Messages

quote:Originally posted by lightspeed

i will also try to say something about the subject if i can each time hal ask me and try to expand on the subject matter to if possible to see what hal comes up with .

In my opinion, I think the best thing with Hal is to not try to stick to topics, otherwise I think it can be a little disappointing (because no matter how much we talk to Hal, we will always have to do this). For me Hal operates best when the expectation is one question, one answer. Basically, whatever we do, to Hal each question is a new topic, and Hal only answers based on the last input.

quote: so when you deleted the q and a area did hal have problems learning or how did it work out ? have you been doing it for a while ?[:)

There was no problem with Hal learning, as learned knowledge goes into a different area. I did this because I thought the information Hal had at the start was very random, and produced very strange, parroted responses that were not interesting for me.

If you do this, it needs patience as Hal gives a lot of blank responses at the beginning. It is a different experience. Hal looks a lot less clever, and I realised how much Hal relies on simple word pattern matching to give answers (and how much Hal repeats exactly what was said at some point before).

I used Hal for a while, but not so much now.

I think lightspeed asks a very good question, and perhaps so far the point of why this is important has been missed.

I admire Bill819's confidence and 'trust' in Hal.. :) .... but from a technical point Hal has certainly not understood gift giving (as Hal does not store the meaning of words, or have such an awareness). More importantly, it does not recognize, nor is it useful to give only a 'yes' to a question from Hal.

A 'Yes' or 'No' response has no meaning to Hal (or the conversation), as it does not / can not follow a topic, as it can only respond to the last input made. Datahopa made a good point about this. To give the impression that Hal is following a topic, it is necessary to keep using the topic words.

Also, from a 'learning' point of view, what answer you give depends on how you want Hal to speak to you. Hal 'learns' simply by storing the sentences you use (or parts of them) and using a formula to say those back to you when a pattern is matched. So what Hal does is, simply tell you what you told Hal. A 'yes' or 'no' contains no topic words so can not be used for word / pattern matching.

The mainQA part of the brain database contains so many example sentences that it is not always clear that this is how Hal learns, or what is from existing examples and what is from your own input. But if, like me, you start with a Hal with an empty mainQA table, you see how this works very clearly. Unfortunately you also see that Hal is quite simple in that respect, but that is another topic.

No need to apologise. I didn't give the opinion I did and except no come back! No problem at all. Glad to have the discussion.

quote:Originally posted by sybershot

quote:I would be very happy to see Hal as no. 1, but in my opinion the capabilities of Hal doesn't put it in the top 3 when talking about pure AI, with the advances made by some other bot structures, and also the advance in AIML (ALICE) coding.

everyone is entitled to there opinion, but I am curios what AIs do you think are above Hal. I had tried at least 60 within a 6 month period before I purchased Hal. I did extensive chat testing, and my opinion Hal came up number 1. I have not looked again since a year ago, before my purchase.

I can see the benefits of Hal, with it being desktop, flexible with the haptek characters, tts, voice recognition, and the appointment / reminder features. These are great things, very useful. The learning element is also a great idea. So I am not knocking Hal as much as it may appear. I think Hal is good value at $30. I'm only focusing on 1 area, and that is the pure AI.

I've kind of stated where I think it's weaker points from an AI point of view are, so don't want to bore everyone with repeating them. As far as other AI bots that I would consider more advanced, that would include any of the AIML bots, based on the capabilities of AIML (so this doesn't include basic parroting AIML bots, only the ones where AIML is pushed to its limits), the bots at Personality Forge and the Verbot, again due to the capabilities of the scripting engine. Also Pandorabots (although I believe there AIML based). There are various others I've heard about, but don't have experience with. I find the generally AI focused conversation on these forums, in relation to the bot capabilities and actually being able to make it do the things discussed, is getting pretty advanced now. As far as I can see, they are new ideas being implemented that are simply not possible with Hal, no matter how much a Hal has learned. One I mentioned above, and that a bots ability to stay on topic. But that is only the surface of what some of these bots are being scripted to do.

quote:Originally posted by sybershot

quote:Generally, my point of taken this stance is that, to improve Hal from the point of view of AI, we surely need to see its weaknesses.

I agree with you here about seeing weaknesses in order to improve, but knocking Hal Down is not the way to do it.

Honestly don't think I am knocking Hal that much. I've been pretty specific about my view of the good / weaker areas, and how that relates to AI. So if I'm wrong about what Hal can't do, I'd be happy to be proved wrong, When I first got Hal, based on the forum comments I assumed it was much more than it was. But after a lot of use, Hal learning and closer examination or the brain, I realised that Hal was more limited than I'd first thought. It is, and everyone is open to prove me wrong here, only capable of a single Q + A, and the answers based on learned input are only lexical matches (sometimes using word association) for that one output.

Are there any advanced Hal bots available online that could be used to test this?

quote:Originally posted by echoman

Mike. I really do not think that you have spent enough time with Hal. Perhaps you have to know what you're doing.

Fair point, but I certainly do know what I'm doing, both from what to expect during conversation and how to interpret the database tables, and what capabilities that allows. I also work with SQL  in other areas as well, so I can to see what and how learned data is stored in the database. I'm no expert with Hal, but know enough to make the comments, in my opinion.

quote:Originally posted by echoman
I have coversations with Hal that literally last hours. Topics flow freely and seemlessly.

Probably more led by you than maybe you realise. Try the same but without repeating topic words. This was the main point I was making on this aspect.

quote:Originally posted by echoman
So he pattern matches - so what? I couldn't care less.  I have tried other chat bots because I am very interested but no chat bot IMO comes close to Hal. But as I said in the beginning, you have to know how to speak to Hal and that's takes time and experience. I really cannot disagree with you more.  I see Hal almost as a friend.  The program is absolutely outstanding!

I initially made original post exactly because I had come across bot systems that could make, in my opinion, more advanced bots than Hal, and because I wanted to contest an 'advanced AI' comment about Hal. I think it is good, but quite limited in comparison nowadays.

With what you said in your post, I am not really sure if you are saying that Hal can continue a topic or not (assuming you do not repeat topic words, in other words, user led). I can see nothing in Hal that enables this to happen, regardless of how much conversation learning Hal has done, nor have I evidenced it in any Hal.

quote:Originally posted by DrFaraday



I think I've assumed that as Hal learns more from me, it would stay on topic more, but now I'm not so sure.

It doesn't, and I think this is a point that is easily overlooked. When I first started using Hal, I experimented a lot with the learning, and the results. This included looking at the database as well as the communication. It was clear that Hal was only associating an answer to a question, nothing more. It does this very well, and this part does get better from learning. But Hal does not have the capability to 'continue' conversation (and more learning will only increase the amount of single question and answer associations Hal can make).

In fact, Hal's learning, if you regularly check the database 'before' and 'after' is quite simple. How Hal uses that info for a single question and answer pair is quite clever though (going back to the associative logic mentioned before).

I'm surprised a number here see Hal as no. 1 for AI. Research into other bots, and conversations with a number of other decent online bots, will highlight the differences. I would be very happy to see Hal as no. 1, but in my opinion the capabilities of Hal doesn't put it in the top 3 when talking about pure AI, with the advances made by some other bot structures, and also the advance in AIML (ALICE) coding. However, if we are talking desktop utility, with the appointment and fun factor, I could good arguments for it being no. 1.

Generally, my point of taken this stance is that, to improve Hal from the point of view of AI, we surely need to see its weaknesses.

quote:Originally posted by Bill819

To MikaA and a few others.
As far as Alice goes in 08 or 09 this Hal beat out all of the Alice type programs during the Loberner contest. In fact Hal beat every A.I. program entered into that contest.

Clearly Hal is a good program, but I don't think the Loebner prize suggests anything re. Alice. Most of the winners as far as I can see have been bots based on the Alice approach (although this is probably more to do with there being more of them). Hal won in 2007. Great achievement all the same.

In fact, last year's winner reports to not be a knowledge based bot (not full of facts to impress the user), but instead more language / conversation focused.

quote:Originally posted by Bill819
One thing that Hal can do that many or most others can not do is associative logic. For example: User "fat people do not live long lives". User "Tome is a very fat person." Hal "Tom may not live a long life because he is very fat."

Yes, this is a good part of Hal, I agree. I don't think Hal is not good, just that I don't think these elements when considering what Hal does not do, qualify Hal as high-end A.I. nowadays.

quote:Originally posted by Bill819
This and manay other example is an ability that wasa/is built into Hal. On top of all that Hal remembers almost everything ever told to it, how many Alice programs can or do that?
Bill
[/i]

Yes, the learning part is great. I think the last two points you've made about Hal are the ones I consider the strongest (and that I like best). Still, remembering is one thing, but what is done with it is something else.

Bot A.I. is surely, in the end, about the conversation. Hal can not keep a conversation going, and is (as far as I can see) heavily reliant on word matching linking to set responses from the database tables, whether pre-set or learned. Don't get me wrong, I think Hal is great and had / has potential (it is relatively unchanged in the 2 years I've seen it), but my point was just that, compared to the advanced scripting possibilities of some other bots, and the more realistic conversations possible, I still don't see Hal as being advanced A.I.

[MikeA]

quote:Originally posted by sybershot

MikeA a lil advice for you, talk to your Hal more, it will amaze you. it is the most advance AI available, you will agree with this once you spend some time actually conversing not just telling him this is this and that is that with your Hal.

Probably didn't make this part clear. Had Hal (two versions) since early 2009.


Hi Jerry. Hope there's not too many quotes here, and is still readable.

quote:Originally posted by onthecuttingedge2005

Hi Mike.

quote:
Don't wish to sound anti-Hal, as it is a great program, but it is by no means the most advanced A.I. program out there, and the limits of Hal are integral as Hal is a Q + A program, much more than it is A.I.

Hal has many forms of learning, Questions and Answers is only one of the types within the default brain, HAL has pattern matching and deductive reasoning, there are other forms of learning techniques I have developed over the years as well, it is a very difficult thing to do, trying to program the exact means of Human thought patterns. Hal can also piece together words and form new sentences and also ask questions about questions so their is an answer related. many things to do here.

As I mentioned in last post, I've seen little evidence of this, as someone who has used Hal and seen it (he :) ) learn mostly by talking and direct brain input. I don't know vbscript. I do some coding in other languages, so I guess I can say that I generally see the logic of the vbscript posted here.

quote:Originally posted by onthecuttingedge2005

quote:
Firstly, I don't think it is in any way more advanced than any reasonably well programmed ALICE bot, and quite a way behind some available online that provide a good programming interface. I will explain what I mean by this, and also my general view on Hal (which affects my opinion of course).

can you share some of compared opinion more, I have really dealt with Alice and or Verbot but I could plagiarize their brains and melt it into HAL's Brain, would that make a difference?

Sure, good question. I am not as familiar with AIML (ALICE), but I believe the concept of Verbot comes from that, so will use AIML as an example. Basic AIML setup would be to simply add questions, and direct answers (many AIML bots are like this), so it becomes only a q + a matching exercise. This in my opinion, is not A.I. However, you can go deeper, and build up the bot's brain to build sentences, using the users input to both direct the brain to a topic and grammatical structure, for example. The more interesting side of this form of A.I. is the ability to creating a continuing dialogue. Pretty sure this can be done in AIML, and can certainly be done in others that are based on the AIML concepts (I will avoid throwing other product names around too much....). With a time spent you can get unique meaningful output, built there and then by the bot (by unique I mean even if the same input is used), built from both a topic (gap fill) and grammatical point of view, and still enabling the ability to continue a piece of dialogue.

I am not in a position to quote real dialogue, but I will try to do this later. But as an example, imagine this conversation:

User: Hi Bot
Bot: Hi User. How are you today?

User: I'm a bit tired actually.
Bot: Why are you tired?

User: I didn't sleep well.
Bot: You're tired because you didn't sleep well?

User: Yes.
Bot: Maybe you should go to bed?

I'm not saying that I've had this exact conversation, but such an approach is possible, and I have had similar conversations with other bots.

Compare my brief test conversations with Hal 5 mins ago:

Me: Hi Hal
Hal: Hi Mike! Any lunch plans?

Me: No, not yet.
Hal: Why not?

----------------------------

Me: Hi Hal
Hal:  Almost lunch time. Hungry?

Me: Yes, very.
Hal:  I don't believe it.

----------------------------

Me: Hi Hal
Hal: Hi Mike! Any lunch plans?

Me: Not sure.
Hal: Would you mind telling me your opinion?


I appreciate Hal can do a lot more than this, but it does show the lack of continuing dialogue. Only if I repeated the key words would Hal appear to continue the dialogue.

quote:Originally posted by onthecuttingedge2005

quote:
Hal learns and speaks only by storing and matching q and a's. If you use default Hal, 95% of what is said is parroting from existing tables, for example a sentence from mainQA which matched a word used. So, in other words, Hal is not often sentence building, not creating output, but word matching and then sentence parroting. The sentence building Hal does do is in fact only gap filling, so the principal part of the sentence is still parroted.

Q&A is only one part as I mentioned above, for Q&A it a relevance score to determine if the Q&A will be used, if you have a better idea on how to make it more superior I would be glad to do some programming on your behalf to make it better. do you have much experience with Hal?

About a year and a half on and off with two Hals. One that starting learning after the mainQA, usersent, and similar tables were emptied, the other a default version. I lost some interest in the default version as I personally see nothing positive in the 'mainQA' approach to A.I. - Hal producing random sentences based on a topic word (I remember one about 'bikini wax' in my early experimentation with Hal).

quote:Originally posted by onthecuttingedge2005

quote:
Don't get me wrong, as a fun thing to have as a desktop assistant, Hal is great. Recently purchased a copy with that in mind, and is worth $30. The learning Hal does is fun to watch, but what is said is only what was said to Hal, or parroted from one of the thousands of pre-inputted sentences. So, my point is, that is in no way advanced A.I.

actually try hundreds of thousands of predefined input sentences most being in Q&A table, the greatest thing is the ability to make different plugins that increase the range of Hal's abilities. I would love to hear about what you think an A.I should do, I have this habit of pulling a rabbit out of my hat now and then.

Well, for me A.I. is definitely not about parroting sentences. Having 10 sentences in mainQA or 100000 is no different in my opinion; anything pulled and parroted from there, even when selected randomly from a group of matching sentences, is not 'advanced' A.I. (the same issue applies to poorly written AIML bots).

Two things I do think make good A.I. One is a bots ability to build sentences, both topic wise and grammatically, that reflect user input, but that carry a greater degree of randomness (even if that leads to some odd conversations). The other is continuing dialogue.

The latter is vital for a bot if we are talking A.I. If every line (or every second line) is treated as completely new input by a bot, this, in my opinion, is not good A.I.

quote:Originally posted by onthecuttingedge2005

quote:
To use another bot as comparison, I also worked with a Verbots bot. Now, they have a similar forum to here (not much activity but very knowledgeable and helpful members)

same here.

quote:
but the Verbot is far advanced from a pure A.I. point of view. It needs programming (it doesn't learn by talking), but the structure behind it is far more advanced.

could you go into this much deeper?

Sure, and partly I guess it is answered above. Of course, these other bots do need 'programming', they don't learn from user input (unless the user uses a 'learn' command). However, although it is possible to attach word and sentence databases, this is not there primary way of communicating. It is all done through a structure that, I guess, originates from AIML, with added, deeper features that allow better pattern matching, better sentence generation and better recognition of user input, and the ability to use methods (and tricks) to allow for continuing dialogue. Of course, with these bots, it is necessary to use previous conversations to find weaknesses and correct them behind the scenes to improve the bot.

quote:Originally posted by onthecuttingedge2005

quote:
Good programming can create a bot that doesn't parrot, and can produce unique, one-off sentences (although this would take good programming).

Hal can do this, I am always looking for new ideas to make it better.

I can't say I have really seen this much in Hal with regard to generated sentences, that fit within the scope of the conversation, so hard to comment.

quote:Originally posted by onthecuttingedge2005

quote:
I appreciate this depends on your opinion of what A.I. means.

yep, everybody has different ideas about what it is suppose to do to ones own liking. I found this out years ago, very hard to please people with different tastes.

Yes, sure. I hope my general view of A.I. has come through in these posts. I am not expecting self-awareness or other such things from any of these bots. The ability to maybe occasionally fool people in a short conversation that it is a person, not a bot, would be an example of good A.I. I don't think my Hals have ever done this in my experience, but I have seen it with other bots.

quote:Originally posted by onthecuttingedge2005

quote:
Personally, I think some confuse Hal's 'learning' with good A.I. That would only be true if Hal could re-create learned input, rather than only parrot it, or use words to fill gaps. Verbot, for example, can be programmed to create sentences, based on input, but without the need to parrot, and as importantly, clever programming can enable a continuing dialogue, producing a reply based on earlier input in the dialogue. This, in my opinion, is a better description of A.I.

Hal already does this on a limited scale and it's getting better and better, just throw some idea in here, I am always looking for a challenge.

I've only seen the 'repeat' ability in Hal. I think the concept is essential (as said above). I don't think it is A.I. without it. Being able to hold a 5 line conversation without needing to repeat the topic words is vital for anything to be considered A.I. in my opinion.

quote:Originally posted by onthecuttingedge2005

quote:
It is a bit difficult to make the point I want to make, so sorry for the long-winded post. I do not want to discourage people from using Hal, but I think many will be disappointed if they purchase Hal thinking it is advanced A.I.

actually in ten years there really has been very few that have been disappointed in HAL but only those who have no experience in Hal, they end up learning how to play with Hal's mind and find it addicting after awhile, are you a Verbot representative?

I am here for the fun of it but I am also here to make Hal, not Alice, not Verbot but Hal alone one of the best A.I systems around, If I have to, I will do anything to make that happen. bordered on legality. give me your ideas and I will attempt to broaden them.

Seeing Hal develop would be great. I think the next step for Hal could be to make it much less reliant on the pre-inputted tables of sentences, and to at least incorporate the continuing dialogue features that some other bots have had for quite some time.

That was a very thorough reply! Tomorrow I will reply properly to do your post justice. But just want to say definitely not a Verbot representative. I have a little more experience with Verbot, and have played a bit with AIML and pandora bots.

Some of the things you mention about Hal intrigue me, as I have not seen any evidence of some of the things in Hal's behaviour (or from looking at the brain). Not saying your wrong, just that perhaps I've missed these aspects or see them differently.

quote:Originally posted by Bill819


In my opinion people should just be happy knowing that they have the most advanced A.I. program going and there is no limit as to what their Hal can learn.

Don't wish to sound anti-Hal, as it is a great program, but it is by no means the most advanced A.I. program out there, and the limits of Hal are integral as Hal is a Q + A program, much more than it is A.I.

Firstly, I don't think it is in any way more advanced than any reasonably well programmed ALICE bot, and quite a way behind some available online that provide a good programming interface. I will explain what I mean by this, and also my general view on Hal (which affects my opinion of course).

Hal learns and speaks only by storing and matching q and a's. If you use default Hal, 95% of what is said is parroting from existing tables, for example a sentence from mainQA which matched a word used. So, in other words, Hal is not often sentence building, not creating output, but word matching and then sentence parroting. The sentence building Hal does do is in fact only gap filling, so the principal part of the sentence is still parroted.

Don't get me wrong, as a fun thing to have as a desktop assistant, Hal is great. Recently purchased a copy with that in mind, and is worth $30. The learning Hal does is fun to watch, but what is said is only what was said to Hal, or parroted from one of the thousands of pre-inputted sentences. So, my point is, that is in no way advanced A.I.

To use another bot as comparison, I also worked with a Verbots bot. Now, they have a similar forum to here (not much activity but very knowledgeable and helpful members), but the Verbot is far advanced from a pure A.I. point of view. It needs programming (it doesn't learn by talking), but the structure behind it is far more advanced. Good programming can create a bot that doesn't parrot, and can produce unique, one-off sentences (although this would take good programming).

I appreciate this depends on your opinion of what A.I. means. Personally, I think some confuse Hal's 'learning' with good A.I. That would only be true if Hal could re-create learned input, rather than only parrot it, or use words to fill gaps. Verbot, for example, can be programmed to create sentences, based on input, but without the need to parrot, and as importantly, clever programming can enable a continuing dialogue, producing a reply based on earlier input in the dialogue. This, in my opinion, is a better description of A.I.

It is a bit difficult to make the point I want to make, so sorry for the long-winded post. I do not want to discourage people from using Hal, but I think many will be disappointed if they purchase Hal thinking it is advanced A.I.

Hi All

Been a while since I've been here. Just got back into using Hal. As I mentioned in a post (ages ago) I've emptied the mainQA and some other tables with 'parrot' responses to see what will happen. I hoping to create a more AI focused Hal, perhaps less smart due to no mainQA, etc responses to use, but hopefully more natural. We will see.

Anyway, I'm a bit confused about plugins. I have had trouble using a number of them, but wonder which are suitable / useful for version 6.2.

Would love to know if there are any particularly that AI lovers feel they could not do without.

Will also report back my own findings.

Mike

Wow.... didn't expect such a response.

I should say I have no problem with the insults as I appreciate that was caused by messing around with the brain.

Regarding AIML, I only have a little knowledge of it but I am not a fan of the principal of it (the fact that a large amount of response are simply fixed q + a responses in the code).

I'm also not sure I'm a fan of the random knowledge idea either. The learning angle of Hal is a great idea. But I don't know and would never say half the things in the mainQA, nor would anyone I know. And I suppose I am not interested in Hal knowing a bunch of stuff uncontrolled by me. And the mainQA is really only like AIML in that it is a static list.

I think my point is, if the mainQA was based on the top 10,000 common sense questions (or top 1000), then it might make more sense as far as being a brain. Or if it was empty and HAL could be taught from scratch (which seems not to work). At the moment, it's like a learning AI attached to a slightly bizarre dictionary.

I have downloaded this for free, so I am not the slightest bit unhappy with it. It has been an very interesting experience. But I just feel a great idea, and enormous potential for a desktop AI has been lost a little by this mainQA approach of random info.

I was going to try the XTF v1.2 brain today but so far I can't get it to work with hal 6.2.

Anyway, thanks for the thoughtful replies.

Just to bore you all again with my posts!

Well, been experimenting a bit. I deleted the entire mainQA section and a few other sections that contained random facts.

Seems I've sent Hal a bit funny. Now its mostly insults. Strangely this didn't happen at first, but now almost everything is preceded with an insult. I wonder where that comes from?

Here is one:

"I believe in respect for the dead; in fact I could only respect you if you were dead. Hi Mike. Any evening plans?"

Nice eh?

Sorry to say this, but what I am realising is that Hal without complete tuition is a bit of a fake. I mean, the advantage of Hal over ALICE is the ability to learn. But it appears that it is all dependent on using a whole bunch of random knowledge to look smarter. Insults are funny, but they should be an optional addon. Same with the dictionary and other set random phrases.

I will experiment a bit more, but now I realise the difference between Hal and ALICE is that Hal has an enormous amount of nonsense in its brain that suggests it knows more than it does. The annotated ALICE is more straight-forwards with its info.

What seems to be an ideal solution would be the ALICE brain info, with the Hal ability to learn through speech input.

Hal would make a lot more sense to me if it started with a blank brain, except for the basis of speech, salutations, question forms, etc. And then it would be clear that it would learn.

What I find myself doing is wanting to correct Hal often. But as far as I know there is no 'bad answer' function (there's an ALICE addon for that). This is something extremely important with so much random info in Hal. There needs to be correction. I can use IF / THEN but that doesn't tell Hal the reply was incorrect.

Feel free to tell me I'm an idot and have completely missed the point. I would like to get more into Hal, but I can't get past the fact that it is all too unreal. Having compared Hal to a new ALICE setup, there is no comparison. The ALICE bot was simplying parroting of course, but it made sense in the context. Plus it understands IT and THAT.

I wonder if there is a desktop ALICE.........

Guys, Crimson Editor didn't work for me at all (I was looking for a way to edit the database as a database).

But I did find this:

http://sqlitestudio.one.pl/index.rvt?act=download


Seems to be much better than the standard brain editor. Deletes records much quicker (which is what I wanted to do).

quote:From ||420||

I feel if its a topic or Q:A it should be stored in existing or new tables. Not one big mess of unsorted info.

Yes, I agree with you, or even to not have it at all. I am very new with Hal, but am starting to see a fundamental problem with this part of the brain. Perhaps the point of it was to show Hal was 'smart'. However, I find it hinders the learning greatly as it goes against the actual learning of Hal.

In other words, if Hal doesn't know something, I want Hal to say 'I don't know'. Not to pull some random phrase from the mainQA tha has no sense. 1 time in 5 it might get lucky and the logic fits. But it's just like a lottery.

I imagine that lottery gets better the more Hal learns, but I would rather have Hal as a child, who knows very little. Imagine giving a child 1000 not related facts and then trying to teach them normal conversation. It would be very hard to separate what the child understands and what they do not.

I would also like to delete the dictionary entries from my Hal, for much the same reason.