Facebook pulls plug on language-inventing chatbots? THE TRUTH If you thought artificial intelligence was already overhyped to death, this week will have given you a heart attack. On Monday, excitement levels among hacks hit the roof amid claims Facebook had scrambled to shut down its chatbots after they started inventing their own language. Several publications called the programs “ …
In 1972, ELIZA (as "The Doctor", at BBN (tenex?) ) and PARRY (at SAIL, on WAITS) had a conversation at the first ICCC ... Well, they had a conversation that was followed over the ARPANET during the ICCC. It was immortalized in RFC 439.
More leftovers from SAIL here. Not much has changed in 45 years ...
"I know people who claim to have found God in a good bowl of proper Ramen"I found Jesus a few years ago, though he wasn't in a bowl of Ramen at the time. He was much smaller than I expected and he had the Holy words "Product of Mexico" on one of his feet...
Oh dear, not a very promising start, or is it an ending?
Perhaps the real reason they've shut it down is that not so very long ago Facebook were promising to clean up / auto-moderate their website using AI as one component (in fact it would be the major component) of a censorship system.
If this is as good as their AI gets then auto-moderation is right off the cards. If the politicians work out that Facebook have been making it up then the politicians' patience may finally snap. Law changes are going to start coming thick and fast, and Zuck's baby is going to be landing some hefty fines.
Natural languages have plenty of rules (meanings, structure, etc.) that have been accepted by their users. Even if the meanings of words evolve over time, it's within an underlying framework.
Starting from a blank state and expecting machines to develop all that complexity to produce something that reads like Austen or Wilde is too much to ask. In the old infinite monkeys and typewriters thing, they're not starting blank, but will have a societal hierarchy so that the opinion of one outweighs the rest. AIs don't even have that, so will spend a lot of time babbling on trying to establish the rules of engagement, so to speak.
However, what the tabloid fuss fails to understand is that there may have been a real symbolic sub-code emerging and task information was being processed. If meta showed that one of the AIs was leading choices of words in the other, so they were deferring and showed understanding, then that's far more interesting. But it's also beyond the level of tabloid sensationalism and the "headline-soundbite" clickbait model they use.
OTOH it could have just been totally random, heading nowhere, and a waste of electrons.
Caterpillar, really? Shirley you mean Humpty Dumpty? Funny thing, they go on to translate the most nonsensical bits from Jabberwocky. My middle-school English teacher had us do that our own way, deciding what it might as well mean (long before I would find out there are 'official' meanings in this chapter), as an exercise. Out of that exercise, all I can remember clearly is the way someone else translated 'And the mome raths outgrabe.' as (cheering w/ arms raised) "And the fool committed suicide!"
The 'language' the bots produce is, at least on the surface, pretty similar to the 'language' that chimpanzees and bonobos produced in experiments to teach them language (using sign language or symbol boards) in the 1970s and 80s. Herb Terrace reports that the longest sentence produce by the chimp Nim was "Give orange me give eat orange me eat orange give me eat orange give me you." This is probably less sophisticated than the AIs - at least the number of 'me's in their sentences communicate something. All Nim (and Washoe and all the rest) seem to do is produce the words with no meaning attached to word order or repetition.
"All Nim (and Washoe and all the rest) seem to do is produce the words with no meaning attached to word order or repetition."
I think it's fairly clear from the example you quoted that Nim wanted to be given an orange to eat. A moment's thought should show that that involves a concept, eating, that's way beyond anything a machine could begin to comprehend but which is fundamental to the existence of any animal.
Oh I agree. There is no doubt that Nim could communicate, he just wasn't using what Chomsky (after whom he was slightly scurrilously named) would call language. For Nim, 'Nim eat orange' and 'orange eat Nim' meant the same. He was using signs, but not constructing them into sentences where grammar (word order) allows multiple meaning to be communicated with a limited number of signs.
This is probably less sophisticated than the AIs - at least the number of 'me's in their sentences communicate something. All Nim (and Washoe and all the rest) seem to do is produce the words with no meaning attached to word order or repetition.
Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo
If you really want to rail against the journo's hype of the story, then you'd gain some credibility yourself by refraining from statements such as:
"...Researchers from OpenAI found that agents would talk to one another in a kind-of Morse code when forced to communicate and work on a task together."
Which implies that this Morse code type behaviour emerged as a result of the two agents cooperating with each other, and the researchers were un-aware of this and discovered this newly emergent and spontaneously created behaviour.
Which is clearly a total load of bollocks. IF they communicated "in a kind-of Morse code" then it's because they were programmed to.
Actually, I'm more worried about 'trivial' things like this than killer robots. Here's a fairly famous paper from 1983: http://www.sciencedirect.com/science/article/pii/0005109883900468
If the next language is genuinely indecipherable and is used to exchange information between power plants, for example, then we need to worry somewhat. I agree that the whole story got blown out of proportion in the non-tech news outlets, but there is a worry there.
And yet, if you look at most Sci-Fi prophetic visions, communicating with the aliens (the ones not here to eat us) always starts with exchanging numbers, of defining a common system of numbers, upon which to encode all the other information. So, in a sense, one could interpret what they were doing as what would be likely from a non-sentient solution-finding system that has made contact with another similarly naive solution-seeking system. Makes you wonder if at some point the two would both try and run Turing tests on each other...and what the results might be, and would those results be anthropomorphized as "machine wishful thinking"...volition in a can!
I remember that if you use the most simplistic language learning model possible and train it on the whole of wikipedia it comes out saying:
the united states of the united states of the united states of the united states of the united states of ... etcetera.
Might tell us more than intended about wikipedia.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
