LLMs in the Enterprise, Can They Trust Them?

[0:00] But it's like, "Hi, I'm Nick, here's my little badge," you know? This whole time I haven't known your name and I've been afraid to ask because I forgot it, and that name tag has saved me. So thank you, I'm really appreciative of that name tag. I am Nick Wu of Illini AI - that's - we got that one today. Illini AI. Close, close. Not quite. Well, it's almost there, yeah.

[0:21] Ah, how's you guys' day going? Back to backs, back to backs, baby. It's Tuesday, it's how it goes. Stuff though, it's good stuff. Yeah, this conference was cool. Lots of really interesting speakers. This guy, the last speaker, is Harvard Data Science Review, which I've read Harvard Business Review - I didn't realize there was a Harvard Data Science Review publication as well. But super relevant content. I - I'm really looking forward to reading a lot of the articles that they're putting out like this month too that he was talking about as well. A lot of it's around especially the role definition and data scientists, data analyst, data engineer, and really like doing industry studies around that too. So it's good, it's really good.

[1:05] I saw there was a presentation on the Data Connect Conference years ago of a data scientist doing a language analysis on roles in data science. Yeah, that's about as meta as it gets.

[1:18] It was. And they said he himself said this one was a bit of a data science project where they were scraping LinkedIn, they're scraping job descriptions, and everything related to that, just trying to get a sense of what is statistically significant around these roles, around definition of them.

[1:35] One, two, three, four. It was intended for the human to support the machine, the machine to support the human network. An AI means something different to anybody you talk to, which is wild. This is AI or Die. And it was interesting too because of course at an analytics conference, there's going to be a ton of talk about ChatGPT and generative AI. But it seems like the genera- like the overall takeaway that folks are having and they're speaking about this is making sure that people say that they used it, but being fine with it being used as a creative tool. Like, "Hey, give a shout out that you use it as an aspect of what you generated here, but we're not going to stop you from it. We're not going to hold you back like they held back calculators back in the day in terms of being used in the classroom."

[2:16] Yeah, I mean, at the enterprise level, people - and you're talking about education, right? Or no? I'm talking about specifically education, academia, the folks who spoke about it today. Yeah, enterprise - the enterprises are locking that down for sure. Yep, definitely. I know we're probably going to touch on that, but...

[2:37] I don't know where he went. Yeah, see you later, man.

[2:42] I had to get my coffee. Oh yes, yeah, this is gonna be a late one.

[2:49] Yeah, I know. I - I need to - I'm on my third cup today actually, just to get charged up for this and the drive from Cincinnati tonight too. Yeah, oh crap, that's right - drive back. I know, I'm kind of pissed I'm missing free drinks right now. There's an open bar across the street for everyone here at the conference, so I'm doing the drive anyway. I gotta drive anyway, so it's a very good point. So he's over there at the open bar obviously, getting free drinks, and then I'll pick him up here in a little bit. It's great. [3:20] Yep. So how's your guys' days going? So I'm in Cincy at an analytics conference, obviously supporting Ali out here. She had a great talk with Rachel too that was really good. I can't wait for them to publish more information on that. But what are you guys doing today? Where are you at? I had a lot of good heads-down time today. Oh my gosh, hence down time today. Also contract reviews. Heavy, heavy reading then. Big - yeah, big literature day for you for sure. I've got my passport renewed, so that was my big administrative accomplishment of the day. Exciting! We're going international, folks. Illini is going to go overseas here in a little bit. It's exciting.

[3:57] Yeah, we actually have our first international customer already, so we're technically international, but now we'll be resourcing international as well. It's a big deal. That's awesome.

[4:09] Yeah, these - these conferences are always - it's such a drag to drive down to Cincy, but I always - I'm happy I did it. Just re-energizes me - the topics that are talked about and just so people are here too, and just reconnecting with old folks I know as well. It's really nice. Yeah, it's - it times is really fun just to get in a room in person, feel the energy, you know, have the random conversations with people instead of these like ultra-scheduled virtual meetings - hard start, hard finish - yeah, over and over and over again. A lot more whimsical when you're just running into people and having conversations like that too.

[4:43] Yeah, I agree. Add some randomness into the mix. Agree, totally, totally. Well, should we start getting into some of the hot news and trends? Let's do it. Kind of ready, kind of excited. There's some interesting stuff that's happening, especially at this conference that folks are speaking about too. And I think the biggest thing that I know came out yesterday that we want to talk about was Jeffrey Hinton, you know, big deal at Google, wants to obviously quit. I see it is really retirement - the guy's 75 at this point - but seems like he really wants to speak openly just about his concerns. And there was a large New York Times article that got released around really his thoughts and his concerns related to that. So do you guys get a chance to read that article?

[5:25] Yeah.

[5:30] It's so interesting because I've talked to a lot of people in Canada about like the AI space, and like one of our partners is in Toronto, another partner of ours is in Canada as well. In their time, you know, they're always talking about like, "Well, you know, the birth of AI - it was in Canada!" Come on, really? They're trying to be like "First in Flight," but Canada for - for AI. Okay, I mean, you know, so this guy's - guys in Toronto, did well, did his research there, right? Lives - is from Toronto. So anyway, and just kind of resurfacing on this like advancement in the AI field around neural networks that this guy did, and the fact that he's been at Google for such a long time, and some of the work that he and others had contributed to the field - I think is very fascinating. And just looking at the advancements of the techniques over time and the - the use cases that it's unlocked, I think is super interesting.

[6:28] So I don't take - I mean, he's been very clear, and I can't tell if it's like more of a publicity thing or not, but been very clear about and not being associated with Google, and he just wants to be able to freely talk about some of the implications, which is such a fair statement because I think everybody's contributing to that now, at least trying to publicly contribute to that conversation, which is definitely a good conversation to have. So I don't have any sorts of feelings about it. I just think it's interesting. He would - he was very specific to say like, "This is not anti-Google." Like, Brendan, you shared that as a follow-up that he had.

[7:12] Yeah, I was gonna say that was my main reaction at reading the article. I was like, "Wow, this is really -" It's when I saw the tweet first, then read the article, so I want to clarify my frame of reference, because the tweet basically said - and I can quote this - like, "Google has acted very responsibly. The article implied that I left so I could criticize Google, but I found that there - I - I actually left so I could like talk about the dangers of AI without considering the impact on Google." So it feels like it was a lot more of like a neutral move or like a really realist - like reasonable thing to say, like, "I don't have to consider Google when I'm talking about this," where the article definitely had more of that spin. And I see we're - we're getting a lot more of that like hype, that spin, that "how do I make this into this big AI is scary and companies are doing whatever with it?" And like, I think that's just kind of the - the days we live in right now is it's in the New York Times, so it's a big deal, and there's a lot of like narrative being drawn around this as well.

[8:09] So I just thought it was an interesting like counterpoint of like seeing his actual tweet and what his stance was coming out of that article.

[8:15] Yeah, and like, you know, obviously there was a lot of hype, and there still is, around the implications of the social media algorithms and the impact that's having on society, and that continues to be, you know, brought up at a - from a legal perspective, from a government perspective. And so I think this is just like fallout of the fact that we don't - we can't catch up on standards and policies and - and expectations - societal expectations. And so we're at the whim of these companies and the decisions that they're going to make in the meantime. And so I think that puts the fear in a lot of people, as they - as it should. I mean, that is true. So there's probably like that heartstring that people are really trying to tug on to with some of the journalism happening around this is just we are at the whim of the decisions being made by these companies until we can catch up, which I know we're planning on talking about some of the stuff that is happening in Europe too, to try to combat that.

[9:07] But it's just - I mean, that's where we're at, you know? Self-governed at the moment, and with the pressures of people trying to hold these companies accountable or at least whistleblowing or, you know, whatever, right? In hearing you say that, does almost strike me as again anti-big tech companies, like the narrative here being, you know, "Don't trust these large tech enterprises in Silicon Valley. We need to figure out how to regulate them." But a lot of conversation today at this conference was essentially regulators move very slow, so they're - we're in that in-between phase where almost the industry and academia at times wants to really emphasize the speed to actually get something out there and the speed to actually get insights from it, but at what toll from an ethical standpoint around if we're leaving it up to the individuals actually building and maintaining these models?

[10:01] There's obviously no central governance around it, and there's groups that are trying to stand that up and really define it, but in the meantime, how long is it going to be? I don't know if anybody can say.

[10:13] Yeah, and I think there's been a lot of conversations about this publicly for a while. It just that it hasn't taken over a common narrative for the everyday individual. And so I think that's been a really interesting transition that has happened. But, you know, we are starting to see those conversations come into place. But I will still say like so many companies that we're talking to at the enterprise level are saying it's way too early to even think about AI governance, which is so crazy because it's like I get it from a build perspective - there are people who are building simplistic kind of machine learning models internally, have been doing that for years, and then have been using them and putting them in production systems. And maybe they're low risk because they're not affecting like all of society - they're affecting things that are kind of operational internal to their company.

[11:07] And so there's not as much of an ethical pressure there. But now there are all of these external systems that they will inevitably interface with. And so - and even more so now because there's going to be a flood of tools utilizing them, and they have to have some sort of stance on it. Which I know OpenAI had just released the fact that they were going to push out this kind of business plan, and they - they pushed out this feature in ChatGPT around in - in the settings that you can actually have it not utilize your chat history into its training mechanism. So, you know, it's - it's interesting to see where that's headed. They're reacting to it pretty quickly, I think, just given the fact that it's May and they're already talking about releasing, you know, functionality for the enterprise.

[11:54] I just think it's kind of crazy to hear companies saying that it's too early for them to think about what their governance and control system is going to look like internally.

[12:01] I chuckled today because somebody said to me - we don't have to - today at the conference - "We don't have to worry about governance because none of our models are in production yet, so we don't need to think about like the ethical side of it until it's live in production," where it's - guys, you got to get ahead of it. Like, it's - it's not just wait and see what happens. You have to think about the framework today, or they're gonna try to avoid other so...

[12:23] Apply for XYZ. So regulators and compliance - I mean, they're coming at us for some stuff - use of data, obviously data privacy, data usage, whatever - but not yet knocking on our door about the techniques that we're utilizing in the AI space or the use cases that we're leveraging AI for, which I think is kind of interesting.

[12:44] Yeah, I actually noticed that was an interesting thing back when we worked with some of the banks. They would avoid like the traditional predictive stuff inside of the sensitive data - like bad example, but like whether or not to give somebody a loan, right? Because that's highly regulated space. We actually heard this recently too, where they're avoiding those high regulation - like scrutiny areas. And so what actually opened up was more of like internal compliance stuff of like checking people's emails, checking their message history, things like that, or any insider trading or like anything that's uncompliant. So it's actually like AI for compliance use cases was actually more approachable. So it's kind of this interesting like, "Oh, AI - is it going to make us like more biased or more unethical?" And it's like, "Well, it's actually helping implement and scale up what you can't really do usually around ethics that exist today." It's one of those interesting catch-22s.

[13:32] That's where we're getting a ton of interest from like cybersecurity teams approaching us just around, "Hey, can you help us navigate AI and how it's being integrated into the enterprise? One, cyber-specific tools that we should keep in mind as a central team that could help us and optimize our performance, but also other tools that business users maybe in sales and marketing might be bringing in that introduce vulnerabilities to our organization as well." So it seems like top of mind for them especially.

[13:56] Yeah, I was just - say cyber, infosec conference I was speaking about data and AI capabilities inside of the enterprise and - and why cyber teams shouldn't care about the procedures that their associates are - are leveraging, because right now they're just - they're not centralized, they're not consistent, and they're not documented. So from a cyber perspective, that's - that's a risk that there's even variability across how people are doing what they're doing. And there's no reinforcement, and there's really poor adoption of a lot of those policies and procedures. So, you know, brought that up because I think that's really interesting. The cyber world leans really heavily on frameworks. NIST is one of them, and they obviously posted or published that AI kind of governance or risk framework.

[15:09] So I brought that up as kind of a parallel for them to - to look into and consider. Even looking through some of the, you know, legal documents that I look through so that we can work with some of our customers - they'll explicitly call out data security and adhering to a specific NIST framework, just to ensure that they're - vendors and their suppliers are compliant with a lot of these regulations basically, or - or they're compliant with the frameworks that they've considered to be kind of their, you know, stamp - formal stamp of approval. So anyway, it was just - it's just really interesting because now we're getting into this territory where cyber teams are being asked - a lot of them - being asked to weigh in or provide some stance for the organization on what the guard rails are going to be and what guides, you know, the company should really follow.

[15:47] And I think at first, which most companies have done, is just kind of lock everything down and try to control access and control the way that people are interfacing with some of these tools, and figure out, you know, just buy them some time basically to figure out what their approach is going to be. But it is a fast-approaching issue. And to your point, Nick, like a lot of these cyber teams are just trying to really understand the level of - or the layers of complexity in this space. Like, what should they be thinking about? Are they thinking about everything in the landscape and the ecosystem? And - and that's a huge challenge. Like, we've been approached, as you mentioned, to provide some even foundational kind of workshops or education for cyber teams to give them a one-on-one-oh-one on the basics of AI so that they can feel like they have a good foundation coming into these conversations. But it's definitely coming for a lot of these teams.

[16:35] I know you mentioned like people aren't starting with this up front. I think a trend or an opportunity that we see is there's new ways that are working in that are coming into the team, right? And there's the piece where you need to like teach them, upskill them, have them be able to do that. Some teams are prioritizing that, some teams are not, right? And then the next piece is this definition of process, documentation of that process, really like structuring clearly how they should approach that. And then the third layer is kind of that governance piece of "we need to generate audits, we need to have procedures documented so that when we go talk to regulators, like we have the data we need to have for that conversation." And I know some teams don't see the need for the upskilling as much, and they might not see like the governance need yet, but I think that middle layer is very important because that's usually what's going to help you with like troubleshooting the operational pieces, the efficiency of your team, like a lot of those critical components of when you're going from one to ten models in production, right? Where you are trying to make sure everything's put in place the same way, you understand how things are structured, right?

[18:09] So that middle layer and that documentation, how-to articles, things like that that can kind of go by the wayside - really the design of those processes and less so the capture - that's important too, but really just having a clear process that your team is following, I think is really critical right now. So even if teams are not seeing the governance stuff come up yet around AI, having a solid foundation that helps with the operational piece is still a good use of time right now.

[18:32] Yeah, we're hearing such interesting terms in a lot of the conversations we're having. Like, I heard "job aid" today. Like, I know that's not a new term, but like it was thought about in the assistance of doing my job, and it's not - we don't have to like, you know, get lost in the - the details of documentation per se, but just think about what is this job to be done? Like, what are we trying to accomplish? And I think that is rapidly changing. I mean, talk to any company that's doing a modernization effort around their data - you know, they're - they're basically saying that all of their associates are going to undergo this like fundamental paradigm shift on how they work with data, right? And so that means that they have to do a new job. Like, they are doing something differently.

[19:19] They're used to working with data - they don't need to learn what data is - but they need to learn this new paradigm shift, and that is non-trivial, especially as you're migrating to a new tech stack that they have to interface with. They're going to try to apply the same paradigms to the new tech stack. This is what a lot of companies did during the cloud, you know, migration era. They were trying to apply their configurations for on-prem to the cloud, and it was just a fundamentally different paradigm. And so they have to - they have different ways of doing it, and they have to be familiar - familiar with those weight or what those new ways are and how to interface with the technology. It doesn't mean you don't know what you're doing or you don't know the core concepts in the space. It's that that workflow is so drastically changing because you're trying to move towards something that's more effective.

[20:08] So I think documentation gets lost in the lots of all of that, which is - which is sad because it has such a bad connotation associated with it because we're all used to sitting down and like typing everything out all the time. But that landscape is changing, which is so fun and interesting. Like, we're getting to that point where that's not how you have to do it anymore. So I just - I find that to be really fascinating. I think it's all related. And I - I'm the new phrases that people are saying in these conversations are really triggering new ways to communicate at anything.

[20:57] Yeah, I heard a term today - "practice-centric education" - which I absolutely loved words. Here's our practice of data ingestion in a way, and documentation being such a key aspect. Maybe education is not the best word, but it's essentially here's our practice, we documented this practice in a way for folks to basically reference it and practice it consistently in their space too. But I know we have Mr. Gordon calling in maybe. Excellent!

[21:22] Hey, what's up, man? Hey, hey, hey, wow! So I'm gonna get a very special guest today. Pete, thank you so much for joining our podcast AI or Die. Let's just spend, you know, the next 20 minutes just talking about what you're seeing. Just want to thank you again for taking time out of your day. Could you first start off just by giving an introduction of yourself for the folks who are listening today?

[21:45] Yeah, I'd be glad to. It's great to be here. My name is Pete Gordon, and I'm a software guy in the central Ohio area and have been for quite some time. You might be able to tell by the little bit of gray on the tip of the beard. But yeah, software development, software architecture, and deep into data engineering, and got dragged into the AI machine learning world about eight years ago to kind of build on kind of statistics that I had in my couple college degrees - a little bit of, but not fully. And I've been doing that kind of on and off in addition to the software development and architecture for quite some time. And then more recently, which is the reason I'm here, got dragged into these large language models. I was doing NLP back in 2016 and kind of traditional chatbots and Google Dialogflow and IBM Watson, which was all the rage - Watson Assistant - and more recently, OpenAI is - is where it's at, and we'll see if Bard comes along. So there's a little bit about me.

[22:37] Nice. And I feel so rude having you go right into an intro. Where are you at? How you doing today, Pete? What's new with you? Thanks, thanks for asking. Now I'm doing great. I'm - I'm in a hotel and about to freeze to death next to the air conditioner in Chicago. I'm actually up here for a conference put on by the Institute - International Institute of Analytics, the IIA group - in Chicago. They have a symposium this week, and I'm - I'm up here for that.

[23:07] Yeah, are you just attending? Are you speaking at all? I'm just attending, yeah. I'm like - I'm part of their expert network and have done some consulting in their expert network related to large language models and up here and other things - identity resolution and visualizations and such. And then I'm up here to meet some of those people I've talked to over Zoom just like this, so hopefully in - in real life.

[23:29] So I'm curious - what kinds of questions you're getting over and over again from folks around LLMs in general? Like, we talk a lot on this podcast about some of the like societal implications, the way people are thinking about it, like, but I'm - I'm curious - are you getting more technical questions or people trying to like wrap their brains around what it is, or are they more so thinking about kind of the user experience side?

[23:49] Yeah, actually that - that's great. I heard like kind of three parts to your question there - like technical like versus like business versus like, you know, UI and user experience. I - I think they're almost all like business goals - like what is this? You know, how am I going to use it? Some questions even though specific down into just like the ethics of it, you know? Can we trust it from a large language model perspective? A lot of people are starting to ask about policies, right? Like internal to companies and - and issuing those. But most of the conversations I've had with larger enterprises have been around what is this thing and - and really giving them context, kind of like I did a little bit in my intro there, to traditional natural language processing and how it changed with Transformers and BERT and "Attention Is All You Need" over the past handful of years and what it means today like with GPT and - and others.

[24:33] Would you mind explaining what "Attention Is All You Need" to the listeners to just so they -

[24:38] Yeah, yeah, sure. "Attention Is All You Need" is - it is - it - um - paper that was written by some people like Google and I think Stanford too, but mostly Google. And it started this Transformer revolution. So before, in natural language processing, the kind of state of the art was a thing called LSTMs - Long Short-Term Memory - and it was two recurring neural nets encoding, decoding, went back and forth to give us better. And it was really translation - like English to Japanese, Japanese to English - was like top-notch like 2012 kind of time frame, you know, 10, 11 years ago. Like this kind of really started I think from, and it took a hold like 2014 time frame somewhere around there. And then I think it was '16 or '18 - I can't remember - the "Attention Is All You Need" came out, and that introduced the world to instead of this LSTM and two traditional or recurring neural nets with encoding, decoding, it introduced Transformers and doing neural nets based on Transformers. And that just changed the game, you know? And I'm not a mathematician, so I'm not gonna like get - even pretend that I can explain it.

[25:46] I have had conversations over beers on what are lie groups and functors with - with professors in mathematics, and - and I'm still trying to grok it. I'm still trying to understand it. But Transformers was kind of the - the big thing for these huge parameter-sized models that were seen with GPT-2, 3, and - and now released to the world as ChatGPT.

[26:10] Yeah, I think - again, I'm also not going to try to explain it because I'm not a mathematician - but some of the things that I was reading in terms of the big paradigm shift that happened with Transformers was along the lines of the recur- the recurring neural nets having to do things sequentially, whereas the Transformers did not have to do the processing sequentially. So that opened up the ability for a better compute - like more efficient compute with the algorithm - and more dimensionality or capabilities around that as well. Which I think is really interesting because when you look at like, "Man, that's a big shift," and like we all just like - well, most of society got exposed to it like in November, yeah? And - and like you said, this has been happening like - the research for this specifically on Transformers has been happening since like 2018.

[27:52] And so there's a massive iteration happening in a short period of time, which is I think where most people in the field are getting a little uneasy. So what are your kind of like general opinions on that, or like have you heard - have you heard people asking about that specifically? Like, is this going to get out of control really quickly?

[28:09] Um, well, it depends on what do you mean "out of control"? I - I've seen the rumors, right? Like the discussion amongst the field in terms of data science and AI machine learning and everybody trying to kind of catch up like with the large language models and how does this impact kind of the traditional supervised learning, right, approaches, which I - I mean, are still good. They were always good on a kind of a case-by-case basis like applied - applied statistics. I don't know, it depends on where you're going with that. Like, are people scared of their jobs? I think every programmer and writer is right now a little bit.

[28:47] Yeah, it's kind of interesting how it goes back and forth. Like, some people say it's - it's gonna have little impact, and you can't trust all the codes that's out there. I was just having a dialogue on LinkedIn with programmers talking about it taking over, and others are just like, "Programming is no longer a career I can do until retirement." So but that's more on the like the program - program side, and I've seen some worries and concerns expressed on the data science, right, side and statistics side. I think they're both kind of unfounded. I think that - I do - I say that, but I - I want to say that I do think that this is - this is a shift like equal to mobile applications in like 2008, 2009 kind of time frame - kind of BlackBerry to touch like full touch with iPhones and Android devices.

[29:29] The other one before that I think was kind of like a cordless wireless revolution was like 2003. That everybody started - I remember I had this like Apple like spaceship, which was my Wi-Fi router in the house, and it was like everybody got Wi-Fi in their house in like 2003. And then you look at the internet as I like to put like 2005 kind of on the internet as far as I think Netscape went public, I think, in that year. I'd have to double-check, but it's kind of like internet, cordless, mobile, and then now I think we're talking about another like revolution just because it changes the way humans interact with knowledge in a way that I think is fundamental. I know I'm not - I'm kind of going off like philosophically.

[30:05] No, it's great! I think because that's the major concern too that we've seen a lot of people talk about is this "how humans interact with knowledge" piece, yeah? Like is - I mean, the misinformation has the ability to - to generate more misinformation, which we've already been struggling with societally. So I think that's definitely one of the bigger concerns that people are - are fearful of. And then knowledge workers who always consider their jobs to be secure are probably now also fearful about what gets replaced. And there's this idea of like replacement versus acceleration and enhancement and how many people can make that shift and what is that enhancement level us into.

[30:34] And so as you're talking to some of these companies and they're asking like, "What are the implications on us as a company?" like what's your response typically, or what is a general research?

[30:42] Um, so I think the implications are - are kind of jumping the gun a little bit. Like, "We don't know" is - is the right response. They're like - I think the initial response like that needs to happen from these companies is kind of a typical like piloting approach - like where are the potential uses? Where are these - where is this capability able to - to better, right, the way that we work and the company works? I did notice in - I had one large kind of group session across multiple companies, and I was a little surprised that there was - it was about about a dozen companies represented, and only a couple of them - like one, maybe one and a half kind of - had - had exposure. These were permanent analytics leaders too, and they had exposure to the natural language processing kind of previously, right? So Dialogflow and chatbots and Watson Assistant like I mentioned earlier, and other Spot framework for Microsoft - it's been around a long time and such like that.

[32:32] And they seem to be, right, at that stage of needing that piloting, right? Like, "How do I find the right business cases to pilot this and begin to taste it?" You know, begin to get a value out of it like that I can say, "Okay, I understand it, I understand the parameters involve the scope, and what - how it might be useful and where it might go off the rails." And - and that seems to be, I think, where they're at right now, and not even yet taking the steps - like we'll see - like I'm expecting them to start taking steps, and I'm seeing a little bit of it, but it's embracing it to - to pilot it and how it can be effective. Startups are going nuts like embracing it like this whole hog, right? They're like, "Give me more AI!"

[33:49] Yeah, there's so much pressure from the venture world on startups too to become more efficient, especially in the macroeconomic environment, and to reduce OpEx wherever possible. So I'm sure there's a knack for curiosity there, but there's also somewhat of an external pressure. And with these enterprises waiting to pilot, Pete, like are you getting a sense of like what are the key hurdles that they're seeing that's stopping them? Is it like an ounce paralysis - "We don't have the team ready to go, we don't have the stack that we need to do this"? Like, are you hearing common themes around "we want a pilot but we are waiting to because of this"?

[34:15] Yeah, I think - I think, well, I think there's kind of like two responses of that, and one is I think that like the one and a half, you know, kind of the - the two, the ten percent, right, roughly, right, but that had already ventured into NLP, right, from an analytics perspective. I think they kind of are getting their world shifted because now it's like, "Hey, wait a minute, what's this large language model stuff and how do we exercise that?" The others I think are - are just yet to take the - the steps required, and it might be like technical limitations is kind of what I'm picking up on. Overall though, I would say that they're all very excited about the opportunity to leverage this for their group, right, within the company. So the idea that they can like carry this flag, right, into executive leadership meetings, right, and - and be involved in this is - is definitely nothing but up in terms of the opportunity, right, for them to - to leverage it for their own groups and to start to apply it to other business units, right, where they can show their value.

[35:10] So that was probably the biggest takeaway, right, from the group was, you know, there is nothing holding them back, right, from taking this flag and running with it, right, in their organization. And I think that right now it's just a - a limitation of "okay, let me figure this out exactly and - and how do I approach it in terms of a business sense and do I have the technical capabilities to like leverage this?" I think that's really interesting because like a technical capability right now, and there's been some shift like over the past, especially with GPT and the chat API versus the conversation APIs, which have kind of separated and moved away from, but there's really only like two APIs an OpenAI that really matter, right, today. And that's the - the chat API and what's called the embeddings like for fine-tuning - it's not fine-tuning because I don't want to confuse that with the fine-tuning API because that's an older one where you actually upload your data into OpenAI.

[36:27] But the embeddings and the chat API are like the two that matter, and embeddings is "hey, I want my data to be able to look at it for similarity across the - the vector similarity database," right, which is everybody's all excited about like this whole vector similarity search. Everybody's going nuts on it with embeddings. And I tell you what, like I never - I - I was first introduced in embeddings in natural language processing like more like Word2Vec was what I understood. But I was introduced to the concept of embeddings with a co-worker four years ago, I guess it would have been around that 2018 time frame, maybe 2017 even, and I never truly grokked it. But now like with this API available and using embeddings with the large language model, like it's starting to like finally made sense like how you do cosine similarity and leverage your own data and then use the prompt engineering techniques like with your similarities to like actually get natural language interfaces that are a really productive for your own specific area of work, your business domain.

[37:44] And Reagan mentioned kind of user interfaces, I think is - is what she said, or conversational interfaces. I think that hasn't even been touched yet, and I'm super excited about that. We have - I have one project - a group I work with called Tuulian, and they help like emerging developers like junior developers, people that want to get into software development, and they do GitHub issues in relationship to your developer profile. So they find open source projects that have outstanding issues that your profile matches to such that you could contribute in those open source, right, and participate, right? Yeah. And these are for people globally that want to participate in software development and - and increase their software development skills. I think that's a great like use case, but I bring that up because like we are literally right now having that - or I say right now, right now we're wanting to have that conversation about the user interface and what is - you know, when I - when I stick up - you want to stick a chat in front of somebody, how am I really interacting like with their goals in relationship to the system, right? Is it just open-ended entirely, or is there a way to direct?

[39:00] And I think what's so interesting is that and it has been open-ended because they're trying to figure out what sorts of clever patterns are people coming up with to try to get it to do what it wants to do. So if you - if you could just explain embeddings like at a very high level for folks to give them the paradigm to think about like why this is important in terms of engaging with...

[39:25] Sure, sure, yeah. So embedding - the embeddings API, the one API that is called embeddings at OpenAI, is effectively "I have a lot of my unstructured documents or unstructured text, and I'm going to be able to logically like divide it in some way." So it might be like the GitHub issues, right? Each of those issues, right, has some logic to it, and we might even split those apart, right, by a certain number of sentences or paragraphs or whatever, and then each one of those he's thrown up to OpenAI, and with their large language model, it will give us back a vectorized value with each of those. And that allows us to like start looking at similarities with other natural language. So when you type in a search or you type in a paragraph or whatever it might be in terms of natural language, then you go up and you get the vectorized value of that, and now you get cosine distance between those - those then bubble up to the top based on the - the similarity it's called. And that bubbling up allows you to use that then in your natural language prompts against the chat API, and this is where you get specialized domain knowledge, right, your knowledge base into the large language model so that it starts to know how to respond to things.

[40:51] So it can actually respond with your - your business knowledge, your policies, your - in that example I gave, the - your GitHub issues the are most similar to your profile, right? And these vectors Oracle representations of that language in the space of the language that the model was pre-trained on - correct? That's right, that's right, that's right. That's how it creates it. That Sim - similarity that you're talking about is how close in this vector space is this language to all of the language and provide some mechanism for it to compare similarity of language. That's right. And - and usually use it with the question of like surge, right? So then you have some language associated with search, and then that's factored in as a vector like space - it gets a vector for where it is, and then now you can look across yours and see how similar it is, right, your data to what is being discussed now or being asked now or being searched for now. And you can do that with anything - it doesn't have to be necessarily something to search for. You can even do it amongst itself, right, to find out like maybe even the scope of a certain domain of policies, right, within an organization.

[42:12] Yeah, and I'm curious if you see that's kind of the biggest opportunity, right, as enterprises are looking to use this? Because obviously ChatGPT, it's, you know, generally been pre-trained, right, on the broad internet because that would be a lot of what your enterprises are asking about - like how do we make this more specific to our organization, or kind of what requests are you seeing on top of that and general foundational model?

[42:33] It might be a - yes, it's coming. It is more me making them aware of it than them asking. I mean, they absolutely are asking the question like, "How do we approach this and how can we leverage this?" right? And then making them aware of maybe areas like, "Hey, imagine like all of your business processes, right, were actually in this vector space, right, so it's vector - vectorized such that you now are able to search across them and compare between them in terms of similarity, right? You can start to like see potentially the codification, if you will, right, the quantification, codification of your business processes in a way that you couldn't previously. So you - you could start to see where ambiguity like truly is, right, and where are more I'm concrete in terms of their relationship and the similarity." So they - I think they're - they're asking, but it's still at a very high level, right? Like, I mean, I know they have asked me like how they can apply this, but they're still just trying to get their heads around what it is, right? Let alone how to apply it. And that's where I think I expect the piloting to like really take off probably like the second half of this year.

[43:45] I think this first half of the year is more of them just - the first two quarters - just trying to get their head around it.

[43:51] Yeah, and what I love about the way you're talking about this too, it's like what it is and how it works at the highest level without diving too far into the weeds on how it does what it does. I think is so important to understand what use cases can be applied to it because then you understand the general function that you're working within. It's so hard - I think that's what - what's so hard for a lot of these groups. Like, you've got innovation teams that their companies thinking like, "Oh my God, we gotta jump on this. Like, what do we do?" And, you know, it can be kind of technical to - to explain how it works right now. There was a really good article from Cassie Kozyrkov recently that was basically talking about how AI used to be buried in the guts of a product, and this is the first time that people have been exposed closer to the like engaging with the model in a more transparent way - sort of - or a more direct way, I should - I guess I should say - which is what's driving so much excitement because the possibilities are endless for people thinking about having how they could leverage this in the use cases for it.

[45:23] So yeah, I do think it's super important to understand, to your point, what is it, how does it work first, maybe at a deeper level than just a generative pre-trained transformer like as the language itself. But yeah, it's so interesting, and I agree the pilots are going to be really fascinating too. It was - I worked - I did a initial testing with GPT-3 in I guess it would have been mid - early 2021, might have been late 2020, and, you know, it wasn't like - using the playground from OpenAI, it was not effective. We're trying to accomplish at that point, and I heard a quote from Sam Altman about - the CEO of OpenAI - about how nobody really produced the interface, right, that they ultimately did produce with ChatGPT, and they were kind of like waiting for somebody to do it.

[46:35] And I don't know if I totally agree with that because I - I know there was one - I can't remember the name of the company that does it for writers - it starts with a J, I think - and it seemed like people tried, but it seemed like the playground just wasn't there like either like with GPT-3, like, and I - I don't know what the disconnect was for the - the kind of like the two and a half, roughly, years - the two years, right, like between like kind of 2020 and 2022. But I tell you what, this - this like putting in front of people like you said so that people could just touch it and experience it at every level, right, not just AI people but anybody, was a game changer. And game changer, and I like to describe it in how we interact with knowledge like on the internet - like it's a game changer for that.

[47:17] I heard so many discussions today around prompt engineering, which is I'm sure you're familiar with as a term from years ago, in the general public now of like, "Okay, our folks are using this. How do we teach them foundational prompt engineering principles so they can use and integrate with it correctly?"

[47:31] Yeah, yeah, I've heard - I've got mixed emotions on both sides. I've heard one side - I was sitting with some people in - in discussing this, I don't know, a month ago or - or more, and one of them said "Google-fu" as - as kind of described prompt - prompt engineering skills are. On the other hand, when you - when you really go through the - the literally infinite - the - the as expressive as the English language - human language is, right, not only English, but as expressive as English language is, so is expressive the ability to prompt engineer, right? So it's - it does seem like something a lot more than how to use quotes and minuses and sight, right, on - on the Google-fu side, right, versus actually being able to just whatever I can express in English, there is a - a probability that it will direct, right, the - the outcome of interacting with a large language model.

[48:27] And yeah, it's a - it's definitely - I don't know that it's neces- I don't - I - I haven't really landed yet if it is - I don't necessarily believe it's a role personally - a prompt engineer - just because I think software engineering or software development or, you know, code scripting or script kitties, you know, like - you know, like - like the way we like nuanced distinctions like in building software and creating with software has existed for a long time. I don't know that they're needed that much because there's like there's a fundamental like logic software engineering - breaking tasks down, having the tool structure and the repeatability to build, right, in tasks that it's just always been there and always needed. And that's kind of how I view this kind of prompt engineering - it's just - it's with language rather than code.

[49:29] One more question for you before you drop off about Auto-GPT - thoughts on like the - the productivity gains or the fundamental difference with Auto-GPT and these agents and - yeah - against what we were exposed to before?

[49:44] Yeah, so the question is - is like what do I think about agents and this kind of bringing together like the plugins and Auto-GPT to do like scraping? I mean, really what this is is just the expansion, right, of this large language model into the internet itself to be able to pull more and to do more. I'm - I'm excited about it. I've been working with someone who's working closely on the LangChain side and likes the flexibility of like swapping things in and out if anybody's familiar with that SDK - LangChain - but and the Auto-GPT is interesting, and I'm huge - I - I think a couple of you might know I'm big into the - the internet and scraping and things like Puppeteer for like driving automation like on the internet.

[50:27] I actually see that as just kind of expected. It's exciting, right, but I think this is more like the traditional just even automation like out here. I'm gonna use that crazy - I kind of - I was about to say stupid like enterprise buzzword like RPA, like, you know, in a robotic process automation. You know, like is it Auto-GPT, or is it - is it RPA? I like - it kind of feels like I don't really care, like - like the big piece for me is the language and the - the - the - the power to create that we have now with language. And if we want to call that prompt engineering combined with automation, if we want to call it Auto-GPT, right, the ability to use language and combine it with this automation is an absolute game changer, no matter what words we use to describe it or no matter how we - we wire the - the sources together.

[51:27] Yeah, that's so great. I - it was just so interesting because when people are talking about the catastrophic effects of this, they're talking about the access that these systems can get or people can hook up access to these systems that can create, you know, catastrophic because it's just a model until it can actually action on something and create consequences. Yeah, it actually gets - and that's - that's our next - our next dialogue we're going to talk about the control on the internet when it comes to automation and when it comes to - when it comes to knowledge and automation control on the internet. Like, there's another piece here that I'll say I've been expecting for a long time - a very long time - to enter into an API economy. I feel like we've actually just cracked the door to that probably in the past few years, right, because people have finally gotten to the point that it starts to be regular in terms of API. [52:11] But there's tons of issues around control and ownership and - and who - who's calling what or who is getting called, right, with web hooks versus like API and web services. I think the issue of control is yet another thing that we see, right, with now we've got language, we've got language interfaces, we have automation - automation - we've always had this issue of like control - like where is the data? Can I get to it? It just moves to a much better and higher level of abstraction. It's no longer your, you know, Postgres database - it's - it's the actual information source or whatever that entity is that's out there that owns it.

[52:49] Fascinating. Pete, thank you so much - I - for your time today, for obviously your insight, and you're just so well-spoken. I feel much smarter just through sitting through this conversation with you, and - and you're a very first podcast guest, so thank you so much for just helping us get something going. I am honored. It's my pleasure. I love seeing you guys here and love working with all of you, so it's been a blast, and I - thank you so much for having me. Of course, yeah, thank you. Appreciate it. I'll see you guys. Bye-bye!

[53:20] All right, gangster. That is episode four. Any additional thoughts, Reagan, Brendan, that you guys want to discuss before we wrap up today? That was great. Yeah, yeah, I mean, I obviously - we got into a little bit of the technical beats, which was great because I actually think that that's so necessary. You know, there's - there's definitely a huge need for a general understanding of what's happening and a breakdown of how - how this thing works and how you can engage with it. I think AI is more expansive than just this too, and, you know, obviously it's being talked about in this vein. But yeah, it was good to get like a high-level overview of what kinds of conversations he's having with some of - some enterprise leaders about this.

[54:05] Green. Brendan, any additional thoughts or reflections? Yeah, I'm just really curious to hear coming out of that conversation more about like prompt engineering because now I've been like hearing it a lot as a buzzword, but I know like Deep Learning AI just released a class targeted towards developer that's available for free. I'm also reading a book right now from O'Reilly around how the generative and NLP models work of the LLMs work. So I think it's just going to be really fascinating over the next couple months here to see how this really firms up of like how enterprises are going to use generative AI - like what are they going to try to pilot first? I think is going to be a really interesting piece that we find out in the next couple months here. And then also, how are they going to like get their heads around it? How they're going to teach there's, you know, their workforce how to use this stuff? I think it'll just be a very interesting next couple months to a year here around how generative AI kind of comes into the market and into the world.

[54:48] Grace and wild times to start a podcast. I think this was a fantastic episode for - thank you everybody for listening. Obviously, you can listen or die by going to our website illiniAI.com. You can subscribe or die on any of the places that you listen to podcasts, and we will see you again on episode five. This is AI or Die. Reagan, Brendan, thank you as always, and that's a wrap for episode four. Thanks everybody.

[56:06] [Music]