How should we talk to AIs?

http://blog.stephenwolfram.com/2015/11/how-should-we-talk-to-ais/

by champillini • 11 years ago

52 44 11 years ago

blog.stephenwolfram.com

bashedly2• 11 years ago

I'm hoping someone closer to the WolframAlpha team and work can shed some light on this - but is the extensive use of the first person warranted here? There's an awful lot of "I" and "I've".

Is the work being done really so directly attributable to Stephen Wolfram or is there a army of hard working individuals behind the scenes not being referenced here? I'm not suggesting they list everyone by name or anything, but a simple shift to something like "our team" would seem more generous. Of course this is all moot if he is indeed primarily or almost solely responsible for the progress being referenced.

Zikes• 11 years ago

The gist of the article is that he sees AI as being so staggeringly complex that English will not be sufficient to communicate with it clearly, but that Wolfram Language might work. Since Wolfram Language is largely his project, a few first person pronouns are to be expected.

TeMPOraL• 11 years ago

It's Standard Wolfram Style. You can see the same pattern in his "New Kind of Science" book - an interesting piece of writing, but the amount of "I", "I've", "my", etc. is plain distracting.

gooseus• 11 years ago

I understand so little of the topics of that book or this review, yet still very much enjoyed reading the takedown:

http://bactra.org/reviews/wolfram/

"So much for substance. Let me turn to the style, which is that of monster raving egomania, beginning with the acknowledgments..."

gohrt• 11 years ago

Wolfram employeees sign contracts giving Dr Wolfram credit for their work, in exchange for their pay.

jjaredsimpson• 11 years ago

https://www.youtube.com/watch?v=77Y6CIyyBcI

That's what the money is for

bashedly2• 11 years ago

I know this is overly idealistic, but it seems to me that paying for "credit" (especially in terms of research) seems "wrong". I understand paying for ownership of the results of someone's work, or paying to state that the broader organization takes credit for the body of work produced by the employees. But for one individual to pay to personally receive credit for the output of another seems blatantly dishonest.

carlob2• 11 years ago

Wolfram Research has several hundred employees. Obviously he doesn't do all of the work, but he is involved in a lot of the design. Not that I necessarily agree with the style either.

eykanal• 11 years ago

So, the best way to talk to a computer is... a programming language?

Lets ignore the fact that, for some reason, he thinks his particular programming language is better than all other ones for interacting with an AI. Am I missing something here? This seems kinda "duh" to me.

0xdeadbeefbabe• 11 years ago

Maybe this is a case of having a hammer and then seeing everything as a nail. But I’m pretty sure there’s more to it. And at the very least, thinking through the issue is a way to understand more about AIs and their relation to humans.

Looks like he's pretty sure there is more to it.

faleidel• 11 years ago

Aren't DSL already doing part of this "bridge" between human and computer languages?

bobm_kite9• 11 years ago

What's the difference between Wolfram Language and just any other programming language?

Also, one of the advantages of natural language is it's imprecision: it makes communication more robust, and leaves room for interpretation.. in much the same way as imprecise duplication of the genome allows for evolution, imprecision in language allows for alternative interpretations of the results.

Plus, if you want to, you can check interpretation and understanding just by asking some questions and creating feedback.

Actually, I just don't get how this is advancing the art in either direction.

cormullion• 11 years ago

Stephen Wolfram has designed (or overseen) the Mathematica Language and all its many APIs for over 25 years, so its consistency is possibly better than other languages. He probably sees the language as being close to human thought -- it's certainly close to his. Judging from my struggles I'm not sure I'd agree...

beepbop• 11 years ago

With a stern, calm voice.

skybrian• 11 years ago

It might be interesting to see what language evolves when you train two or more AI's to communicate over a connection to complete some task.

hacker_94• 11 years ago

Interesting that the Wolfram Language is very much a LISP dialect, with the syntax swizzled a bit here and there to be a bit more human readable.

Also his Riffle example is thinking about the problem the wrong way round. It looks like he tried to turn code into english, when it should be the other way around. Here is an actual english version:

  1. Start with a purple square
  2. Clone the square, rotate it by 0.1 radians and scale it to fit the previous square
  3. Also alternate the colour between purple and yellow.
  4. repeat clone process until square is just a dot in the middle of the screen

Note how easy it was to understand what I meant, and yet a computer would fail at so many of the instructions: scale how? how to test if it fits? alternate whose colour? repeat what steps? 'clone process'? clone same square or newest created square? what defines a dot? screen? middle?

All these ambiguities we can solve, but a computer cannot hence why our programming languages are so specific to the point of cognitive overload. Also none of his later examples show to me that Wolfram is better than English, unless of course you use wolfram daily (as he 'the creator' likely does/did), in which case you're brain is already optimized to read it...

This isn't even mentioning that language is only part of how we communicate. We can often leave so much out because our visual system, body movements, previous experiences, memory and other senses fill in the rest.

joesb• 11 years ago

All these ambiguities we can solve

We can solve it by specify what each word means. That's no different than programming.

Also, your step 4 has bug because "repeating the clone process" does not include "rotating and scale it to fit" part. It'll just keep cloning the square all days and never get the square to be the size of a dot.

Also, If you screen is not a square and you did not start in the center of the screen. Then the square can be as small as a dot, but it won't be "in the middle of the screen".

hacker_94• 11 years ago

Joesb if anything you are proving my point! As another human you saw large ambiguities in my instructions but could still figure out 1. the problems (you knew no one would clone the square infinitely as it wouldn't 'make sense') and 2. the required solutions (i probably meant scale+rotate as well to be included in the cloning process, otherwise no interesting visual change would happen).

Tell me what programming language / compiler can do that by just specifying word meanings? Context, sentence structure, your understanding, etc all play into it as well.

AnimalMuppet• 11 years ago

We can solve it by specify what each word means.

The word that's going to cause the most trouble is probably "it". What does "it" mean? In this specification, it means the cloned square. But in a longer specification (written in this style), the meaning of "it" will depends on which instance of the word we're talking about, because what it refers to keeps changing...

zamalek• 11 years ago

He goes on ad nauseum about how fitting the language is. He's right.

Back in university I'd frequently help seniors out with projects. The one project involved battleship AI in mathlab. In maybe 3 hours I got it as good as it was probably going to get pitted against a truly random opponent. However, PNRGs aren't truly random. It was so quick and so effortless to grab stats on the biases the PNRG had (in order to optimize the guess selection process). It was the first time I ever used the language and I can honestly admit it had very little to do with my capability and so much more to do with how damn fluent and intuitive that language is.

It's inconsistent, messy and evolutionary but, wow, what an absolute masterpiece.

hacker_94• 11 years ago

Any specifics as to what you mean? How it was more elegant that a more mainstream language?

carlob2• 11 years ago

matlab or mathematica?

rm445• 11 years ago

According to Wikipedia, Mathlab was a computer algebra system (written in Lisp) from the late 60s, whose author went on to write Macsyma, which inspired Mathematica.

Like you I'm still not sure whether parent poster meant that, or is confused about Matlab vs Mathematica.

winter45• 11 years ago

Concerns about consciousness, intelligence,and Stephen Wolfram's use of the personal pronoun aside, I think it reasonable to presume that communication with early stage AI systems will desirably include the development of a vocabulary and syntax structured for 'precision'.

It does seem reasonable that if AIs evolve they will, in their evolution, develop constructs of 'reality' that Wolfram calls "Post Linguistics Emergent Concepts"(1) and that, if they do, until human languages develop deeply "precise" 'words' for each of these PLECs, Spock will have to translate for Kirk; Spock, talking to an AI, will be (mostly?) unintelligible to Kirk.

This raises the question can one reasonably believe that humans are capable of more than 'bare bones' precision in words? Can one be both Kirk and Spock?

On that note it is encouraging that Wolfram appears to think our systems may be (become?) a driving force in our evolution; we may evolve to a point the use of a precise or imprecise language is a moment-to-moment choice.

It may be that these precise and imprecise languages will develop into a single language.

Before that happens, if you were such a dual-linguist, which language would you prefer? Under what conditions would you switch? What would communication be like in a world where the commonly spoken language was a meld of precision and imprecision? Which of the two would have the most influence on the other? Past a certain stage are there generally distinctions in character?

(1) see: https://www.youtube.com/watch?v=TMviBl46dXg

laotzu• 11 years ago

This idea of Post Linguistic Emergent Concepts is an interesting one which Marshall McLuhan alluded to in his 1964 book The Gutenberg Galaxy: The Making of Typographic Man. He suggested that the sheer speed of electric processing made the alphabet obsolete and that in order to cope with this drastic speed up in communication we would have to deprecate our slow spoken languages and adopt an entirely new method of communication:

Now, in the electric age, the very instantaneous nature of co-existence among our technological instruments has created a crisis quite new in human history. Our extended faculties and senses now constitute a single field of experience which demands that they become collectively conscious. Our technologies, like our private senses, now demand an interplay and ratio that makes rational co-existence possible. As long as our technologies were as slow as the wheel or the alphabet or money, the fact that they were separate, closed systems was socially and psychically supportable. This is not true now when sight and sound and movement are simultaneous and global in extent. A ratio of interplay among these extensions of our human functions is now as necessary collectively as it has always been for our private and personal rationality in terms of our private senses or "wits," as they were once called.

kruhft• 11 years ago

I've always wondered where languages like Lojban[1] would fit into conversational AI. Designed to reduce ambiguity, I would think they would be a major stepping stone to bridging the gap between conversation and logical correctness needed by a truly intelligent system.

[1] https://en.wikipedia.org/wiki/Lojban

sidcypher• 11 years ago

I'd say a truly intelligent system doesn't need Lojban, it can learn about the world and its languages from observations and feedback, like children do. A pseudointelligent system, however, would probably be much more useful if you could talk in Lojban to it, with less communication errors and all.

dkural2• 11 years ago

Wolfram is the ultimate company failing in product thinking. For decades they yell "You can do anything with this!", but fail to identify, specifically, one thing it does better than other products that are designed to specifically do that thing. Thus for anything it could do, other products spend more attention and do that specific thing better/cheaper, with a tighter UX and more content.

It turns out most people don't care if their car is also a surface-to-air missile, or if their dishwasher can also compose poems.

taliesinb• 11 years ago

Yeah, but it's still fun to actually bolt the missile fins onto the car, even if it's being driven by an unstoppable maniac :)

AndrewKemendo• 11 years ago

It's the ultimate example of a technology in search of a problem. They got it pretty close to right with adoption of mathematica in academia but has yet to seem to really take off in industry.

Watson has the same problem - and when they released the SDK last year it was kind of like, well, just try it, but they seem to have at least determined to focus on health care.

veidr• 11 years ago

But you can only really say they are failing, if you are looking at them through a short-term lens. Interviews with Stephen Wolfram give the impression that he is thinking long-term (as in, his entire lifetime and beyond).

Mathematica does actually seem to have dominated its niche for decades. People did and do all manner of things with it.

Wolfram Alpha, their "computational knowledge engine", has only been live for 6.5 years. It's sort of baffling who it is for (but I do, very occasionally, find it extremely useful), but it seems what they are trying to do is bigger than a normal startup that helps you find a taxi or rent out your extra bedroom or search the news.

They may well fail, but I don't think you can say that they have failed yet.

dkural2• 11 years ago

This is what Wolfram says to their engineers who question it. Wolfram / Mathematica has been around for a while. What's the long-term idea? What exactly will it do better than some other tool in that category? You can't reach there if you don't know where you're going. It's worse than failing - they don't know what they are supposed to be failing at.

oneJob• 11 years ago

I misread as "How we should talk to AIs."

Thought this was the AI (aka Stephan Wolfram) telling humans how it'd prefer we interact with it.

lovboat• 11 years ago

If AIs are not able to communicate we us properly then those are programs but not real AIs. I didn't find any interesting idea here and the sheer act of proposing that program as a language for AI is really humorist to say the least. People are using R and Python to construct a solid foundation for machine learning, the NLP and the IA will evolve to communicate with us by its own means, that's the real AI.

CM30• 11 years ago

It's a good point, but if we have to basically 'program' them to get them to understand anything, they're not particularly good/useful AIs. The vast majority of the population won't do this, and any decent AI needs to be able to understand them.

caligastia• 11 years ago

Wolfram is going off into the weeds. The whole point of NKS is that the universe is digital. We live in a computational universe, therefore we are living and breathing and interacting with natural (not artificial) intelligence as a function of merely existing.

Assuming that by AI, Wolfram is referring to a computational device built by humans, the communication has already happened in the act of designing and constructing the machine. Real communication is over once you flip it on.

What Wolfram is proposing is to then run a Wolfram Language interpreter on top of this new machine and use that to communicate, presumably to ask it for the meaning of the universe. Instead of creating a lingua franca between man and machine, this will prove only that the machine can emulate a von Neumann computer and run the Wolfram interpreter on top. You are NOT communicating with the machine at that point, you are communicating with a parser and a nifty collection of mathematical routines.

Think of a parrot. You teach it to say "Polly want a cracker!" - and the parrot may even learn a few human words - but are you really communicating? Can the parrot express to you the thrill of flying, can it explain to you what a total mind-fuck it is for it to be stuck in a cage begging for crackers when it should be flying with its flock and living its life? No, the parrot can only ask you for a cracker, just the way you taught it.

Man will never have a relationship among equals with a machine, regardless of how much software he pours on top - the machine has a frame of reference that can never be communicated to a human, it's different at the electronic level, never mind the symbolic or linguistic. The machine will have a reality so different from that of the human that even if it could communicate something of substance, it would be as futile as asking your parrot to help you with a math problem - the parrot is not stupid, and in fact can do many things you would find highly mathematical, but it cannot help you with your homework because you are in two different mental universes, no human-invented language can bridge them.

joeevans1000• 11 years ago

"And by now people routinely ask personal assistant systems—many powered by Wolfram|Alpha—zillions of questions in ordinary language every day."

Is this true? I didn't realize this.

ddingus• 11 years ago

If it's an AI, then I should be able to talk to it like I would other intelligence.

That is how we talk to an AI. This article is framing a set of powerful computational entities as an "AI", and that is not yet defensible.

So much of what we communicate among humans depends on implied, common context. We know what would make sense to other people, or have a good idea of that, based on our understanding of ourselves and other people.

When I think, "AI", I think about something that knows me other than a set of attributes and rules, and I can know it in the same way. Imagine a little kid type intelligence connected to those powerful compute entities and with a great memory. Better than our memory. And it's fast.

It's that "little kid" type intelligence that is missing! We have the fast, we have the great memory, we have lots of powerful compute entities too.

What we actually don't have is the "intelligence" to complete the idea of "artificial intelligence"

How should we talk to AIs?

Related Stories