I think the days of Google Moonshots are coming to a close. In general Google hasn't been very good at shooting for the moon. They've lost their leadership and vision and all that's left is a giant advertising behemoth.
Did any of their moonshots really hit the mark or even come close?
I think the main problem is they lost the edge on execution. Traditional "vision" is not strictly needed for moonshot projects; the vision for moonshots is provided by a small team working on their idea. But should the technology start panning out, the company needs to focus aggressively on finding a commercial home for this technology.
And this is where I think google (and many big companies) fail. Technologically, Loons worked. I did not believe it when I read about those, but looking at the trajectories in public ADS-B data (those balloons self-advertise) it is clear that they move as designed, doing 2D positioning using only vertical control and can stay afloat for months. Wow! Mission accomplished. But commercially, nothing happened.
I do not know if it was because business types were not sold on it and did not build a commercial case; did the technical team simply live in their own bubble not talking to the outside; something else? But a 10-year "moonshot" project that pans out technologically and fails commercially indicates that the right hand ("internal VC") does not know or care what the left hand (techies) is doing, which is the problem at the company. My 2c.
Selling stolen technology you tried to patent from under the company you suckered into believing into buyout, and then having those patents stripped by the Judge might have something to do with it.
Well, their target market is very poor people, which is not normally a great segment to expect to be early adopters of expensive new technoogies. Also they're competing against Starlink, which is basically the same concept but in space, which is probably impossible given that LEO is well understood and Musk can launch tons of long-lived satellites quite cheap.
All good points. I am just saying that moonshot projects are supposed to be "very high risk, high reward" projects -- unlikely to work, but if they do would provide big benefits. Your points argue that this is a "high risk, no reward" project, which is not the way to run a 10-year moonshot.
Occasionally starting moonshots as "high risk, uncertain reward" is OK for a short time, but if there is no clear reward on success after a year or two it might be continued as a PR or a charity, but not a moonshot. My 2c.
It's kinda weird to think of brain as an X project. DistBelief was developed out in the open in the main source code repo by people you think of as googlers, like Jeff. So what makes it "X"?
My own personal image of X was that if you had to drive to some old Air Force base to see it, it was X. Since brain was software, it doesn’t fit the model.
Waymo may / will earn them some money, either they sell it to a crazy big investor, or they license the technology out to a company backed by a crazy big investor.
alright I'm gonna ask the naive question: how exactly does Brain make money? is it offered as an independent product? or is it's incremental contribution to the rest of Google that high? what even is Brain, to Google devs?
To me it's an umbrella for all of the projects that fell out of an initial realization (by Jeff Dean and his circle) that Google's infrastructure was large enough to put into practice a class of older ML ideas that had spent years being thought infeasible. They developed the means of training huge models, then useful models that resulted from that, and follow-on project like TPUs.
Is Brain doing great? From a few things people have said, it sounded like Brain was struggling so Google bought Deep Mind instead. All of the really cutting edge AI stuff has come out of Deep Mind, not Brain.
Just to clarify here, Deep Mind in no way replaces Brain. The AI space, like any other space, isn't a single unified landscape. There are areas where Deep Mind is a true standard setter--particularly deep reinforcement learning (AlphaGo, for example)--and areas where Brain has been one of the standard setters--natural language processing springs to mind, where Brain has historically pushed the field forward with projects like word2vec, BERT, and the Transformer architecture generally.
Brain is also a much more general organization. They do things like develop Tensorflow, which is one of two most popular ML frameworks in the world (and until recently was far and away the most popular), and TPUs, which are ML-dedicated ASICs that have a huge impact on training and inference.
Both orgs are world class and historically important, and they certainly overlap, but they aren't replacements for one another.
I strongly disagree. Without a shadow of a doubt, the most successful/important AI model of recent times is the Transformer (from Brain). As another example, the fastest accelerators, TPUs, are from Brain. If you only look at "amount of output" and use # publications at e.g. NeurIPS as proxy, Brain vastly outperforms Deepmind. Deepmind really exceeds at flashy and showy PR releases like AlphaStar, not so much at laying the groundwork for AI breakthroughs. (Though of course AlphaFold or AlphaGo were great).
Brain is not a pure research group. It builds infrastructures/platforms for machine learning related tasks and its research part is relatively small compared to the entire organization.
That's the total cost over the decade+ the program was in existence. Divide by 10 and it becomes roughly 10% of Google's revenue.
So it wouldn't be insubstantial, but they could fund a similar effort if they really wanted to. Shareholders would have a thing or two to say about that though, which is why Elon Musk is keeping SpaceX private.
That's not what the poster meant. It wasn't counting rockets, but the fact that the cost was spread out over multiple years.
Yes, if it were possible to go from zero to moon in one year then that would be all of Google revenue. But spread over 10 years that's just 10% per year.
The parent comment's point is that the money spent on the space program was likely very highly weighted to the first years building that first rocket. So dividing by 10 because it was roughly a decade is probably underestimating the budget needed to get to the moon.
At the peak of Apollo program spending in 1966, Dreier says, NASA accounted for roughly 4.4% of the federal budget — 6.6% of discretionary spending — more than the Manhattan Project that developed the first atomic bomb.
We don't have the technology to get a person to Mars or to land them when they get there. We probably have the technology to build a spaceship that they could die in on the way there - but that's the closest that we realistically have.
There's much greater radiation exposure away from Earth's magnetic field. I believe this is still an unsolved problem. Proposed solutions for shielding are heavy and expensive.
You just have to keep the lights on and not die of cancer for six months. The radiation isn't that bad.
Edit: The specific number is under 2 millisievert per day. And "One sievert carries with it a 5.5% chance of eventually developing fatal cancer based on the linear no-threshold model."
> We’re currently at a few billion to put a whole horde of them on Mars I believe.
Uh... don't count your Mars shots before they're hatched. Even getting robot probes to Mars is very very hard. Only about half of Mars missions have been successful. In fact, since the fall of the Soviet Union only one organisation has succeeded in putting working probes on the Martian surface: NASA/JPL.
I don't think it's impossible that SpaceX will get there, but certainly not soon and not for "a few billion". If they succeed at all, it will require drawing on years on investment and expertise by NASA.
I don’t think anyone was alleging that google should be literally landing people on the moon, so invoking the economics of moon landings specifically is hardly relevant.
> Seems like if you really want to make it possible, you can.
More like "if the hard groundwork has been done 60 years ago". It's not as if SpaceX had to quite literally invent the orbital rocket.
So after 60 years of manned spaceflight, 9 manned lunar missions, and 30 years of flying reusable crewed vehicles with a space station that has been permanently crewed since 1998 someone better do it MUCH more cheaply.
The culprit isn't just "space is hard" - it's decades of cost-plus contracts and heavy government involvement that kept costs high and results low.
Not really; a moonshot is a huge investment to advance the state of the art in something where we know it's possible. It's not a wild shot in the dark.
Well if we're nitpicking, that was the payload, not the Saturn V booster...
But unfortunate though that was, it didn't represent a fundamental uncertainty in the achievability of the goal - simply a fatal engineering mistake, which was immediately rectified.
The point is, how do you know your path is to the Saturn V and not the Soviet N-1? In the case of Google Moonshots, nobody even know if they're going to use a rocket.
I'm sure the N-1 would have worked as well, if the project had been funded (in both time and money) properly. The Soviet failure to reach the moon was a project management failure, rather than the failure of a blind risk to pay off.
I think it's generous to say it's in its final stages. They haven't even left Phoenix have they? And even in Phoenix, it's still largely a beta project in a few predefined areas.
It certainly seems like Google's most likely moon shot to take off. But considering only a few thousand people have used it, it's not exactly a big success.
Waymo, like any and all software, will never be “complete” - but I assume you’re referring to Waymo being ready for it’s first “real” release somehow - but we still don’t know how exactly Waymo will be integrated into cars or which automakers will go with them.
That point of mine now has me wondering what Waymo will look like in 10-20 years’ time when autonomous vehicles become both ubiquitous and un-cool because it’s nothing new anymore (just like how Facebook is un-cool today). Will Waymo’s maintainers be adding new features that benefit the consumer, or will there be pressure to “monetise” it from every angle? And to what extent will brand-conscious automakers go to hide the fact they’re using Waymo, e.g. will they insist on rebranding it - not just for themselves, but between their own sub-brands (e.g. Ford vs Lincoln, or VW and Audi)?
As an outsider, the reason why Waymo is way ahead is clear. Extremely talented team, access to Google infrastructure and expertise, deep pockets and a culture of prioritizing safety over everything.
This ignores the cost of sending data wirelessly over cellular networks so remote operators can see the vehicle, the cost of maintaining the software and hardware on the vehicle itself, the cost of maintaining the vehicle fleet, the cost of creating and operating a consumer-facing support system (!!! for Google), the cost of dealing with the enormous amounts of data those vehicles create (even for Google this is not trivial), the cost of maintaining the special HD maps necessary for the vehicles, not to mention the sunk cost of billions spent developing it.
That's not to say you shouldn't dream big (it's a moonshot after all). But there are plenty of reasons to think it won't be viable even if they can solve the technical challenges, and that much still isn't even clear yet.
The phrase "value of full autonomy" is excluding fleet costs on purpose. It's about drivers.
Customer support is also going to be far far fewer than one person per vehicle.
The enormous amounts of data? If it's not valuable they can just discard it! Having sensors attached to something doesn't obligate you to store it forever.
Everyone knows the software/configuration costs are immense here. But that's the lion's share of the difficulty, and there's no reason to act like minor hurdles are bigger than they are.
> The enormous amounts of data? If it's not valuable they can just discard it!
I'm curious what the legal requirements for this will be. But I imagine they'd want to hold on to data of driving scenarios for at least a month, in case they get accused of wrongdoing by other drivers. If they had no data to back up their case with all those sensors, it would look awfully suspicious and essentially one witness against nobody - so they'd have to hold onto the data for however long the legal teams deem is okay.
This may sound crazy, but it's already happening at the scale of testing with just a few dozen cars.
They can be accused of a crash even if there isn't a crash though. And if they have no data to back up their side of the story, why would any court believe them? In other words - people can just randomly accuse them of hit-and-runs and they'd have nothing to say otherwise.
I don't see how it's any different from accusing random people of a hit and run.
Though enough data to disprove a hit and run wouldn't actually take up very much space. Medium-resolution camera views and some acceleration data? Sure, pop a single SSD in there and it'll hold more than a month's logs.
Most importantly, their fleet could be made obsolete by personal ownership of assisted driving vehicles. So by any of car manufacturers... (Better version of Tesla autopilot for instance.)
Barring some radical breakthrough, progress usually follow an asymptote. Waymo has reached their asymptote after perfecting their autonomous driving for over 10 years, and is still not confident enough to launch it even in their spherical-cow geofence.
The real world is full of edge cases - cameras fogging up, proximity sensors confused by ice buildup, thick fog limiting visibility, sun low on the horizon blinding the cameras, badly marked construction sites, black ice, line markers under snow, potholes, slush, contradictory traffic signs, deep puddles, worn out ruts, suicidal wildlife, road debris, etc. etc.
The full self driving product that people expect isn’t a beta product that’s only available in a special region. To be accessible as running water, decades away is probably accurate estimation.
In the mean time we can all drive around in our self driving cars while we wait for them to become a reality? It doesn't make any sense. Either they exist, albeit in a limited form, or they don't. Does it get you from a to b in most normal daily use cases? If yes, then it counts.
I would apply the 90/90 rule here [0]. It looks ready, but I don't think their safety is up to scratch yet and I'm almost certain the cars aren't safe in adverse conditions.
Depends how you define things but their work with Deepmind ('Our long term aim is to solve intelligence') is interesting. I'm glad we get breakthroughs in protein folding rather than just ads.
They bought Deepmind though, they didn’t dream it up as a moonshot. Plus you can argue that keeping the talent at Deepmind accessible to Googlers contributes to the development of Google’s internal models, which does mostly contribute to their bottom line.
they missed cloud not because they were too busy. They thought of their internal cloud infrastructure as a competitive advantage to be kept to themselves. Very similar to the Intel's thinking of their own fabs which has brought Intel to the current failure.
Wrt. moonshots - one need to have huge political capital inside the company to carry the high risk of large investment. I.e. one has to be Jobs or Musk. Even Yang failed. How Nadella succeeded it is kind of a miracle. It looks like there are no people of that caliber at Google these days, and without it you have a CFO touting "investing for long term" which in CFO mind obviously means investment like buying real estate and not a technology moonshot.
I think so. I used to be a big fan of Google, they made a great search engine, gmail was ahead of the competitors. Even the moonshots were pretty cool/ interesting. Not sure when things went really sideways.
Did any of their moonshots really hit the mark or even come close?