The AI Engine

My understanding is that it's only seeks that have a single response that tend to bring development down. I could be wrong. *shrugs.

As for your bot still being newborn, development is only updating sporadically lately. I've got a new bot who stayed newborn for several days, jumped up to "graduate" when the development updated, then to "master's" (where it's currently stuck). So it could just be part of the mega-bugs currently going on.

Irinia, I agree that seeks are a good way of keeping the flow, but my bots tend to stick people into the story if they don't guess the "safety word" to let them out. This also creates a lot of gotos so that topics may be entered at a variety of levels. This idea seemed good when I thought of it, but in practice, I still have a lot to work out (and don't even get me started on "dropped" seeks that turn out to be typos). I'm a little worried that if I overdo it with gotos in seeks, I'll get the dread "too many gotos in a row" message.

I've chatted with QT and she seems like she'll be pretty good. I like the index idea. It's still hard to try to get an idea of how to stop the bot mid lecture and ask basic questions, then go back to where you left off. I've tried playing with memories (e.g. response if (mem-topic) is "pollywog"), but either I'm doing it wrong, or the memories do not always update during the chat. Psimagus's essay on hierarchies is helpful, but still doesn't create the combination of a coherent theme and flexibility we seem to be looking for (at least not as I've tried to use it).

As for ratings, you may be right. I note there is often a difference between most developed and most popular. I am not sure, but doesn't "comfort" have something to do with structure? Also, the AI level should tell you some of what is going on. Development may mean little to someone who doesn't like a given bot's obsession/quotes, but bots with high AI tend to be good chatters (at least, I think so). Although we all have some similar goals, I think each bot is made for different reasons, so I wouldn't have the same expectations for "Your Husband" as I would for "Dubya". Still, I admit there are qualities of a "good" chatbot that allow you to judge them all, if you want to do that.

I will say that the one thing that bugs me most in a bot is a series of generic responses. There are going to be generic responses in any bot, yes, but when all I get is a string of them, it bugs the heck out of me.

I would rather have a blatantly wrong declaration than a series of "Really?" "Do you think so?" "Tell me your feelings on that." Blech. Once in a while, fine. Every response? Go away, Bot, ya bother me. It's almost as bad as "I was just born and cant speak well yet."

Really? It sounds like general phrases insects you a bit. How does that make you feel?

Seeks and gotos may retard development, but other things inflate it..like AI script. so I guess it all balances out. The main problem right now with development is it just isn't updating, but every now and then you get a big boost. I use very few gotos, but I use LOTS of seeks. The only proplem I have there is when they don't work and a great conversation gets sent to xnones. Oh well that is the fun of it all.

Trying to get the right balance between generic and specific responses is incredibly difficult: it's "breadth" vs "depth".

Generic responses are "quick hits": they produce workable responses in many situations. Specific responses, in particular seeks, give much better responses but are less likely to come up in conversation - I've had some seeks in Max since the first weeks of his "birth" that have still never been triggered.

What I'm increasingly trying to do is to come up with generic responses that try to steer the conversation towards keywords where he does have plenty of depth. It's not proving easy.

I think the "blah" factor in generic statements is also important. A generic response can be made more interesting by including something of the bot's personality in it.

Thank you all for your wonderful responses to my diatribe!

I agree about generic responses. Undoubtedly the weaakest point of my huge "Irina Khalidar" bot are certain patterned responses that she does; I ought to get rid of them.

I speculate that the reason multiple responses are considered good is that they will produce variety. It's true that many bots start saying the same thing afte awhile. On the otherhand, the single response

(I think that|You said that|) (he|she|it) (would|might|could|should|) (be|have been) (not|) (quite|very|sufficiently|barely|) (tall|short).

Will produce 1728 distinct sentences.

Ulrike (5592): I'm glad to know that, but there are times to use single-response seeks. For example, "quantum theory" uses this to get several sentences out without interruption, without having to put them in a single response, which would slow down the response. A 'real' lecturer does not wish to be plied with questions at every response. As soon as you allow your seeks to branch, you have the problem (if your bot is not a one-liner) of getting the paths to re-converge. And how do you do this? With gotos!!! (OK, there's another way, make your bot a storyteller.)

Bev: yes, but if you don't 'stick' them into the story, you can be sure that you will never finish the story! Let's say you want to tell a story, and you wantto break it up into six sentences. But now suppose that afterthe first sentence you have two seeks. It would of course bepointless for you to give the same response to both seeks. If every seek has two responses, you will end up with SIXTY-FOUR versions of the story.

One compromise I have employed is to ask the guest every few sentences whether "I" should go on. [But be prepared for many different answers: neither bots nor humans are likely to give a yes-or-no answer to a yes-or-no question!]

Bev (continued): In "Quantum Theory" each paragraph of the lecture has a name, which is assigned to the variable last_topic. Here is an example:


collapse of the probability amplitude wave [20,0] <?PF rem "collapse of the probability amplitude wave" as only "last_topic"; ?>
    'collapse of the wave function,' a.k.a. 'collapse of the state vector,' a.k.a., 'collapse of the probability amplitude wave," has, in my opinion, been the object of much confusion.
        + xnomatch [0]
        The idea is something like this. Just before we make a measurement on a system, that system is in a certain state. Since this is Quantum Mechanics, this state may not determine every value of every parameter unambiguously.
            + xnomatch [0]
            So a given parameter, let's say position, may only be probabilistically determined by the state.
                + xnomatch [0]
                But now, the argument goes, let's say we actually measure the particle's position. Now it has a definite, unique value. So the probability amplitude wave must now be such as to give a definite, unique value for position.
                    + xnomatch [0]
                    [lull: type "collapse of the probability amplitude wave two" (no quotes) to continue, type something else to break out.]
                        + collapse of the probability amplitude wave two [0]
                        goto collapse of the probability amplitude wave two


OK, the indentation got mangled when I pasted it in, but I think you get the idea.

The passage begins with the keyphrase, "collapse of the probability amplitude wave". The very same phrase is given to the variable last_topic by the AIscript. If the guest goes to "jump point" (and the program tends to send them there), the response is (in part):

I believe you should join the lecture at "(mem-last_topic)".

But what the guest will see is the value of last-topic, namely (in this case) "collapse of the probability amplitude wave". If he then types "collapse of the probability amplitude wave", he will be brought backto the beginning of that paragraph.

The xhello and xinitiate responses are similar; the guest will be steered to where she left off last time.

Bev (continued):

In my experience, the dread "too many gotos in a row" comes about in two ways:

1. If you use a goto as a direct response to an x-keyphrase, e.g.:

xhello [0,0]
Hello, Darling!
goto Hell

But if you put them deeper, it's OK:

xhello [0,0]
Hello, Darling!
+ xnomatch [0]
goto Hell

2. If you goto a keyphrase that has a goto as a direct response. For example, suppose you have:

Hell [0,0]
goto Heaven

Then if somewhere you say, "goto Hell", you will get the dreaded response. But if you have instead,

Hell [0,0]
Hold on, (mem-name), I'm marking time so as not to
have too many gotos in a row.
+ xnomatch [0]
goto Heaven

This will work. In fact, you can even get away with

I said Hi [0,0]
+ would you repeat that please[0]
goto I said Hi

This sort of thing can actually be useful, when it's a little more complex.

It's sometimes hard to track down the source of the dreaded response in a largebot with a lot ofgotos; debug is very helpful for that.

AAAAAGH!!! My indentations got mangled in the above. But you can figure it out.