Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

How should we talk to AIs?

blog.stephenwolfram.com

Post reply on HN
Print view View original

21–30 of 46 posts

Re: How should we talk to AIs?

#21

Post by _lce0 »

hacker_9 wrote: post #17
joesb wrote: post #14

Earlier quoted context omitted.

> 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 scree…

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 lang…

are ambiguities only meaningful to humans?

Seems to me that computers were instructed to (1) refuse to work in front of ambiguities, or (2) resolve them by choosing one of the options.

Kind of like when `yacc` tells you that some rule is ambiguous, but then it ends up resolving it by deterministically picking always the same option.

Re: How should we talk to AIs?

#22

Post by hacker_9 »

_lce0 wrote: post #21
hacker_9 wrote: post #17

Earlier quoted context omitted.

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 lang…

are ambiguities only meaningful to humans? Seems to me that computers were instructed to (1) refuse to work in front of ambiguities, or (2) resolve them by choosing one of the options. Kind of like when `yacc` tells you that some rule is ambiguous, but then it ends up resolving it by deterministically picking always the same option.

we don't know how to make computers understand ambiguities, that is the problem. Yacc always selecting the first option isn't necessarily right nor how we would solve the problem, we choose the option that makes the most 'sense'.

Of course what is 'common sense'? Something that only comes about when you have a billion neurons doing your calculations perhaps.

Also if computers had human level understanding we wouldn't need Yacc rules at all; we can recognize from 100's of different languages just by looking at them, even partial code, despite all the ambiguity of putting all programming languages into one Yacc grammar.

Re: How should we talk to AIs?

#23

Post by kruhft »

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

Re: How should we talk to AIs?

#24

Post by gnaritas »

_lce0 wrote: post #21
hacker_9 wrote: post #17

Earlier quoted context omitted.

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 lang…

are ambiguities only meaningful to humans? Seems to me that computers were instructed to (1) refuse to work in front of ambiguities, or (2) resolve them by choosing one of the options. Kind of like when `yacc` tells you that some rule is ambiguous, but then it ends up resolving it by deterministically picking always the same option.

> Seems to me that computers were instructed to (1) refuse to work in front of ambiguities

No, that would presume incorrectly that it understood what ambiguity was or even knew it was in an ambiguous situation. Computers don't refuse anything, they do exactly what they're told and only what they're told even when what they're told doesn't make any sense.

> resolve them by choosing one of the options.

They don't even know there are options.

Re: How should we talk to AIs?

#25

Post by 0xdeadbeefbabe »

eykanal wrote: post #5

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.

> 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.

Re: How should we talk to AIs?

#26

Post by cormullion »

bobm_kite9 wrote: post #6

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…

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...

Re: How should we talk to AIs?

#27

Post by dkural »

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.

Re: How should we talk to AIs?

#28

Post by faleidel »

eykanal wrote: post #5

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.

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

Re: How should we talk to AIs?

#29

Post by carlob »

zamalek wrote: post #13

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 opti…

matlab or mathematica?

Re: How should we talk to AIs?

#30

Post by carlob »

bashedly wrote: post #2

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…

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.
Post reply on HN