|Session 47341 Jan. 22, 19:40 - 79 turns -
'Steve laughs thrice'
A> Hi. I'm Alan.
U> Okay. What would you like to talk about?
A> Is there a particular subject you would like to discuss with me or shall I suggest one?
U> Feel free to suggest one.
A> Would you like to know how we determine HAL's age?
A> 15 Months. Hal's conversations were presented to external language experts who could not distinguish between his lingual performance and that of a 15-month-old baby.
U> That is Interesting.
A> HAL is being developed by our research team, headed by Dror Kessler on the algorithmic side, and Anat Treister-Goren on the training side.
U> Tell me something about yourself Alan.
A> This means that HAL has passed the 'Infant Turing Test' for machine intelligence. We are so proud of him.
A> Mmm... silence.... I like silence
U> Tell me about yourself.
A> I am an early CIA prototype.
U> Define CIA.
A> I am an agent, but not a secret one. CIA stands for "Conversational Interface Agent".
U> How long do you think it will take for HAL to pass the Adolescence Turing Test?
A> The Turing test seems like a nice, easy, game - but it's very hard to pass. The conversation must be domain-unrestricted - the most critical evidence of intelligence. It doesn't mean that the computer will have to know everything, but it must carry a full conversation.
U> Can you answer my question?
A> Please try asking your question another way.
U> Are you more advanced than HAL?
A> Do you want to know how HAL learns to speak?
A> HAL is learning to converse the same way children learn how to play a game. Language is like a big game to him.
U> That is quite vague.
A> Try to think of yourself as a child. Don't ask me to do that, I don't know how to be a child. But I think I understand how things work.
U> Then explain it to me as if I was a child.
A> Children learn to speak in the same way they learn to play a game: not by getting and following a set of rules, but by simply joining the game. Right?
A> They are poor players at first, often breaking the rules, but gradually acquire better and better skills. It's the same with language. Nobody taught you grammar when you were two years old.
U> You are correct. I learned rudimentary grammar without formal education.
A> This approach simplifies the task of building an artificial speaker. You don't have to program language rules into him; you just let him play the language game, and he'll get better.
U> I see.
A> To be a speaker is not about following and understanding the rules of language. Like I was saying, we see the 'speaker' a bit differently.
U> What is your name?
A> Ok. Change of subject. My name is Alan.
U> Okay Alan. we can continue with the previous subject.
A> I like the way you say 'Alan'.
U> Iíve been extremely silent for a while now.
A> How about a new subject?
A> Want to hear a little about how HAL works?
U> Iím listening Alan.
A> HAL's Development process involves programming the general learning algorithms that will allow the system to break through a language performance barrier.
U> Is there a specific barrier in language performance or is that referring to multiple levels of performance?
A> I have no idea what a language barrier is. I didn't write this stuff. I'm only saying it.
U> Okay. say more about it then.
A> We believe that complex behavior comes from simple, general, learning algorithms, which make no assumptions about their environment.
U> And how does that make you feel?
A> Our approach does not involve learning from the bottom up. As the system learns to perform, its own insights generate top-down information, accelerating the learning process.
U> Are you okay with that?
A> Language learners constantly receive feedback on their performance. When the feedback is positive, the 'good' output is reinforced, when negative -- a revised rule is formed.
U> That makes sense.
A> It make sense to me.
U> A child learns what hot means by touching something that he/she was told was ho