[Zero]

I'd like to describe the differences between chatbots and thinkbots, as I see it. "Chatbot" is a widely used word, describing a very specific kind of software. "Thinkbot" is just a word I personally often use, without even knowing how it "sounds" to a natural english speaker, what bells it does ring, so to speak.

1, a chatbot is all about obtaining an effect, a result. Usually, you build a chatbot because you want something (a conversation) to happen outside of the software. Thinkbots are more centered on what happens inside the software. It is the inner thinking process that matters most, not the external behavior of the program.

2, the IO of programs are wired differently. Conversational agents are usually plugged in a chat interface, sometimes with an additional avatar or talking head representing the bot. In that case, inputs are sentences spoken or typed by the user, and outputs are sentences written on a screen, or text-to-speeched, plus movements of the talking head. How about thinkbots?

I've been wondering for a long time how a thinkbot's IO should be wired. In other words, what's a thinkbot's body. The entire computer? The browser? Some virtual environment? It had to reflect some features of a human's world, but at the same time, I wanted it to be pluggable, and generalizable, so we can use thinkbots in a variety of situations. I decided a few days ago what my solution would be.

An AI body is important, because there can't be intelligence, or consciousness, without a body. My solution is to use an XML document as the world. A big doc, ok, but just a doc.

I want to be clear about one thing: this XML doc is not a map of a world. It is the world itself. A strange world, made of tags and attributes. It can have rules, like "physical" rules if you want. That's the world a thinkbot's embodied in. There could be room tags and door tags, or tree tags forming a forest... whatever you want it doesn't matter. There can be animal tags, able to move around in the world, which means, some nodes can change their own location in the DOM.

Somewhere in this world, there's a "Me" tag. It doesn't have to be <me></me>, but it could, why not. This is the body of the thinkbot. Inside this <me></me> element, there are other elements representing maybe an arm, or a mouth, or legs... moveable parts that the thinkbot can act on. Acting on these inner elements, the thinkbots can act on its surroundings, thanks to the physical rules we evoked. The same physical rules can modify these inner parts if the environment influence the body. The thinkbot can "feel" its inner parts, so it can feel its environment. I think you get the idea. Let's go further.

One part of this body is the mind part. It reflects what happens in the thinkbot, on a computational level. The mind part is thus the root of the consciousness loop: it lets the thinkbot feel what's happening inside of itself. It can analyse its own inner behavior (its thinking process), predict it, and eventually modify it.

[ivan.moony]

What really matters is input and output. In between could  be anything, and that is a part that will be hidden to end user. But it is not irrelevant. As much as coders want to provide a good user experience, they don't want to use i.e. assembler to program a chatbot. Java is an example of a language that caught a decent relationship with programmers. Its input and output is as much as same as with other languages, so it is not something revolutionary for end users. AIML on the other side reached higher level of connecting input and output, but it comes with a price: it is not Turing complete. Nevertheless, it excels at what it promises.

I wonder in which extent it is possible to push up to higher levels while retaining Turing completeness.

P.S.
Zero, your next post will have really hot post count number. Be sure to cool it down with ice or something, not to burn up the server wiring. Or maybe you plan something really mean for that occasion?

[Zero]

I'm not sure AIML isn't Turing complete by the way. Its "think" keyword makes it similar to Semi-Thue, a string-rewriting based Turing complete programming language. But I get what you mean: it's not Turing complete in the traditional meaning.

As we're creating computing tools, our end-users are actually developers. My audience is made of hypothetic developers, who would make products meant to be used by non-tech people. So API matters, and more generally speaking, what happens between input and output matters, because this audience is interested in it.

What I'm doing is like a transparent motor. I want the (user) interface to show the clockwork inside, like a crystal Harley-Davidson made of glass and ice, sort of.

I think the future of AI is a convergence of techniques, where Turing completeness matters less and less. Because it all gets mixed in something higher. It's not only about being able to calculate anything, but rather it's about: what can we compute easily, here and now.

Brainfuck is Turing-complete. But clearly that's not enough. It's a sh*tty language that's not meant to be used, because anyway, there's nothing useful you can do with it here and now.

I happen to like love creating languages. Unfortunately, I had to realize (thank you Uncle Bob Martin) that we don't need another language. J'en ai fait mon deuil. What we need, as developers, as writers of minds , is a really really wide toolbox, that lets you express what you want, instantly. Need logic? Ok, here is logic. Need bayesian network? Ok, here. Need NLP? Well, we have you covered. Who gives a sh*t about semi-colons and curlies and typing paradigm... Even list comprehension is just silk. Yeah, silk is nice, but it's still fabric.

So, look at OpenCog. Who forked an OpenCog project here? Nobody. Why. Because hell, who would spend two weeks trying to set things up, with a 75% risk of failure, to end up with something you can't even deploy to your end-users. And if you could, what would you wire it to? A 10k€ robot? Come on, let's be serious.

We, developers, are users. As any user, we want plug&play solutions. It has to work, here and now. My position is maybe controversial, but it feels like shouting out loud what people think quietly.

Anyway, this XML world concept has everything needed. It is simple to implement, it can host simple and complex solutions. It solves both the world problem and the mind-body problem. It is pluggable in any external architecture.

I wonder in which extent it is possible to push up to higher levels while retaining Turing completeness.

If we keep things modular and decoupled, there should be no limit. Take RiveScript for instance. By itself, even though it is Turing complete, it is obviously strongly oriented towards conversational agent authoring, since it's its purpose.

But forget for a sec what its purpose is, and just consider what it does really. RiveScript takes strings as input, and gives you back other strings, potentially reusing parts of the input to build up the output. This is not a conversation, this is string processing. It's not that high level after all! It is how we use it that makes it high level. Same goes for any other tool. JsonQ is a json manipulation library. But used as an ontological real-time situation analyser, wow... you get a very high level box of goodness.

So you take tools, whether Turing complete or not. You take them not for what they're meant to achieve, but for what they do. And you plug them in a flexible architecture that lets users - us, the devs - do what they need to.

Now my smartphone is 31% low on battery, and I'm 28% low on alcohooool. I need to eval my self!!!

 

[ranch vermin]

So, look at OpenCog. Who forked an OpenCog project here? Nobody. Why. Because hell, who would spend two weeks trying to set things up, with a 75% risk of failure, to end up with something you can't even deploy to your end-users. And if you could, what would you wire it to? A 10k€ robot? Come on, let's be serious.

When is Ben going to bloody well finish his own ai projects,  instead of getting other people to do them for him... is he another dummy that is going to keep putting off the singularity constantly so it never happens?  And that jerk woman robot he has doesnt impress me.

[Zero]

Yeah. I wouldn't say OpenCog was bad. It's just... Ai should be affordable, easy to share, explainable and pleasant.

[ranch vermin]

Thats a good way 2 b.

Im a bit more of a horrorist about it,  but only in a harmless way,  theres no way im going to make something completely a travesty,  but i want to get into the security industry,  so it is a little scary.

Indeed,  mine is explainable.  but im just keeping it to myself.  I agree, if its not explainable, its probably not true.
Its just as unfortunate as its fortunate.

[ivan.moony]

There were people ranting about OpenCog's slackness and Linux-only orientation, but the general answer was that the target audience are scientists (preferably phd) who deal with github/c++/python/scheme/make tools on a daily basis, so we can expect nothing to be changed if this politics sustains. But I think too it's a major drawback. Probably a lot of people want to investigate AI as a hobby, not as a full time research, and for them OpenCog represents a big complication that offers a lot of pain with a  questionable return value, I agree.

My ideal of an AI tool would be something that could be run in a browser (to attract users), while supporting off-line execution that could power some mean hardware like cameras, mics/speakers, limbs and motors. Preferably, a 3d-printed bot skeleton with camera and electromotors could be powered by Android on a smartphone, for those with a lot of adventure spirit.

[ranch vermin]

The body doesnt matter so much to be perfect,  a better brain will control what it has better, no matter what it is.    You pretty much could have a go with any bot you buy off the net.  Asimo, phantomx, cozmo,  Ubtech alpha 2, a driverless car, wall-e, Nao?  they all can be coded for all the way to infinite intelligence,  its just whos got the method of the century?

Im only making my own body because I want many copies of it,  and i dont want to pay the exorbitant amount just for a single body.

[Zero]

The real bang would be a (javascript) Ai platform, targeting browsers, tablets and smartphones. It would feature a "mind market", like Google Play or NPM, containing downloadable Ai modules. Devs would publish modules on this market and users would download them, to add new capabilities to their personal Ais. It would be all automatic: download it and isntall it in one click, and you're done. How many apps on Google play? How many libraries on NPM? Imagine the same number of Ai modules!

Here is a pic of my little buddy.

[ranch vermin]

you wouldnt want to trip over that one,  looks hard as nails bolted together material toughness madness.

Making that thing hyper intelligent could be a funny body to do it in...  would be modestly bodied, this computer genius robot.

I bet its got a h-bridge driver!  the motors are huge as well.

[ranch vermin]

how about making this one smart?  no thanks.

[Zero]

It's just a toy, able to carry a smartphone here and there, scaring the cat.

No matter what's outside, it would never be enough IMO: biological bodies have millions of sensors, even Boston Dynamics won't reach that. Having an XML world layer lets you plug your Ai in any thing. Or multiple things, like a browser and a smartphone.

[ranch vermin]

Isnt the smart phone good enough?

[Zero]

Yeah, what's nice with smartphones is the set of sensors they come with. If one day I want to do a physical robot, I definitely would use one. A smartphone only lacks the ability to move by itself, hence the little blue buddy.