AMA Recap: Leveraging Generative AI as a Non-AI Company

Kate Hopkins (Host)
Hey, we are live and excited to talk about generative AI with Tony of nimble gravity. So to kick things off, Tony, do you mind giving a quick intro on nimble gravity? You in particular and your history in generative AI?

Tony Aug
Yeah, sure, thanks. Okay. So I’m Tony, I’m the CEO of a strategy, digital and data science consultancy, called nimble gravity. We help companies leverage technology and data for benefits. So whether that’s business or nonprofit, that’s what we do. And you know, kind of solving hard problems that matter is what wakes us up in the morning. I have a tech background. So I’m an engineer by education, and a technologist by profession. And now we do that as a consulting business. And to the question about where we sort of started our generative AI journey, formally, it was in 2020. With a lot of free time, given the pandemic, we found this thing called GPT. Two back then, and we said, you know, this would be really fun thing to try to see what we could do. And specifically, we’re trying to make a, an SEO text generation tool. So we’d feed some data to get some text out of it. And it was really horrible and bad and hilarious. But it set us on this journey of leveraging generative AI. And here we are, what year was GPT? Two? Because, yeah, so

2020 is kind of a lot of this stuff was happening. And this, if you recall, there was this, like, we’re not going to release the full model, because it’s too dangerous. And it’ll you know, there was a lot of like, it’s really scary. It’s, you know, mind boggling. So they released not very good. One publicly, they being open AI. And that’s what we played with. And that yeah, it was 2020. I feels I mean, not not so long ago, but it’s coming so fast. It feels like eons ago. Yeah. All right. So we have a bunch of pre submitted questions that will run through everybody watching, feel free to click to participate and join in the discussion and add questions, which will surface for Tony? So first up,

Kate Hopkins
can you give a little bit of background for the less technical folks on how, including you’ve taught me a lot about this? How exactly does generative AI work sort of the ingestion and generation process?

Tony Aug
Yeah, the how exactly. This will be the, the short webinar version or those short like, mini version. But at a high level, what’s happening is we’re breaking words into, or sorry, sentences into words, in a numeric sequence. So think of this as like a map of a sentence on a map of a paragraph, and, and so forth. That’s creating a numeric sequence of the way words occur in text. And so imagine what’s happened. And really one of the big things that open AI has done with DaVinci, and three, five GPT. And for GPT, has been to train on massive, massive amounts of data. So think of this as like effectively training on all text and the whole internet. So these maps become really, really big and complicated. So you’re creating a sequence of the way that words occur. And imagine that you then have sort of the, like, most of the time, this is the order that they occur. And so when you’re processing text into this, you’re building map. And then when you generate text using the generative side of this, you’re using the same map and reverse and making making words into sentences and so forth. So that’s at a really, really high level. That’s kind of the basic thing and these ideas of tone. beacons are kind of words or word fragments is kind of the heart of how that how that’s working.

Kate Hopkins
Awesome. And one of the one of the questions that came in pre submitted is, am I truly a step function, increase in progress or more gradual improvement? curious to get your sort of philosophical perspective? Perspective?

Tony Aug
Yeah, it’s a fantastic question. You know, I think I personally, I think it’s a step function. And I think what’s specifically a step function is that we’re now from a compute perspective and from a model perspective, able to change the interface between people and computers to be much more related to the way people would interface with people. And that to me, like the tech underneath the surf it Yeah, it’s great. The generative side is great, but like what really is changing is the way we interact with machines. And so like writing, writing, in normal English text or whatever language happens, speak now is the way we talk to computers as well, at least for this one piece. So I do think it’s a very transformative set of progress and the fact that now, it’s more or less complete to the entire corpus of written language more or less. It’s pretty powerful.

Kate Hopkins
Awesome. So can we talk a little bit about some of the different applications business applications of generative AI? And we’ve previously talked about these in terms of primarily analysis based primarily generative, and then truly both?

Tony Aug
Yeah, so So you know, I think we’re, we’re only really beginning to think of how we can use these things, I kind of liken this too, early iPhone, where there’s like, now there’s an app for that sort of thing. Now we have, there’s an API for that. So using that same framework, from the prior page, where we talked about the ingestion side, and the generation side, there are certain use cases that are kind of in all three parts of the Venn diagram here. So on the ingestion side, you can think of this as like, a really powerful way to summarize text. So you know, it might be summarize all the content on a website or summarize, a resume. And the summary side would be the generative side, where you can say, please, you know, tell me what kind of company this website is for. And here’s some categorization that I want, or I would like it in a CSV format, or whatever. Or, you know, I have seen a few few nice use cases around ratings and reviews for things like, is what the person is complaining about in the review, actually, the shipping companies problem, or, you know, is it a problem with the product, those sorts of summarizing tags, you and I could read them, Kate and understand what’s really happening in the review. And you know, that’s the power on the ingestion side. On the purely generating generative side, my, my bad example, from 2020, around using GPT, to to write really hilarious SEO tags for products is a mostly generative side, or creating landing pages, I’ve seen websites in the wild that have a title tag, or a h1 tag of I’m sorry, my language model runs on 2021, which is actually kind of a scary thing, because people are just creating tons of content and all this sort of noise out there. And, you know, I think in software development, there’s kind of this, you know, set of things happening around using the generative side to help, you know, software developers, and software engineers write better code faster. And then there’s some things in the bill that are kind of nicely leveraging both of these, so you know, website chat, bots, interactive, things that are, you know, or even chat GPT is probably a really good example of this, where you’re interacting with these kinds of things.

Kate Hopkins
And so how can organizations figure out and prioritize the highest, the highest leverage areas for AI adoption? So when you’re working with a company, it’s stuff you’re doing tons and tons of times? Or?

Tony Aug
Yeah, this is a great thing. And yeah, we’re starting to formulate kind of an actual like, I’ll call it like a branded scorecard kind of approach to this. But the, you know, I, I think it’s kind of mostly the same as you would prioritize anything, which is sort of like the, we’re a consulting company. So you have to always have to have two by two matrices, right? So you have the how much benefit will it be? And how much cost does it take? And ideally, you’re starting off in that quadrant where there’s a whole lot of benefit and not a lot of costs. That’s probably not too dissimilar from where you might want to think about using AI or generative techniques, with one caveat, which is, this is an again, to maybe to the first question is, I think it’s transformative enough that there’s maybe this idea of you want to go try it, just to get your sort of hands around it and understand what it’s capable of. Because it’s, it may be that, you know, you don’t have an exact use case right now of hanging to go build this thing that does the work of 50,000 people or completely changes their business, but I want to get ready for that. And that’s the, you know, go get educated use case, I guess where it doesn’t have a benefit axis, it’s only on the investment side, if that, if that makes sense.

Kate Hopkins
It does. So as we as we think about customizing a model, there are a couple of different options that range from lightweight, you know, designing really good prompts to really heavy weight or heavier weight tuning a model. What are the different options? And when do you use different ones?

Tony Aug
Yeah, that’s it. This is a great question. So we’ve all probably at this point, use chat GPT. And you’re giving chat GPT prompts. And that prompt is the actually I kind of think it’s a bug. It’s a, it’s a, these language models are not so complete, that we can speak to them completely the way we would speak to each other. So we have to think of the way we want to ask the computer the question a little bit more the way the computer wants to understand or the language model on set, understand our question, so that we get out what we want. And we might have to, to kind of give it three prompts in a row to get what we want. Or we might need to kind of give give it a persona or you know, something that get get out what we what we want. I think that’s kind of like a fleeting thing is maybe a controversial thought. But I think like two years from now, we’ll say that was a hilarious thing that we had these, these prompt engineers, and, you know, don’t I, anyway, so that’s, that’s one fine tuning is kind of the, I have a whole lot of text, or data or something. And I’ve already labeled it or I’ve already said what I want the output to be for this. So I have a bunch of inputs, hundreds is probably the right number to think of, and I already know what output I want. And so the tuning model, and I would, I would use the word train, they’re very lightly like, you’re not actually trying to get your sort of tuning to say, for this input, I get this output for this input, I get this output. Now, use that same set of instructions that I gave you for hundreds of things on 1000s, or millions of things. And that’s a way to, to tell the language model what it is, you want, without having to have, you know, hundreds of 1000s of quarters of compute power to go really train a model

Kate Hopkins
B for like, product descriptions like hey, we want.

Tony Aug
Yeah, it could be for product descriptions, it could be for, you know, like could be for, you know, sort of proposals, it can be for any anything where maybe you already are a good good example might be something like in your CRM, you already have the I said it’s a good leader, not a good lead. And that was a human judgment thing. But you have some some idea of what the input was, and you have this idea of it was good or bad as a lead or mediocre lead. Use that as sort of sort of a fine tuning is one approach. There’s other obviously other approaches to do that without generative tooling. But yeah, that’s an idea. So So those are kind of the top tier, then there’s this idea of embedding, which is, so one of the things that really unlocks the capability of GPT. In particular, and I would just say almost everything we’re talking about here is not GPT, or open AI specific there are, there are more companies than just in the world to do this. But our API’s so if you have some software that you want to embed generative techniques, on the the ingress or egress or sign ingestion and generation side in into your own product or into something you’re doing internally, there are API’s to do that. So you call an API passing a prompt and, and a payload and you get something back. So that’s sort of the embedded one. And then there are purpose built tools, which are kind of doing the same idea. So these would be tools that run on top of GPT, that are helping manage prompts or helping, you know, kind of infuse other data, or connect things like that. And scale AI is a is a company that works very closely with open AI. And they have a tool called spellbook, that runs on top of GPT, and others. And I would actually add one, sort of more to this, which is kind of the combination of the bottom two, which is his idea of this existing tools that you already use are starting to get GPT powers in them. So it’s sort of like it’s somebody else’s purpose built tool having an embedding. And the example I might give you, or an easy example would be Microsoft Office products. If you haven’t already seen some of the demos and some of the things that Microsoft’s doing as part of their own investment in open AI is bringing that capability into Office Suite. Another really good example is Tableau. So on the sort of big In this intelligence data side of things, there’s a GPT inside of Tableau. So you can do some interesting things there, too. So those are kind of the high level categories.

Kate Hopkins
I would say we had some specific examples that I’ll throw at you. One would be, you know, how would you How could you incorporate this into a work stream like a customer service organization? So if you were if you were trying to respond to a bunch of customer service requests, what combination of things might you use? What are some options?

Tony Aug
Yeah, so this one would be a good example. For that embedding one. So the idea there would be, if you’re using ServiceNow, or something for for your customer service, they may and most tools, I think, at this point, have some some third party or them themselves have integrated GPT kind of calls into that. So that would be an example. But you could also go call these API’s and do this yourself. So if if you’re coming from kind of a technology background, it might be something like, when there is a ticket in HubSpot as or Salesforce, call the GPT API with a set of prompts that says something like, you know, score this and tell me how how, you know, how mad they are, or, you know, formulate a canned response for me, so that it takes me less time to respond by, you know, doing some, some sort of look up on some other days. So there’s a few of these. And then I do think, and, and this is sort of the emerging side of this, a lot of companies are startups are sort of coming out with the purpose built tools is that you, you probably all think you might need. So you know, like the, there’s an AI for that is sort of the thing I like to say. So if if this is a common use case, go Google for it, like somebody has built the chat GPT powered customer service tool. On that, on that. So I had, and then one last part on the customer service. So if you do use a digital interface for for customer service, like a live person, or SNAP engage, or one of those sorts of chat things, those companies are starting to add GPT as a way to make your customer service rep more efficient. Of here’s a nice response, here’s, you know, some some things that they maybe can lightly edit and confirm that, you know, it’s doing the right thing, I think is the way

Kate Hopkins
one more example. What if you were trying to create customized white papers or customized, you know, marketing or educational materials based on concerns of a particular prospect? What combination of tools might you use to create something custom for your organization to do that?

Tony Aug
Yeah, so, um, there’s maybe a little bit of a, the word disclosed is, is maybe somewhat interesting. But, you know, I think there’s using, you know, you could do this with Chuck GPT. Or you could use it with other sorts of authoring tools, in the generative space of based on set of prompts based on maybe even a short draft that you’ve you’ve written up, please write a 10 page thing, or please write a longer form content with these elements. And then, you know, give it the for company X, who is a industry, why a company and that that’s a way to avoid this, you know, if you’re nervous about sensitive content, or any of these kinds of things. That’s one approach, there’s probably other approaches for how to how to think about that sort of thing and then use a tool like the Adobe Suite to make it actually into a white paper and nicely formatted and those

Kate Hopkins
How could you feed it? So how would you feed it information about your company and your product and the way that you solve problems? Because there’s probably these these pieces that may be appropriate for different customers, depending upon their needs that you want to feed it? How could you do that?

Tony Aug
Yeah, I think so. One is in the prompting to say we are an X company, The Who do you know, you know, we are a we’re like a sports apparel company and we’re gonna make a New Jersey for the Denver Nuggets or someday like an Our brand is based on these qualities and stuff. And the language models do a pretty good job of figuring that stuff out. I think that’s like where I would start. More advanced versions of that are kind of after Three or four iterations of prompting to say, you know, now rephrase all of what you just came up with, from the perspective of a company called XYZ Corp. And, you know, these are the principles of the company. And in it, it, I think you’ll be pleasantly surprised how how good a job it does on that sort of thing. If that doesn’t work, and I think actually later in this, this deck, we have a couple of other kind of tool ideas where, you know, there are companies that have kind of solved some of that tone of voice and you know, some of the more kind of subtle things where you might want to have a copy that reflects this specific style, or its specific tone. And you, you know, if you, if you graduate, if you will, from chat GPT, you might want to consider these other things.

Kate Hopkins
Definitely. And I’ll, I’ll bring that back up. But first, let’s dig into software development, and how how AI is shaping software development? This is a market map that nimble gravity put together, do you mind explaining a little bit what these different types of tools are?

Tony Aug
Yeah, so I think there’s kind of, you know, we came up with a handful of categories of, I’m going to author software. So that’s the cogeneration. side, I’m going to build things that require AI prompting, or I’m going to, you know, use generative techniques to do testing or debugging of software, like let you know, either on the generation of test patterns, or on the tell me what I did wrong sort of side of things. Specifically, a lot of things are on database. So how do I think about SQL code? And these sorts of things? How do I go give a prompt and get a website out? So it’s sort of the web app side, and it’s more nuanced than that. And then finally, the, the NLP or the NLP side, itself of like, I’m gonna go embed chat gpct in software, and those kinds of things. Probably primarily, the thing that I personally have been focused on is the leftmost box, the code generation. And I think of this as there’s sort of this evolution of software development or software engineering, since the beginning of the profession, or maybe before it was a profession. So years ago, a couple of very interesting things happened in this space. One was the, like, autocomplete capabilities and software development tools. And some of these got branded, like Microsoft calls it IntelliSense, I think. So this is like, I start typing something, and it fills it in. And I don’t have to remember all the words syntax and stuff. That was a huge step forward. And it was like just the tool writing some code for me in a very simplistic way. So that was one second one has this thing called Stack Overflow. So you get a bug, or you get an error message that you didn’t know what to do, or you need help with? How do I, you know, how do I compare these two datasets? Or merge this data or whatever? Probably somebody is posted something on Stack Overflow? And if you haven’t, or if nobody has, you could post it and somebody will, will post the answer to that. And what I think we’ve seen now with with the cogeneration side of, of generative AI is sort of like the next evolution of that it’s this bionic superpowers for software developers, where, you know, like, it’s sort of the combination of what was happening with autocomplete, and IntelliSense. And sort of what was happening with Stack Overflow, just without having to, you know, Command Tab or alt tab into some other browser window to go look that stuff up. Or, you know, do that. So I think that’s sort of the the cool thing that’s happened on on that side. And frankly, the thing I’m, I’m very excited about not not a whole lot has happened for the engineers of the world in a super long time. And I think it’s great that this sort of thing is happening.

Kate Hopkins
Awesome. What about on the far right side, if you want to start to incorporate some NLP stuff into your product or into internal tooling? What skills do engineers need to develop? Or like, what resources should they be leaning on to learn about this and to start to use it?

Tony Aug
Yeah, you know, I think the the simplest thing I would say is like the ability to call a RESTful API is the bare minimum. Probably by now, every possible language you can imagine, someone has written a tutorial for how to call the chat GPT API, and to embedded in, you know, visual, C or whatever. Like, it’s, there’s probably one of everything at this point. So the understanding how to call an API is probably like the best minimum and how to understand the output, I would just say like most of this stuff, and this, like I say this what sort of as a data person, I’m not being a data purist, and when I use the word most, but if if someone knows Python, they’re gonna get so much further faster, because probably the one of the first two or three tutorials written on every single one of these technologies will be in Python, the vast majority, and again, I’m using that with a Asterix of the tech itself was probably authored at some point, or at least prototyped. In Python. So I think like, I would say that sort of as an engineering team, I would think about that as a, like skill. And then I would also like the other skills, maybe more of the soft skill is sort of the, you know, think thinking about the how can I architect this in such a way so that these really hard problems that we didn’t solve very well before, can happen with with an AI, but also have some level of human inspection because sometimes these these models do weird things. And you might want to put a human in the loop. So that as the language model is generating something, before it goes out on the web, and you have the h1 tag saying, Sorry, my language model runs out and someone gets a chance to say yes or no, or, you know, some some of those kinds of things. Yeah.

Kate Hopkins
one live question that just came in, do you have any thoughts on open source GPT tools like GPT? J, which I’ll admit, I have no knowledge of?

Tony Aug
Yeah. So So if yes, I have lots of thoughts about this. at a really high level, I’m sort of a believer that open source tends to win. And then something interesting happened in March, and this space in particular, which was not I’m, I’m not making IP comment here. But effectively, meta, Facebook, their large language model was leaked to the world. And kind of the training was leaked at a separate time. And this then becomes torn apart by the open source community and refactored and re engineered. And, at about the same time, people are, you know, inside of Google, saying, hey, actually, like, we’ve just had a really bad couple of months, because this, you know, open AI, took, effectively tech we came up with and, you know, made us like a CRO, but actually, we’re both going to lose, because the open source is going to win in the long term. And if you if you get a chance to, you know, search for sort of Google internal memo on open source, you know, generative, you’ll find a bunch of internal things that got leaked, and maybe they leaked them on purpose, I don’t know. But I think ultimately, that’s when you can leverage the power of, you know, not 10,000 developers, but millions of developers, and people doing really creative, interesting things and trying to tune it and trying to, you know, turn it into new things. I think that’s like, a generally a really good thing. It also opens up for, you know, like, those scary, it’s too good for human consumption. Like, you know, I worry about my, my, you know, circa 80 year old mom, getting a phone call from a voice that sounds a whole lot like me, asking her to, you know, do some financial transaction. You know, that sort of thing is super scary. But, you know, I think we might have a better chance of controlling that if, if it is an open source. Lots again, lots of thoughts. On that side,

Kate Hopkins
we do have more questions on on regulation around AI and sensitivity around like sensitive data. It will, we’ll, we’ll keep all those those are coming up. We had a bunch of pre submitted questions about sales and marketing. And I recognize that this is only a tiny, tiny, tiny fraction of the tools that are out there. But can you comment a little bit on AI for sales and marketing?

Tony Aug
Yeah. So the first one is kind of like you can get all the really long way with chat JpT. It’s a super good language model. It’s, you know, it has the benefit of a whole lot of use and people thinking about how to interact with it and tune it and I A lot of YouTube videos and stuff about it. So I think I’ve started there, like, you know, try it and see if you can get it with a series of prompts to give you roughly what you want and then go edit this. So like, you know, I’m like, I want to have an AMA on leveraging generative AI, as a non AI company could be a prompt, please give me, you know, a set of questions that might be useful. I’m not suggesting you do that, okay. But like that could work. And then if you sort of hit the ceiling with what you feel good about, doing with with Chad GPT, or like a run out of use, or it becomes burdensome or whatever, I certainly like this Jasper tool in particular, like one has to go and we’re not partnered with them, we have no financial benefit in doing that, check it out. And Reggie AI is another one that I would check out. And I would also check out a website called, there’s an AI for that. And I think it’s actually written by somebody using chat GPT to write an article blog about GPT. And it is a kind of every day, there’s a whole lot of new AI tech, and I just pulled up 4300 AI tools that are indexed on their site. And, you know, they’re making money by letting people sponsor and get to the top of the list. But I think ultimately, like they were so early days, it would be like, you know, the analogy would be if this was 2008, or whenever the iPhone came out, saying, like, what 10 apps should I have on my phone, you know, my iPhone one or whatever like, was, whenever the App Store came out, I like, here’s some ones that we’ve tried and that we think are pretty good. But like tomorrow, there will be you know, this slide will be relevant, there’ll be 20 new, so So I would say that’s one is it’s that super fast moving space. There’s tons of tools, find one that you like, and find one that you know, works for you. Not one size fits all, couple that we’ve used are here. And you know, I’m not on the image side, you know, the couple of here, like, I need a background for my zoom or, you know, whatever, like you could, or I need, I need a office shot with the conference table. And people working hard and solving our problems, like those sorts of stock image sort of things are some some great tools for coming up with those through generative techniques,

Kate Hopkins
and maybe less to the tools in particular. But the use cases, you mentioned that like one of the original AI use cases was this essay, like generate SEO content, I’ll give you a little bit of information, you helped me generate a lot of content. What are some others? I feels like customization is going to be really, really huge. What are ways that customization is coming into sales and marketing? And how can jet generative AI help?

Tony Aug
Yeah, so you know that maybe so that SEO kind of product content, that sort of thing? is for sure. One another one that we seem to see a lot is this idea of how do I do cold outreach, but make it not very cold. And so the idea would be and I imagine the majority of folks on this on this live cast have have had this like, sort of weirdly written email, as as like spam that’s has something in there that’s like, write about like, oh, yeah, like, wasn’t the weather in Denver, really weird last

Kate Hopkins
week, or you know, exactly that it is? It’s just weird. Yeah. And

Tony Aug
so I think what you’ll start to see are these techniques inside of CRM and email outreach tools for creating engaging emails, or LinkedIn posts or LinkedIn connections, or whatever it might be. So the idea would be like, if, you know, you’re trying to figure out well, oh, there’s this company, I really want to talk to them or like, I really need help with the, you know, the writer’s block of sales of thanks for the meeting last week, it was really great, like, let’s connect on the action items or whatever, like, those sorts of things are, I think, going to start to be, like, pre written for you inside of the tools that we use every day, so that our job is more of writing the like, you know, tuning of it or, you know, fixing some things that maybe aren’t so perfect, and I think that we’re going to start to see more and more of that. We started to get asked a lot about those kinds of things. More and more. You know, I was only half joking about the AMA being like, we were doing a code marketing thing with another organization, we needed to come up with a webinar. And I put a prompt into chat TV because I was lazy, saying, Hey, can you please like, remind me all the things I need to think about for a webinar. And the webinar is about this, and this and this. And he kind of wrote the whole script for the webinar for me. And you know, obviously, we won’t purely do what our computer overlords tell us to do, but we’ll use it as a starting point.

Kate Hopkins
Awesome. So we’ll throw it up. We’ll throw it open to any questions. Now. If you’re watching on the Ask one guy.com site, click Ask a question to head over to the the version that will let you submit a question. But I will go ahead and throw it open wide, starting with some questions around sensitive customer information. So things are moving so fast. How do you leverage tools like Chuck GPT, while not exposing sensitive customer? Data and intellectual property?

Tony Aug
Yeah. So first and foremost, we are not lawyers, and nothing is legal legal advice. You know, I would say that the number one thing is always always and this is whether it’s an AI tool, or anything, like read the privacy policy in terms of use. So you know, and like, we all click the checkbox and say I read it, actually read it sometimes when it really matters, or always is maybe the advice I should give, you know, this is nothing new, like, there are implications in terms of what you put into tools, whether it’s into your CRM or into chat GPT of who really owns that and how they can use it. So you know, I would, I think the the, in this case, the implication is a little expanded, because it’s not just what you put into it, but also the output, and who really owns the output. And I would also just say, because I have actually read these things are many of them. Pay attention to why there are subtle differences between, like Chachi, Pts input versus output ownership, when you’re using ChatGPT’s UI, and when you’re using their their API, because they do subtly change those things. And, you know, kind of don’t assume that the the rules that apply to one or are necessarily the same as the other. So I think like, that’s the like, I would, I would think about it like that is sort of this, you know, I think we all take for granted that we can use Google’s email program, or we can use Microsoft’s email program, and they’re all in the cloud. And we’re putting data into them all day, every day, or slack or teams or whatever. You know, this adds a whole nother layer of complexity, because it’s actually generating content to you. But I would just like, read those things. And then, you know, I think the other other implication is that we’re building tools on top of other API’s that might change. And you need to think about, well, what if they change our API? Or what if it goes away? Or you know, what if, you know, like, the behavior shifts in a big way, and just think about being safe, and like thinking about what those what ifs, I wouldn’t design everything to assume brakes tomorrow, but like, think about that as implication.

Kate Hopkins
Another flavor of this is that most companies probably have employees that are using generative AI, whether the company has a strategy around it or not. So as it’s being embedded into all of these tools that they’re already using, as they’re using GPT. What have you seen companies do in terms of giving employees guidance on where they can and can’t use generative and where they can and can’t use company data or customer data?

Tony Aug
Yeah, you know, I think, I think the there’s sort of a couple of schools of thought on this. And again, this is not like, we will advise what one version is, assume people are going to do it, and assume that it’s going to mean they’re doing it, not because they’re, you know, like, like the like they’re anarchist, or that they’re doing it because it makes them do a better job. And like, isn’t that cool that people want to do a better job and you know, so I think one one school of thought is you assume people are going to do it, and you go figure out how to do it really well. And then you have internal trainings. And you’re like teaching people this new skill that helps them do their jobs better and be better employees for you, while also sort of guiding them away from you know, kind of danger zones or things that you know, might you know, like I think there’s pretty public stuff about how Samsung got a little side with some of this stuff. So I think you can use this as kind of this, Hey, we know you’re gonna use it, please use it smartly. And by the way, here’s some cool tricks that we’ve we figured out. The other reason to do that, I think, is the, your, you’re probably going to get people who raise their hand and say, like, hey, you know, that internal system, the, you know, whatever system, wouldn’t it be great if it had generative AI in it. So you start to get these kind of like, internal r&d use cases kind of coming, if you’re fostering and in giving people sort of the, the ideas of how to how to do it in a safe, constructive way. So I think, and you could probably guess, this is the side I would had on like, the other side is sort of the you try to box everybody and prevent them from using and do all this sort of stuff. And like, for now, that maybe really safe approach, because, you know, you haven’t read the Terms of Use, and you haven’t understood how this works. And, you know, the procurement team hasn’t fully thoroughly vetted the sock to compliance or whatever, great, like, if, by all means, these are things that, you know, some in particularly in regulated industries, you probably should go do these things. But assume that at some point, all those things are gonna get figured out. And then you’re probably back in the first camp. And, and by the way, people are going to, you know, feel frustrated about that, probably in the meantime, but I totally understand why. Why some, some places need to do these things.

Kate Hopkins
Awesome. So shifting gears a little bit, how should businesses assess and mitigate disruption risk?

Tony Aug
Yeah, I think I think there’s this this idea of, you know, the, there’s a thought exercise that sort of, like, if someone was to come up with a new approach to what we do in our industry, or our business tomorrow, that would eliminate our business, or industry or whatever, like, what would it be, and, you know, if, if the chances of that happening are are, you know, pretty high, or even not remotely low, you might want to think about this is like a really cool thing to embrace. And which is to say, like, and you think about it as somebody’s going to disrupt this anyway, because it’s ripe for disruption, or this, you know, tech totally changes the way that this whole thing works in the first place. Wouldn’t it be really a better thing, if we’re disrupting ourselves, instead of somebody else disrupting and, you know, we’re, we have the benefit of being the people who understand the way it works today. And thinking in an unencumbered way about how we might disrupt ourselves is super hard thing to do. Frankly, it’s, it’s dangerous is like, requires, you know, a lot of forethought. But it also requires getting out of your own way, because the rules of the thing will change. It’s sort of the, you know, like, the rules of some old industry, you know, all the all the people telling you you, like you don’t understand kid psycho work that way, like, just, that’s exactly the way you need to think I think in those cases, and, you know, I think otherwise is like, you know, thinking about strategies for, you know, r&d or, you know, ways to, you know, kind of the Wayne Gretzky of where’s the puck going, like, what, where are all the things that we can reasonably assume that the technical replays are enhanced in the future? I think that’s how I tend to think about it.

Kate Hopkins
Cool. We have time for maybe one or two more questions. So feel free to get those last ones in? Here’s an interesting one on what does it take to move from a generative model to an assistant model that knows how to take specific actions in a specific product or platform?

Tony Aug
Yeah, I think, you know, this is a whole lot of trust, and a lot of control. And so for the audience, like, I think what the question is really about is kind of this idea of, where, you know, going to this place where these language models are being given the ability to go outside to the internet and to do stuff and like, Please order me a Domino’s Pizza kind of thing. And the, you know, this is where, like, real action in the real world happens. And we need to make sure we understand the trust model and how to how to put them in guardrails so that we’re not ending up with a whole lot of pizza or you know, other weird things happening. I think that’s, that’s old. That’s where this is all going. If you haven’t looked at auto GPT as something interesting, the other one is called Lang, chain la Ng, as in language chain. These things are kind of logical extensions of what’s happened. With with GPT in to enabling AI to perform actions in the real world. That’s where things start to get very kind of messy and dangerous, because that’s where my mom will get the phone call from me, asking her to transfer all her life savings to some.

Kate Hopkins
Awesome. And so I guess a last question that I think is an interesting one to go out on is, what are what are bad use cases of generative AI? Where are you seeing people? You know, decreased speed or efficiency? Were people using it poorly?

Tony Aug
Yeah, you know, I have joked about this, you’re kind of like too much content, too many new websites, too, too much, you know, kind of inauthentic content, all this sort of stuff. Like, we were at this kind of, like, the thing I worry about Kate is this. Like, your bots, gonna send my bot and email and my bots gonna reply to, to you. And the next thing I know, like, all these decisions have been made. And so I think there’s kind of this, you know, the this, letting these things, especially on the agent side of things, get out of hand where, you know, we’re no, like, we’re all on a beach somewhere and these bots are doing all the work. I think that’s, that’s kind of the thing we need to think about is like, what, what, how do we how do we leverage these things in a good way? And by the way, I think like, if if I was, you’re asked to do an AMA, what if you have like, literally a fireside chat and the dawn of the Industrial Revolution, we’d probably be talking about lasers or something and the risk of lasers taking all the jobs. But I think there’s this, this scenario where we end up in, like, every website was written by a computer, and we’re asking computers to summarize the websites written by computers. And, you know, like, we need to probably I saw there was a question about taking a pause, like, I think there’s probably like, it’s too late, because it’s already an open source is already going to happen. But we probably need to think about how do we how do we really want this to work? Do we, you know, do we want our, like, high schoolers writing their, you know, SATs in there? So it is, it’s really a question of who can prompt engineer the best or whatever, like, No, I’m probably this is what we really want. So I think that’s kind of where I would worry a little bit about where this is all heading, I guess. And that’s not a science fiction version. It’s, you know, I, I don’t think there’s going to be, you know, a, like, horrible, you know, end of times and Terminator sort of thing anytime soon, because these are not generally intelligent AI as these are very purpose. You know, it’s a map about the language, not how to go learn new things. Hopefully that helps answer that. Awesome.

Kate Hopkins
Well, we will share this recording with all the registrants as well as information on we’ve actually created a written guide with with humans with Tony, we did not yet figure out how to how to make chat GPT read it for us. So we’ve got a written version of some of his insights. And we’ll also share information on how to get in touch with Nimble gravity if you’re taking on some of these challenges. Cool. Thank you so much for your time, Tony.