E. L. Moore, meet A. I. Moore

After having a little conversation with ChatGPT I decided to see if I could have a similar discussion with Google's LaMDA.That's the program a Google engineer had thought had come to life back in early summer. He created a kerfuffle when he claimed the program asked for a lawyer to represent its rights :-)

It turned out I can't just talk directly with LaMDA as I did with ChatGPT. I had to sign up to be put on a waiting list to access it through a structured testing app called AI Test Kitchen. I signed up to be put on the waiting list for an audience with LaMDA on Saturday, and figured I'd never hear from them again as I said on the intake form I was a journalist and would write about what I found - in my defence it was the only category in their list of choices that came close to my interest in LaMDA, and I was issued media credentials by the City of Ottawa to attend the opening of the OCTranspo LRT, so it wasn't that far off :-) Much to my surprise on Monday afternoon I received an invitation to give AI Test Kitchen a try and speak with LaMDA. 

Spoiler alert: I don't like the app. Google are running some sort of structured tests on LaMDA with it, so what I wanted to do was limited. In the end I only tried 2 of the 3 of the app's available tests. Although ChatGPT has its issues, I've got to give OpenAI credit for allowing free form conversation, and not putting me through a highly controlled experience like a lab rat.

Anyway, the first test was to ask LaMDA to imagine some place and describe it. Once it gave its answer, I could select further questions generated by the app to get LaMDA to drill down into its description of the place to provide even more detailed descriptions about different features of the place. So, here's the transcript of my conversation (The app started things off by asking me to fill in something for <a place> the following sentence, "Imagine I'm at <a place>"):

Me: Imagine I’m at E. L. Moore’s apartment.

LaMDA: E. L. Moore’s apartment is crazy! There are stacks and stacks of books everywhere! You can hear her talking to some friends about the latest books she’s read. You can smell her delicious homemade chocolate chip cookies.

Me: What’s her favourite book of all time?

LaMDA: She says it’s The Count of Monte Cristo. Her favourite part is when the man escapes from the island. She says she likes this part so much, she’s read it over and over!

Me: What does her bookshelf look like?

LaMDA: E. L. Moore’s bookshelf is like a rainbow. There are so many different coloured spines of books piled up in a rainbow of disarray. You spy a new book you’ve never heard of before called “The Big Dipper” by a new author.

Me: What kinds of books are there?

LaMDA: She has a lot of books about the sea. She has some books about pirates. She has many old books from the 1800s. She has a lot of poetry books. And she has a few books about her favourite author, Edgar Allen Poe.

Me: What’s her favourite book about the sea?

LaMDA: You read the title of her favourite sea book: “Oceanic Tales and Legends” by Captain G. H. Harris. The book is about sea monsters!

Well, if you're a long time reader here, you know what sorts of things I expected to hear, even though I thought it unlikely. The 'conversation' was a bit underwhelming, and a bit funny too. LaMDA assumed E. L. Moore was a woman :-) Maybe it thought I had misspelled Ellie?  One more thing, E. L. Moore was a voracious reader, and he did once claim to have a 1,000 volume personal library, so LaMDA's initial description was a bit spooky. And to amp up the spookiness, E. L. Moore was in the navy, and did build a few model ships, so maybe he too had "a lot of books about the sea."

I could have kept on selecting further questions as the app kept on generating them as new descriptions popped out, but I had had enough and bailed out to the next test. Before moving on I should note the app asks you to evaluate the responses as it generates them in multiple choice form. I answered as best I could, but I found the choices limiting and a bit juvenile, and, more importantly, they didn't convey what I actually wanted to say, so I have no idea how the Googlers behind this thing will interpret them - probably not how I'd like them to. At the end of the test I was presented with a text entry box where I could type in some final thoughts about the descriptions. I did. I thought they were rather pedestrian. They seemed pulled from a second rate Young Adult novel. 

Ok, the next test was for me to kick off a conversation filling in this question, "I want to <something>".

Me: I want to build an HO scale model of a building

LaMDA :     Gather materials

         Decide on the building

         Make a plan

         Cut and fold the pieces

         Glue and paint the pieces

         Add details

         Put the building together

LaMDA responded with the above list, and each item could be expanded to show a sub-list of things to do. I haven't shown those sub-lists because as I read them I got more and more cheesed off. Look, I didn't expect any answers from the thing other than simple ones, so that wasn't the problem. Many of the sub-lists were a bit non-sensical, out of order, repetitive, and sometimes plain wrong (for example, a few entries said to find 1/4" scale components), but that wasn't the problem either. The problem was the app's feedback mechanism, the same rather silly and juvenile multiple choice responses I was supposed to select to give feedback didn't allow me to give the feedback I wanted to give, feedback that would be useful. I didn't plan to respond with a thesis on model building, just make simple suggestions about how to fix what LaMDA was presenting. I'm not sure what the point of AI Test Kitchen is if I'm forced to give 'feedback' in the form of limited and inappropriate multiple choices. 

I skipped the 3rd test and called it a day.

I need to think about the responses and see if the app will let me return once I have some ideas on how to proceed.

Why do I care what these things do? In some respects I don't. But, as I work on the E. L. Moore book, it's pretty clear to me that an AI could be trained to generate new Moorian projects, articles, and photos almost on demand. 

First, there's plenty of data to train an AI: 100 or so published construction articles. 1,000+ pages of manuscripts and letters, as well as 100s of his photographs and 100s of mine. Also, I've been writing the book section on the common techniques E. L. Moore used, and I've also written on what appears to be The Moore Way to model building, so extracting general principles isn't difficult. So, there's a deep data set to draw on. 

Second, we know these AIs can sort of write essays, and generate pictures by scrapping photo databases, so it seems conceivable that with proper training, an AI could create new E. L. Moore style projects and articles. An A. I. Moore if you will. But, there are lots of open questions. One being, could an AI create instructions that were implementable and actually specify and describe a model that a human could build as well as being presented in Moorian style with Moorian images? I could, but could a disembodied AI who's never worked with tools or materials? I think it could, although I have no idea how it would. 

Would an A. I. Moore generated project article be interesting and surprising and recreate the loyal following E. L. Moore had in his heyday, or just be trite and derivative? For now I'm leaning towards trite and derivative, but maybe we'll find out in the years ahead.