It’s always been a battle between generalists and specialists. Despite folks like Don Norman making the case why we need both, we’re seeing this play out again. A post by Tom McDowell triggered this recognition, so let me talk a wee bit about what I see going on.
Tom’s post was about how you shouldn’t automate a process until you understand it. The issue is that until you map it, you can’t say you truly comprehend it. Further, if you don’t comprehend it, you can’t decide what the real tradeoffs are in automating it. This triggered a realization in me.
So, I’ve argued in the past that we shouldn’t digitize processes until we’ve looked at them. If you automate before reviewing them, you merely solidify your existing paths. Which, if like the Humanities & Social Sciences building at UCSD, you’ll find that people cut paths cross the grass, not stay on the concrete. That’s because the design wasn’t considering human behavior. (I think this example was in one of Don’s books; I can’t find which, but I lived it.) Digitization instantiates your processes, so make sure your processes are good first!
Which brings me to the current situation. Large Language Models (LLM) are in use more broadly than their real affordances. The underlying technology is more general, but these have been optimized for language, that is to make good sounding language. And they’ve been trained (illegally) on a vast corporate of (English) language. Consequently, they make very convincing statements about the world, whether correct or not. They make it easy to do more tasks than just language, so people do that. There are consequences.
For one, we’re losing the development path. Doing low-level work established the foundations for high-level work. Er, if you’re human. If you don’t get that indoctrination (not hazing), you don’t become an expert. Outsourcing that work means you don’t have a way for your new workers to develop the necessary expertise, which will walk out the door. And you can’t just hire experts. Knowing what you can outsource, then, becomes critical.
Another, of course, is that outsourcing the wrong things can cause trouble. We have seen bad processes reinstantiated in generated content, instead of actual learning experiences. I’m sure you’ve had the dubious pleasure of interacting with a chatbot instead of customer service (there can be good ones, but it takes a fair bit of work to make it so, contrary to the claims). Then, even if you have oversight and responsibility, you can fatigue if there are expectations to oversee something that’s mostly right. You learn to trust it, despite a conscious awareness that’s wrong.
The problem that struck me is that it’s likely the case that folks are over-trusting these generalist systems that seem intelligent. (They’re really just probabilistic engines, but to some extent so are we.) They do have lots of knowledge, but it’s context free, and trying to get them to understand context is what is known as prompt engineering (a term that’s disappearing). Still, it seems like a quick and cheap fix to get these generalist engines to do specialist tasks. I’ve been arguing that rather it makes sense to make specialist agents. The recognition came that the reason is because we really don’t have AGI (artificial general intelligence) yet (and we’re a long way from doing that).
So, yes, I have some bias. I’ve been and am pro-AI, but I don’t like the way the big companies have built their systems, they’re overhyped, and the costs are yet to be seen. Take all this with the proverbial boulder of salt. Still, I reckon there are better alternatives, at least in the long run. That is, small agents you build and train, so that you know what they can, and can’t, do. Little systems are cost-effective and more secure, too. It’s being smart about sensibly distributing roles for generalists and specialists across technology. That’s where my thinking (and recommendations) are going. What are your thoughts?
Leave a Reply