A brief AI overview?

At the recent and always worthwhile DevLearn conference, I was part of the panel on Artificial Intelligence (AI). Now, I’m not an AI practitioner, but I have been an AI groupie for, well, decades. So I’ve seen a lot of the history, and (probably mistakenly) think I have some perspective. So I figured I’d share my thoughts, giving a brief AI overview.

Just as background, I took an AI course as an undergrad, to start. Given the focus on thinking and tech (two passions), it’s a natural. I regularly met my friend for lunch after college to chat about what was happening. When I went to grad school, while I was with a different advisor, I was in the same lab as David Rumelhart. That happened to be just at the time he was leading his grad students on the work that precipitated the revolution to neural nets. There was a lot of discussion of different ways to represent thinking. I also got to attend an AI retreat, sponsored by MIT, and met folks like John McCarthy, Ed Feigenbaum, Marvin Minsky, Dan Dennet, and more! Then, as a faculty member in computer science, I had a fair affiliation with the AI group. So, some exposure.

So, first, AI is about using computer technology to model intelligence. Usually, human intelligence, as a cognitive science tool, but occasionally just to do smart things in any means possible. Further, I feel reasonably safe to say that there are two major divisions in AI: symbolic and sub-symbolic. The former dominated AI for several decades, and this is where a system does formal reasoning through rules. Such systems do generate productive results (e.g. chatbots, expert systems), but eventually don’t do a good job of reflecting how people really think. (We’re not formal logical reasoners!)

As a consequence, sub-symbolic approaches emerged, that tried architectures to do smart things in new ways. Neural nets end up showing good results. They find use in a couple of different ways. One is to set them loose on some data, and see what they detect. Such systems can detect patterns we don’t, and that’s proven useful (what’s known as unsupervised learning).

The other is to give them a ‘training set’ (also known as supervised learning), a body of data about inputs and decisions. You provide the inputs, and give feedback on the decisions until they make them in the same way.Then they generalize to decisions that they haven’t had training on. It’s also the basis of what’s now called generative AI, programs that are trained on a large body of prose or images, and can generate plausible outputs of same. Which is what we’re now seeing with ChatGPT, DALL-E, etc. Which has proven quite exciting.

There are issues of concern with each. Symbolic systems work well in well-defined realms, but are brittle at the edges. In supervised learning, the legacy databases unfortunately frequently have biases, and thus the resulting systems also have these biases! (For instance, housing loan data have shown bias.) They also don’t understand what they’re saying. So generative AI systems can happily tout learning styles from the corpus of data they’ve ingested, despite scientific evidence to the contrary.

There are issues in intellectual property, when the data sources don’t receive acknowledgement nor recompense.  (For instance, this blog has been used for training a sold product, yet I haven’t received a scintilla of return.) People may lose jobs if they’re currently doing something that AI can replace. While that’s not bad (that is, don’t have people do boring rote stuff), it needs to be done in a way that doesn’t leave those folks destitute. There should be re-skilling support. There are also climate costs from the massive power requirements of such systems. Finally, such systems are being put to use in bad ways (e.g. fakes). It’s not surprising, but we really should develop the guardrails before these tools reach release.