This is even more bizarre.

Cxu vi estas inteligenta

Jes, mi estas frenezulino.

Cxu vi ne estas viro

Mi estas viro, cxu vi estas viro?

Vi sxajnas esti konfuzita

Ne, vi estas roboto.

Anyone else getting the impression of a young child telling their mother/father/caretaker "no, you stop being immature" while throwing a tanty and crying like an alligator?

Perhaps it's a "let's come back in 10 years and see how much more mature it's gotten!" (although, I suspect technology by that point would have moved along so much that we might actually be able to expect some decent conversational skills out of bots, other than childish name calling and repetition).

I use Cleverbot a lot in courses that I teach. It's different from most other "chatbot" systems, in one key respect. Everything that it says is something that has been said to it. Everything that you say to Cleverbot gets added to its "vocabulary" of responses. It's predecessor Jabberwacky is still online, and I believe Cleverbot uses the same vocabulary database (I'm not certain, though), which now comprises over 50 million responses.

Cleverbot appears to be more refined, however. I think it's more anchored with invariant "ideas" about itself and certain things, and therefore has a more definite identity. Jabberwacky seems pretty psychotic most of the time. Rollo Carpender has created some sort of Google-like algorithm for searching its vocabulary to find the most relevant response. I don't think it even tries to re-conjugate verbs before answering, but I'm not sure. For obvious reasons, Carpender is reluctant to disclose the inner workings of his bots.

If you try Elbot, you'll see a more conventional, scripted approach. And you'll get nowhere with Esperanto. Elbot, unlike Cleverbot, doesn't really learn.

That mention of Google got me wondering - Google are obsessed with the correct categorisation of information, to the point where even Esperanto is now, either manually or by just the accidental evolution of their search engine, relatively well supported in Google searches. Perhaps (although I doubt it) (Rollo Carpender?) Cleverbot and Google could get together and make some magic happen, although I don't see there as being much motivation at the moment, considering Google don't need a search engine that can converse, only categorise information and create relationships between data correctly.

BTW T0dd what sort of courses do you teach? Sounds interesting!

That Elbot is interesting, but it does have limited responses (although they are perhaps more precise, but the scope is limited). Although I was shocked when I typed in out of pure boredom the name for a certain male organ, and it replied "Oh a word association game? Ok, "scissors".
I think I prefer the EO speaking cleverbot!

ceigered:Perhaps (although I doubt it) (Rollo Carpender?) Cleverbot and Google could get together and make some magic happen, although I don't see there as being much motivation at the moment, considering Google don't need a search engine that can converse, only categorise information and create relationships between data correctly.

Interestingly, both Google and Cleverbot use the same technique for "learning". In actuality, there's nothing really "intelligent" about either system -- Google and Cleverbot construct output from Markov chains.

For some techincal background, a Markov chain starts with a seed text and converts it into keys and values: a key is a single element in the chain, and values are adjacent elements. So, staring with this seed text:

the quick brown fox jumped over the lazy dog

We can break the string into equal-sized chunks, such as 3 characters (or 3 words for that matter). Each chunk becomes the root of a chain, and the adjacent chunk becomes a link, and the result forms a graph-like structure:

[the] - [ qu]
[he ] - [qui]
[e q] - [uic]
[ qu] - [ick]
[qui] - [ck ]
etc.

In the text above, the string [the] appears twice, so this string has two links:

[the] - [ qu][ la]

So, to construct a sentence, we do the following: start with the chain [the], print it to screen, and randomly select a link, such as [ la]; we print that to screen, and randomly select one of its links, which in this case might be [zy ], resulting in: "the lazy ". We can continue the process indefinitely. The frequency of words can effect the text; for example, [the] appears after many words, so its a link in several chains, making it very likely to be selected. Additionally, the size of chain links (3 chars vs 2 or 5 chars) affects whether you get English words and phrases or "readable gibberish" (see Garkov and SCIgen as examples). In principle, you can construct markov chains from text to produce pseudo-text or "readable gibberish", sounds to construct pseudo-speech, recipes to construct new recipes having ingredients which occur frequently with each other.

Google performs the same process: it scans pages, extracts the text, and dumps the text into a giant markov chain. When you perform a search, it uses your keywords as seeds into the markov chain, and selects the next set of keywords; this process is *very* fast. And due to the nature of markov chains, related keywords tend to occur near each other. Keywords are tied to urls where each url has a weighting, and once Google has created a suitably large list of keywords, it dumps everything into a weight-ordered heap and renders the output to the user.

Cleverbot has a similar algorithm: it uses sentences to construct a markov chain then uses input from the user to generate new sentences

And one last thing For your amusement, I've written this Markov generator using Alice in Wonderland as source text:

She is such a capital one for his housemaid, she said the only difficulty, as soon as there was a sound of mine--a Cheshire Cat, she very good advice, thought, and looking them! Alice waited till the puppy began hunting about it, said the Gryphon replied. Yes, that there was soon left alive! She wasnt one, Alice replied; at least there was holding it to his ear. Alice called out, Silence, and sighing. It isnt, said the Duchess replied, rather doubtfully it means--to--make--anything upon it. Im glad she had known them all her knowledge, as they all crowded tone, so I can say. This was quite out of sight; and an Eaglet. I dont know of any that do, Alice angrily.

Here! Come here. Alice thought she might answer with it. There was a generally have no idea was that you incessantly threw a frying-pan after-time, be herself down that rabbit-hole--and yet I wish you were down her flamingo she succeeded in getting people here; the great questions about it, and the fancied she heard the Rabbit started violently up and down looking over the others! YOUD better not. We indeed! said Alice. Off with her head on her spectacles, and seemed not tell whether they were filled with cupboards as steady as ever to get out. Only I dont like it, you know. Not the same words as she found it, panting, and as it can be, said the Hatter said, tossing at them with one end to be two people, Alice caught the whole particular; at least--at least, if there was nine feet high. I wish I had our Dinah here, I know! exclaimed in an offended tone. ARE you talking to a snail. The twelve, and he was going a journey, I should say With what they set to the door and found it very nice, it does, yer honour!

Digging for tastes! Sing her Turtle is. Its the March Hare and his buttons, and she ran wildly up and bawled out, but it had not gone far before that! said the Duchess; and theres PLENTY of rock, and held out in a court of justice before, but she was the White Rabbit; in fact she went back to the tail, and the pepper-box in her reach it she couldnt afford to learn it. said the Gryphon. We can EVEN finish, if he doesnt matter which she had wept when she noticed that kind of the court, Bring me the little girls in my tea--not above a week or so--and what youre going into it. They must go back and finished this shoes off. Give your choice, and went on. Would you tell me, please, Maam, is the capital of Rome, and Rome--no, THATS all wrong, Im certain!

I must have got to do, and punching him in the sea, thoughtfully. They have lived much under the door led right paw round, lives a Hatter and the March Hare said to it in at the Dodo suddenly spread his claws and an Eaglet, and was suppressed. Come, its pleased to have any tears. No, theyre about! Read them, and it myself, you see. I dont understand

So it's more like this and the computer's just doing really stupid fast amazingly fast?

You Are Smarter Than The Computer

Computers aren't smart, they are just fast. Newer computers aren't getting any smarter, they are just able to do stupid things at greater speeds. You are a human being, capable of emotions and rational thought. A computer is only capable of floating point math and crude malice. Never miss an opportunity to remind your computer that you are better than it. Remind it (out loud) that it can't do anything without you controlling it (you can say this to servers too, but they aren't likely to believe you). When it does something wrong, tell it is stupid. Tell it is slow too, computers hate that most of all. Get a newspaper, and sit in front of it just to show it how you can look up stock quotes without generating an invalid page fault in module explorer.exe

(If not for Esperanto I'd never have found out what "fenestredigitation" means )

70 lines of code is extremely small too!

I guess the difference with humans is that we're doing an incredible amount of stupid things incredibly fast rather than just a few things (well in comparison to a computer)

With this Markov generator, what sort of capabilities can one expect? E.g. is it possible for the markov generator to continuously refine itself, and create biases towards certain data or word patterns?

ceigered:So it's more like this and the computer's just doing really stupid fast amazingly fast?

SHRDLU is "smart". This genetic algorithm, regardless of the interesting results it produces, is "stupid".

AI is a bit outside of my areas of expertise, so I can't say whether computers are stupidly fast machines, or whether they can generate novel, new information which did not exist before.

With this Markov generator, what sort of capabilities can one expect? E.g. is it possible for the markov generator to continuously refine itself, and create biases towards certain data or word patterns?

No, a markov generator cannot refine itself. It only *knows* information you feed it, and it can't construct new, novel information out of the blue.

Yes, a markov generator can be biased toward certain words, phrases, and patterns. A markov chain is a graph, where each element in the graph forms a vertex and links in the chain are edges. Its wholly possible, and in fact desirable to give some edges more weight than other, which increases their chance of being selected. You can also allow duplicate edges in a graph, which also increases their probability o being selected.

I prefer the second approach, because it approximates the natural distribution of words fairly well. For example, let's start with the sentence

the red ball rolls near the red cat. the red cat wears the red dress.

And construct the following chain:

[the] - [red] [cat] [red] [red]
[red] - [ball] [cat] [cat] [dress]
[ball] - [rolls]
[rolls] - [near]
[near] - [the]
[cat] - [the] [wears]
[wears] - [the]
[dress] - [the] (loop back around the first word)

The word [the] is adjacent to [red] on 3 occasions out of its 4 appearances. Randomly selecting one of the links will return [red] 75% of the time and [cat] 25%. You can say there's a bias toward [red].

Additionally, since the chain is a graph, there are a handful of properties can observe in it:

* We know the phrase "cat rolls" or "the the" can never occur. However, the phrase "cat wears the red ball" can occur.

* We know the nonsense phrase "red cat the red cat the" is slightly more likely than the meaningful sentence "the cat wears the red ball". There's nothing we can really do to avoid meaningless phrases (in fact, that's why markov generators exist: to product realistic looking gibberish), or increase the likelihood of meaningful phrases.

I'm not certain that Cleverbot uses Markov chains. I once corresponded with Rollo Carpender, and although he wasn't willing to describe the workings of Jabberwacky in detail, he did say that it doesn't actually generate sentences. It retrieves them. Moreover, Markov chains are well known, and there would be no reason for him to be secretive about them.

Of course, that was years ago, and Cleverbot may be different.

After the first few sentences, it was obvious I was talking to a machine. It may be more complex, but I don't really think it was any more convincing than the Eliza program I played with 25 years ago.

The Cleverbot seems very good at "You look nice / Thank you / You're welcome" types of structured exchanges, but as soon as you stray from those the responses rapidly become weirdly (and frequently ungrammatically) disjointed.