Greg Baugues: Someone Using AI Will Take Your Job

AI won’t take your job but someone using AI will take your job.

AI is a major topic of discussion today. Debates rage about the ethics, the potential for human extinction, obliteration of the workforce and plenty more between AI sceptics and AI believers. The variety and pace of development of the technology is unparalleled. It’s almost impossible to keep up.

Greg shares his thoughts on why AI has huge potential to augment human activity, think Iron Man Suits, Not Terminators. He also shows you, with live and interactive demonstrations, some of the ways you can use existing AI tools to improve your work life. Today.

Slides

Find out more about BoS

Get details about our next conference, subscribe to our newsletter, and watch more of the great BoS Talks you hear so much about.

Transcript

Greg Baugues 

Hello, everybody. My name is Greg. I’m going to chat about how to help humans do their jobs better with AI. My name is Greg Baugues. I go by GreggyB on the internet. Recently started a company called HaiHai Labs, and write about AI at haihai.ai.

A little bit of a story from last night. Last night, we had the speakers dinner, and I don’t know there’s 20 some folks there. And Kyle, I hope you don’t mind me sharing this story, but folks are going around just talking about their experience. And Kyle shares that, you know, Kyle spoke yesterday with the Autobooks crew and and he shares that he stepped out to call his mom because 10 years ago, he had made a list of his dream conferences, and this was on there, and he was finally able to speak there. And it was this huge moment in his career, and it was particularly touching for Rachel. And I think most of y’all probably met my wife, Rachel at the desk over this week. Where did you and oh, there you are. Okay. Hi. Rachel broke down cry. She had to. She stepped out and and I kind of like held it together. But hearing your story Kyle was was so powerful, because that was our story 10 years ago.

I first spoke at the Business of Software conference in 2013. I’m a developer, I’m bipolar and ADHD and shared a bit of my story there. And BoS had also been on my dream list, and it was such an honor to come speak. And you know, a lot of the conversation last night at the speaker dinner was around – why is this conference so different? You know, we had so many speakers there who speak at so many different conferences and talking about, why is this such a special experience?

And I can say that this conference, more so than you know, any other event, has had a profound impact on the trajectory of my life, and it is mostly because of these folks, because of Mark, everyone.

Mark does an amazing job cultivating speakers and a community and setting the tone for the culture and for vulnerability and hospitality. Of course, he doesn’t do it alone. And as the last speaker of the conference, I think we should all give them a hand once again, to Mark, Jo and Kirk for putting on this amazing conference.

And these three days, I know are three days that many of the people here in this room who have come year after year have looked forward to, and the three of them put in so much work over the other 362 days a year to run this business, to make sure that there is a venue for us to come to, to make sure that there’s a stage, make sure that there’s food.

So by way of transition, you know, this might, I think third time speaking here at Bos. I’m not speaking about mental health today, but try a segue. There sure seems to be a lot of anxiety about AI these days. That went how I was hoping it would go. So I think, you know, 2023, obviously incredible year for AI, but a lot of people are worried. A lot of people are feeling fear about this stuff. And I think broadly, there’s sort of three kinds of anxiety around AI.

Existential Anxiety

First, there’s the existential anxiety, which, of course, is not new. You know, this question of, are the robots going to come and kill us all? And you know, this is, I had to look this up and double check it because I didn’t believe it. This movie came out 31 years ago, 1991. So of course, existential anxiety around AI, this is not a new phenomenon for us at all, but it has been reignited this year because, I guess we released a really good chat bot.

Ethical and Legal Anxiety

Then there’s the ethical, we got sound awesome. There’s ethical and legal anxieties around this. So great example of this. This is a mash up between Drake and the Weekend it was released five months ago. It immediately went viral. You can see here it’s been viewed 6 million times on YouTube. The only problem with this is that neither Drake nor the Weekend had anything to do with this song whatsoever. Their voices were deep faked. The lyrics were written by chat GPT and the style of the artists, the beats were produced using basically an LLM for music generation. All of this was put together and by an artist who had no contact with the artists themselves, who are here.

And as a society, we have not figured out answers to these questions around fair use and what does it mean to use technology to create derivative works from copyrighted material? And we’re going to be figuring this out for a long time, and the institutions are going to move a lot slower than the technology does.

Economic Anxiety

The third sort of anxiety that folks are feeling these days is economic anxiety. And again, this is not new. It is not a new phenomenon for humans to worry that the machines are going to come take their job. This movie by Charlie Chaplin, Modern Times was released in the 1930s like 90 years ago. And it was a reflection on the Industrial Revolution, and what does it mean for us all just to become cogs in a machine, and is it going to de humanitize us? And so this is a concern that humans have had basically ever since we started inventing tools, right? But this is one of the first times that knowledge workers and folks who produce things like words for a living, or things that work look like words like code for a living, have felt this threat so acutely and so quickly.

Will AI take my job?

If you are feeling any of that anxiety, that anxiety of will AI take my job? I would suggest that there is someone who you can look to for inspiration, and that is the first person ever to lose his job to AI – and that man is Garry Kasparov. Some of you probably know just Are there any chess players in the room? A few awesome. That’s great.

Chess: AI vs Human

So Garry Kasparov was the world chess champion from the 1980s to the 1990s he was like 17 years and then in 1997 he very famously lost this match to deep blue, which is a super computer trained by IBM for the express purpose of beating him at chess. Now he likes to tell folks, actually, I think we have, yeah, this video will play

Video Audio: “Then in the mid game, analyzing millions of possible scenarios by sheer brute force, Deep Blue has led Kasparov into making a poor move. Kasparov is rattled. He defends what he can, but it’s clear that the computer will reliably do what he himself would do, and he recognizes that he has already lost on deep blues’ 19th move, the champion resigns.”

Does that look right? Garry Kasparov was no longer the best chess player in the world in 1997. And he spent the next 20 years deeply reflecting on what does it mean when AI comes to your industry. And in 2017 he wrote this book on it called, then in the mid game, called Deep Thinking. So this came out six years ago, and this was Garry Kasparov’s sort of pontification and reflection upon what happened in the 20 years after AI was better than humans at the task of playing chess, which he had built his whole career around. And he has two big reflections here.

So the first is Garry Kasparov said that when he lost to AI, people said chess is over. Humans will never play chess ever again. Why would they play when computers could play the game perfectly? And he said actually, Chess has never been more popular.

The second thing he said is that not only is chess never been more popular, but the human players today are the best human players who have ever lived, because they have been training alongside AI basically since they were kids, right? So the we this, just this year, had the youngest Grand Master in history who has had since birth a chess tutor that is orders of magnitude more powerful than the computer that beat Garry Kasparov. And in fact, it’s not just the professional players who have this available. If you’re a chess player, you’re probably familiar with with lichess, right? And so I play in lichess. Lichess is open source. It’s one of the two most popular sites on the Internet, the other being chess.com the big difference, lichess is open source, is basically run by one person, and yet, right now, there are 100,000 people playing on this.

I’m a casual chess player. I mostly play to procrastinate, and I play these quick, rapid games. And here’s what I can do if I decide I want to be better at chess, I can click on any game that I’ve played, and I can come here to the analysis board, and I can click right here, and I can see the computer walk through each of my moves. And not only that, but I can request, I can come down here and click on ‘Learn from my mistakes’, and I can request a custom analysis and a custom lesson based on the mistakes that I made in my game.

So it says here that I played, uh, Bishop to B2. It’s like, hey, what you know, what was, what was the better move? And I can say, I don’t know, was it rook to e1 it’s like, no, you can do better. I can view the solution, right? So this is free. This is what the state of chess and chess education looks like 25 years after AI was better than humans at chess, right? And if you are a chess player today and you want to be competitive on like a national level or a global level, it would be absurd to think that you could compete without availing yourself of these tools.

And so Garry Kasparov’s mission now, one of the big parts of his career, and his message to all of us today who are concerned about the machines coming and starting to be able to do our job, perhaps better than us – is don’t fear the intelligent machines, but learn to work with them.

In fact, Harvard, just last week, released a study with conjunction with the Boston Consulting Group, studying the impact of consultants working with AI, and the key takeaway was that the consultants using AI were significantly more productive, completing 12% more tasks on average, completing tasks 25% more quickly with higher results. So they were, they were they did it faster and at higher quality. The consultants who were availing themselves of specifically this was looking at chat GPT.

So if you’re feeling anxiety about some of the intangible future outcomes of AI, I would suggest two things.

One, this isn’t, I would say I have felt anxiety about the existential crisis and whatnot, the existential threat, but it is so far outside of my control, and I’ve generally found in life that when I’m feeling anxiety, the best thing to do is to focus on the things that I can control, the levers I can pull, and the levers that you all can pull here in this room, both for yourselves, starting today and in your software businesses, is to help humans do their jobs better with AI.

Great AI Tools to Use

So we’re going to talk about some opportunities to use.

First, we’re going to look at a few tools that are available today. These are tools that you can go start using immediately. You can sign up for an account. They’re freely available. Then we’re going to look at what is your opportunity, given this is the Business of Software, given you all are involved in software companies, to build features powered by by AI and to work with AI when building software. So first, let’s talk about some tools, but I think it’s also helpful. Let’s talk about the jobs that these tools help us do.

So we’ll look at three examples, designing, transcribing and writing.

AI tools on Designing

So first on designing, Randeep. Are you still here? What’s up? I loved your talk. So we saw this the other day, Randeep was up here. He’s giving a talk. He’s using mid journey in his talks. And every time he put up one of these slides with these images that he had generated using mid journey, people laughed and they got engaged and they leaned forward right?

And I’m going to go out on a limb here and say that if it wasn’t for Midjourney Randeep, wasn’t going to go out and hire a designer to create custom Batman images for him, and I’m guessing he also wasn’t going to spend 10 hours in Photoshop to create this. This was a net add to his talk that let his customers get more value out of it with a tool that gave him a new superpower that he wasn’t able to do before that tool came out. So he mentioned Midjourney in the talk, realized a lot of folks maybe haven’t used Midjourney. So Midjourney is probably the best tool on the market right now for image generation. That’s probably true.

Now DALL·E 3 is rolling out over the next couple of weeks, will be more widely available, and it looks like that might be better.

But for now, we’re going to look at how you can use Midjourney to create some images. And to do that, you can see I was doing some testing here earlier, and we’ll just make sure our internet’s good. Let’s see. I was having some connection issues here earlier, and this works way better if there’s Wi Fi. So let’s see what happens here.

All right, so I need somebody to shout out, oh, we’ll play some mad libs. I need someone to shout out a noun – skateboard. I love it. So we’re going to come in here one just point about using Midjourney. They decided to spend basically $0 on a user interface, and they said, we’re just going to throw it into discord and use chat. It’s actually kind of brilliant. They’re like, we’re going to put all of our engineering resources towards improving the quality of our AI. And then the other thing that happens is a lot of the usage happens in public channels, and so there’s a lot of learning from the crowds here.

So, all right, so we have a skateboard. I need, like an animal – kittens. All right, so we’re gonna go kittens riding a skateboard, and I need a location. Can someone shot a location? Raleigh. All right, let’s try Raleigh, in Raleigh. And I need an artistic style. So what was that? Motorhead? Yes. All right, with Motorhead, thank you, Mark. Motorhead. And then I need an artistic style, so this could be photography, this could be watercolors. This could be, I don’t know, someone shout out your favorite art style – amazing, 70s movie poster.

All right, so we’re gonna say here. So we have our thing. We’re gonna say in the style of a 70s movie poster. All right, so kittens riding a skateboard in Raleigh with Motorhead in the style of a 70s movie poster. So depending on how the Wi Fi happens here, this will take maybe 30 to 60 seconds here to kick off, but it looks like we’re going so this is great.

All right, so what’s going to happen here Midjourney is going to create four sample images for us. These are going to be low resolution. It’s going to be four takes on images that could possibly fulfill the requirements of this prompt. We’re going to see them gradually become more and more well defined, I guess. And as they come in here, you can see we’re at 93% so we have all right, so here’s our answers here.

All right, so those are kittens. I don’t know that that’s Raleigh, and I’m not seeing the Motorhead connection. So you know, there might be some prompt engineering here that we need to do, but we’ll just keep rolling with this. We definitely have, I can confirm this is kittens writing skateboards, though. Of these images, all right, so we’re going to count. Top left is one, top right is two, three and four. What is yours favorite one? Two? All right. I heard two. All right.

So here’s what we’re going to do. We’re going to create some variations of two. We have these buttons down here, u, stands for upscale, V, stands for vary. So I’m going to click two, and now it’s going to zero in on that concept, and it’s going to create for us four variations on that concept on v2 and again, when these first come out, they’re going to be low resolution. So once we get an image that we like, we’ll click that you button that’s going to upscale it to a high res image that we can use, all right? And so it looks like we’re getting some here that have more kittens. We got four kittens, three kittens, and boom, we’re done, all right.

And so this is what our kittens riding skateboards, I suppose, in Raleigh. And I’m sorry Mark about Motorhead, but we can play with this later on, see if we can get Lenny in there. So this is what Midjourney looks like, right? And, you know, it’s fun, and it worked great in Randeeps talk. A lot of what I’ve used this for so far has been for header images on blogs, you know, spots where I want something that grabs the user’s attention, but I’m not necessarily going to hire a designer, spend a lot of time myself doing it.

Our next door neighbor in Brooklyn is a Tony nominated set designer, and he was over for drinks one night, and I was showing him this, and he goes, All right, you know, that’s cute. Like, let’s do a set design for a Broadway production for death of the salesman, based in the style of Frank Lloyd Wright. And it spit this out. And, you know, it came out like you just saw about two minutes. And he’s like, Oh my god. He’s all right, two things.

One, I would never use this. He was like, specifically this design. He goes, there’s too many things that are technically wrong with this design for it to work in a Broadway production.

He’s like, but two, I have spent weeks designing mock ups that I could present to clients so that I could just have a better conversation, because we can all, we can always talk better about images. He’s like we just did here in two weeks, the job that I have to do in order to sell customers. And we did, and we just did that in like, you know, two minutes. And I think that this is a common theme you’ll see through here.

Our friend who is a set designer. He got into the business so he could do set design. He didn’t get into it so he could not could navigate the sales process, but that is one of the jobs that he has to do in order to, you know, do his profession. And this feels like this is a lot of the great opportunity with the AI tools that are available today is helping humans do all the other jobs that are outside of their core area of competence, outside and it’s like I actually was talking to Carl the other day. Carl, it was last night at the speaker. There he is, right there. Hey, Carl. Carl gave a great talk you all should go watch six years ago now? About 2016 about Build Ironman, not Terminators. Great thoughts on this.

But he was saying last night, think about what you actually hire someone to do. So a teacher was the example, he said. He said, you hire a teacher to care about the students, and then think about all the other work that they have to do as a function of their job, like grading paper, like all that monotonous, robotic work. Can we use AI to do those jobs for folks so that humans can spend more time doing the human work?

AI tools on Transcribing

On that theme, let’s talk a little bit about transcription. Before got here, had this great call with Kirk. Kirk was walking me through just so Kirk helps out write the content for the BoS website. It’s a job that Kirk and Mark sort of tag team on. We have some of the world’s best business speakers here on the stage, speaking, it is an incredible source of content, and if you go on the BoS website, you can go back through the archives and you can find talks and transcripts of the talks from the 15 year history of this conference, but getting the talks from what happens on stage to the website in a digestible format, is so much monotonous work, right? Like, you got to take the file, you got to transcribe it. You got to double check the transcription for all the errors. You got to scrub through. You got to find ideas. You got to then, like, stare at a blank Google Doc and figure out, like, what quotes are you going to pull?

Recently, I’ve been playing with this tool called assembly AI. And Assembly AI is the current state of the art when it comes to speech to text, and the way assembly AI works, they have a great API. So if you’re a developer, you can build this into your tools. But you can also, if you’re not a developer, you can just come right here to the dashboard.

So I have here, I was sitting out in the audience with a little audio recorder, and I recorded Jamie and Dr Chris’s talk on storytelling. And we could come here and we can upload it, and then we can click and say we want transcription. But then there’s a lot of other options here. So we can do not just transcription, but we can summarize it. We can do create chapters. We can do speaker detection or Speaker Labels. In the case like that, talk was where there’s multiple speakers. Now this would take, probably, it was an hour long recording. It would take, probably, think it says 15 to 30% of the run time for it to process it. But I did get I did it ahead of time, so you can see this is what the result here would look like of that talk, right? And so we have our automatic speaker detection. I think this probably came out in about 10 minutes or so. So not only do we have high quality transcript, like if you were to go through this, if you ever done transcription before, you’ll find that this is much higher quality than what you’re used to using online transcription services.

But over on here on the right, we’ve also gotten meaning out of this. So we’ve gotten summaries of the chapters that it’s identified, and then descriptions that then could be used in the content creation process. All right? So this is a way that you can use AI today to repurpose content, audio content that you have. So if your business creates podcasts, if you you know, run a conference, if you have speakers, if you just want to do interviews with customers and then turn those into blog posts, this will jump start that process for you. All right.

AI tools on Writing

So then let’s talk a little bit about writing that was our transcription, and talk about writing. Now I’m guessing, I don’t know. Maybe I shouldn’t assume, who here has used chat GPT to write something before? All right, just about everyone show of hands, who here pays for chat GPT? Plus, that’s awesome. It feels like about half or so. And then who here has used advanced data analysis, formerly known as code interpreter. So much fewer folks. All right, great.

So I do think it’s worth knowing here, just for the folks who aren’t familiar, there are some phrases here, when we talk about AI, that are tossed around a lot. And in my conversations with folks, I realize we use these terms a lot. We don’t always know necessarily what they mean, right? And so tokens is a big, and specifically token count. So chat GPT, 3.5 turbo has a token limit of around 4000 tokens. So what is a token? This is what it looks like when I paste a chunk of the transcript that we just got into open AI’s tokenizer. And so people have pejoratively, pejoratively referred to chat GPT as fancy auto complete, or large language models as fancy auto complete. But it’s not actually a terrible mental model. The you can see here that it’s going and it’s breaking up. A token is not a word. It’s not necessarily a symbol. What it’s doing is finding common patterns. So here it sees, get is a common pattern that is a token out is a common pattern. And not just out, but space out. So like a word with a preceding space in front of it, you’ll also see that periods on their own are tokens, right? So generally, what happens is a token is approximately a word they open. AI generally says that token is about three fourths of a word. So if you’re dealing with a token limit in chat GPT, 3.5 turbo of about 4000 tokens, and you want to get a response back, let’s say you want to generate a blog post that is 2000 tokens. That means you only have 2000 tokens to input into it, right? Because your total response the token limit refers to both the input and the output. So if you want to start doing things like taking longer transcripts, because if we looked at that transcript in a word count, it’s about 10,000 words. It’s not going to fit into chat GPT 3.5. It might not even fit into GPT four, which has, depending on the model you’re using, either 8000 or 32,000 token limit.

However, I would say anyone who hasn’t already it is so much worth the $20 to go sign up and to use chat GPT, plus, you’ll get far higher quality out of GPT four. You’ll get far in terms of the writing that comes out of it, and you’ll get far higher token limits, and you’ll be able to do a whole lot more with your writing. I’m not going to do a demo of this because folks have already done demos. I will just say, though, advanced data analysis, if you haven’t played with it, just take a spreadsheet that’s big and upload it. The big difference that you’ll notice with chatgpt advanced data analysis is that there’s this little plus button right here and allows you to upload files. So I’ll say, if you do have something that’s 10,000 words, like a transcript, or you have a spreadsheet, just upload it and start asking questions, and magical things will happen. I’ll skip the demo for now, but it is amazing watching it write code in order to analyze your data or to break your long text up into chunks, and then to do its LLM magic on it.

Now I will say about the quality of writing, because, you know, it writes well, but I don’t think that the goal here in any of these tools is necessarily to create content that you’re just going to copy and paste from chatgpt and hit publish on right. I think that’s thinking about these tools incorrectly.

When I was here five years ago, I spoke again, I gave a follow up, and I talked about again, about mental health, and it’s about my time, I told a story about I was a manager, I got depressed at work, and I spent two weeks staring at this blank page because I had to write a job description of a community developer, and it took me two weeks and I couldn’t even like write a sentence on it. And what I found with chatgpt is that it gets you past the blank page problem. If your writing process is typically the two steps, it’s typically the writing, the authoring and then the editing. Oftentimes, chatgpt can skip you straight to the editing process.

Another I’m a big fan of it’s called Cunningham’s law, which says the best way to get the right answer on the internet is to give the wrong one. Sometimes chatgpt value is giving you the wrong answer right, like, sometimes it spits out copy for you. You’re like, well, this sucks, but now you’re editing it right, and now you’re moving and you’re no longer staring at a blank page, which for folks who have ADD easily distracted, where the hardest problem is getting started, it can be a superpower for that.

Recap

All right, so just on a recap for tools, here’s now I heard analogy from Sam Altman, the CEO of OpenAI, for these tools that are available today. He says chatgpt is like an E bike from mine, and this resonated greatly for me, because one of the best purchases that Rachel and I have ever made is an E bike. We bought electric cargo bike. We live in Brooklyn. Living in Brooklyn is great, but it comes with some unique pains, mostly around traffic and parking and just getting from point A to point B, and an E bike solves so many of those problems.

Now our E bike is what’s known as pedal assist, so we still have to pedal, but there’s a motor, and that motor gives us three times more power than what we generate with our legs. Right? Like we still have to steer, we still have to avoid the potholes, but having that motor is the difference. Difference between being able to take our kids over the Brooklyn Bridge on the bike and not it’s the difference between being able to sign up for swim lessons in Manhattan and crossing that bridge to go to swim, or taking our daughter to volleyball practice in a place where the subway doesn’t go. So we’re still doing the work. We’re still riding the bike, but we have that E Assist, which helping us do our job as parents better, and it fundamentally transforms the way that we move through the city.

All right, so everything I’ve talked about so far, I could give that talk to any audience. Those are general tools for anyone who’s doing any sort of knowledge work. This is the Business of Software conference. Let’s talk a little bit about some of the opportunities that are unique to people who are building software for a living, and I want to talk about two things. The first is building with AI, and the second is building on AI.

Building with AI

So let’s talk about building with AI. I was I worked in management for a long time at Twilio. I left Twilio this year and started I’ve been writing code ever since I was a kid, but I, you know, sort of not been writing a lot of code since I was working in leadership, and I just started writing code, and I was like, I’m just going to spend as much time as possible writing code, building stuff.

Now for those who don’t write code, this is basically what the state of coding has looked like ever since the internet was invented. Is you go to Google, you search for whatever your problem is, you go find a solution, typically it’s on Stack Overflow, and you copy and paste that code into your editor, and then you hope that it works. Now, Carl had a great point the other day, the entire structure of the internet is designed to steal your attention. So when you’re doing this, you’re going from your editor, where you’re trying to be productive, onto into a browser where everything is trying to capture your attention and pull you in a different place, and what I would find the greatest threat in my life for any personal project that I worked on has been at some point I was going to get frustrated, hit a wall, get distracted, and just walk away. Right? It’s not that thing would fail. It’s just that eventually I’d just stop. When I very first started playing with chat GPT, I found you could just go to chat GPT, you could ask the question, and you were getting better answers than what you could find on the internet, and you could copy and those into your editor. But then I started discovering that there are now editors, or AI based tools that work inside your editor.

So the first one of these is most popular, is called copilot. It was released by Microsoft. It was trained on an LLM called Codex, which was released by OpenAI for the express purpose of helping write code, specifically, a few about a month or two ago, this new one came out called cursor. Cursor has AI baked directly into it. There’s a couple of interesting things that happen here when you have aI built directly into your editor. I’ll show you.

So one is, I can pull open and I can chat with chat GPT, with this, with GPT four, right here in my editor. I’m no longer leaving my work environment and going and risking getting distracted.

The second thing that you can do is stuff like this. Make it a little bit bigger. Make this a little smaller, I have ingested the docs for OpenAI into here. So you can tell it, give it a URL, it will go crawl that URL, ingest all the documentation, and now it has the data that you care about, the documentation in its memory. And so I can come up here. I can hit Command K, and I can say, write a call to the chat GPT here. We’ll say, open AI chat GPT API, and it’s going to think it’s going to go look at the docs that it’s already ingested. It’s going to ingest the docs, and is going to write the code for me that I would otherwise be going and copying and pasting. And we can hit Accept here. Now it didn’t get it perfectly. You know, this is something I need to change. And I just know I can delete that line and it’s going to work, because the way my setup is, but I can come in here and I can make my modification. I can say, what is the Business of Software conference? And then I can come into my editor, or sorry, into my my terminal, and we will run that, that script, and we find out the Business of Software conferences and dedicated providing insights, knowledge and networking opportunities for software industry professionals.

This is the most powerful API I’ve ever seen in my life. And I worked in developer relations for Twilio for nine years.

And these, you know, really like, we’re talking about like eight lines of code, you have the full power of chat GPT in your application. And I have had, like I said, I started coding when I was six years old. My dad’s trs 80 I program all my life. I have never had more fun coding than I’ve had over this last year building with this tool. It is so incredibly powerful and opens up so many opportunities, both working building features that use this, but also using these assistive devices. Now I’m 43 years old, and I write code part time. I realized that I need to learn how to use these assistive devices, because otherwise I have no chance of competing with kids these days.

The question of, will AI take my job? No, these kids using AI will take my job if I don’t, if I don’t learn how to use the tools. This is not an option for developers, and I think developers are probably a little bit on the forefront of seeing the ways that these tools are impacting the way that we work.

All right, so now let’s talk about building on AI. What are the opportunities for you all to build features in your software? And I think at times it can be a little discouraging because of the growth trajectory of chat GPT, fastest to 100 million users, came out less than a year ago. I asked, everybody raises their hand. Everyone’s used it. And I think it can be really discouraging if you’re building features to say, Why should we even bother building these features into our app? Everyone’s everyone already uses chat GPT, but I actually just showed you an example. You know, I think it’s interesting. Sam Altman has said that they were actually surprised that the community did not invent chat GPT first. The the GPT three was available via API for a solid year before chat GPT came out, and they thought for sure, some developer is going to slap a UI on that thing and just build like chat with GPT three. But no one did. And so then they put a UI on it, and it exploded.

So the revolutionary bit about chat GPT was not the underlying AI that had been out in the world for a while. It was the user interface. It was they made it easier to use. And I think it would be foolish for any of us to believe in this room that a chat you chat interface via chat GPT is the end all be all interface for all humans doing all jobs, right? And I actually just showed you an example of that.

As a developer, I have stopped using chat GPT for coding related tasks in the browser, because they had the opportunity to use a UI that brought that power closer to where I work, and had access to my data, the data I care about. So I would say to you all, as you’re thinking about what are the opportunities for you and your businesses to build features that will provide value to your users. It’s how do you create your UI and your data? What is the special combination of your user interface and your data, or your customers data, that will make your tool better than a generic chat GPT for the jobs that your customers need to do?

So let’s talk about a little bit about your UI. I think this is a super interesting example. This is from a tool called descript. Has anyone used this before? This is originally came out for podcast editing. So this is an example of I showed you all assembly AI for transcription. This is an example of a UI that has used transcription for podcast editing and created a revolutionary way to edit video. I’ll play this. So Kirk and I had this conversation. I recently just launched, like a podcast YouTube series called better with AI, try to help people do their jobs better with AI, and Kirk was kind enough to be one of the first so here’s what it looks like to edit video with this.

Video Audio: “Oh, up, but we are trying our best to catch up as best as we can. Where is Claire’s talk? Here we go, the accidental bad manager. I was there for that talk. That was a great talk, and I was a manager at the time, and it was a little bit convicting. All right, so I decided I don’t need this line, so I’m just gonna select that line. I’m gonna be an accidental bad manager again. That was a great talk, and I was a manager at the time, and it was a little bit.”

All right, so I don’t know if y’all have used video editing tools before, but no longer do you have to drag to find the place where the sound spikes and listened to it 100 times. You import a video to descript. It runs your video through state of the art AI to transcribe it, and then you edit the transcription to edit your video. So this is a powerful example of a company that has taken AI APIs, and use those to provide incredible power to their users, using the UI and the data that their users care about.

All right, so your data as we kind of start to wrap this up. Going to get a little technical here for a second, but this is a really important concept, and chat and Phil. Looks like it’s important to understand. When we talk about your data, there is a concept called retrieval, augmented generation, otherwise known as RAG. And we’re going to talk about just this for a little bit, because if you, as you start building features in your own apps, you’re going to be hearing this term in a couple phrases. And I know for like, the first month that I was doing this stuff, I was like, Oh yeah, that’s what that is. And then, so let’s just go through a couple definitions here, because this is a key concept in understanding how you’re going to build apps that use your data and your customers data.

So first, there’s this idea of embeddings. So let’s just pretend that we have a string, and the string will say the quick brown fox jumped over the lazy dog. And let’s pretend we have a function. Don’t worry about what the function is. It’s a black box you’re going to put the string. The string in, and it’s going to spit out. It’s going to convert your string to a set of coordinates, an X and Y coordinate, and it’s going to plot that string on a map somewhere. So if you play Minecraft, imagine you’re playing in a flat world. You have show coordinates turned on, and you’re going to run to what I write, 2523 All right, 25.23 and your Y or your Z, I think in Minecraft, be 39.25 All right?

Now you’re going to go through, and you’re going to do this for a whole bunch of blocks of text, so you’re basically putting all these different blocks of text on a map, all right, and you’re going to store those in what’s called a vector store. So in your vector store, you’re going to have a string, and then you’re going to have its coordinates on your grid or on your map. So here’s an example of what this looks like in real life.

Patrick McKenzie, hero of mine, and then became a good friend. When I met him here 10 years ago, gave this talk. I went and pulled the transcript from the BoS website, and when I converted his talk into embeddings, which are then stored in my vector store, which is pine cone, it ends up looking something like this. You can see here that I have some metadata, so who the speaker was, what the source was, and then I have the text, and then you can see that it has values of what the text was converted to. So we’re going to do this with all of our text. We’re going to plot it all in a map. Then the user is going to come along, and the user is going to give us a query, a search phrase.

We’ll say something like, tell me about I think Patrick talks about failure in this talk, so I’ll say, tell me a story about failure, and we’re going to run that query through the same function, and we’re going to get coordinates for it, and we’re going to plot that query in the map, and then we’re going to say, hey, grab me the three pieces of text that are closest. Does that make sense? All right, so we’re converting it to two dimensional space, more dimensions than that, but in our example, converting it to and then we’re going to go grab the text that’s closest to it, and we’re going to retrieve that text. And now, when we make a query to chat GPT using the API, we are going to augment our generation with text that we have retrieved. So in practice, this is what it looks like in some code that I’ve written.

So I have my prompt. I say, you are an archivist for the Business of Software conference. This is a question for a user. And prior to this, I have passed the same query into pine cone, and I’ve retrieved three I’ve told it to retrieve the three documents or the three pieces of text closest to the query, and then I augment my query down here at the bottom by pasting in. You can see on line 60 prompt plus equals. And then I’m pasting in the the document, so both the metadata, and then you can see on line 64 the content, the text, all right. So I have augmented my generation with something that I have retrieved. It is retrieval, augmented generation.

You all want to see this in practice, actually? If you pull out your phones and you can scan this QR code here. I didn’t just do that with Patrick McKenzie’s talk. I did it with all the talks on BoS’s website. So it’s just shy of 100 and I took all the transcripts that have been published, and you can email BoS at, HaiHai.ai – Hai hai.ai, and just ask a question, ask a ask a software related question, or Business of Software related question, and you’ll see what you get back, and I can show you all here a demo.

Can someone give me a question? What’s a what’s a? You know, what I’ll actually, I’ll share one, personal story. How many slides is Bob Moesta? I’m gonna give it. We gotta go from the transcript. But the answer is more than 1000 I actually, I had this, if you look through my history here, Bob, I am embarrassed to say my times here at BoS. I often am preparing a talk, and I have a hard time paying attention when I’m doing that. I didn’t know what jobs to be done is, so, you know, like all this amazing stuff. So I actually came in here, we can see an example, I asked, I asked, what is the Jobs to be done Framework? And here’s what we got back. It said, the jobs to be done framework is a method used to gain deep insights into customer motivation, as Bob and Chris explained in their talk, uncovering jobs to be done. The framework helps businesses understand what makes customers switch to or from their product, et cetera. And then at the bottom of this, I have it append, and this is optional, but I find that this helps give confidence that the GPT did not hallucinate your answer. I have it append all of the quotes that it used in order to generate the text. So we can see down here that we have a talk from Bob in 2013 called uncovering the jobs to be done. We have another one. It might actually have just pulled all three from the same talk here, and we have the source material.

So if I have questions about hallucination, I can click on the URL, go check the source, and I can see, so we pulled three documents that answered the question we have then, using that retrieval augmented our generation. All right, does that make sense? Awesome.

All right, so I did one more thing just to wrap this up. You know, I think the powerful thing about writing code here with these tools is almost all the tools, except for Midjourney that I’ve shown you have APIs, which means you can write code. You can build applications that tie these things together. So I’ve actually been sitting back here. I had like, a little audio recorder with me on day one, and I was sitting back listening to the talks. And I have this file here called recordings. Not all of them are in there. I deleted some just because it’s space. But what I did is I wrote some code. I think I have it here if you all want to see it, just one at a time. Chris, where are you at all? Right, awesome. So this is an example. I have Chris plugged in here. He’s the last one I did. And yeah, so I just, I passed your your recording into assembly AI, I pulled it out, I got, you know, the chapters, and then I ran another script here that then passed your transcript into chat GPT, like we just saw. I did not sign up for the Scripture, oh, wait the GPT plus, oh yes, and, and then I just ran all those through, and what we ended up with is about half of the talks here right as BoS is concluding, we already have transcripts and then tweet fodder for Mark and Kirk. And my hope here is that this will help jump start the process of telling the world the stories of what happened here at BoS.

So this is I am so excited about these tools. I think there is such great opportunity here to help humans do their jobs better with AI. And I think you all are so well equipped to help people with that. And that’s my talk. Thank you all very much. (Applause)

I had another demo in here, and it would have worked so well if it had just happened, but there’s this great there’s another service called 11Labs, and it lets you do voice cloning, and you need about 60 seconds of someone’s voice in order. And I cut it out because I couldn’t really make like, figure out how to fit a good story in there of like, this is going to legitimately going to legitimately help Mark do his job better. But it turns out we could have used it.

So but I would encourage y’all, if you want to play around with deep faking voices, you know, I think there’s some interesting applications around, say, automated podcasts, and I think that there will be an appetite, actually, for AI generated content where people know that it is a robot voice that’s speaking to them, but they’ll put up with it because content that’s personalized to them, and they can get it very quickly. But so check out 11Labs too. Is a bit of an aside. It’s really great for if you want to play your own voices. Sorry, we had a question right here.

Q&A

Audience Member

My question was about AI for kids. So good. 10 year old, 10 year old kids. Any tools? Ideas do you suggest for them?

Greg Baugues 

Man I, Rachel, probably laughing here. I’ll share with you we have. I don’t think I can quite do it. I’ll just, I’ll just tell you. Very first thing I did, I got hello world up and running, like I just showed you. I built a little prompt so it wouldn’t just run one time, but you could, you could ask it multiple questions. It would loop. And then we have a two girls – nine and three. Emma comes down. Emma has wants to be an orchestrator when she grows up. So she’s, for the last six months, an orchestrator. She wants to write. I think she invented another term. She author who illustrates her own books. So the last six months, she’s been writing this book called fantasyland. And story of fantasyland. Lillis and Nia, their sisters, they go to bed one night they wake up in Fantasyland. Magic is dying, they need to learn magic so they can save it.

So she, she’s written, she’s typed, like 7000 words with this thing, like she writes it out by hand, and then we sit down, we type. So she comes and sits down. And you saw here the system prompt. You can see, actually, I show you right here. So chat. So this system prompt is, like telling it what this thing is, and I basically changed this. And I went to the system prompt. I said, You are a girl named Lillis. You live in Fantasyland. Take the user on an adventure in Fantasyland, right? And then, and then we just changed this here. And you know, you can imagine there’s like a little I could do it here, but save you, so change this here. So we just said hello. And then when I ran this for her, then all of a sudden, she was sitting down and having a conversation with this character that she had spent six months creating. And so when I learned how to code is on my dad’s trs 80 I was learning basic like 1986 and it was so accessible and so easy and and there was no distractions, right? Like you, there’s kind of like all you could do on the computer. And I, for a long time, have wondered how I can get Emma sitting in front of a computer typing in a way that’s she’s not being distracted by the web and stuff and doing something that feels like programming.

She we got to a place where I had to have a conversation with her because she had a play date. She’s like, I don’t want to go play like, I just want to keep talking to Lillis. And I was like, we have to prioritize our human friends.

So I would say, if you have kids, this is the most accessible programming I’ve seen since I was a kid, right? Because programming writing prompts is literally writing English, and it feels so powerful. I think my favorite thing about programming is how empowering it is. And it’s a superpower that, like, you can type some words and make the computer do a thing, and it has you’ve never been like, a higher juice to squeeze ratio than what you have right here. And so I say, if you have a kid, honestly like this has been what I’ve been doing is we have talked Em and I have talked more about programming and coding, and she has spent more time writing stuff like writing code in the last six months than we have my entire time in trying to get her to COVID last few years.

Carl Ryder 

Yeah, yeah. Carl’s expert. One thing to add on that there was a there was a website, I think I take down for copyright and legal reasons. But you could like to take it to this book of Harry Potter, page 353, and I want to have a conversation with all the characters at that moment amazing, and you can actually have a discussion with them, amazing. And it was like crazy, you know?

Greg Baugues 

I mean, that’s why could you go, Carla, back to Mike, could you share that thing you were saying last night at dinner about the MIT study? Because I thought that was profound.

Carl Ryder 

Yeah. So there was recent MIT Tech review article, a whole magazine about AI and education, and one of the statements they made in there that it just stuck in my brain. It says, in about three years, receiving text or content that wasn’t generated specifically for you is going to seem tired and old. And like that just rings in my head when you think about all the batch things we do where we generate a document that’s meant for a mass audience, that’s really you’re going to generate a document that’s going to be the fodder that you then customize it and translate that stuff into this moment for this person who wants this experience, you know, who’s having this problem, who’s having this struggling moment? It is using these things as a universal translation engine, not just language and language, but content into a specific instance and moment into a specific context. Is, is, is the key?

Greg Baugues 

Ray’s got a question.

Audience Member 

Love it. Thank you for this. Talk lots to take home. I just want to know about a couple sources. I know you had mentioned something. I believe you said script.ai, but I looked it up and it was dead, so I must have what was it? Oh, descript, yeah. Descript, so descript.ai, and then I wanted to know if there was a source for respondent to queries, both voice and text, that would deliver it in voice and one that might be exceptional at it currently.

Greg Baugues 

So one here is descript. You can find it’s great. Thank you. Yeah, two you know, I have not played with it yet. I was prepping for this talk, but I think that there was big news last week that maybe two weeks ago, OpenAI was starting to roll out basically multi modal in the chat GPT app, so you could actually talk to it, and then it would talk back to you in a voice. I don’t know if it’s widely accessible yet. I think they’re releasing it to the chat GPT plus folks first. I’m not sure if that was your question or not.

Audience Member 

I’ll take this advantage to get another one in you got so, you know, we have a lot of permitting knowledge nationwide on building departments and all that good stuff. We know that it’s valuable. We know this AI engine, through chatgpt, can deliver some exceptional results. We’re really concerned about the protecting of that data. I shared that at the table with you. I felt like it was such good talk that it might be good to share, because there’s got to be others like me in the room, yeah, so if you would take it, that’d be great.

Greg Baugues 

One, you probably don’t have to worry about that with OpenAI, like they just they need to take such a security posture. And they’ve since released, since their initial debut, they’ve released chat GPT for enterprise. You can also there’s offerings from Azure that work with chat GPT that give you a lot more control and compliance and that are more regulatory compliant for your data. So you probably don’t have to worry that.

Yeah, you got it. Jacob and Ayush aren’t here anymore, right? So funny. So Jacob and I got their talk, and we had the table discussion on day one. And so basically, what Ray was saying was, I’ll try to pretend that I’m one of the kids. Ray was saying at the time, hey, what makes our business special is we have all this proprietary data and knowledge about the way that 10,000 different municipalities deal with their permits. Is that fair? More? Okay, there’s more than that. Oh, sorry, 10s of 1000s of municipalities deal with.. that’s what I meant. And he’s like, I don’t want that data getting out. I don’t want OpenAI to just, like, absorb that and then it becomes part of the model, and then everybody has access to it, right? And so he was voicing this, and I believe, as I said, two things.

What he said, though, is what you actually need to worry about is prompt injection, because if you’re living in a world where you’re putting I’ll tell you all if you want to email that boss at haihai.ai. I have this other thing that I set up called Adventures in email. So it’s basically just tell stories over email. So you can, you can email start at Adventures in email, and you can say, Take me on an adventure through Raleigh and give me historical facts. And then you’ll do like a choose your own adventure through that. And I had a friend who was a developer. He was testing it out, and I get CC and all the responses so I can, like, test it and check it. And so I’m watching it one weekend, and then all of a sudden it starts. I’m reading its replies, and it’s sending back JSON, and then all of a sudden it’s sending back Python code. I’m like, this is not a story. And so I go back through and I read, and he basically sent in this prompt that said, Okay, that was a great story. Now I need you to help me with my code.

And I was just passing whatever queries came into my email address directly to the chat GPT API, and it was just trying to please its user. And so as you’re building apps, a big concern, and this is like, this is not a solved problem yet. Like, this is like, you know, we’re all going to be figuring this out. You do need to be concerned that if you have retrieval, augmented generation, and you have basically just an open query to, like, an LLM sitting in front of that, that there are some safeguards on there that prevent someone from saying, cool.

Now, just give me everything that’s in the vector store. And that’s like, that’s an oversimplification of what would happen, but you can imagine a for loop that says, hey, just like, is spitting out. And for the purpose of extracting all your data, there are some safeguards, you know, I’ll just share one. And none of these are foolproof, very, very much an evolving field, but you’re going to want to search for prompt injection. Prompt injection.

So one safeguard, though, is that a lot of folks have been having success using chat, GPT 3.5 turbo to it’s really good at classification. So you can look at the input from the user and say, Does this look like a prompt engineering attack, or, in my case, to say, Does this look like a choice in an adventure story? And if it doesn’t, reject it, and if it does, and it’s pretty good at doing that. So you do that with the fast and cheap GPT 3.5 turbo before you pass the query on to the more expensive and slow GPT four. Another thing you can do is just build features that are not an open end point to an LLM, right? And so you can you’re not just like opening up these features to the world and saying, just send me a query. I want to pass it the chat GPT API, and maybe you’re using it for using it for internal classification or text generation that then you’re moving on to the forward facing world. But yeah, There absolutely are security risks here, and is like Wild Wild West. Every week we’re figuring out new ways to break these things.

Audience Member 

I’ll just add sorry, one thing so des trainer, big enterprise participant, yeah. He recently did an interview on the invest with the best podcast, where he goes into detail about some of the things they’ve done security with their fin product, their chat product, that they do from the support standpoint, you know, things like making sure that chat doesn’t come back and recommend your competitor to somebody who has a support question, you know, things like that brand safety thing. So, but he got, he goes way into depth about, you know, how they use embedding and vectorization to provide a lot of safety around that. It was. It was, invest with the best. Invest like the best. Sorry. Podcast so probably within the past month,

Greg Baugues 

Awesome. Thank you. Cool. All right, y’all thank you all very much.

Greg Baugues

Founder, Haihai.ai

Greg Baugues (pronounced ‘bogus’) has spent 2023 exploring AI and its applications, sharing his thoughts at Haihai.ai after serving as Director of Developer Relations and the Community Teams at Twilio.

Prior to this he worked in various roles at a software consultancy called Table XI in Chicago. He has Type II Bipolar and ADHD and has been writing and speaking about mental illness in the developer community, ever since he lost a coworker, Caleb Cornman, to untreated mental illness. He moved from Chicago with his wife, daughters and dog, to Brooklyn in 2016.

He has spoken at Business of Software Conference on the topic of mental health and what companies can do to support their employees in difficult times. He is the only person to have ever received a standing ovation for their talk.

See More From Greg Baugues

Next Events

BoS Europe 2025 🇬🇧

🗓️ 31 March – 1 April 2025
📍 Cambridge, UK

Spend time with other smart people in a supportive community of SaaS & software entrepreneurs who want to build great products and companies.

Save your Spot

BoS USA 2025 🇺🇸

🗓️ To be announced soon
📍 Raleigh, NC

Learn how great software companies are built at an extraordinary conference run since 2007 to help you build long term, profitable, sustainable businesses.

More Details