On Wednesday night, I gave the first presentation in what will become a three-part series on Artificial Intelligence. This series is part of a larger series called Theology on Tap, organized by Christ Church Anglican in Phoenix. The event was held in the First Draft Book Bar which is in the independent bookstore, Changing Hands, in Phoenix, Arizona.
The night started with my struggling to get my iPad connected to the screens. Fortunately, I had brought just about every cable I could find to the event so that was resolved in time. I was rather nervous to speak in public since I hadn’t given a presentation to a room of people since I was at university. Also, it wasn’t obvious to me whether I was fully qualified to speak on the topic of artificial intelligence. In 2000, I graduated with a Computer Science and Artificial Intelligence degree and a few years ago I wrote an AI as part of a Machine Learning course though that hardly qualifies me as an expert.
Preparation for the event had been challenging. My goal was to try to demystify a topic that, to understand the basics, one would have to have a mental picture of a computational process that is so counterintuitive as to render it near impossible to analogize without obscuring fundamental concepts of how it functions. I was to present to a mixed audience and attempt to neither bore them or confuse them. In reality, the AI training process (which is what makes it so powerful) is a set of mathematical operations that incrementally adjust billions upon billions of numbers until the system is “trained”. I based the bulk of that part of the presentation on the existing metaphor of navigating a foggy landscape by blindly feeling around for a slope – the metaphorical goal being to reach the lowest point on the hill. I even had some nice clipart of boots, a compass, an altimeter, and a notebook in the hope that this would enhance the picture I was building.
So on the night, after I had gotten connected, I was supplied with a glass of wine and sat back to watch my friend Walter introduce the topic and then introduce me to the audience. There were between forty and fifty people, which felt like a lot. I was in front of a screen with my Google Sheets presentation, my iPad was on a stand and showing me my notes and I was nervous. The AirPlay system that would have displayed my slides on the two available screens had not worked so I was stuck with one screen that seemed oddly dim for some reason I was too panicked to investigate. I had previously practised my slides on Jess and the cats and they had found it mostly illuminating. I had even brought a ludicrously powerful laser pointer that can reach far enough into the sky to blind light aircraft pilots. No feeble equipment for me (except the 4K video camera that I had bought that is so cheap that its brand is not worth mentioning). I was prepared as I could be.
Well, the presentation, to cut a long story less long, went well. I started out with a preamble on the topic and then a very basic explanation of what programming was. I thought it best to show a simple program so that I could compare a program that is written from start to finish with an AI program that is trained. Next I posed questions about the nature of language rules. The audience furnished us with examples, such as grammar and spelling. I countered with a theory that politeness and even the answer to universally known questions, such as “What is the capital of France?” were following kinds of rule.
Then the difficult bit. I showed an Artificial Neural Network which, if you haven’t seen one, looks like the workflow diagram from hell. Massive complexity had to be reduced down to an input, an output, and a, um, middle bit. I think I had made the numbers large enough for the audience to see. We stepped through the training process that had been peppers with images from Midjourney until we had fed a sentence “Tomorrow is Wednesday” through all the stages and shown some numbers turn into some other numbers. I think I most people followed. I showed that the training for GPT-2 took around as many calculations as there are teaspoons of water in the Pacific Ocean. It occurs to me that in the future we might categorize an AI’s power in Pacifics. Or kiloPacifics. Perhaps not.
Next, came the fun bit which got some laughs. I showed some ChatGPT conversations I had to show its power and limitations. I started with language abilities. In my broken French, I got it to converse and correct my questions. Then I switched the conversation to Polish which on reflection, when I read it out loud, I forgot to translate for the audience though there were some English translations shown in the slides. I then presented some amusing interactions with project planning (a rip-off of the Harry Potter series, a day-long fishing experience for cats). The last example was coaxing ChatGPT into describing how to perform a minor crime. The audience found this bit the most entertaining, mainly because ChatGPT is, if nothing else, a tool for reliably producing absurdities.
We rounded off with questions. I had a pre-questions section where I effectively questioned the audience. I didn’t want to call it Check your understanding so I settled for Reflections. I’m still not sure of a better word for it. Several people in the room came up with interesting ideas and questions around AI. Some had been personally affected already by the changes – mostly teachers coping with cheating students – and some had used it with mixed results. Then there were the formal questions posed to me. I had to take a pause and a swig of Merlot to gather my thoughts before answering. This wasn’t just a conversation. I was being recorded and I needed to give something more concrete than “yeah robots, right?”. Some of the examples, such as “Can an AI hack into a bank?” had me a bit stumped but I suppose time will tell, sadly.
In the end, it was a really uplifting experience. Several people thanked me and I answered a few more questions when I was sat back down in the room and then in the parking lot and then I was done. I barely slept that night.