The Google X moonshot factory is struggling to get products out the door
recode.netI have been wondering if Google hasn't already lost the self driving race to Tesla. Google still drives a few test vehicles around and has totalled some 2.7 million km:
https://en.m.wikipedia.org/wiki/Google_self-driving_car
Tesla has more than 100,000 model S driving around and in just 7 months their auto pilots have already surpassed the km driven by Google for the past 7 years by almost a factor 100 (roughly 200M km). That is real use in many places, situations and weathers, not just test drives in California.
I realise it's not apples to apples and that Google's cars may be more autonomous for now. But with the numbers stacked against it like that I doubt it will be long before Tesla's auto pilot is vastly superior.
As I see it, given the lack of a Google car, they will have to team up with a major car company to get enough cars out there. And that requires a more elegant hardware solution than what they currently put on the rooftops.
http://www.extremetech.com/extreme/231097-tesla-records-its-...
> I realise it's not apples to apples and that Google's cars may be more autonomous for now.
I think that is an understatement. This video shows why staying in your lane on a highway (what Tesla autopilot does) and dealing with city streets are two just completely different things: https://youtu.be/tiwVMrTLUWg?t=8m49s
I think Tesla has a ton of catching up to do to compete with where Google is right now.
Tesla's auto pilot isn't at all comparable to what Google are doing. It essentially the same as the adaptive cruse control (cruse control that adjusts to traffic in front) and lane assist (auto stearing to stay in lane) you can get from most other car manufactures as an option.
That's not to say that tesla isn't developing full automation it's just that what they sell now is just clever branding on something you can get elsewhere.
I believe that at some point in the future they will launch a joint-venture for a fully automated Uber competitor with Tesla handling designing and building the vehicles and Google providing cash, self-driving software and building the actual app.
Musk is literally too good friends with Sergey Brin and Larry Page for this not too happen.
http://motherboard.vice.com/read/elon-musk-and-larry-page-ha...
http://www.businessinsider.com/googles-secret-apartment-elon...
Data is the ammo that will win the war of self-driving car companies.
Tesla, with a large, update-able fleet in the field, has a much larger data-collection platform than Google. Even if they're not using it for full automation now, the collection itself is very valuable.
This is almost certainly why Uber is putting self-driving cars into the field now as well, even though they are clearly not ready for prime-time. They need those sensor packages to be driving around on real roads and in real traffic.
But in terms of collecting training data to develop the algos that will eventually be fully autonomous, Tesla is way out in front. That's what the OP meant: every Tesla is collecting information that gets sent back to HQ to train the neural nets (or whatever) that will eventually drive the cars themselves.
200M km of adaptive cruise control & lane assist on a highway is very different from 2.7M km of fully autonomous driving on city streets & highway.
Think of the data collection aspect of it, though. Tesla's driving AI will know what happens in the real world better than Google's driving AI does.
I think the general case—navigating unusual terrain, intersections, traffic patterns, changes in aggressiveness tolerance, construction, adapting to "normal" problems like needing to turn left illegally at a stoplight when the oncoming lane is continual—is almost entirely unrelated to where Tesla has an advantage. This entirely autonomous general case seems to be where Google is attempting to dominate.
How much data are each of Tesla's cars really collecting vs each of Google's cars? If the data is not useful, it won't matter how many miles they've driven. Do they have the same sensor capability and collection?
Imagine the difference between, say, trying to determine a user's video preferences based upon their hit statistics on IMDB vs direct viewing data from Netflix that includes everything they've watched, how long (1 minute vs the entire show), when, the order they've watched (eg, mood predictors), frequency, etc.
I'm not saying that this is the difference between Tesla and Google, just that quantity and quality don't necessarily equate.
Not really, because by comparison Tesla cars are relatively blind. Google's cars are loaded up with all kinds of sensors collecting as much data as they can. And they have the advantage of driving back to the plant to hoover off the data with as fat a pipe as they want.
Tesla is far more limited in both what it's collecting and how it's collecting it as it needs to go over a cell signal. For example they are using a camera for a hefty part of the autopilot system. They obviously can't stream every frame of that camera up to their servers for deeper analysis. Besides being infeasible bandwidth requirements it would be an insane privacy violation of the owner.
Does Tesla collect data when Autopilot is turned off? Otherwise I would be very curious to know how many miles of non-highway driving it has logged.
I don't know, but they do have a history of logging a lot of data. They've used some of the logged data to rebut high-profile critics in blog posts in the past, in cases of accidents or low-mileage claims.
Data is a commodity.
Tesla may have an order of magnitude more kilometers, but the quality of data that Google is collecting is (IMO) more than an order of magnitude superior per kilometer.
Tesla (currently) has one radar, one camera and 4 short distance ultrasonic sensors. Google has LIDAR plus a lot more.
Tesla's suite may or may not be sufficient for operation. But for training, good data is critical.
The requirement of what you need to compare with that data though, is important. Tesla's cars are training off less detailed data, yes. But that also means they're designed to work with less data available. Google's cars require $150,000 worth of sensor hardware. Will that price go down, sure? But Tesla's will be even cheaper.
Google could easily build their software model so that it only uses a subset of the available sensors. They can then use the full set of sensors to judge the effectiveness of that model.
Tesla can't do the inverse.
I think that in 10 years time, time of flight lidars will become ubiquitous, just as common as digital cameras are now. And this will usher a new revolution of interaction between reality and the digital world.
Teleportation, obviously.
.
.
.
Oh how sweet it would be
for it to be me
that's the one to reveal
in days before
the IPO of many of these
(and the main other 3)
that what the world
has been waiting for
is sitting with me--
with open arms--
desperately ready to be--
with you--
always.
.
.
.
Some say with wonder
other with haste
all that's to be said
but even in this case
where is it again
what makes it all
seem sudden
a change
leaving the world a part
tense at the seams
knowing even then
aside from the waves
crashing toward the bitter end
it was dealt
squarely
only pausing to be
as there it is again.
It's not just going to be about straight miles driven, it's about the technology being focused on and what's possible with it. Both Tesla and comma.ai are focusing primarily on what the car can see and do with it's own sensors... Google's cars are only functional on roads excessively mapped far above the normal Google Maps level. Google's existing strategy for self-driving cars isn't practical at a national level because of that extensive mapping requirement, and it possibly never will be. (Google's cars may have driven x number of million miles, but it's all the same very small number of roads.) And while prices drop as technology develops, bear in mind that in addition to all of that, Google's sensor platform is the most expensive out there.
Additionally, I've read some really interesting articles about research other car manufacturers have done. For instance, Google has never tested in bad weather, but Ford has been working on self-driving cars that work in snow. And while Google just assumes the humans are meant to be 'along for the ride', Volkswagon did some really good UI work, in terms of figuring out how to make the car's actions predictable, and hence, less scary. (Essentially, the car indicated to the driver what it was about to do before it executed a maneuver.)
Google is really good at capitalizing on their self-driving car project for marketing purposes, but it's extremely unlikely it'll ever be a market leader.
>> Google's existing strategy for self-driving cars isn't practical at a national level because of that extensive mapping requirement, and it possibly never will be.
>> Google's sensor platform is the most expensive out there.
>> Google has never tested in bad weather
What about a self-driving cars as a service ? they can be the first to start a very profitable service that is limited in area and in weather even thought the sensors are more expensive(and they can claim "we aren't cutting corners like everybody else!")
And that could be a great place to be in, strategically.
A few years ago I was 100% gung-ho about Robotaxis, and I went through a mild depression last winter when I gave way to mounting evidence that it's a really hard problem and the 'it's 30 years out' naysayers are probably correct. It just sucks being wrong.
Google's Koala cars are functional under only the most idyllic, constrained and carefully monitored conditions. There's a huge laundry list of unsolved, and unknown problems between what Google has demonstrated so far and where they need to be technology-wise to run a robust, reliable, profit generating service at the scale needed to cover their R&D.
The casual thought experimenter generally fails to recognize the frequency with which they utilize higher level reasoning when driving that's well beyond the limits of the current state of the art in AI. Nobody has the slightest idea of how to solve this, let alone dig into all the as-of-yet not understood logistical problems inherent in commercializing the technology, an unexplored realm rife with any number of unknown unknowns.
The real world is a very messy place. Unlike Google, Uber is eyeballs deep in the messiness of the real world, so they're probably better poised, though a lot can change in 5 or 10 years. The competitive playing field has been so dramatically altered in the past 2 or 3 years that the days when Google was the only company anyone took seriously feels like ancient history.
With regards to the sensors, my bet is that by the time AI's capacity to reason is where it needs to be, the sensors and software needed to see and interpret the dynamic driving environment will be dirt cheap. Probably all you'll need is cameras, their cost keeps going down and the state of the art in image processing is progressing and will continue to progress.
Sure, they could definitely do similarly to what Uber just announced, in a small scope area like the Bay Area suburbs. But that's not likely to be a high margin product, certainly not something Google would want to hang onto long term. It's far smarter to be selling software licenses or cloud access to millions of units built by other car manufacturers and operated by individuals or other companies at scale, that's where the money is.
But you also have to realize that you'll also highlight the weaknesses of the technology. People may not be able to easily specify to the car where they'd like to disembark. What if people want to be taken just outside the service area? Uber is including a human with their self-driving project for now, which doesn't save them (or you) any money.
Google learned from Glass that a small number of users and a lot of public attention and hype about a product can quickly eviscerate it. The technology was good, but people without hands-on experience misunderstood it, and a couple small incidents became national news. A small rollout can just as easily kill your project as kick it off.
Even if Google decides to sell access to other companies, at scale, they can still say: "our service only works in summer , in that list of expanding areas". And of course if they sell access on a per-trip-basis, it's almost as if they own the service themselves.
>> What if people want to be taken just outside the service area?
Google is currently trying to be the comparison search engine for people who want to order rides - via their Google Maps. If they sucsseed, they'll just fit you with the right service according the limitations of the self-driving car ,etc.
And regarding Google Glass - IDK. Even Uber is marketed on a city-by-city basis,
Tesla is either already mapping roads through the sensors on their cars, or is only one software update away from doing so. Their mapping data will be a lot fresher than Google's.
This is true but mostly irrelevant, because mapping data is the least valuable self-driving data. Google focuses on it because Google has it already. Google Maps is one of the things Google has that few else have, so it's obvious for them to build their technology on it... nobody else can copy them.
But if you have the most cars collecting the freshest data, you still have to bear in mind, the car collecting that data, is currently relying on the old data. Which means your cars can't trust the map data. And the reality is, with how much things in the real world change, your map data will never be trustworthy. You can have it, but you can't rely on it.
Which is a better self-driving car? The car that can look at a cloud-based 3D map of every object in the entire town, assuming it's current, and decide on a route? Or the car that can ping Google Maps, get told "turn left on Main, right on Washington" then entirely from it's live surroundings, decide how to drive?
Tesla cars don't have enough sensors to capture great data. It will require a lot more than a software update for then to record data on par with googles quality
They claim autopilot uses "cameras, radar, ultrasonic sensors and data."
I think even just the cameras alone would be valuable. It seems a Tesla has a better idea of what's going on around it than a human would, so these sensors should be at least close to sufficient for training an autonomous driving agent.
> Google's cars are only functional on roads excessively mapped far above the normal Google Maps level.
Since this is a core part of your argument, I'm gonna go ahead and [citation needed]
I've not heard anything of the sort and I've seen Google's cars on all sorts of roads. Plenty of their talks have been about the cars detecting anomalous situations and reacting appropriately, as well.
There's no reason to believe they've exceeded this limitation: http://www.slate.com/articles/technology/technology/2014/10/...
If Google is really aiming as high as it says it is, you would expect the rate of successes to be fairly low. A true "moonshot" success - one that creates a whole new industry - every decade would still be a phenomenal success. Maybe we just need patience?
Many people (myself included) would compare X in intent and setup to PARC from the 1970s. The latter managed the extraordinary feat of pushing out a major innovation on average about once every 2.5 years from 1970 to 2000. That's things like laser printing, OOP, Ethernet, etc. that went on to build industries or take existing ones in new directions.
X, on the other hand, has been around for over 6 years now and as far as I know, its only marginal success to date has been Glass. I don't think it's unreasonable to say that X's success rate so far has been lower than many people expected.
I dunno, given the public backlash, far from kickstarting a new industry, Glass killed face-mounted wearables stone dead. The word "glasshole" will probably stick around for longer, so there's that...
It may be true that it's harder now to make those kinds of leaps than it was in the 1970s though.
Why?
I mean, everything is obvious in retrospect. I'm sure there were some PARC guys bemoaning that all the fundamental discoveries were already made in the 1950s... Google today probably has bigger budgets than Xerox had then (tho' that's just a guess on my part)
A lot of discoveries get exponentially harder. Compare the LHC to experiments people were doing 100 years ago for an extreme example.
Evidence of this assertion?
It's a speculation, not an assertion. But it's a speculation many well-informed scientists have made in the last few years.
For example: http://blogs.wsj.com/ideas-market/2011/02/07/the-difficulty-...
The fruits might be hanging higher nowadays though.
> Many people (myself included) would compare X in intent and setup to PARC from the 1970s.
That was probably true originally. Lately, (especially post-Alphabet) it seems like the pressure is on them to produce commercial products, albeit ones that might be many years out - my impression is that researchers at PARC were not, realistically, under the same pressure (see: the many innovations that were never commercialized).
That makes sense, though - Google is actively in search of new business models, while Xerox enjoyed such enormous market share while PARC was well-funded that I doubt they felt that kind of pressure until later.
PARC may have pushed out innovative tech every 2.5 years but didn't it take much longer for most of those technologies to change the shape of or come to be foundational to tech?
Their biggest success is probably Google Brain.
Google brain was an acquisition, not a Google X project.
Google Brain was an internal development, while Google DeepMind was an acquisition. What the conceptual difference is supposed to be between the two is hard to say.
Yes you're right. How confusing!
This depends on the goal of the projects. If it's pure R&D that doesn't have a requirement to transition to commercial use, then you are absolutely correct. I don't think that's the case here. Some percentage of projects has to transition it's technology or research into commercial use, otherwise what is the point? Google is a publicly traded company, I wouldn't be a happy stockholder if I knew they were blowing $1 billion a year and with very little to show for it.
One thing did jump out at me in this article.
Mike Cassidy, who stepped down from Loon, ran the team “like a fire drill,” a former employee said.
I read this as, "Leadership likes to change its mind about direction and focus ... constantly". That's a recipe for disaster. Focus is key. You need to find the right direction as quick as possible, then execute.
How? Google is not built on fundamental research like Xerox was. It's essentially a consumer company. Yes search and ad tech at scale is hard engineering, but it's the same as the physics of light.
Or maybe you can't just put people in a room and say "invent stuff".
It may sound hackneyed, but I believe necessity is the mother of invention and it doesn't seem there is that kind of motivation here...it feels forced.
Yes - this is the VC model. Hitting 100-1000x results every once in a while, as opposed to lots of 3-10x results.
Yes, but eventually the VC needs to hit that 100x result. GoogleX has not. Heck, nothing's even really made it out the gate.
The self-driving car is the project that seems to have the biggest potential impact, but the string of high-profile departures over the last year or two is worrisome.
Also, the leader wearing Rollerblades to meetings seems far too much like a parody of "quirky tech visionary" for my comfort. But maybe I'm more traditional/close-minded than most here.
Yes. VCs who don't have the big return don't get to have a 2nd fund. :-) I'm bullish on self-driving cars. Even if their cars don't make it, I think the AI behind it will.
The title feels like an oxymoron. I would assume the reason for spinning out a "moonshot" division is that if some projects fail, at least they are not part of the main division. Perhaps some of these projects are not the type that can ship an MVP?
One of the problem might be the incentive/motivations are not match between Google and people who works in those moonshot projects.
* Even if the projects get 10, 100 millions $ in revenue / profit, it is meaningless compare to billions in profit from search. The folks in those projects are not likely to benefit significantly from it.
* The smart folks probably know it. If they join the project (self driving car), some of motives are to learn as much as possible using Google's resources, name, connections and set it up for their own next venture.
It seems obvious that it would be a struggle for them to ship products - they are "moonshots" after all.
Reading this article makes me wonder if it was a good idea to put all of them into one division though. Even when you go in knowing a project is a long-shot it can be demoralizing when it fails. I can't imagine how hard it would be to work in a whole division of mostly failed projects.
Isn't a venture capital firm (even a good one) a business investing in mostly failed projects?
The difference is that those projects are spread out across different companies, as opposed to a single division in a single company.
I think it would be so fun to work there, even with all this political stuff happening. I guess it's sort of like Google's internal "Y Combinator", where they try out a whole bunch of startup ideas and have practically unlimited funding.
I think the fact that a project would receive unlimited funding is a negative, not a positive.
Instead of getting lean startups that have to move quickly because of resource scarcity you have bloated startups that feel no pressure to move quickly because of the unlimited resources they are receiving.
Is not as bad as article suggests: https://www.solveforx.com/graduated/
It just operates on hype curve. GoogleX takes projects from "technological trigger" to "peak of inflated expectations": https://en.wikipedia.org/wiki/Hype_cycle
"Graduation" doesn't mean success. It seems to just mean "it moved out of X", even though usually that's just to somewhere else in Google it makes no money and sells no products.
- Verify/Life Sciences is failing pretty hard and has failed to develop any actual products. https://www.statnews.com/2016/03/28/google-life-sciences-exo...
- Glass is an absolute failure. I own one, I loved mine. It's still a failure.
- Gcam went to... Google Research? It went from moonshot R&D to normal R&D. That's not success.
- Google Brain... went to Research as well. And really, it is likely just a rebranded version of whatever DeepMind was already doing when Google bought them.
- Project Tango went from Google's moonshot R&D to Motorola's R&D (ATAP) which Google kept ownership of. That's also not success.
Android Wear, Flux, and Project Insight (Indoor Maps) probably all count as actual successes for X, but that's about it.
> They say the issues at X aren't technical hurdles, but a combination of red tape and knotty internal politics
Unsurprising. The valley (and tech in general) is forgetting it's power doesn't come from management and politics.
This article frames it in part as due to the Alphabet reshuffle, but hasn't it always been this way? Has X been decreasing in effectiveness recently, or were they just always there to be Sergey's batcave?
This feels like the type of operation better suited for a tier-one educational institution like Stanford or MIT.
In a academic context, there is no subtext of creating products. Yeah, churning out papers is as contrived...but I still feel like the timelines in academia are more generous.
Google talks a good press event, highlighting their embrace of failure, their desire to take on crazy moonshots etc...but its all in the context of quarterly reports. Something has to give.
> The Google X moonshot factory is struggling to get products out the door
Isn't the whole idea of a "moonshot", in this context, a project that has both a long window before any payoff, and a high risk of failure. So isn't this a very much "water is wet" story?
Meh.
If you're familiar with the internals of Google (or have ever worked there), then you wouldn't even be slightly surprised by the content of this article or the ever present stream of product failures from Google. In fact, if you're aware, then you know things are likely to remain this way until something is done about all the BS present internally.
You see the "difficult to work-- FOR" and almost (or all out) sociopathic leaders being outed to some extent in the last year or so: commonplace. That is, Google likes to position itself as being above such stupidity, but if you listen to what they say more carefully, you see this is their default/go-to strategy. That is, they explicitly seek out such personalities (similar to many VCs), as they believe (based on "data") that it's what's more likely to lead to success.
Also, while it would be easy to believe Google is all about the "moon shots" they like touting/hyping (especially given how much money they are dumping in that direction), if you look at who they are putting in key positions, and how everything is setup, you immediately realize that (regardless of what they are doing) it will all likely go nowhere.
Something is very strange about that place. It's like they have no brain. It's like they are just outright stupid. Which I suppose is hard to say (and have believed), especially after years of hype (and supporting anecdotes) about its exceptionally talented pool of employees.
Essentially, it often seems like a sea of INTJs who like to parade around as though they know what to do with data (and are above bias and feeding into their own BS, because they are "data-driven"), but that at the end of the day are just going based on whim/gut, one which is more self-centered and out of touch than "in tune with" and reflective of the world at large (or where it's going).
I suppose if you spin around in your own shht enough, and surround yourself with more of-- yourself-- then eventually, you'll fall into line believing your own BS and that you must be right.
I remember when I was at Google for a short time in 2010. The place was a source of endless annoyance and irritation. The field was wide open, and it was all there for the taking, but then they just consistently and continually kept making the dumbest decisions. And they'd defend those decisions as though they were God almighty and immune to being wrong. It was the greatest consistent stream of stupidity I had ever seen. And by the looks of it, nothing has really changed. It's just been shht, then more shht, then more BS trying to explain away the shht, as though shht isn't what it is.
It's not that they "fail often due to releasing more and sooner" or "see something beyond the field of view of many;" it's that they just plain failed and it was most likely due to stupidity/credentials (you heard what I said) being heard over repeated statements of what made more sense (or of what would be more likely to go somewhere). Also, that failure that looks like a half-assed piece of crap likely was likely 2 years (or more) in the making, rather than the 6-8 weeks it seems went into it.
It's just shht every time, and as soon as you step into realizing it, you'll see that it's always just more shht from them. Their only successes (even in their "main" business) have come from the competition "falling off" (F'ing themselves over), rather than from them releasing things that are worthwhile or better.
The place was extremely infuriating to me, and I couldn't wait to leave. I was in silent shock the majority of the time I was there, and it seems that even though 6+ years have gone by, not much has changed!
I'll say it outright and in plain English: the "almighty" Google-- the almighty enterprise of innovation-- the almighty force for pushing the web/world forward-- is completely full of shht! They couldn't put out an innovative (or even just quality) product to save their lives! If anything surfaces from them that's not more garbage, then it was likely from an acquisition. And even then, it seems they are F'ing even those avenues up more frequently as time goes by.
Google Photos: acquisition!
Google DeepMind: acquisition!
Google ATAP: acquisition!
Etc.
They've got nothing.
why does it matter if the moonshot factory is getting products out the door or not?
Because everyone always has to be shipping and disrupting! Big Business! Product! Sales!
Who's to blame?!
Attention, Decision, Interest, Action. AIDA.
We're adding a little something to this month's sales contest. As you all know, first prize is a Cadillac Eldorado. Anybody want to see second prize? (second prize is a set of steak knives.)
Third prize is you're fired.
These are The New Leads. These are the Glengarry leads. And to you they're gold. And you don't get them. Why? Because to give them to you would be throwing them away. They're for closers.
I realize that this comment is tongue-in-cheek, but the reality is that Alphabet is a corporation, not a research university. Corporations by definition need to ship and make money.
It doesn't, but if Google keeps feeding moonshot stories to the press, they should expect the press to circle back and see if there is any follow-through.
If Google X is really just a nerd-PR exercise, let's call it that.