Eliza’s children

These last few years there’s been all this excitement about artificial intelligence, machine learning, and “algorithms”, but at the end of the day I am rather disappointed. It seems the current approach to AI is all about frequency analysis, feeding in billions and billions of pieces of data and working out the most common correct response to each possible input. This basically means the method will be simply be reproducing what is fed into it.

Back in the day, the idea was to develop learning systems with a world model that they could reason about, but that is hard. The Cyc project has been the most high-profile one and still does not seem to have much overt use. (I find it somehow ironic that a site dedicated to the best in AI uses CAPTCHA to avoid spam in their contact forms, but maybe it’s sufficient to weed out the most stupid scripts, and the smarter ones might be worth conversing with.)

The original conversing agent is ELIZA, still available at a text editor near you. As was observed at the time, people who should have known better were taken in by the seemingly intelligent responses, and the same seems to have happened to Blake Lemoine at Google who got tender emotions towards a piece of software. Now, I have not specifically tried the LaMDA system, but others that are built on similar principles and what has struck me with them is that they do not seem to have a very deep knowledge representation of the conversation itself, but exist in the here and now of the current utterance, much like a senile person might do, reproducing patterns, but not connecting the pieces together.

Bits from a recent conversation with a chat bot that pretends to be a person:
I have a dog and a cat
Simulated ones?
No, they live with me
I work from home, so I have plenty of time for them
[ … ]
Do you have any pets?
Yes, I have two cats and a dog
They keep me very busy. How about you?
Ah, are you making this up as you go? Earlier you said you have one cat and one dog.
I am not making it up
I do have a cat, but she lives at home with my parents
So how many cats in total?
There are four of them
They are all rescues
Exponentially increasing.

It’s an interesting bug that the last sentence in each utterance misses the full stop, but not other punctuation.

I write like lots of people

I Write Like uses statistical analysis to compare your writing style with that of well-known English-language authors. I tried different entries from this blog and each of them was apparently reminiscent of a completely different author. Maybe a longer text would give more dependable results? I tried feeding in my PhD thesis. It was most similar to Agatha Christie. I hadn’t earlier noticed the extensive note apparatus of Murder on the Links. Actually, I didn’t notice it now either. For comparison, I also fed in my licentiate thesis, by an even younger me, presumably less mature. That was most similar to the writings of Gertrude Stein. A matrix row is a row is a row is a row?

But, Siggy explains why the algorithm is so often wrong.

Do I make sense?

I’d had a video automatically transcribed. I presume all kinds of AI jiggery-pokery had gone into turning audio frequencies into text—it was clear that the transcription engine was aware that a lot of hemming and hawing could be deleted, on the other hand not-quite-silences could be extrapolated into what probably was being said. A few times this worked amazingly well, but all too often the resulting text was nonsense. I was going to blame this on the lack of deep understanding on the part of the transcription engine, but as I replayed and replayed the same five second clip, trying to hear what was being said, by myself, I came to wonder: Is this actually what my students hear when I lecture—words that may be grammatically correctly connected, but rarely make any sense?