Locus of intelligence

I’m not a curmudgeon, or even anti-AI (artificial intelligence). To the contrary! Yet, I find myself in a bit of a rebellion in this ‘generative‘ AI era. And I’m wondering why. The hype, of course, bugs me. But it occurs to me that a core problem may reside in where we put the locus of intelligence. Let me try to make it clear.

In the early days of the computer (even before my time!), the commands were to load memory into registers, conduct boolean operations on them, and to display the results. The commands to do so were at the machine level. We went a level above, with assembly language, that was a translation of that machine instructions into somewhat more comprehensible terms, assembly language. As we went along, we went more and more to putting the onus on the machine. This was because we had more processor cycles, better software etc. We’re largely to the point where we can stipulate what we want, and the machine will code it!

There are limits. When Apple released the Newton, they tried to put the onus on the machine to read human writing. In short, it didn’t work. Palm’s Pilots succeeded because Jeff Hawkins went for Graffiti as the language, which shared the responsibility between person and processor. Nowadays we can do speech and text recognition, but there are still limitations. Yes, we have made advances in technology, but some of it’s done by distributing to non-local machines, and there are still instances where it fails.

I think of this when I think of prompt engineering. We’ve trained LLMs with vast quantities of information. But, to get it out, you have to ask in the right way! Which seems like a case of having us adapt to the system instead of vice versa. You have to give them heaps more context than a person would need, and they still can hallucinate.

I’m reminded of a fictional exchange I recently read (of course I can’t find it now), where the AI user is being advised to know the domain before asking the AI. When the user queries why they would need the AI if they know the domain, they’re told they’re training the AI!

As people investigate AI usage, one of the results is that your initial intelligence indicates how much use you’ll get out of this version of AI. If you’re already a critical thinker, it’s a good augment. If you’re not, it doesn’t help (and may hinder).

Sure, I have problems with the business models (much not being accounted for: environmental cost, IP licensing, security, VC boosting). But I’m more worried about people depending too much on these systems without truly understanding what the limitations are. The responsible folks I know advocating for AI always suggest having a person in the loop. Which is problematic if you’re giving such systems agency; it’ll be too late if they do something wrong!

I think experimenting is fine. I think it’s also still too early to place a bet on a long-term relationship with any provider. I’m seeing more and more AI tools, e.g. content recommenders, simulation avatars, and the like. Like with the LMS, when anyone who could program a database would build one, I’m seeing everyone wanting to get in on the goldrush. I fear that many will end up losing their shirts. Which is, I suppose, the way of the world.

I continue to be a big fan of augmenting ourselves with technology. I still think we need to consider AI a tool, not a partner. It’s nowhere near being our intellectual equal. It may know more, but it still has limitations overall. I want to develop, and celebrate our intelligence. I laud our partnership with technologies that augment what we do well with what we don’t. It’s why mobile became so big, why AI has already been beneficial, and why generative AI will find its place. It’s just that we can’t allow the hype to blind us to the real locus of intelligence: us.