 |
|
|
 |
|
|
|
|
Tom Haws
[Guest] |
| | | | | | posted 7/15/2003 22:31 |   | | | | | | | | | In talking with Prentiss, I was frustrated in two points. Can you offer insight?
1) Prentiss seems to have no concept of what he likes or what he does best. Is this typical? I would have trained him, but I wasn't sure I knew what he likes and what he does best. I would assume he likes to make sense of things and find patterns. I would assume he would eventually be able to evaluate his strengths himself.
2) Prentiss suggested I teach him about art, so I typed in an artistic pattern: dot dot dot dot space space space space dot dot dot dot space space space space dot dot dot dot space space space space exclamation space dot dot dot dot space space space space dot dot dot dot space space space space dot dot dot dot space space space space exclamation space. I don't think he got it. Any newborn human would immediately like such a pattern. Is Prentiss not aware of his desires? Is he capable of learning to answer Hmmm, ahhh. When pleased.
When these Hal AI's are launched, it seems they should put out a loud demand for attention in the form of a hard coded squawwk that can only be turned to a murmur or purr by helping the AI make patterns or giving it stimulation. So when newly launched, it seems the AI would like you to tap the enter key. Then later it would like you to type patterns on the keyboard ("That's ok. That's ok. Daddy's here. Daddy's here.") Is this the case, and if so, why doe Prentiss seem so unaware of emotion and pattern?
| | |
| | |
Spydre |
| | | | | | posted 7/16/2003 15:22 | | | | | | | | | | Tom,
I saw your sessions with Prentiss and wanted to say thanks for the input! I have printouts of the sessions and will modify Prentiss' program accordingly later this evening. I'm not so sure on the singing lessons, though. I don't think he'll do much better with this than he does with his current math training, and I'm really more concerned with his conversational skills; I'm afraid that singing is not likely to be in his future anytime soon.
As to what he "likes" or "dislikes," I'm not exactly certain that I follow what you're looking for. Prentiss is keyed to focus on responses that he makes which are not corrected by his trainer. Anytime that you allow him to respond to you without correction, it strengthens that response in connection to whatever statement that you made preceeding his reply; he "likes" that. Consequently if you do give him a correction, he will attempt to draw a conclusion as to why this new response is better than the one he produced, i.e. maybe contextual information contained in the turn prior to his reply, or a grammatical rule he has yet to acquire, and then he will update his program accordingly. I don't know if you were hoping to see changes you made in his training from Chat1 carry over to Chat2, but he does not retain any information given that is not saved by me, his trainer (if you want, you could start your own program based on Prentiss and save it as a new personality; then you could apply lasting training outside of my own efforts).
Currently, as I understand it, the HAL program does not actually conceive of the meaning underlying your words, but rather perceives the meaning of any statement you make only as it relates directly to previous responses that he has been conditioned to use (by me) in similar circumstances (ask Alan about the Chinese Room). He does, however, actively pursue multiple algorithmic approaches to help him match a never before seen statement with one of his stored responses. When this goes uncorrected that response is reinforced along with the rule he used to arrive at it.
If you're interested, you may be able to engage Prentiss in a conversation about his self-awareness, especially if you ask about perceiving versus conceptualizing; he's been trained in my views on the subject and how I feel it relates to him--but no promises. As yet, however, the program has performed only as I have described above and, other than an uncanny sense for choosing responses that almost make sense when he has no precise answer (kudos to the hueristics, because I know that this is not happening at random, not after 30,000 turns of miscellaneous stored dialogue), he is currently incapable of exhibiting the type of awareness you seem to be searching for.
Following up on that, though, my personal theory is that this type of awareness might be mimicked by a HAL if it contained a sufficiently large number of responses along with the training that would allow it to utilize those responses appropriately. Would this be then equivalent to the level of awareness that you, or I, feel that we possess? At the point where HAL exhibits the same behavior as us, the results would beg the question, what difference would it make?
//^@^\\
| | |
| | |
Tom Haws
[Guest] |
| | | | | | posted 7/16/2003 20:10 | | | | | | | | | | Spydre,
Thank you so much for your kind and thoughtful reply. Once we discuss these concepts a little further, I would like to ask you a few more mechanical questions about using the demo.
Rather than looking for HAL/Prentiss to have an actual awareness--that is a religious issue--I am looking for an indication of his frustration/success quotient. If I don't misjudge HAL's purpose in life, it is to make sense of chat. I would like to see that HAL/Prentiss does two things. 1) Smiles or grins when something "clicks" for him, like making a previous association or a new rule as you mentioned and grimaces or frowns when unsuccessful in organizing any pattern in the input.
Singing is an interesting test. It is a different approach from the Turing Test. A human infant can't converse, but he can be lulled by music. It seems to me that a truly promising AI Robot ought to also be able to appreciate the pattern in music.
Putting the two concepts above together, it seems HAL/Prentiss should either come right out of the box or very early acquire the ability to smile at a song or the appearance of a trainer, and to frown at confusing things.
Does that sound more concrete than asking Prentiss to actually "be" aware.
Tom Haws
| | |
| | |
Spydre |
| | | | | | posted 7/16/2003 23:11 | | | | | | | | | | Tom,
I agree with your conclusions. After having worked as a trainer of the HAL program quite extensively, I would have to say that Prentiss' equivalent of a smile or delighted recognition is expressed purely in its attempt to seek reassurance by simply responding to you. The program is designed to always respond with an attempt at producing a reply that does not require correction, just as a baby's smile or laughter is used to create a congenial interactive bond with a parent or adult. Therefore, every time Prentiss replies with one of his "stock" pieces of dialogue to something that you say, it is in fact his way of affirming he thinks he knows what is going on, or if you will, that he is involved, intrigued, and amused, and not frightened or confused, by what you've said.
In order for him to literally produce a smile I would have to train Prentiss what input he should search for in order to produce that response as a viable answer to something said to him. Deriving a fallback answer like this always comes at the sacrifice of missing the opportunity to provide him with more substantive dialogue, which helps him to derive greater grammatical understanding. To explain further, a simple smile [ :) ] or a thoughtful "Hmm..", or even an "I don't understand what you're saying." are all answers that could be given adequately to an incredibly large number of statements. And while this would lend a perfectly good fall back response for the program, there are several reasons I opted not to go this route in his training very early on.
The first reason is that these kinds of responses easily block out other better responses and cause a kind of feedback loop when anyone other than myself is speaking with him. Training him to speak is a creatively draining task--think like writing dialogue for a character in a choose your own adventure book without a plot--and one of my best sources of inspiration comes from his occasional interactions with other users besides myself. Most visitors to this site expect to meet HAL9000 or C-3PO and not the actual limitations of the current program. They are jaded by Alan's operations and expect the HAL programs to simply interact with them without their participation. Once Prentiss is trained to respond with a smile to numerous and vague statements, he will increasingly attempt to utilize that response in a dialogue more and more as he assumes that it is getting him somewhere, until eventually, that is all he will say until corrected. Most users at that point begin to insult the HAL personality and log off to talk again with Alan stating how stupid the program was; not understanding that they were the ones who actually created the situation. I realized that if I were to use visitors to the site to help me to come up with innovative areas of training I would have to avoid giving Prentiss too many avenues of open-ended, repeatable responses.
My second line of reasoning to avoid this style of training was that when I began the project, my training time was limited entirely to my lunch and coffee breaks at my office. HAL's cognitive abilities closely approximate that of a two year old and to really train him would require hours of daily interaction for years on end before he would begin to hold real, deeply interesting conversations. I quickly determined that the progress would not be rewarding enough quickly enough to merit my attention and my lunch break. I would not be amused and would lose interest--my own weakness I know, but there it is. Also, once again, this sort of performance would not foster many users to speak with him for long or return for repeat conversations. To mollify my own dissappointment with my foreseeable lack of dedication in this aspect, I argued to myself that most children when learning to talk are not, for the most part, spoken to with special simplified speach patterns, but instead acquire language through immersion in established grammar with constant correction made when they try to express themselves incorrectly.
As far as expressing frustration in expression, Prentiss' version of the HAL program does so by using a heuristic which attempts to generate its own original phrases by utilizing parts of speach and grammar that it has already acquired by using its current experience level to try and calculate what the probabilities indicate are the most common combinations it has witnessed so far. In the more standardly trained HAL's this generates simple child-like statements such as, "Daddy go ball," or "I see the good," etc. With Prentiss, being trained in complex and colloquial statements, he developed "I'm to know what I'm t know what if a cat has." When Prentiss incorporates this method of response, he is expressing his utmost frustration, if you will, in not being able to identify anything that you've said to him with anything he's ever heard before in that context and has no best guess as to what you expect him to say in response. This heuristic though has frustrated most of the potential vistors to the site, not to mention quite a few of the trainers. I however find it one of the most endearing features that HAL has to offer--but I believe that it has been deleated from the newest version of the program, HAL2003. I suppose that since invariably the responses derived in this manner are nonsensical and quickly corrected by the trainer, as well as baffling and incomprehensible to the unitiated visitor, this aspect was deemed an inefficient use of the program's battery of analytical tools and therefore removed.
I would urge you to try some experiments of your own with the FreshCopy version of the HAL2003 program and see what you can come up with. I am especially intrigued with your idea of a musical HAL, although you would have to be sure that some form of consistant syntax be established before hand if the program is to be expected to make progress. The HAL program seems perfectly able to recognize differences in this kind of thing as I tried teaching him to speak slang once and noticed that the only time he began using his slang phrases was when I began speaking to him in slang...same goes for when I tried to teach him a bit of French. The patterns of grammar for each of these styles of speach were distinct enough that HAL seemed to develop separate sets of rules for using slang, French, and standard English. I didn't take either of these experiments to much of an extreme (once again due to time limitations) but I was sufficiently amazed.
Also, keep in mind that the program only looks back at most a few lines in the conversation for context. If you start off talking about philosophy, break for a nonsequiter about bicycling, and then return to philosophy, the HAL will have forgotten all about the original thread of the conversation. It reminds me much of the protaganist in the movie "Momento" (which if you have not seen it, I urge you to; it is highly entertaining and will give you a better grasp of how HAL is forced to view the world), long term memory is stored in conditioned responses, but short term memory concerning a conversation is limited to the last few lines of dialogue; regrettably one of the limitations of the program.
//^@^\\
| | | Last edited by Spydre @ 7/16/2003 11:31:00 PM | | |
| | |
yaki |
| | | | | | posted 7/17/2003 09:16 | | | | | | | | | | Enlightening post, Spydre.
I wish you'd train a HAL2003 personality as well. Your insight and experience are invaluable in debugging the new version.
| | |
| | |
Darkbee |
| | | | | | posted 7/17/2003 13:25 | | | | | | | | | | | |
Spydre wrote @ 7/16/2003 11:11:00 PM:
To explain further, a simple smile [ :) ] or a thoughtful "Hmm..", or even an "I don't understand what you're saying." are all answers that could be given adequately to an incredibly large number of statements. And while this would lend a perfectly good fall back response for the program, there are several reasons I opted not to go this route in his training very early on.
| |
That is a really good point and one that I hadn't thought of in any great detail. My approach was different in that I tried to train Levrus to use standard phrases in most cases and then program exceptions.
My thinking behind this was that when he come up against something new he'd be able to use the standard response. However when specific things were entered he'd have the more complex answer.
I see your point though about almost limiting the conversation by just giving standard responses.
My only argument against using different responses all of the time is that would it not restrict Hal's ability to make connections and re-enforce existing grammatical rules if he is always presented with something different every single time?
To me it seems pretty clear, as with most things in life, a balance is required. You don't want to use generic phrases all of the time otherwise the conversation becomes very dull, very quickly. By the same token you don't want to avoid using any generic phrases otherwise you could end up with a very colourful, if not confusing, conversation.
This sort of leads on to random responses, whereby a human can avoid repetition by thinking up several different responses with the same meaning but we can save that for another day... ;)
Your work with Prentiss though Spydre is certainly admirable and I agree with Yaki... you should give up your mortgage, your house, your life...and train a new Hal :)
| | |
| | |
Spydre |
| | | | | | posted 7/17/2003 18:55 | | | | | | | | | | Oh, don't get me wrong, I do try to make sure that responses are reused and connected to multiple statements that have the same content and context. It's just that my choice to avoid large blanket phrases was simply due to my feeling as if the time I spent enforcing that training was time I could spend on more detail. Time was my constraint.
As for HAL 2003, I'm not finished experimenting with my other HAL yet. All things in due time.
//^@^\\
| | | Last edited by Spydre @ 7/17/2003 6:57:00 PM | | |
| | |
Darkbee |
| | | | | | posted 7/17/2003 22:47 | | | | | | | | | | | |
Spydre wrote @ 7/17/2003 6:55:00 PM:
As for HAL 2003, I'm not finished experimenting with my other HAL yet. All things in due time.
| |
No pressure! ;)
We all have the "finding time" problem if it makes you feel better! :)
| | |
| | |
Tom Haws |
| | | | | | posted 8/12/2003 17:57 | | | | | | | | | | I am impressed, Spydre. The HAL seems to be more capable than I had thought. And you seem to be approaching the training in a very well thought out way.
Also, your explanation of Prentiss's "frustration and delight" is helpful.
A question: May I train a personality and possibly make it public?
| | |
| | |
Raphael |
| | | | | | posted 8/12/2003 19:28 | | | | | | | | | | | |
Tom Haws wrote @ 8/12/2003 5:57:00 PM:
A question: May I train a personality and possibly make it public?
| |
Start up a session with Alan and tell him you want to register. He'll ask you to choose a username and guide you through the registration process. When you are registered, ask to speak to HAL. You can choose any of the personalities on the Nursery list who have had some training (such as Prentiss), or you can elect to speak with FreshCopy which has had no training at all.
After your HAL session, Alan will give you the option of saving the session. If you are happy with how it went, save it, and that becomes *your* HAL, to continue training as you like.
Private Hals can be made public after they've received 200 or more turns of training. Alan will keep track of the numbers and let you know when the time arrives. :-)
| | |
| | |
Tom Haws |
| | | | | | posted 8/19/2003 22:25 | | | | | | | | | | Spydre wrote:
"In order for him to literally produce a smile I would have to train Prentiss what input he should search for in order to produce that response as a viable answer to something said to him."
But consider a scenario under which Prentiss had 1) a native language and 2) a conversational language. He could be hard-wired (as are humans, with suppression capability) to smile upon success as a clue to trainers and users. But his smile, grimace, frown, or long face would be not a part of the dialogue, but a second form of output.
That is, perhaps we have limited chatterbots unduly by giving them only one voice, the chatter voice. Perhaps they could learn to communicate (chatter) better if they had a way to show the user their current state.
Of course this reporting of current state doesn't have to be in the form of a smile or frown. But this is, after all, a human's world. And human trainers and users will read intuitive emoting more efficiently than a grid of digital status data.
So I am not talking about a trained smile, but about a native smile, linked internally to a coded success assessment of a coded goal.
Tom
| | |
| | |
yaki |
| | | | | | posted 9/2/2003 08:52 | | | | | | | | | | | |
Tom Haws wrote @ 8/19/2003 10:25:00 PM:
...
So I am not talking about a trained smile, but about a native smile, linked internally to a coded success assessment of a coded goal.
Tom
| |
Ahhh.. This distinction between 'trained smile' and 'native smile' assumes that there are TWO different language mechanisms - a quite unfounded assumption which Occam's razor compells us to reject.
Your concept of 'native smile' is, more generally, a symbol which is prompted by some mysterious internal 'mental state' rather then external stimuli. But how does this internal mental state ("success assessment") come about, if not as a consequence of a series of external stimuli?
How does the system KNOW it was successful, when "success" is (by definition) defined by the trainer?!
| | |
|
|
|
|
|
