I Was Wrong

On Jan 7, 1991 Josh Storrs Hall made this offer to me on the Nanotech email list:

I hereby offer Robin Hanson (only) 2-to-1 odds on the following statement:
“There will, by 1 January 2010, exist a robotic system capable of the cleaning an ordinary house (by which I mean the same job my current cleaning service does, namely vacuum, dust, and scrub the bathroom fixtures). This system will not employ any direct copy of any individual human brain. Furthermore, the copying of a living human brain, neuron for neuron, synapse for synapse, into any synthetic computing medium, successfully operating afterwards and meeting objective criteria for the continuity of personality, consciousness, and memory, will not have been done by that date.”
Since I am not a bookie, this is a private offer for Robin only, and is only good for $100 to his $50. –JoSH

At the time I replied that my estimate for the chance of this was in the range 1/5 to 4/5, so we didn’t disagree. But looking back I think I was mistaken – I could and should have known better, and accepted this bet.

I’ve posted on how AI researchers with twenty years of experience tend to see slow progress over that time, which suggests continued future slow progress. Back in ’91 I’d had only seven years of AI experience, and should have thought to ask more senior researchers for their opinions. But like most younger folks, I was more interested in hanging out and chatting with other young folks. While this might sometimes be a good strategy for finding friends, mates, and same-level career allies, it can be a poor strategy for learning the truth. Today I mostly hear rapid AI progress forecasts from young folks who haven’t bothered to ask older folks, or who don’t think those old folks know much relevant.

I’d guess we are still at least two decades away from a situation where over half of US households use robots do to over half of the house cleaning (weighted by time saved) that people do today.

GD Star Rating
Tagged as: , , ,
Trackback URL:
  • Oliver Seiler

    We’ve had Roombas in our household for a number of years now. Some observations:
    * They would clearly benefit from having a map
    * They require a lot of regular maintenance to keep the functioning
    * My wife has the tendency to micromanage the Roomba, largely defeating the purpose of having it

    These days I think sweeping is ultimately faster and (for my wife) less time consuming.

    I like the way that bet was phrased; makes me remember the optimism of the early ’90s around nanotech and the like.


    Are you certain this is not 20/20 hindsight? It may still have been a good prediction with the knowledge you had at the time, of course you had to have consulted others if you were to widely publicize your prediction as truth, but you didn’t publicize it widely as the truth, did you? Did you know back then that older AI researchers predicted more conservatively? Do these more experienced researchers perhaps mistake an exponential curve for a linear process (it’s the relative progress over the previous 20 years that counts when you make a prediction for the next 20 years)? Do we really know that the reason there aren’t yet the cleaning bots you described in 1991 is of a technological nature and not for example a perfect storm of legal wars over patents, economic crises, etc…?

    It’s commendable to admit it when you’re wrong, but on the other hand 20/20 hindsight is a form of cognitive bias.

  • Daran

    My first thought on extrapolating the requirements to 2020 was that I can see how the sensors and AI will be sufficiently developed for cleaning tasks, but the limiting factor for wide spread use will be the (robustness of the) mechanical system. 3D printing to the rescue?

    • VV

      > but the limiting factor for wide spread use will be the (robustness of the) mechanical system. 3D printing to the rescue?

      What? Why?

  • consider

    Another 20 years seems much too far away Apparently Hans Moravec thought so too in the 1990s… He seemed to think that this will happen around 2015 with robot servants by 2030:

    Robots will demonstrate world modeling: a general understanding of objects and what they are for, and of living things and how to interact with them…Simulations will allow robots to practice and perfect new tasks before attempting them. Robot servants will be able to “read” the moods of the people around them.

  • TheBrett

    You might see them used for cleaning larger commercial buildings first, ones that might otherwise require a fair amount of labor and time to keep clean. It would probably be cheaper to have a host of Clean Bots to do the basic cleaning work, and then keep a single janitor on staff to deal with the weird stuff like clogged toilets and leaking fixtures.

    Home cleaning seems like a difficult market to really crack unless the robot is a very capable cleaner and very cheap. You’re competing with the home-owner/rentier spending 45 minutes a week sweeping the floor with a $5 broom, or vacuuming the carpet with a $120 vacuum that last for years – not a lot of difficult labor is being saved. And richer households can afford to hire a cleaning service.

    • IMASBA

      Most people who have a maid don’t have an exclusive contract with that maid, so a more fair comparison would be with a cleaning robot that can be rented for a few hours per week. That would definitely make it several several times cheaper to use cleaning robots, of course that may still not be cheap enough in 2020, but somebody had to make this point.

      • TheBrett

        It depends on what the cleaning robot can do, and how cheap it is in comparison to a maid service. The cleaning services out right now are pretty cheap on a per-room basis, at least where I live.

  • jhertzli

    This might be a sign that economics is getting ready to climb the science hierarchy.

  • Pingback: Editorial Intelligence()

  • Mark Bahner

    In January 2010, microprocessor costs were about $5 per gigaflop. A cleaning service could reasonably be assumed to require a human brain, or about 5 petaflops. So the cost of the microprocessor alone would have been $5 million in January 2010. Given that, it’s not surprising that such a robot didn’t exist.
    It’s also logical that the first examples of such robots will be in the form of a cleaning service robot. That is, a robot that is rented like Molly Maid, rather than a robot that is owned. It doesn’t make sense for anyone but the richest people to own a robot that would operate for 1-2 hours per week. It would make much more sense to rent. Rentals could even occur late at night (e.g. 11 PM to midnight) at reduced rates.

    • IMASBA

      I see we made the same point about the possibility of renting cleaning robots for a few hours, but you were first.

      • Mark Bahner

        Great minds think alike.
        Here’s how I view future of robotic cleaning. First, a human brain is somewhere in the 0.5 petaflops to 20 petaflops range. (I previously estimated 5 petaflops.) Since the cost in January 2010 was $5 per gigaflop, the microprocessor cost for doing maid services was somewhere in the range of $2.5 million to $100 million.
        However, the microprocessor cost in 2020 will be in the range of $2,500 to $100,000, and the cost in 2030 will be $2.5 to $100.
        The cost of a body capable of doing a maid’s job is probably in the range of $2000 to $20,000.
        So the capital cost for a human-maid-equivalent will be in the range of $4500 to $120,000 in 2020, and $2002 to $20,000 in 2030.
        So it seems unlikely that anyone but fairly wealthy people will own robot maids before 2030, but people who own maid services are likely to have robot employees even as soon as 2020, and almost certainly most maid services will employ robots by 2030.

      • IMASBA

        Estimate of dozens of petaflops are still based on a simple neuron-switch model, while it was recently discovered dendrites also process information, neurons are probably not simple switches and who knows how much more complicated the brain really is. Fortunately a cleaning robot only needs a tiny fraction of the capabilities of a human brain, probably only something like a primitive fish brain.

      • Dan Browne

        The FLOPS as a measure of human brain processing capacity seem to be a little far-fetched. The reason is that it’s skewed towards high flops for specific areas such as pattern recognition. Other areas such as math or search capabilities are much much less than even hundreds of FLOPS never mind billions. How many humans can do several floating point MATH operations per second? None. Likewise, how many humans can do millions of searches per second on patterns in a library? None.
        I suspect that as soon as we get reasonably hefty pattern recognition neural nets trained up for a variety of tasks, we’ll have human equivalent skillsets in narrow areas of functionality. We don’t need a human level of intelligence to perform a maid’s tasks, simply the automation of that limited set of tasks.

      • IMASBA

        Actually humans perform the equivalent of many FLOPS in processing emotions and social interactions. There are savants who can harness this computing power to be human calculators and that ability can be temporarily triggered in many ordinary humans using magnets, this does not lead to an inability to recognize emotions or display social behavior so apparently only a small fraction of computing power otherwise destined for those things is rerouted for mathematical operations and this small fraction is probably very inefficient at math (because the brain isn’t build up around addition, multiplication, etc… circuits) so it becomes all the more impressive that such a small fraction of brain computing power enables human calculators.

        And yes, naturally a robot maid only needs to have a fraction of human computing power, I said as much earlier

      • Dan Browne

        The areas that you have described as requiring many FLOPS fit neatly into the group I also described as requiring many FLOPS (pattern recognition). If you’d read further down the post you’d have noticed I started talking about neural nets to do those types of tasks (pattern recognition). Though you can indeed construct a complicated iterative algorithm to “process” individual parts of patterns over many repetitive operations until the entire pattern is processed I would suspect a trained neural net optimized for recognizing such for recognizing the
        specified pattern would be much faster and require much less in the way of
        processing capacity needs. That is precisely why I am skeptical of the
        “human brains process at the equivalent of X gazillion FLOPS”.

        Human brains DON’T
        PROCESS like one step at a time algorithms AT ALL and thus the comparison is apples and oranges.

  • Mark Bahner

    Oops. That would have been $25 million for the microprocessing, given $5 per gigflop and 5 petaflops of processing. (I started out with 1 petaflop for the processing speed of a human brain…in which case the cost in January 2010 would have been $5 million.)

  • Philip Goetz

    Why don’t dishwashers and washing machines count as robots? The work carrying the clothes to the washer is trivial compared to the work of the washing. In the 1970s I had a home with a laundry chute, which delivered clothes into the laundry. If instead of a chute that works by gravity, I have a hamper I must put dirty clothes into, which a robot carries downstairs, does that count as a robot, while the more-efficient laundry chute doesn’t?

    Household work can still be more-easily reduced by altering the house than by making robots smarter.