Hacker News new | past | comments | ask | show | jobs | submit login
Total Horse Takeover (greaterwrong.com)
82 points by reedwolf 5 days ago | hide | past | web | favorite | 34 comments

This analogy does a surprisingly good job of highlighting what is wrong with robotics right now as well. Robots are hard to control at a low level, and typically you don’t want to. But sometimes you have to. We don’t have very good ways of specifying tasks or behaviors at different levels of granularity and having the robot figure out the right affordances to that tasking.

Huh. I may have just figured out my next research proposal.

Interestingly, we do exactly that when controlling our own bodies: You're typically able to delegate the exact mechanics of walking, holding things, speaking, etc to your subconscious "muscle memory" (what ever that is exactly) - however you'll take "direct control" when performing an unfamiliar movement (e.g. dancing or doing Yoga for the first time).

Interestingly, the "direct control" is also gradual: You're able to position your hand very precisely without meeting to know where your arm is going.

The system is even flexible enough that it can eventually learn unfamiliar movements and allow you to "delegate" them as well. (E.g. an experienced dancer or Yoga practitioner)

You might be interested in the Three Layer Architecture approach, e.g. http://www.flownet.com/gat/papers/tla.pdf

What do you mean "Robots are hard to control at a low level"? That's the only level where they're easy!

At the lowest granularity, they have servomotors, encoders/resolvers, and kinematics that convert motor rotation to a motion in 3D space. Robot manufacturers have made the code to make a joint or linear move to a predefined X/Y/Z/i/j/k location in space using those tools really quite simple. Inputs and outputs to control grippers and other end-of-arm tools are also straightforward. As long as you keep the process low-level, they're incredibly reliable and easy to program: a clever maintenance technician can pick up a pendant for the first time and have a well-scoped task programmed in a week. A competent sales engineer can clamp a small robot to a conference room table and demonstrate picking up and putting together your product in a one-hour sales demo.

It's when you expand the scope when it gets really complicated: ask it to pick up a variety of objects, located by vision algorithms in arbitrary orientations and locations, that are moving in real time, in a changing environment, is simply a hard problem - and I think it always will be. Robots really only know how to move their tool center point from one location to another, and asking them to "do what I mean" and "pick up one of these and put it on one of those without making me worry about low-level stuff" is a level of simplification that has never really worked in any other domain.

It’s hard in two senses: one, it’s hard to translate low-level commanding into meaningful work in all but the most controlled of environments. That’s why robots that are economically viable today are nearly all on factory floors: that’s basically the only setting where you can control the environment well enough for them to be useful.

And two, for “advanced” robots (quadrupeds/humanoids) it’s very hard to get them to even balance, much less walk. Yes, there are robots that do this today — surprisingly well, in fact. But there is a tremendous amount going on algorithmically under the hood to make it happen. Figuring out how to make them work took decades of research in controls.

This is largely a restatement of what you said above, so I suppose it depends on what you mean by “low level”. My intent above was low level in horse terms, which includes standing and walking. Things humans think of as atomic behaviors.

Well, we do have a body of knowledge on those problems, as a sister-comment points out, Biology, and Cybernetics which abstracted the lessons from biology and generalized them to all machine-like systems.

It's just that this body of knowledge is not really compatible with the kind of stuff that the hyper-rationalist cult of lesswrong are into, and so tends to be ignored.

In fact, the author here is re-inventing Cybernetics from whole cloth (especially that of Ashby and Beer, in particular, Ashby's Law of Requisite Variety and Beer's Viable System Model), but without any engagement with the mass of existing knowledge.

[EDIT: I've just realized that the original author is being re-posted on lesswrong in this case, and isn't necessarily part of that milieu. I'd still encourage them (if they're reading) to check out Ross Ashby and Stafford Beer, because their intuitions align very strongly with the kind of research done by those Cyberneticians!]

I’m sure it’s just because I’ve never owned one, but I have often thought horses were the first “self driving” vehicle, and could still have some features superior to automobiles. You can feed them hay, they occasional apple and sugar cube, and they are largely carbon neutral. Even producing natural fertilizer as a byproduct of combustion. Built in cruise control, obstacle avoidance, and voice control, and a limited ability to self navigate and self-park.

Cars took off because, among other things, they require considerably less maintenance (especially cleaning!) and are more reliable (horse illnesses can be a serious problem). While they are self-replicating, it takes human time to train a foal to usefulness, and bloodline maintenance can get complicated and expensive if you're after something more specific than just "a horse". The methane emissions aren't great for global warming either.

Seaweed is efficacious in cow digestion to lower methane emmissions.

Perhaps it works in horses too ?


Not just methane. The automobile had the advantage of removing millions of pounds of horse manure from city streets.

Aside from the so-called "great horse manure crisis" story being somewhat vaguely sourced (horse manure was of value back then and was traded as such), even if it were true or at least becoming a growing problem, I feel this problem is likely overstated. If cars had not removed it, we would've solved this problem by other means.

Human "manure" was also as big a problem: sewers were relatively recent and still being build, and toilets were a brand new invention. Yet we've largely solved this issue without getting rid of people from cities.

with horses in particular, you don’t “control”. You establish a relationship and negotiate. Goats are like this too, as are huskies and malamutes and some other dog breeds.

Relate and negotiate turns out to be a much healthier metaphor for getting shit done, too.

Every domesticated animal is like this. (I'm not including animals like e.g. goldfish because ISTM there important ways in which they are not domesticated.) The other important aspect is that members of any particular species or lineage within a species have largely the same behaviors. Training and experience might add useful behaviors, but they don't completely eliminate existing behaviors.

For instance, cattle have a way of existence in which they are comfortable. (They really don't like anyone directly behind them, but off the rear quarter is tolerable. Structure like a fence or a road makes it easier for them to walk in a straight line. A single animal might be comfortable in isolation most of the time, but will seek to join a group when stressed. A path that is comfortable at a given speed will not be followed at a higher speed. Etc.) One person who understands that way and is willing to work within it can move a large herd more quickly and easily than twenty ignorant people.

ISTM human beings are also "domesticated", although they display a broader range of interesting behaviors than other domesticated animals do.

> The other important aspect is that members of any particular species or lineage within a species have largely the same behaviors

Re: Why we domesticate horses, but not zebras.

Zebras have a different herd mentality to Horses, pretty much they don't have one. They do move around in groups for mutual protection, but they live more as individuals and do not follow a single alpha male herd leader as horses do. So with horses you can set yourself up as the alpha leader and they will follow you, but that doesn't work with Zebras. This combined with their more aggressive nature has made them impossible to domesticate.

What does ISTM stand for?

"It Seems To Me" is my guess based on usage.

All things considered, I’d say neither the horse, the goat or the husky are good metaphors for getting shit done. The bull is ;)

Perhaps suited for some sets of things, but certainly not for the acquisition, display, or sale of fine china ;) There are many sorts of doings to which the bull approach is terribly ill-suited.

> Flaccid horse mass isn’t that helpful, not even if we throw in the horse’s physical strength to move itself according to your commands, and some sort of magical ability for you to communicate muscle-level commands to it.

Flaccid Horse Mass is my new band name.

Original link on Meteuphoric: https://meteuphoric.com/2019/11/04/total-horse-takeover/

I'm glad to see this blog active again after a year of quiet.

Thank you. Please read directly from the source instead of participating in anything to do with LessWrong.


For one thing, this page should have just linked to the original rather than replicating it in full. HN doesn't replicate work done by others, and is superior to "Greater Wrong" in that respect.

Just FYI, the page replicates the work in full because the author has set up automatic crossposting from her blog to LessWrong (and GreaterWrong is a mirror of LessWrong).

Yes, but the GP was asking to avoid linking LW altogether — not just when there’s a better original source —- which seems ... overbroad (and flamebait).

Full disclosure: I organize for LW Austin group.

LessWrong can be reasonably characterized as a recruiting ground for pyramid schemes and/or a cult of personality. It is an attempt to reframe libertarianism as apolitical and as the only rational, correct course of action. Effective altruists are urged to donate to unaccountable AI research projects controlled by LW essayists.

For many reasons, I don't believe it's responsible to link people there.

Some starting points below, but please do your own research:



I'm surprised to see that there's no other commenters who read this as a metaphor for managing a company.

Communicating the main goals to engineers (take me to London) is essential. The engineers might not know where, or even what "London" is. Pointing the horse's head in the right direction and saying "go" is all you need to do to make progress in the right direction. Micromanaging every leg movement, or telling the horse to use a bicycle is counterproductive (Agile! Jira! Confluence! Monday! Whatever the latest trend is!) If the horse teaches itself to ride a bicycle (engineers choose a system because they need it) then that does reduce the control that management has. Is that a bad thing if it still gets you to London faster?

Forget a horse, the same argument can be made on a person's own self. Our conscious mental efforts cannot even "total takeover" our own bodies.

Horses (and other animals in training) are problem-solving agents. To teach it something, you pose problems to it that it can figure out.

Especially for dogs, you pose a problem in the form of "How do I get a reward from this Hooman?" And by adjusting the difficulty of the required solution a dog understands what you want.

This article reminds of me of this game that demonstrates how difficult it is to operate a horse: http://www.foddy.net/CLOP.html (enabling flash required)

My read on this was that it's a useful metaphor for the economy, with the state as the rider and the horse as the industrial base. The original communist idea was to direct industrial output in some detail, which worked about as well as trying to control the individual muscles of a horse. The US, meanwhile, directs the economy by deciding "I want $X billion dollars worth of helicopters" and then letting the private sector figure out the low-level actions required to get there. The Chinese are probably an even better example in this regard.

This is really the argument for abstraction

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