So, I’ve been pretty harshly critical of the current generative artificial intelligence (Gen AI) situation. Again, I have no problem with the tech itself, but the big providers I have several problems with. Not to rehash those objections, but I’ve been investigating alternatives. And, I think I’ve now converged on a solution: small hybrid multi-agent systems. I owe thanks to several informants, including Elizabeth Dalton, Markus Bernhardt, and Scott Parker.
First, despite my other objections, what I see is that the companies are pushing these changes as ‘inevitable’ as well as desirable. And, sadly, way too many orgs have jumped on board. Which, while understandable, isn’t desirable. What we’re now seeing are the results. Companies are hiring folks back that were fired. We’re also seeing that we’ve hollowed out the pathways for new hires to acquire the necessary skills. Which is part of the larger picture of negatively impacting learning.
Yet, I’ve been a proponent of AI in the past. The technologies on tap have real benefits, whether they’re AI or something that we now understand so we can’t call it AI anymore. So I need to converge on an integration. And, that utility for specific needs is the key to my reconciliation.
In interpersonal situations, I remain a strong belief in ‘partnering’. That is, recognizing that others have complementary knowledge to your own, and different perspectives. So, why have one supposedly all-seeing all-knowing solution, instead of smaller more targeted ones?
Ok, yes, there’re issues of cost. Right now, it’s seemingly cheap to get an AI to do your rote tasks. However, two things are relevant here. For one, the costs are going up. Companies are going public, and that means that they’ll have to start generating revenues to repay the venture capital investments. And we’re seeing that, as companies realize that their AI expenses are growing, to the extent in some cases where they’re more than the people they’re replacing. We’re also seeing that they haven’t yet faced all the incumbent costs: IP, environment, labor. That’ll change the game too.
The flip side is that smaller language models are more cost-effective. There are open source models that you can obtain, install locally (or on private servers), train on specific documents, and use for specific purposes. Yes, there’s the development costs, but they’re more secure, faster, and less costly. As well as addressing the other concerns.
Moreover, you can have more than one. You can have specialized agents to handle X, Y, and Z. If it’s a language task, you can use a language model. If it’s another type of task, you can use another model. You can even mix in symbolic with sub symbolic, and programs versus AIs. If diversity is good for people, it also appears to be useful for systems. A mix of custom agents with an agreed upon interchange protocol is, it appears, more cost-effective, private, maintainable, and more.
Look, if you don’t mind the costs, and the questionable ownership of data, and security, and dumbing down your workforce, and … go with the big ones. They certainly will benefit. But if you want to be secure and sustainable, small hybrid multi-agent systems makes more sense to me, at least in many instances. Sure, language models trained on all the language out there is useful. Er, for general purpose language tasks. So make that one agent in your solution, if necessary (tho’ aren’t your language tasks typically more constrained?). Which, I’ll suggest, are a small subset of all the things you need to get done by technology. As I learned from the Aussies (originally a Brit saying), “horses for courses”, in this case meaning use the right tool for the job. Am I missing something?