Ken's Blog

ChatGPT and similar Large Language Model engines are not actually intelligent. They are mindless drones that do data-driven pattern matching. Not that there’s anything wrong with that.

Today a colleague told me of a long term research plan. His group sins to create AI engines that would be much closer to what we think of as intelligent. They would learn and self correct, and have the ability to develop over time, somewhat analogously to how the mind of a child develops over time.

I was very impressed. I told my colleague, “You know, this may go down in history as the moment when Frankenstein became a documentary.”

I can’t be sure, but I got the feeling that he’s ok with that.

During a visit today to my doctor, we got to talking about the Super Bowl. Turns out he is a really big Chiefs fan.

He told me that two days before the big game, he’d told a friend over the phone “I predict the Chiefs to win by three points.” It was, he said, a missed opportunity.

“Do you mean,” I asked, “that you thought you should bet on the game?”

“Well yes, except that I never gamble.”

“Probably just as well,” I said. “You know how it goes. If you had actually bet on the Chiefs, they wouldn’t have won. That’s how the Universe works.”

“Yeah,” he said, “I agree. That is indeed how the Universe works.”

So I said “Just think. If you had gone through with that bet, the Chiefs would probably have lost, and many people would have been sad. And it would have been your fault.”

“Yes,” he agreed, “it would have been my fault.”

“But look at the bright side,” I added, “nobody would ever have realized it was your fault.”

“Except you.”

Today was n the middle of a Zoom meeting, and I needed to email a note to somebody about a follow-up meeting. Instead of typing the email into my computer, I picked up my phone and I dictated the email.

The other people on the Zoom meeting thought it was kind of amusing, just watching me talk to my phone like that. I had forgotten that not all of my colleagues do that.

One colleague pointed out that with her Israeli accent, it probably wouldn’t work on her iPhone. Which is rather unfair when you think about it.

People have told me that when I dictate into the phone, my tone of voice changes entirely. Instead of speaking in a friendly way, it’s as though I’m giving orders to an underling.

What is really happening is that I have gradually learned what manner of speaking will produce the fewest errors on the phone. But it is kind of interesting that this replicates an unequal working relationship.

Which makes sense, because people and computers are not equal. Let’s just hope that the inequality continues to be in the same direction.

This first generation of Chatbots usually writes in a very generic way, unless told to do otherwise. By default, results are grammatically correct, sentences are well balanced, and paragraphs do a good job of introducing and structuring ideas.

But Chatbots usually don’t sound like any particular person. I suspect that this is deliberate, based on their intended use. If you are going to use a Chatbot for information retrieval, maybe you don’t want it to sound like your Aunt Edna from Brooklyn.

But the uses of Chatbots will inevitably expand, and sooner or later it may become more common to have them imitate personal styles of speech. The result might not be as grammatically correct, but it will seem a lot more human.

The question is not whether this is possible (it certainly is), but how it will be received. It is not clear how people will respond to being spoken to on a regular basis by computers that imitate the real people they know.

Will it be the next Google? Or will it be the next Clippy?

One measure of the gradual advance of media technology is how we think about the phrase “virtual representation”. There was a time when a book was thought of as a virtual representation of reality — fictional or otherwise.

Then we stopped thinking about books so much, and we started thinking about photographs as virtual representations. Eventually photos became too familiar to think about too much, and we started to think about movies and then television as virtual representations.

A while back, long after we stopped thinking of movies or TV as novelties, we started applying the idea of “virtual representation” to computer simulations of reality — fictional or otherwise. But now we are about to move on to the next phase.

The resynthesis of words or images performed by ChatGPT or Bard or MidJourney or DALL-E is the current focus. “But wait,” people think, encountering those programs for the first time. “They are just a kind of virtual representation, not the real thing.”

Which you could also say about books.

Every day is so intense
A random series of events
Does it all make any sense?
Frankly I am on the fence

Nearly every time I see a Powerpoint presentation, it is filled with bar charts. Bar charts seem to be the lingua franca of conveying information to a group in presentation form.

Clearly bar charts appealing to both presenter and audience. They convey essential information in a bold graphical form, they show trends in a way that is easy for the eye to follow, and they abstract away details the neither presenter nor audience cares about in the moment.

There will come a time when the visual equivalent of ChatGPT or Bard will make bar charts for us on the spot. We will be discussing some topic, and one person will say “You know, there’s been an increase in … over the last decade,” or “The number of women enrolling at MIT recently, compared with the number of women at NYU, has been …”

At that moment, a bar chart can optionally pop up in the view of both people, presumably mediated by our smart glasses. We probably won’t even think much about it, other than to wonder how anybody ever explained anybody before we all had conversational bar charts.

I was surprised at the highly negative reaction to the factual errors in Google’s answer to ChatGPT. It should not have come as a surprise at all.

Those of us who have been playing with ChatGPT know that it has at best an indifferent attitude toward whether its highly confident answers are correct or not.

What Google is demonstrating with Bard was that they are at the same place as their competition in the development of chatbots for Web search. We already knew that much work needs to be done to marry the verbal facility of chatbots with the degree of trust that we place in search engines.

Yes, it’s a work in progress, this effort to infuse chatbot answers with true provenance. But it’s a really fascinating work in progress, isn’t it?

Yesterday I showed a demo to a colleague, in which I spoke to my web browser, and the text that I was speaking showed up in a 3D scene. He was astonished. His exact words were “That’s amazing!”

A few minutes later I pointed out to him that both he and I do this all the time on our phones. We talk into them, and our speech gets converted immediately into text.

“Oh, right,” he said. Until then he hadn’t made the connection. Seeing text show up in a 3D scene in a web browser in response to somebody talking at their computer was out of context, and therefore somehow startling.

I wonder how often this happens to us in our lives. Something that in one context would seem perfectly prosaic, in a different context appears, at least at first, to be amazing.

I am sure that it won’t come as a surprise if I tell you that I wrote this post by talking into my phone. Nothing really amazing about that these days.

Like many people around the world, I am completely horrified by the earthquake this morning in Turkey and Syria. Many thousands of people have lost their lives.

I don’t think it appropriate to write about anything else today. Let us just give our thoughts and prayers, and help if we can, on this day of terrible tragedy.