The Speed of Wisdom in the Age of Artificial Intelligence

This week OpenAI revealed GPT4, its much anticipated successor to ChatGPT (technically ChatGPT has a far less famous GPT3.5 cousin called InstructGPT, which would also be a predecessor of GPT4). An overview can be found here.

In addition, Microsoft recently began describing how it's integrating GPT3.5 and GPT4 capabilities into its Office 365 product line with its Copilot announcement. You can learn more about Copilot here.

What's interesting about Copilot is that it seems to represent the beginning of something that Sam Altman, CEO of OpenAI commented on a few months back when he was asked about the importance of "prompt engineering." (If you're unfamiliar with ChatGPT, stop reading this post, and spend the next hour playing around with it. Go here, get yourself up to speed. No matter what walk of life you're in, you need to be aware of what's going on here, and that's not hyperbole.) Sam Altman was asked if "prompt engineering" would become a skill that people will become adept in in the future. For the uninitiated, prompt engineering is a fancy term that you're probably already somewhat familiar with if you've ever used a search engine. When you slightly change the search criteria you give Google Search for example, you'll get a slightly different list of results. Something analogous is true of GPT3.5 and its predecessors. Depending on how you ask the question, you'll get different results, even if the question is essentially the same. This was tagged early on with the name "prompt engineering." When asked about this new form of "engineering," Altman replied that he believes that prompt engineering will go away. So, since the CEO of this firm believes that, it's reasonable to assume that the engineers working on each iteration of this LLM family will work toward making this go away, and in Microsoft's Copilot we may see the nascent beginnings of how that might happen. Copilot introduces something called "grounding" to the GPT tech stack. Right now what grounding does is it "reformats" your prompt in a way that Microsoft's engineers predict will be easier for the LLM to understand and make sense of. So, in essence, they're taking some of the "prompt engineering" burden off of you and giving it to this layer of code. You might say "big deal," but it doesn't take much imagination to figure out what's probably going on here. If you think in terms of "what is the fitness function for this grounding capability?" it's something like "try to make sure that the LLM does a good job of providing a reasonable answer, and limit how often it hallucinates." Hallucination is a tricky problem for LLM engineers, because it has multiple causes, one of which is prompt content and bias. You could imagine how someone working on solving for that fitness function might build up a "connotation dictionary," and design it to intentionally substitute words and phrases with less bias that have equivalent meaning for example (though I have no insight into what it's actually doing, you could imagine how something like that might improve output quality and reduce hallucination). How else might you design this grounding process to reduce hallucination? Well, one way would be to steer the questioner subtly toward subjects where OpenAI and Microsoft know it is stronger and less likely to be confused by conflicting source data for example. How would you do that? Well, you might begin to make the grounding process more interactive. Ask follow up questions. Clarify confusing phrasing. In short, all the things we do in conversational interactions with fellow humans when we're not sure we're on the same page with one another. And once you've achieved that, what exactly is the value of "prompt engineering?" You're no longer in control of the "mental model" the stack is creating based on your prompts. The stack is instead "figuring out" what your prompt is. If you've ever had a conversation with someone who is sleepwalking, you might see the subtle difference. Someone who is sleepwalking will just answer questions, and they might not make a lot of sense, but you can often have a (very strange) conversation with someone in this state. This seems analogous to GPT3.5 era conversations. A GPT4+ stack that uses this more sophisticated form of grounding (interactive) might be more like a conversation with someone who is awake.

GPT4 adds a bunch of other interesting capabilities, not the least of which is that it's multi-modal. The photo at the top of this article was generated by another Generative AI called Midjourney. Again, if you haven't used it, stop reading this article and go spend an hour familiarizing yourself with it. Once again, whether you realize it or not, this type of generative AI has a lot to do with your future, regardless of what field you're in. So what does Midjourney have to do with GPT4? Well, they're both forms of Generative AI with very different jobs ("fitness functions" to borrow [abuse?] the vernacular of the field.) Midjourney can accept prompts, similar to GPT, but rather than replying with a wall of text, Midjourney draws what you described, often in lush detail, and with remarkable degrees of control based on what you ask for. What OpenAI has done is incorporated this type of image based generative AI engine inside the GPT4 stack, along with image recognition. It's likely this trend will continue, with more and more "modes" implemented into the stack. Already, you can do things like draw a web page or mobile app screen on a napkin, take a picture of it, send it to GPT4, and it can write HTML, CSS, and Javascript to produce a working prototype of that napkin sketch. It's probably not great code, but it will be enough that an experienced engineer could use it as a spring board. So if you haven't figured it out, things are getting trippy folks. You need to keep an eye on what's going on here.

And I said all that to get to the point of this article (I know, burying the lead, how awful of me, right?) What's the speed of wisdom? Re-read that. What. is. the. speed. of. wisdom? I'm not talking about knowledge acquisition, because clearly we're heading into a world where knowledge acquisition is a largely solved problem (which seems weird to write, believe me). I'm talking about wisdom. Wisdom is, I think, at the intersection of experience, knowledge, and good judgement. As a little thought experiment, let's ask ourselves "why did people suddenly jump to the idea that prompt engineering was an important skill for them to learn?" Well, experience has taught us that when a new technology arrives on the scene, humans can be good at "plugging its weak spots" and we can thereby be beneficial to ourselves and our fellow humans by playing the "value add" game to the technologies we are just becoming exposed to. However, I'm not quite sure folks are on top of how fast this speed boat is moving. I've been following LLMs for about a year now (going back to GPT2), and I can definitely say it's difficult to stay on top of what's possible with these generative AI tools and their close relatives. So, in this case experience may very well have misled us. I would predict that "prompt engineer" will not appear on too many resumes in the future, because the whole concept will probably be replaced by some more sophisticated form of "grounding," a concept that's literally brand new to us now, but will probably seem like table stakes in a year or two. So again, what is the speed of wisdom? Should I invest time in getting better at prompt engineering? Maybe, but it probably shouldn't become an important element of your career plan I would say. In fact, you should be concerned about your career plan no matter what field you're in, at least anywhere past the 3 - 5 year point. (I know, that's pretty provocative. It will be interesting to reread this article in 2026 - 2028).

And in a broader sense, what can we say about the wisdom of unleashing tools like these on an unsuspecting world? I am a technology optimist, and always have been, but these advances are happening so fast and have such potentially transformative (both for the good and for the bad) possibilities, that even I am concerned about how fast we're moving, and how little our institutions like our legislators are even aware of what's going on. If wisdom is at the convergence of knowledge, experience, and good judgement, and we all collectively have almost no experience with this stuff, and this stuff almost replaces the ability to gather knowledge on demand, what we're left with is good judgement. Who decides what good judgement is, especially for things as transformative as the AGI world we're now clearly headed toward? A year ago if someone had asked me how close we were to AGI I would have said "5 years minimum, probably more like 10 or 20." Now my answer would be something like "I seriously have no idea. It's probably not in the next year, but it's not out of the realm of possibility." I also predict we're going to soon be in a weird space where we're struggling to determine if something qualifies as AGI. There will probably be situations where a generative AI can be tricked into producing a result that demonstrates that it is not in fact thinking, but it will become progressively more difficult to find these things, and so I predict we will be living in a "Turing Test Fog," for a while, and we're already beginning to see the mists of that fog, as people fall all over themselves to anthropomorphize what is essentially "autocorrect on steroids." They attribute intent and will to this algorithm that's just endeavoring to come up with a good story based on a prompt. It's all just math and statistics, folks. However, when we eventually emerge from that fog, we're going to have even more difficult questions to answer... is an AGI that is indistinguishable from a human in terms of how it interacts with the world entitled to voting rights? This was science fiction 5 years ago. It's still technically science fiction, but the wall is getting thinner and thinner, and there's a lot of money lined up to make the wall of fiction go away.

The future is amazing. I hope we have the wisdom to manage this transition. It's upon us now. Perhaps the best we can do is stay awake and attentive.