6 ms·

AI Is Inventing Languages Humans Can’t Understand. Should We Stop It?

41 points by jumpkickhit 9y ago

emodendroket 9y ago
> But at the same time, it feels shortsighted, doesn’t it? If we can build software that can speak to other software more efficiently, shouldn’t we use that? Couldn’t there be some benefit? Then why are we putzing about with strings in the first place? Let them use binary formats if that's the goal.
WalterBright 9y ago
Sounds like "Colossus: The Forbin Project" has arrived!
Vanit 9y ago
Reminds me of the language of the Observers in Fringe, whose language was so advanced it was rare to see repeating symbols.
- tpeo 9y ago
  Somehow this doesn't seem like a desirable property in a language. One of the purposes of (human? inter-sentient?) languages is the reliable transmission of information. Verbosity and repetition in language have a purpose similar to that of redundancy in engineering: they ensure that if someone misses out a part of the message, the whole message isn't ruined. If your message being get across is predicated on successfully transmitting an arbitrarily large sequence of symbols where every one of them might be crucial (supposing that word frequency is inversely proportional to meaningfulness, which is rather reasonable), this "talking" thing would get hard pretty fast.
jpster 9y ago
> “Getting the data into a format that makes sense for machine learning is a huge undertaking right now and is more art than science. English is a very convoluted and complicated language and not at all amicable for machine learning.” Is there an advantage to using less irregular human languages - say German - in machine learning?
- CrystalLangUser 9y ago
  German, where noun gender doesn’t make any sense? and things like dative case have to be specifically marked? German and English have the same lineage anyhow, both have irregular verbs (in many cases English irregular verbs are also irregular in German). Analyzing any real language (not Esperanto) is going to be extremely difficult. I would think English is marginally easier than say Korean where everything is based around context, including dropping the subject itself and particles. Otoh, some aspects of Korean are probably easier than English. And all 3 have irregularities despite Sejong the Great purposefully creating Korean Hangul.
microwavecamera 9y ago
Have we considered that maybe Facebook just isn't very good at this AI thing? Just putting that out there....
- akhilcacharya 9y ago
  They're up there with Google and OpenAI, so no.
  - microwavecamera 9y ago
    Facebook is no where near the level of what Google has accomplished and you're comparing them with OpenAI, a non-profit. A multi-billion dollar company competing on the same level as a non-profit. Facebook own AI research site lists 4 projects, not including a broken chat bot. How could I have possibly arrived at that conclusion? https://research.fb.com/projects/ https://research.fb.com/projects/
alexandercrohde 9y ago
What a junk article. Link bait. "Humans can't understand" is ambiguous, and the article wants you to believe this means "above human comprehension," when really it just means "makes up some new language that we could absolutely reverse engineer."
dan-compton 9y ago
This article is terrible. I still assert that a relative black box will lie at the core of any "strong" AI.
ajaygeorge91 9y ago
who is up voting this shitty articles
- jumpkickhit 9y ago
  I submitted, so it's my fault. I ignored the title and thought the content itself was interesting. That AI appeared to be communicating with a method it figured out itself. That's what interested me enough to want to share it here. I'll use more discretion next time i submit something.
anotheryou 9y ago
Please also stop these fresh people, I don't understand a word. And these protocolls, http and such. I can also not understand them!
IceDane 9y ago
What a load of horseshit. Here's what you need to do before publishing an article on AI: 1. Hire an expert that actually knows this stuff 2. Have him read the article 3. Throw the article out when he tells you to throw it out 4. Feel shame Stop writing these imbecilic, clickbaity shit articles about AI. This leaves people that don't know anything about AI or neural networks thinking that we're somehow developing skynet, instead of just writing a computer program which essentially implements a mathematical model that continually adjusts itself based on some metric.

41 points by jumpkickhit 9y ago

> But at the same time, it feels shortsighted, doesn’t it? If we can build software that can speak to other software more efficiently, shouldn’t we use that? Couldn’t there be some benefit? Then why are we putzing about with strings in the first place? Let them use binary formats if that's the goal.

WalterBright 9y ago

Sounds like "Colossus: The Forbin Project" has arrived!

Vanit 9y ago

Reminds me of the language of the Observers in Fringe, whose language was so advanced it was rare to see repeating symbols.

tpeo 9y ago
Somehow this doesn't seem like a desirable property in a language. One of the purposes of (human? inter-sentient?) languages is the reliable transmission of information. Verbosity and repetition in language have a purpose similar to that of redundancy in engineering: they ensure that if someone misses out a part of the message, the whole message isn't ruined. If your message being get across is predicated on successfully transmitting an arbitrarily large sequence of symbols where every one of them might be crucial (supposing that word frequency is inversely proportional to meaningfulness, which is rather reasonable), this "talking" thing would get hard pretty fast.

jpster 9y ago

> “Getting the data into a format that makes sense for machine learning is a huge undertaking right now and is more art than science. English is a very convoluted and complicated language and not at all amicable for machine learning.” Is there an advantage to using less irregular human languages - say German - in machine learning?

CrystalLangUser 9y ago
German, where noun gender doesn’t make any sense? and things like dative case have to be specifically marked? German and English have the same lineage anyhow, both have irregular verbs (in many cases English irregular verbs are also irregular in German). Analyzing any real language (not Esperanto) is going to be extremely difficult. I would think English is marginally easier than say Korean where everything is based around context, including dropping the subject itself and particles. Otoh, some aspects of Korean are probably easier than English. And all 3 have irregularities despite Sejong the Great purposefully creating Korean Hangul.

microwavecamera 9y ago

Have we considered that maybe Facebook just isn't very good at this AI thing? Just putting that out there....

akhilcacharya 9y ago
They're up there with Google and OpenAI, so no.
- microwavecamera 9y ago
  Facebook is no where near the level of what Google has accomplished and you're comparing them with OpenAI, a non-profit. A multi-billion dollar company competing on the same level as a non-profit. Facebook own AI research site lists 4 projects, not including a broken chat bot. How could I have possibly arrived at that conclusion? https://research.fb.com/projects/ https://research.fb.com/projects/

alexandercrohde 9y ago

What a junk article. Link bait. "Humans can't understand" is ambiguous, and the article wants you to believe this means "above human comprehension," when really it just means "makes up some new language that we could absolutely reverse engineer."

dan-compton 9y ago

This article is terrible. I still assert that a relative black box will lie at the core of any "strong" AI.

ajaygeorge91 9y ago

who is up voting this shitty articles

jumpkickhit 9y ago
I submitted, so it's my fault. I ignored the title and thought the content itself was interesting. That AI appeared to be communicating with a method it figured out itself. That's what interested me enough to want to share it here. I'll use more discretion next time i submit something.

anotheryou 9y ago

Please also stop these fresh people, I don't understand a word. And these protocolls, http and such. I can also not understand them!

IceDane 9y ago

What a load of horseshit. Here's what you need to do before publishing an article on AI: 1. Hire an expert that actually knows this stuff 2. Have him read the article 3. Throw the article out when he tells you to throw it out 4. Feel shame Stop writing these imbecilic, clickbaity shit articles about AI. This leaves people that don't know anything about AI or neural networks thinking that we're somehow developing skynet, instead of just writing a computer program which essentially implements a mathematical model that continually adjusts itself based on some metric.