The Evolving Role of Prompt Engineering with SnapLogic’s Greg Benson

Hello, and welcome to the latest edition of the Techstrong AI video series. I'm your host, Mike Bazar. Today we're with Greg Benson, who is chief scientist for SnapLogic, and also our professor of computer science at the University of San Francisco.

And we're talking about prompt engineering. Greg, welcome to show. Thanks for having me, Mike.

I think a lot of people are trying to figure out, is prompt engineering like an actual job or is it more like a skill that we're all supposed to develop over time? And some of us may be better at it than others, but it's something we should all aspire to know how to do. Yeah, it's, I would, I would say it's very fluid right now.

Uh, you know, when chat BT was first launched, uh, there's, you know, some quick, uh, assumptions that, oh, prompt engineering could be a discipline or major even. Um, I think, uh, I think in my experience, uh, both, both in, in academics and, and in industry, uh, it, it's really gonna be, uh, combination of, of a skill. Like, you know, something like, you know, being able to use Excel combined with domain knowledge, uh, that you're gonna need to use together to, you know, extract the best outcomes or generative results from, from the LLMs.

Is it roughly akin then to, with the rise of sql, we saw people who were called SQL Jockeys, and they were basically folks who knew how to use that language to extract the most insights from a database. And this seems like a variation of that theme. Uh, I would like to be able to use natural language to query it myself, but there might be other people who would just know what's in that LLM more so than anybody else, and how to tease it out.

You know, it, it's, it's, it's very interesting, I would say in, in some ways there that, that's a good analogy in, in other ways, right? The surface area of the prompt, um, is both expansive and also, uh, very approachable, right? Like, uh, it's, it's easy to get started and easy to get results.

It is, as you maybe, uh, uh, allude to it also does require some experience and, and maybe some intuition. Um, and, and I can tell you the, the research in, in prompt engineering is extensive. You know, in the last, you know, three years, there's, you can find over, you know, 2000 scholarly articles on different approaches to prompt engineering and, you know, how, how to, how to, you know, get them to behave and perform, um, in the most, uh, performant way.

So it's, uh, it's evolving. I, you know, um, I, I think so. Do we all need to know how to, how to, uh, write prompts?

I think so. I think also, uh, that, uh, I think also that it's going to, it's going to be important for every job, um, every, every discipline, um, to have experience with it. One last, you know, thing I tell my, my kids, you know, even though this Gen AI thing is, is, is becoming really big, you still wanna stay in school.

It's actually a testament to, um, to the liberal arts education, because oftentimes you need to articulate yourself very clearly, and you learn that through right. Writing and speaking, and, um, and, you know, formally explaining yourself. Yeah.

And to make another analogy, in a lot of ways, there are people who use search engines better than others, and ultimately, will these LLMs kind of maybe coach us a little bit about how to think? And they say that, um, the sign of intelligence is your ability to ask questions is more important than your knowledge of answers for everything. And, um, in some subtle ways, are the machines gonna make us smarter because they'll be coaches us about how to frame questions?

I, I, I would, I would say absolutely. And I have, um, lots of firsthand experience with, uh, with the LMS myself and because, um, it has greatly accelerated my own personal learning in sub-disciplines or areas and topics that, um, I want to learn, but, um, maybe haven't had the time to do the traditional approach to learning. And the really key aspect of learning in my mind with LLMs and helping you formulate the questions is that you can give it a good sense of where your knowledge is.

Like it can meet you where you're at, so it can explain things in terms of what you already know and make those, those smaller steps to get you to the new knowledge. And that has been extremely powerful. And the other side of this thing that I don't think we've all wrapped our heads around is the LLMs have reasoning engines in there or reasoning capabilities.

And we might over time start linking different LLMs that are optimized for different tasks together to, uh, manage some sort of workflow or whatever. Um, won't that require us though, to have a better understanding of the relationships between different tasks and workflows? And we're all gonna be supervising these orchestrations.

It's, uh, it's a, in my role as chief scientist is a question we grapple with all the time because we're enabling our customers to easily build out, uh, these LLM applications or agent like, uh, workflows. And a, a big question on my mind is how do you, how do you, if you do automate something that previously required some sort of human decision making or human um, reasoning, and you, you transfer that to this LLM based, uh, activity, how do you have confidence it's going to do what you expected it to do? How is it going to conform to norms or even not make, you know, really bad decisions or actions?

Right? And currently, you know, a lot of the thinking is there's gotta be a lot of human in the loops still that the, the, these LLMs can accelerate many things, but, um, but there's also going to be at this point in time, um, checks and balances with, with humans involved. But it's an, I'll tell you, Mike, it's an open, it's a fascinating and, you know, vigorously researched open question right now.

Yeah. There's one way I put it, it's one thing to be wrong. It's another thing to be wrong at scale, right?

One of the other things that I think I'm seeing happening is people don't fully realize that the LLM is probabilistic. So it's giving you an answer that is most likely. And we seem to be inserting that into more deterministic processes and not realizing that, uh, there's a lot of wiggle room in there.

So do we understand how to really use LLMs within the context of something that we want to be right a hundred percent of the time and done the same way over and over again? 'cause that may not happen with the LLMI, I think. Yeah, I think that's a, uh, that's a great question.

I think, um, what's, what's a little elusive about it is there's a lot of activities that, um, it's okay to have some non-determinism. Think about, like content generation or, um, even if you're asking for summaries of information that then is consumed by a human right, it can do that stuff very well, but you're not expecting it to necessarily be deterministic in that scenario. But yes.

And then given the types of problem solving that you s you can have it articulate, right? That you can carry out in these prompts, and it's, I think it's, it's very easy to see how oh, we could insert that into making a deterministic decision to, uh, be at the center of some sort of process, right? Um, so that, um, that, again, open question.

I do wanna say that one way that that can be somewhat, um, mitigated is that I've come up with this framework of thinking about these, you know, if you want to call it ag agentic or agents or LLM based processes, you can have things that are query based, right? Things that go out and, um, come up with a plan to, uh, find information, summarize it, maybe answer questions that require if, you know, if you're a human, like you were saying earlier, like doing, searching and gathering information, synthesizing that, and then it gets presented to you like that's an activity that's no risk, because the, the result is something that gets consumed by the human. On the other hand, if you have activities that are state changing, right?

Activities that, you know, transfer money from one bank account to another, or delete a user account, right, based on a decision from the lm, then your probabilistic concern, um, comes into play. And the question is, what kind of safeguards, what kind of me mechanisms both in the prompt, both pre-processing the prompt and post-processing the prompt, what things can we do to get assurance that we would want to put it into that decision making? Um, uh, point, We all have friends who we dearly love, and they will tell us things that are very, uh, that they are certain you're absolutely right in the most convincing manner possible, and we know that they're patently false.

Um, we see the same behavior with the LLMs. It seems like they're designed to be, uh, extremely helpful to the point where even when the thing being suggested is kinda, you know, beyond the pale, is it where it says it in a way that maybe we trust too much. So do we have to either, is it the humans that have to adjust here, or do we gotta tweak the LLMs to be a little less certain?

Well, that's a, that is a good question. It was, and that is a, you know, early on that was definitely a common, and it's still common today. It's very assertive by default, right?

Mm-Hmm. Um, you can definitely, you know, give it instructions to, you know, maybe equivocate a little bit more or be, um, you know, be a little bit more, uh, less assertive, I guess, in, in answering. Um, but, but nonetheless, the, the, the point is like, as humans, you know, um, there's a little bit of, you know, I'm, you know, fairly skeptical person with when I, when I talk to humans or computers, and yeah, I think there's a little bit of, you know, uh, of training of the humans to, to uh, assess, to, to use human judgment on assessing information, whether, you know, you did a girl Google search and you got to some article on that, you know, that could be wrong.

And the information that the LLM gives you could be wrong or misleading or slightly inaccurate. This all feels to me like a giant computer science experiment that we're conducting on the public in real time. And, um, we're all learning as we go along here, but do we maybe need to think about this in a more of a constrained fashion so that we do get our arms around this a little bit better?

Or is it better that we're all just kind of learning at the same time and seeing what happens? Yeah. You know, you know, I tell my students, Mike, that, um, you know, I've, in, in my, my lifetime, I've got to experience lots of, you know, incredible shifts in technology, probably like you.

Um, and I tell my students, you, you know, computer science students, you, you are at a amazing time in computer history, right? Like, this is, um, it's transformative. You know, the next five years are, everything's changing, you know?

Uh, and I think that, um, it is an experiment, um, like a lot of technologies, maybe crypto is a experiment, still is an experiment, right? Uh, uh, but yeah, I, in, in terms of your point or you know, about do we, do we need to put restrictions, limits? Um, you know, I I, I do think that, uh, in, you know, California, where I live is, you know, I think we have something like 30 plus bills either signed or in process on various aspects of AI regulation.

Uh, and, um, I think we're gonna, as a society, we're gonna have to decide, you know, how we want to, uh, put maybe limits or safeguards from a policy perspective. Um, I am a little skeptical of putting, of, of technology limits, like partly because I just don't think it's feasible. The cat's outta the bag, right?

We, it's getting better, faster, cheaper people can build these things. Yeah, I mean, maybe not at the scale of, you know, the big frontier model players, but it is getting easier and more accessible. I think it's fair to say that for all the positives, there are people out there who are a little queasy, especially in the IT field.

They're wondering, are there gonna be jobs for coding? Is that gonna be handled by the machines? And will the it SM function be automated to the N degree?

You're a teacher of computer science. What are you telling your students about? What does the future of computer science look like?

Yeah, it's funny, my icebreaker for this fall semester that started last week, uh, I asked, during the introductions, I asked students to, um, give me, give the class their individual predictions of how gen AI is going to impact their future careers. Um, and, you know, uh, there was, you know, a lot of different answers. Uh, my, my message to them is, you, you have to embrace it.

You have to engage it, you know, there's a proliferation of coding assistance that are changing both software development, but also computer science education, um, and how we engage with, it's not something that can be ignored. It's something that, um, is a tool. It can at times, be a productive tool.

Um, is it our computer scientists or, you know, in any field, but let's talk about computer scientists. Are there, is there a risk of, um, of maybe re reduced opportunity through, um, generative ai? I, my current position, Mike, is I'm, I'm a bit of an optimist.

I, and just from my own current personal experience, I think it's gonna make us all smarter and more productive so that we can contribute to society and to businesses to make everything more efficient and, uh, maybe, maybe higher quality, um, with, so I wanna do more with the same people and not, not sort of the same thing we deliver with less, right? So maybe that's just my, maybe I'm being a little, you know, too optimistic, but that's, uh, I think we're gonna be working with it and it's going to make us more productive. All right, so what's the right path forward?

Should I just kinda continue to experiment with these various LLMs, or do we all need to go to prompt engineering bootcamp to learn how to use these things? Yeah, good question. Um, I'm actually, uh, I'm gonna be teaching a new class in the spring semester, which, uh, I'm calling Gen AI in practice.

And, uh, among other things, there will be prompt engineering. It's also for computer scientists about, you know, how to use the APIs and, um, you know, how to, how to prepare data and how to integrate them into applications, how to do these, you know, multi LLM interactions, um, ag agentic interactions. Uh, so, uh, do we all need to learn prompt engineering?

I think that, um, over time, and you're already seeing this to some degree, is the LLMs and the tooling around the LLMs, um, will help guide you. Like you can ask it today. Hey, I'm trying, you know, the, the, the, quite quickly, what was amazing, Mike, is how, how quickly do we use LLMs to help us with LLMs?

For example, there's tools that you can, you can go to chat GPT or go to Claude, and they both can readily help you write prompts, right? So in a way, they are accelerating the, the, the prompt engineering process, um, themselves. So do we, do we, I think where you need to maybe, um, you know, if you're a developer, you, you need to learn the tools and the APIs.

The APIs are evolving pretty quickly. Um, uh, but I really think it comes down to there's a little bit of like going back to science. So, um, these are, these are massively complex systems and just like the universe or you know, our ecosystem is massively complex.

We have to go back to sort of scientific methodology, right? Hypothesis experiments and validate, you know, our approach. And we'll be doing the same thing with these LLMs.

Alright folks, you heard it here. I would argue that the best way to learn anything is to play with it because, well, there's no substitute for having fun. And as one wise man once said, the only thing we have to fear is fear itself.

Hey Greg, thanks for being on the show and sharing your insights. Mike, thanks for having me. And thank you all for watching the latest episode of the Textron AI video series.

You can find this episode on our website. We invite you to check out other episodes there as well. Till then, we'll see you next time.