CFACT just produced an impressive example of a chatbot emulating several sorts of abstract thought. At their request the bot GROC 3 created three long song poems successively in the styles of Kipling, Lady Gaga and The Rolling Stones. You can read them here.

Chatbots do not think rather they computationally emulate thought. But it is easier to talk as though they actually think so I will do that here. Using emulation language is tedious and clumsy.

The poems are not bad and some lines are really good but that is not my focus. My interest is cognition. To begin with the instruction “write a poem about X” is pretty abstract. It is far more abstract than asking for facts already written about or what someone already said.

Mind you there are thousands of books and articles on poetry and how to write it. A great thing about chatbots is they can read thousands of books in a thousandth of a second so maybe that helped this one.

Asking for a long poem in the style of a specific writer or performer is a whole different critter. To begin with the robot has to figure out what that style is. Then it must come up with an applicable poem in language that fits that style. Being able to do this is truly impressive.

I do not know how to describe a style. Then too I have not read much Kipling or heard Lady Gaga sing, although I am a big Stones fan. So I cannot guess how long it would take me to do what this bot did in seconds or less. Maybe weeks, months, or years. Maybe never because it takes talent or in this case the successful emulation of talent.

I have read people saying that chatbots are just statistical engines that find the text most likely to be the answer to the query. Since these song lines have clearly never been written before that cannot be all there is to it.

Here is a great chorus for Lady Gaga which I am sure is new writing:

“Oh-oh-oh, the data’s callin’,

Oh-oh-oh, the veil is falling’,

Flip the script, we’re taking flight,

CFACT’s the fire in the fight!”

If you think songs are irrelevant let’s try a different case that I saw when ChatGPT first came out. On a science blog someone asked it to summarize an important research paper which it did pretty well, as well as a person might do it.

It likely used words that were not in the article itself because summaries do that. Once again this is far beyond probabilistic search of existing text.

Then they asked for a summary in different formats ending the game with a good haiku version. For those not familiar with this imposingly strict form here is what Google AI just gave me:

“A haiku is a Japanese poem that is made up of three lines and 17 syllables. The syllables are arranged as 5-7-5, with the first and third lines having five syllables and the second line having seven. Haiku are usually unrhymed and are meant to capture a moment in time.”

As with the CFACT songs the chatbot is here doing serious interpretive work. Combine this with the ability to quickly read a million pages and you get a very useful possibility. A chatbot should be able to tell us what is fundamentally going on in a given area of research where there are far to many published pages for a human to read. Even small research areas can generate several thousand journal articles a year.

In any case the GROC 3 CFACT songs are a clear example of how creative chatbots can be given proper instructions. This is a contribution to AI research.

These results are amazing. We have built a computational system that can not only write a long poem about a specified topic but cast it in the style of a given author. The math must be wonderful.

The big question is how best to use this amazing emulated talent? Stay tuned!