months.
That said other parts of the ecosystem are having incredible bursts of innovation and new products: B2B SaaS, Cloud providers, Space (e.g. Starlink), Bio-Tech (e.g. Moderna and mRNA technology), Semi-conductors/Chip companies (TSMC + Nvidia + many many others).
If you're wondering why VCs are eager to back new companies with small revenue but fast growth - it's because interest rates are very low, and the returns to winning are very high.
I wonder to what extent the leaders of these self-driving car projects understand that their efforts to create the technology have almost no chance of success. How many of them understand this but carry on due to either wild optimism, or to just make money, or on the principle of "it's not up to me to decide whether to try, someone else is paying me to try"?
The degree to which self-driving car hype seized even the minds of many otherwise smart people in the last decade has been strange for me to see.
Because no one can solve the tech problem, no one wants to stick around for the point where the failure is obvious.
Especially, for a while, self-driving cars have been at the level that impressive demos are possible but actual deployment isn't. So a fine career move is sticking around for long enough to create a demo and then leaving and blaming the failure on your successor.
Reason the tech can't deployed is that while 99% of driving problems involve just a complex, adaptive system, the 1% or .01% remaining involve "understanding what's happening", a far higher bar, one that requires a system well beyond what exists today.
This may well be true given how fragile a construct most deep learning systems are.
But there are still substantially different deployment issues for different systems. Google can deploy a question-answerer or a search-by-image-description system and lowered accuracy in practice doesn't really do much damage. Google can't deploy a self-driving car with low safety in the same way.
We won't have self driving cars that replace 30% or more of driving on existing roads and highways by 2040.
The problem is too hard and everyone's drunk the ML kool-aid. (Just like everyone is drinking the NFT kool-aid right now.)
Hype, unrealistic dreams, and spin.
We'll be on Mars before we have generally available self-driving cars that do not require humans.
When cars were introduced, people kept getting hit in the street. "Jaywalking" as a crime was created to shift the blame away from the technology (cars) to the victims (pedestrians).
For self-driving vehicles, it suffices that we outlaw the edge cases, and enforce it to make them insignificant. The Uber ATC crash? Blame the jaywalker. Kid chasing a ball in the street? Blame the parents. Treat self-driving cars the same way that we treat trains: get out of the way, or you've no one to blame but yourself.
What that leaves practically is preventing crashes with the static environment, self-driving cars, and "normal" road users (i.e. ones that can't be blamed if a collision occurs). This is much easier than the plastic bags and animals and cyclists and jaywalkers and wacky drivers that form most of the remaining unhandled 1%.
The only country that can pull this off, IMO, is China. The mix of authoritarianism plus technical prowess allows the missing 1% of cases to be bulldozed in the name of progress.
Both seem unlikely.
I’m not saying Waymo is dead, I’m saying creating a sustainable work culture for self-driving is really hard and Waymo evidently don’t have a solution for that. I actually thought Krafcik would be a balancer, but seems things didn’t work. That said, Waymo has made some other major HR mistakes... they hired (and soon fired) Tawni Cranz, one of the most toxic HR heads you could find...
This must be a joke? The multi million-dollar exit with zero actual earnings is an almost uniquely tech phenomenon. There are Wall Street-ers that earned high pay for generating high earnings that many years later were found to be value-destructive, but at least at the time they were paid, they were cash-generating. That Wall Street jobs paid so much was because comp was structurally a function of cash generated.
It was also the reason they filed patents. Like the answering machine patent: to prevent it from existing (commercially) for as long as possible.
I used to live in Mountain View and at one point I saw a few Waymo vans (this was 3-4 years ago when they were more active) speeding on El Camino (just barely, and being manually driven at the time, but even though I work for a competitor I was concerned if an accident happened it'd be a bad publicity event; i.e., Waymo vehicles should always be extra, extra careful, etc.); and it just made me wonder what was going on. So I connected with him on LinkedIn; and then ended up reaching out to some QA folks I believe.
Anyway, John was super responsive and I always appreciated that. From what I could see (just one anecdote), he was a CEO who dived deep into customer concerns. There's a lot to be said for that.
My two cents is it's a very delicate space. One bad event and the whole industry can be set back. So it's an area (despite extremely cutting-edge tech; deep learning, etc.) that has to roll out gradually.
Besides him not leaving a clear successor[1][2], I think his communication is disingenuous. The kind of management-speak that aims to hide poor decision-making behind a facade of cheer and care.
People are scared of "might makes right", but I've found "nice makes right" to be far more insidious and pernicious in our society.
Maybe just retrofitting off the shelf cars made sense if they thought they needed hundreds of thousands of cars ASAP, but it turns out they would have had plenty of time to develop the car and even build factories for it while waiting for the technology to mature. And they didn't really need to partner with any manufacturers just to retrofit existing cars anyway.
The counter point could be that the technology is there, but the regulation is not.
Why must everything be retrofitted versus a complete system overhaul? The initial costs may be immense but the return on investment will be far greater.
Just my two cents.
The actual car companies also outsource lot of stuff to other manufactures e.g lights, wires etc. I would imagine understanding and maintaining those relations woulds also be a hassle for Google instead its just better a car companies does all that.
Building a car is so hard. And Google mostly sucks at hardware manufacturing. There are good reasons why their market share of Chromebooks, Android phones, etc. is vanishingly small.
The koala cars may not have been successful but that was many years ago, years that could have been spent making something better.
The idea that they believe it will never be more then Level 2 is factually wrong. You might believe that, but they don't.
Are you following the car industry? Car startups get crushed regularly, and almost non are successful. Pretty much all of them bleed money like crazy.
Tesla went almost bankrupt over and over again.
I must be the only one who thinks full self driving will not happen before year 2100. Is there an example of anything that's completely automated that's dangerous? I cannot think of a single thing. At best things are partially automated and need humans still.
The thing is, if you have "partial" self driving, you might as well just drive yourself IMHO. The consequence of being wrong with a car is literally your life. I know people like comparing this with airplanes, trains, or the elevator, but cars in the USA is just too chaotic an environment. It's nothing like any of those.
The probability of crashing with a plane is extremely low, even factoring in the higher risk of takeoffs and landings. The entire process is coordinated with Air Traffic Control.
Since planes are so safe already and pilots are relatively inexpensive, there's no strong financial incentive to fully automate planes.
But the main reason we won't see planes without pilots is political and related to trust. I think for the same reasons we are less likely to see truly driverless cars.
> https://en.wikipedia.org/wiki/List_of_automated_train_system...
But i stand corrected. Guided tracks is hardly a good comparison for this.
That really isn't true. Pilots, in particular, have huge challenges in keeping attention up between take off and landing. Likewise for trains, it turns out if you don't require humans to do anything but watch...it is much harder than if they were actually doing the flying/driving themselves.
The airforce is moving to unpiloted drones, though there is someone at a center to intervene as needed. It is only a matter of time before they decide that cargo can be moved in this way (think: fedex).
Nope. Me as well. I think self driving is AGI-complete, and we are as far from AGI now as when we were living in caves, banging rocks together.
Edit: downvoting is fine but please at least add to the discussion by stating why you disagree. I’m speaking only from my personal experience working in factory automation and realize there’s probably a lot of differing experience
Yes, but probably to a lesser extent than driving a car. Everything has sensors that are meant to identify a risk to a human and stop whatever operation poses that risk.
E.g., there may be robots picking thousands of pounds of parts and driving them around a facility but they stop if they sense a human a their path. Same for welding, stamping, etc. All those operations can injure or kill but unlike vehicles they can rely on the mitigation of “stop and wait”. It’s a much easier problem than self-driving
The carpool lanes will be converted to rail tracks. You can pay extra for a rail-compatible car if you want and sleep during your commute.
As implausible as it sounds this is still way more likely than driverless wheeled vehicles becoming both safe and useful.
A US statistic: In 2010, there were an estimated 5,419,000 crashes, 30,296 deadly, killing 32,999, and injuring 2,239,000. It is hard for me to imagine a scenario where automated driving is as unsafe as that.
We will truly look back and think, "Did we really trust others to operate multi ton metal machinery at exceedingly high speeds at each other day after day?"
They already are. I'm talking about full self driving. A car without a steering wheel. Cruise control (arguably the first car automation) has existed since the 1900s. Potentially earlier depending on how you define it.
What if our roads were built to communicate with the cars? What if there were only AVs on the street and they communicated with each other? What if every bicycle, school backpack, shopping cart did so too, identifying itself in the process?
I think the problem of self-driving is harder right now in the transitional period in which they have to coexist with human drivers and essentially trying to "emulate" how a human driver gathers information (not really, but you get my meaning)
I'm not sure it won't take decades, but I feel like there will be a magical few-months period in which we go from "can it be done?" to "oh, ok,we got it"
[1] shoot, I lost the reference. But Cruise apparently assumes "use whatever is available until it works, we'll make it cost-effective later"
That sounds like a nazi-state dream - easily track every backpack, car, bicycle, etc...
no thank you
It's clearer and clearer that self contained 'autopilot' units are going to keep failing in scenarios involving unpredictable people & situations. They don't have enough data and never will (without AGI, anyway). So we need to spoon feed them that data by tagging everything & everyone around them.
I think this is just the usual goal post moving, if a machine does it then we decide it wasn't really dangerous for a machine to do it.
Garmin Autoland will take a plane which is otherwise working and is in the sky and put it back on the ground because the pilot has a problem (e.g. your plane nut husband finally has another stroke† while taking you and the kids cross country, you've never learned to fly but you just press the button like you were taught and the plane will tell you to remain calm while it figures out where to land and gets itself back down). It isn't cheap (the system needs to control flight surfaces, engines, radios, brakes, more or less everything on a plane) and of course it can't fix a broken plane, but a random non-pilot is going to do a much worse job and "land an aeroplane" seems like it counts as dangerous to me.
> The thing is, if you have "partial" self driving, you might as well just drive yourself IMHO.
If you live in the relatively small service zone for Waymo, you just get in the car and it drives you elsewhere in the zone, like a taxi, except the car is driving. Waymo doesn't want you "partially" driving anything, it doesn't want you near the wheel or pedals, it would prefer if you watch a video or read a book.
† In some countries private pilots have to be fit, but the US has gradually decided that eh, it's your plane, your risk, the main obstacle today for medically unfit private pilots is getting insurance.
I strongly disagree. Most machines are not capable of killing you because it makes a mistake.
Do you have an example of something humans did that could kill them that was automated completely that still can kill them, but generally does not anymore? An elevator is one example, but in the case of an elevator its a very bounded problem to solve, and even still hundreds die for unnecessary elevator deaths yearly.
Not to mention an elevator is not an entropic environment. A full self driving car would have to be able to deal with ice suddenly falling on the ground, people crashing around them, etc.
Not nearly as complex a problem as self driving cars, but I'd still rather not get hit by a train
Anyone who does that has either never traveled in an airplane (even just as a passenger), or just never observed and thought about the vastly different levels of interaction planes have with their environment.
> Do you have an example of something humans did that could kill them that was automated completely that still can kill them, but generally does not anymore?
And I pointed out that the parent comment had already provided an example. Now you are arguing with me that the example given, about Garmin Autoland, isn't comparable to self-driving... Which isn't at all the example you asked for in the first place. Maybe focus and read a little more, write a little less. You'll be more easily understood.
My understanding from looking at research is that airplanes by experienced pilots rarely killed people... and they... still don't even when (partially) automated? There doesn't appear to be definitive evidence that autopilot is safer [2] as many of the innovations in safety have been conflated with procedural improvements.
In fact there are studies [1] that suggest autopilot is actually making flights unsafe as pilots are becoming complacent and airplanes do not flight themselves completely from point to point including taxiing.
In other words, no, plane automation is not an example as it does not satisfy the criteria of "generally does not" compared to the baseline of no automation.
[1] https://journals.sagepub.com/doi/abs/10.1177/001872088502700...
[2]https://www.eurocontrol.int/sites/default/files/publication/...
Try not to be so condescending, sheesh.
> The Garmin Autoland is not equivalent to full self driving?
This has nothing to do with your argument criteria of:
> Do you have an example of something humans did that could kill them that was automated completely that still can kill them, but generally does not anymore?
> Try not to be so condescending, sheesh.
I didn't mean that to be condescending out of turn, but if you are going to engage in an argument by posting a strong disagreement and an opinion:
> I strongly disagree. Most machines are not capable of killing you because it makes a mistake.
You cannot change the criteria for an example that you asked for simply because you don't feel like arguing that the example did or did not meet your criteria, nor can you do so simply because you didn't bother to read it carefully.
So no, I'm not being condescending out of turn, you really needed some correction on that. Sheesh.
Elevator deaths are extremely rare in industrialized countries (a couple dozen per year in the USA), and most are of people working on the elevator (e.g. accidentally falling down the elevator shaft), not passengers riding in an elevator. I think construction elevators are also quite a bit less safe than ordinary passenger elevators.
Typical elevators are one of the safest forms of transportation, substantially safer than stairs or ladders.
I've seen a lot of goalpost moving in the other direction on this one, to be honest - Self Driving Cars started as "never have an accident again!" and have slowly migrated to "look, you suck at driving, the robot sucks less, so stop bitching when the robot runs over grandma"
Indeed. For reasonable values of "land" and "on the ground." It's better than a non-trained pilot, but it also relies heavily on the concept of "emergency" in aviation, in which everything else is put on hold and everyone else gets out of the way.
It has basically nothing to do with driving a car.
First, the sky is large - "Big Sky Theory" is the term for it, and it generally means that you can fly around however and wherever you want without hitting anyone. Yes, there are controlled airspace segments, various levels for flying around, but the reality is that if you just ignored all of that and flew a plane from A to B, you wouldn't hit anyone (almost always). It breaks down a bit around airports, and there have been the occasional "plane A not talking to anyone and plane B not talking to anyone colliding" events, but they are exceedingly rare.
Second, large airports are going to be (almost by definition) controlled airspace, with ground response crews. And there are common airbands for radio communication. The Garmin system relies on these things.
If you hit the "Oh Crap!" button, yes. It will control an otherwise operational plane down to landing at some airport (I believe it prefers controlled airspace and emergency services, though I don't know details), and it will clear the road in front of it by setting the appropriate emergency transponder code and broadcasting prerecorded messages on the appropriate frequency that basically amount to "Get out of my way, I'm coming in for a landing at this airport on this runway." Which, for an incapacitated pilot, is absolutely the right thing to do. I'm fairly sure it won't clear the active runway, though. That's a problem for the humans on the ground.
But - literally everyone else in the sky will get out of their way, and ATC will ensure that. If there's an A380 on approach and this system is triggered near the airport, if ATC needs the A380 to get out of the way, they'll tell them to go around and go hold somewhere until the Cirrus is dealt with.
It's really a very, very different class of problem than self driving cars.
There are no roads, no obstacles, usually not even other planes constraining your plane for any autopilot situation. With autolanding, "other planes" might start becoming an issue, but it's still a vastly different level from the momentary and immediate coordination necessary between cars on public roadways.
Autopilot is simple and a solved problem. Self driving is not either.
Garmin autoland seems like a perfect example of something dangerous that's completely automated.
What are you going to say when they go to SF next? Just another Jurassic Park ride?
Maybe the common denominator is just cars. The regulators are fully absent - they care about the shape of your headlights, but do they care that SUVs have them mounted at the height of mirrors for every other car? Do they care that cars have absurdly overpowered engines that have no imaginable use? Do they care that trucks have fronts so high it's impossible to see a kid walking on a crosswalk and when you hit it, it's hit head-height and thrown under the wheels?
None of these things would ever fly on a train or airplane.
Monorail trains (like at the airport)
Combine Harvesters
Airplane Autopilot/autoland
Elevators
And resumption of operation after an emergency stop happens only after a human check and intentional operator acknowledge.
So you're technically right, but the circumstances are vastly different. Nobody will buy a car that does a full emergency brake whenever a leaf falls off a tree. The environment is so much more difficult on the street than in a factory.
HOV and Express lanes make this easy. They have well established markings, limited access, and traffic goes in one direction. So industry and regulators would work together to solve standardization of markings for travel and entry and exiting.
Then you take it to the rest of the interstate system and then to limited access highways. Eventually you get down to the neighborhoods and city driving.
I see real promise in Tesla's beta system which has a few thousand drivers. What it tracks is far more advanced than what my current software does in my 3. They finally went to a persistent model and far better labeling. Now where my car shows cones around a parked vehicle they show the vehicle as well. They show all parked and stationary objects. Something the current software does not always do if at all.
Just within the US the lane markings, signage, and even signaling, is not consistent from state to state. Worse the rules for what is an acceptable lane is not consistent either. So the first step there is to standardize it all across the nation.
As a M3 owner you get to see the feature set growing over time, but the zero-to-one moment is when that FSD is finally rolled out in 9.x
I can't wait. It's going to be amazing.
On the subject of danger. Literally everything we do is dangerous to some degree. Everything. Self-driving does not need to be perfect, but only better than human drivers. We're likely already at that point with Waymo vehicles.
On the subject of automation. Humanity is on track to have the computational power to perform whole brain emulation decades before 2100. Even if we don't solve the general AI problem through other pathways, we will solve it through this path. Ethical problems aside, once this is achieved, everything will be automatable at a human level of competence.
Is this actually true? I haven't heard of this. You're saying we will have an AI as good as a grown educated adult before 2100? I can't believe this at all - do you have a citation?
The human brain works, therefore this technique will work, it's just a matter of having enough computing power.
I will re-iterate though, we are pursuing various alternative pathways to artificial intelligence, and modern machine learning has already demonstrated super-human performance within constrained domains. It's my personal belief that we will achieve human-level general intelligence long before WBE becomes practical.
I don't believe there is anything guaranteed about "whole [human] brain emulation" in this century.
[1] This isn't quite true, because we estimate there are around 10^22 to 10^24 stars in the universe, while estimations for number of synapses range from 10^14 to 10^15.
Let that sink in. We know it isn’t impossible to run a general intelligence computer on 20W because every single one of us is living proof that it is possible. There is no reason why something man made shouldn’t be able to do the same thing. How to do that is of course a different matter, but it isn’t speculative that it can be done, it’s a direct consequence of our own existence.
It's not impossible that some sort of thing would happen, but anyone who tells you that it can't possibly fail to be achieved in the next 80 years is lying, either to your or to themselves.
There is a certain # of calculations/s necessary to meaningfully emulate a neuron. We currently have no idea what level of fidelity is required, but we can make guesses for each of the levels of fidelity that might be required.
The estimates that assume the highest (and likely unnecessary) level of fidelity currently top out at 2100.
For the record, I never claimed that failure was impossible, I used the phrase "on track". It's totally possible that there might be some surprise quantum weirdness going on that would be intractable to emulate with classical computers, but we've found no evidence of that to date.
But 1) that is an estimate of human brain computing power, based on number of neurons and possible connections. We have no idea if that is really a valid unit of computing "power." Do brain only compute things via voltage thresholds across synapses? Then yes, maybe this is an accurate estimate. But there are a whole lot of subcellular signal cascades happening at molecular levels we can't even begin to count and we have no idea whether or not those are also computing something accessible to the rest of the brain. 2) Matching the computing power of a brain doesn't mean you can emulate it. Emulating something requires knowing the target architecture and software. It is possible we can figure out exactly what a brain is doing to an accurate enough level to emulate it by the end of the century, but we certainly don't know how to do that right now, even if we had the computing power. Remote imaging techniques are not nearly good enough (because again, so many of the processes are subcellular) and you can't open up a brain without destroying it. This is actually a much broader problem in biology and we have figured out ways to partially dissect some animals while still keeping them alive long enough to figure out how a system works in vivo, but doing that with a brain is not something we have ever come remotely close to doing and there isn't really any ethical way to even think of how we can try to figure it out.
I don't disagree with your take and I'm a self driving car proponent, but I'm worried about what process we take to get there.
One thing I've taken away from the pandemic is that people seem to have no problem imposing their tolerance for risk on others. Seems like we are on a path to play this dynamic out again in how self-driving cars come to market unless that safety profile is really well controlled and understandable.
Even if at a population-level self-driving is slightly safer statistically than person-driving, there are enough edge cases to give me pause right now, and at the individual-level it may raise my risk either as a pedestrian or driver and certainly changes what is predictable behavior [1].
[1]: https://arstechnica.com/cars/2021/04/why-its-so-hard-to-prov...
Consider Google's (okay, Alphabet's) standards for success though.
My guess is that a metro area with less than 0.5% of the US population does not qualify as a successful product. Maybe as a successful small beta test.
Also, given Arizona's overly cozy relationship with self-driving [1], I would not necessarily trust a program too much on the basis that it's operating there.
-------------------------------------------------------
[1] https://www.theguardian.com/technology/2018/mar/28/uber-ariz...
It's not clear that member of the general public can actually sign up for it. Has anyone done that and taken a driverless ride in Phoenix? Happy to get more updated info / be corrected.
(And if this seems like a unreasonable amount of suspicion, I invite others to go back and read their press releases from the last 3 or 4 years, and tell me if that gives you an accurate picture of where the service is today)
The problem is this statement is written like a technocrat wrote it rather than someone who makes public policy. From an engineering perspective, it’s true that it would only need to be better than a human driver. To be implemented though, it needs approval in the public sphere and not just engineers. This presents a very real publicity problem.
You are correct that everything in life contains risk. Risk is defined as severity x probability. While the severity may be the same, I think humans judge the probability very different between human and autonomous drivers.
I think it’s rooted in the need for humans to understand what’s under the hood (no pun intended) to trust the decision making capability. We already have this with human drivers through the tool of empathy evolved over millions of years. We can reasonably assume we know what humans will do. (Incidentally it’s also why witnessing someone with mental illness puts us on edge). We have no such ability to decipher an autonomous car, especially for the layman. So this distorts the uncertainty in the risk assessment and any accident can disproportionately cause our assumption of risk to elevate.
I think there's another issue here, too. It relates to the observation that folks have a fear of flying, but not really a fear of driving, despite driving having about a 1 in 98 chance of killing you over your lifetime, while for aircraft it's more like 1 in 7178 (2008 data from USA Today...hopefully directionally correct). It's been discussed in research extensively[0], and seems to relate to whether or not the person feels they have a sense of personal control. I think your point is well-taken that the technological risks and the publicity challenge are independent barriers to the widespread adoption of AVs.
[0]: https://oxfordre.com/communication/view/10.1093/acrefore/978...
Flying an airplane. They put pilots in there to make people comfortable, but if need be, a modern passenger plane can fly itself, including take off and landing.
Driving a subway. BART in San Francisco launched as self driving in the 60s. It freaked people out so much that they put a human in front. But if you sit up there you'll see the human doesn't do much. They press a button to close the doors, but that could easily be automated.
Also, I took a ride in a self driving car in SF a few years ago. It had a safety driver, but he didn't do anything (he grabbed the wheel once, but then they looked at the data and found the car was about to do the right thing anyway). It handled things like cars going the wrong way, double parked trucks, trash bags in the street, jaywalkers, etc.
Eg, I wouldn't consider the stuff Tesla is doing as self driving at all. I have significant worry because to reduces attention while still requiring it for safety.
If you're talking "Hail robocab, don't touch a steering wheel, arrive at your destination", then Waymo is already there: https://blog.waymo.com/2020/10/waymo-is-opening-its-fully-dr...
You might mean "does everything everyone uses a vehicle for today", however I feel that's perhaps moving the goalposts to unattainable. Eg, replacing nascar racers with self driving cars would defeat the point of nascar, so that's won't happen. (though it might be an interesting side event.)
Between the two extremes of public pilot in limited area and handling every situation, there's a very large space for tremendous value. Just serving urban areas would significantly reduce the need for vehicles. There are many people who only use public transit, or could if they had more economical ways to handle sporadic trips such as grocery shopping and visiting friends across town.
Here's the interview: https://open.spotify.com/episode/0IuwH7eTZ3TQBfU8XsMaRr?si=6...
We know (from introspection) how trained skills work. Everyone knows the feeling when that bunch of neurons which you trained to semi-autonomously do something for you (e.g. switch gears, write letters, type words, observing traffic, staying in an imaginary lane etc.) raises their hand so that Daddy AGI (our conscious) can come have a look at that weird situation it doesn't know how to handle.
Tesla can't do that, either, and literally no one on this planet has the faintest idea how to do anything like it.
Tesla is focusing on data because they believe they can machine learn through the problem.
Waymo has a lot of data, but is also focusing on better sensors, augmenting them with maps, and ML perception feeding into more traditional programing doing the actual control and decision making.
It's kind of like Tesla and Waymo are both learning basket ball. Tesla believes that if it just plays a lot of games of basket ball, that is the best strategy. Waymo think that playing some games is good, but it should also do things like practice free throws and weight train.
Also, a few misconceptions:
1. Waymo currently only has public service in Arizona, AFAIK. However they are testing in San Fran, Seattle area, and Michigan. Easiest first doesn't mean later is impossible.
2. pre-mapping is NOT a blocker of automated service. Consider how much effort goes into making each section of road. Driving a car by a few times is a trivial cost in comparison. This is further shown by the fact that Google has been doing street view for well over a decade. Also I'd like to see Tesla demonstrate they can drive with the road covered in snow and still know where they are without a map.
3. CEO departures can be a red flag. They can also be a completely benign changeover, or even green flag that the board identified others would perform better in the CEO position. The leaving announcements are always fluff about spending time with family, so it's really hard to tell what's going on from the outside. If this is followed by more leadership leaving Waymo, then I'd be concerned.
The metric to beat is "is self driving on average substantially safer than human driving?". Once you beat that metric, it would be insane not to allow self driving. This won't take 100 years.
Brief summary of his jobs:
- NUMMI (GM/Toyota), 1984-1986
- MIT, 1986-1990
- Ford, 1990-2004
- Hyundai America, 2004-2013, left as president/CEO
- TrueCar, 2014-2015
- Google/Waymo, 2015-2021"The company said Friday that it is promoting its chief technology and operating officers, Dmitri Dolgov and Tekedra Mawakana, to lead a decade-old effort to make self-driving cars a reality. They will share the title of co-chief executive... Mr. Dolgov is one of the founders of Google’s self-driving car project. He joined the program when it began in 2009 and led the development of Waymo’s autonomous system, known as Waymo Driver. He studied physics and math at the Moscow Institute of Physics and Technology before earning a doctorate in computer science from the University of Michigan. As chief operating officer, Ms. Mawakana has led the effort to commercialize Waymo’s self-driving system. She has a law degree from Columbia University and previously worked at other tech companies such as eBay Inc. and Yahoo."
https://www.wsj.com/articles/waymo-ceo-john-krafcik-is-leavi...
My guess it that traditionally the chief operating officer would have been promoted and the chief technology officer would stay as the chief technology officer--but that there was a significant risk of the chief technology officer leaving if he wasn't promoted. So you end up with co-CEOs.
I can’t help but think people conflated what the problem statement of self driving really is. We’ve made huge strides in perception in the last two decades but self-driving is a much more complex problem than just accurately perceiving the world sound us.
There may be many limitations, but self-driving in sunny flat suburban areas would already cover a decent portion of the American market. Sure they might not be ready for chaotic, busy city centres or regions with harsh weather conditions, but that's letting perfect be enemy of the good - those living in areas suitable for self-driving cars would certainly appreciate them, and who's to say there can't be a gradual retooling of infrastructure to accommodate them as they expand to more and more areas?
"Eliminating the safety driver is an important step toward making Waymo's service profitable. But it may not be enough on its own because Waymo says the cars still have remote overseers.
These Waymo staffers never steer the vehicles directly, but they do send high-level instructions to help vehicles get out of tricky situations. For example, a Waymo spokeswoman told me, "if a Waymo vehicle detects that a road ahead may be closed due to construction, it can pull over and request a second set of eyes from our fleet response specialists." The fleet response specialist can then confirm that the road is closed and instruct the vehicle to take another route."
https://arstechnica.com/cars/2020/10/waymo-finally-launches-...
They say that they don't do one very specific thing, and they give one very specific example of something they do. Leaving it open to interpretation.
I'll remain very skeptical about their no safety driver claims.
We'll probably have to wait until the first fatality or serious injury to know in certain. The lawsuits and police/NTSB investigations will expose a lot of the inner workings.
https://youtu.be/D1sZnbORfAE?t=830
Stuck in a parking lot, with an unexpected massive obstacle (Christmas tree display) blocking the "road". "RAD MONITORING" shows up on passenger screen. Objects get labelled and vehicle starts making progress.
It seems to be happening a lot less in recent videos.
It is nice that they have deployed to the public, but I doubt anyone capable of driving themselves will use it any time soon if it can't even make left turns.
If they actually launched in Phoenix that would be encouraging. Instead they're stuck in the suburbs.
1. This is a looooong term bet, so you need revenue NOW to sustain your company until you get to full self driving.
2. The self driving car should be a practical product, to be purchased and operated by actual common people who drive cars today. So you need to avoid $10k LIDAR spinners on top of the car.
3. Self driving is not going to be a binary thing. It will be a gradual increasing capability. You need to let people use it in every intermediate stage and feed telemetry back to you so you can incorporate it in your development.
4. Self driving cannot rely on extremely detailed mapping of all public roads that exist on the planet. The system will have to be smart enough to figure out what to do, without specialised external help.
If you don't follow these, be ready for a long haul (10-20 years) without any revenue, because that's what it is going to take. And google does not seem to be the long-haul type of operation that this needs, judging from https://killedbygoogle.com/.
Waymo is just one big cope party filled with HBS grads who can't fathom that they didn't succeed at something.
I think SDCs (self driving cars) are great, and will be available in some capacity in the very near future, BUT I think we are at least one more paradigm change away from the dream of SDC.
Take for instance liability. It seems quite likely that manufacturers will want to disclaim liability for SDCs. It is easy to see the roots of that in current driver assist features. They are very careful to call out in the documentation (if not in the naming and presentation) that the human driving is responsible for safety.
If this liability was not an issue, SDCs in hands-off, eyes-off, brains-off mode would be available today, right now.
So how will the paradigm of liability change? Or will it?
Why though? They could just accept the liability, charge an insurance premium slightly lower than legally mandated liability insurance for human drivers and then self-insure (and potentially re-insure against bigger risks).
This would not only solve the liability problem, but also allow them to make a nice profit because the car presumably gets into at-fault accidents much less often than the human drivers.
But now that everyone assumes we're in for the long haul (Aurora says 2025 for L4 in some cities), it doesn't make sense for him to want to stay.
I suppose that is a nice way of saying wait out my non-compete....
https://en.wikipedia.org/wiki/Anthony_Levandowski#Criminal_c...
Humans seem to be perfectly capable of doing it though.