How is Tripling Down on Becoming a Large Language Model-based Firm

SlatorPod #164 Krish Ramineni, CEO,

In this week’s SlatorPod, CEO Krish Ramineni joins us to talk about scaling the AI meeting assistant and building on the latest advances in large language models.

Krish starts with his journey to co-founding Fireflies, which began as a drone delivery service and as a result of conversations with customers and investors, evolved into an AI meeting assistant to solve their own pain point.

The CEO shares how they found their product-market fit after focusing on automated transcripts over human-assisted note-taking. He discusses the early days of AI investment and how with the rise of APIs and large language models (LLMs), you no longer need multiple PhDs to attract investors. 

Krish explains how Fireflies leverages technologies like Whisper to improve their language transcription, allowing them to be more accessible to global companies. He talks about their decision to improve their Super Summaries feature through GPT technology.

The CEO shares his excitement about the potential for LLMs and how Fireflies are building a Chrome extension that uses LLMs to summarize any article or video on the internet. He advises that simply building a wrapper on top of OpenAI is not a defensible moat for companies, but rather you should build a unique platform with a unique angle into the industry you’re selling to.

Subscribe on YoutubeApple PodcastsSpotifyGoogle Podcasts, and elsewhere

Kirsh talks about the current fundraising environment where there is a lot of money being thrown around for generative AI companies, but only a few will weather the storm. When it comes to hiring machine learning talent, Krish doesn’t believe in prompt engineering and also holds the view that machine learning companies may no longer need to hire large cohorts of ML PhDs to scale.

The pod rounds off with the company’s roadmap for 2023, which includes creating an ecosystem of extensions on top of Fireflies. These extensions will offer powerful functionalities to users in different sectors like healthcare and recruiting. 


Florian:, give us the elevator pitch for kind of the discerning SaaS user and enterprise customer so we understand the product first.

Krish: Fireflies is an AI meeting assistant. We fundamentally believe most of the work that we do happens in meetings. We have a lot of meetings, especially when I was at larger companies like Microsoft. Meetings was where all of the knowledge originates and my belief was, I can remember an email I sent two years ago. Why can’t I remember a meeting I had 2 hours ago? And that was the simple premise and then we said, let’s build an assistant that helps you remember all of your conversations that you can search back through, and as a function of that, we built Fireflies. It joins your meetings across Zoom, Google Meet, Microsoft Teams, all the major video conferencing platforms. It records, transcribes, summarizes, can also translate. You can do a whole bunch of stuff, analyze the call, run sentiment analysis on it. So meeting notes and search is just one part of it and that’s a nutshell what the product does. It’s used across the entire industry, whether you’re in sales, marketing, customer support, product management, recruiting. Our belief is that we should democratize AI for every person inside an org and we’ve been building Fireflies for the last couple of years to make it super easy so that every person has a meeting assistant.

Florian: Tell us a bit of the origin story. Like you said initially, the belief is it should be easy to remember something of course that happened just 2 hours ago, but you were working, I think, at Microsoft and you were working on something, customer voice analytics platforms there. So was that maybe that triggered a bit of the tech idea or just tell us a bit more about the origin story and how you got together with the Co-founder and kind of started this.

Krish: What I worked on at Microsoft was actually more oriented around customer feedback. So at a high level, when people are providing feedback on all these different applications that you use, like Word, PowerPoint, people leave a little bit of feedback in the system when something’s not working or if they have some user feedback. And we were working on an initiative that would quantify all of that, understand it. Not just look at metrics, but what people were saying. So that triggered a light bulb in me which said, hey, there’s a lot of interesting insights around language and conversational data. How can we quantify that and build some sort of insights around it? So while what I was working on at Fireflies is completely different from a technology product and market use case point of view from what I worked on at Microsoft, but I think Microsoft taught me the importance of natural language processing and the power of being able to gather all of these insights from different conversations. So in fact, when I left Microsoft, the idea for Fireflies was nowhere what it is today. In fact, my journey was actually to go to grad school and one of my friends was at MIT, I was visiting him for the summer before I started grad school and we were just working on a bunch of different things. We started off building drones. Nothing to do with what we were working on in SaaS enterprise. So the actual name for Fireflies comes from us wanting to build a drone delivery service that dropped off food and stuff like using drone delivery and packages. And when you have drones fly around at night with lights, they look like Fireflies. So that’s a funny story with where the name came from and we’ve done a bunch of different projects since then. We actually never changed the name. We just said let’s stick with the same name and continue because we were too lazy at the time. It’s too much of a hassle. But today, right, Fireflies, I think one of our customers said that’s such a perfect name because it’s like a fly on the wall. Usually that’s a phrase that people use. I’ll be a fly on the wall in this meeting. I won’t say anything, I’m just going to listen and take notes. So that’s what Fireflies resembles which is great because that was never our intention. I believe that the journey to what we have as a product today culminated from a lot of epiphanies from having conversations with different folks, customers, investors. And one of the common things we realized was we were getting feedback from so many different people and we couldn’t remember all of that. And we felt like, hey, if just there was a way for us to reference back through all of this feedback and insight when we’re interviewing customers or talking to investors for advice, that really was the thing. We started building something to solve our own pain point and that turned into really building the Fireflies platform as it is today.

SlatorPod – News, Analysis, Guests

The weekly language industry podcast. On Youtube, Apple Podcasts, Spotify, Google Podcasts, and all other major platforms.

SlatorPod – News, Analysis, Guests

Florian: You were basically on the journey to coming up with kind of a product or service idea, you stumbled upon the product that you ended up building, right?

Krish: We built it for ourselves initially, and we would just use Twilio and stuff to help us record our calls because we were having 5, 10 a day and we then just kind of hooked it up into some simple transcription. This was the very hacky early version of it. Then we said, look, transcription is not that valuable, it’s the notes that’s valuable. Let’s think about that and before we even went to a full on product or wrote any code for the summarization product, we literally had Sam and myself offer a few folks within our closed circle saying, hey, we’ll have a managed service where an assistant will join like a human and take notes for you. And we did this with like five people or something like that, and we did it for just a handful of meetings. And then we realized there’s no way you can do this, you can’t scale this, because it’s very tedious to build like a human operated service. Also you’ll need to get like NDAs and all the compliance stuff, so we immediately shut down that idea down and said, okay, we’re at a crossroads here. Is the technology good enough? Because transcription at that time wasn’t that good. It was also very expensive. So how do we build this system out? So we took a bet. We invested in the core product. We made sure to do the few things that were good really well, like the transcription, the ASR. We started building out the interface and then gradually realized we don’t want to ever build these human-in-the-loop sort of assistant-based services. You can hire people off Upwork and stuff, right? And people can build businesses off these. We said we just do not want to do that. There are companies today like Scale AI, which have like the Mechanical Turk for humans as an API, and they’ve built an incredible business. So it is a very ops intensive business anytime you have to manage humans, so it’s a very tedious process. So we believed that automation was the only way, and even if the quality was lower in the beginning, our scale would help us improve and get better over time. So that was the bet we took and then today transcription costs have gone down. The quality is almost 90, 95%, depending on language and accents, and this is for English at least, and the other languages are starting to pick up. So I think it was the right bet we made at that time in 2018. And we ended up launching actually, after building for almost six, seven months, raising our seed round in 2019. All in all, we ended up launching in January 2020 out of Beta and then a few months later, the pandemic happens. Everyone is remote, everyone wants a way to make their Zoom meetings and Teams meetings and Webex meetings a little bit better, and Fireflies just picked up as a result of that and it was like the perfect storm to get things going.

Florian: That was a good time to launch anything that’s remotely connected to remote. But wasn’t it daunting to launch into an NLP, natural language processing domain, not being like an NLP PhD? I mean, there’s hundreds and hundreds of PhDs in that space that are trying to productize something, and you guys are coming kind of more from the problem angle, and then you’re using the tech. Wasn’t that daunting in the beginning?

Krish: I think that there were a lot of investors in the early days that were oriented around, you need to have like, 10 PhDs. This is a hard problem. Bigger companies have tried to solve this and have failed at it. Why is it like two 21 year olds going to solve this out of college? And I believe that there are things with where we were in the industry where the tide was turning. Today, to work on AI, ML, ASR you don’t need 10, 20 PhDs. I actually believe with the rise of even LLMs today, right, and what OpenAI is putting out there, all this traditional ML work is no longer needed, like where you have to do classification and a lot of these things. Now, it’s just as simple as an API call. So I think that we always believed someone would get there and create it in a way where it’s just as simple as an API call. We didn’t assume it would be this fast. We thought it would have to be a long slog. It just so happened that the technology matured fast enough, and we focused on the problem space and winning the customers over, because there are a lot of ASR providers out there that do a great job, so we never wanted to be an API company. The transcription is just a means to an end for us, just like how storage is a means to an end. Like when cloud companies came out, it was all about just storage, but that’s just one level of functionality. But what can you do to build an actual platform on top of it? So, for us, we believed that good transcription was just around the corner, and it took some convincing, but we said, okay, if investors are not convinced, let’s go build it. Let’s prove it and then we ended up finding a few partners and companies to work with, and the product worked. Obviously, it wasn’t seamless in the beginning. It was a work in progress, but there were a few that could see the potential for this and how big the market was, and yeah, it just took off from there. And unlike some other companies that were focused on very much enterprise or call centers, we took a very horizontal approach. There was also a pushback there. They said the technology is not good enough, so you have to find a very verticalized segment so you can train your data just for contact centers or just for the insurance industry. And we said, no, I don’t think that’s how AI is going to work. It’s going to be able to work across any type of conversation. And in fact, OpenAI has proven that with the way they’ve built their models, the large language learning models, it can work across anything. Sometimes overdoing it and creating complexity into ML wasn’t the solution. In the case of OpenAI, they trained on a very general model across a whole bunch of data, so I think we were thoughtful enough. My Co-founder, Sam, our CTO, was also thoughtful about this and we stumbled upon something where that hypothesis turned out to be true. It was as much a technological hypothesis as much as a product and customer based hypothesis, right? You have to find product-market fit. We also had to find technology market fit, and that ended up having to work out. But it was very painful those first couple of years because you don’t know if the technology works and customers are always going to be expecting a lot out of you, so yeah, it was definitely tough.

Florian: You’re saying basically you didn’t kind of invest heavily in kind of super customization of this model and now it’s almost a godsend that these large language models are coming out and are so broad in their capabilities because your platform can integrate this very easily without having to have maybe looking back at all these customization in a particular vertical. Very interesting. I do want to talk a bit more about kind of ChatGPT and LLMs, but before that, one problem, of course, is also language, right? So can you talk a bit more about the multilingual features that you have? I mean, how, where does translation, machine translation kind of factor into the current product, if at all?

Krish: Absolutely, so for us, English was primarily the main focus for transcription for English-based meetings. We’ve had a lot of requests over the months, over the past couple of years, since actually day one for other language support, so transcription support for Spanish and French and German and a lot of these common languages. So one of the things we decided to do was, okay, let’s start investing in foreign language with the idea that it won’t be as good as English today, but it can surely follow the same curve that English followed over the last couple of years. And just recently, we launched to almost 30 different languages, and that’s something we’ve been continuously improving on. We’ve been leveraging technologies like Whisper for foreign languages, and then we have some other plans to really make it have a huge jump in the quality of the transcription that comes out of these languages. And then we have to also build our system so that it can understand those and process those different languages. So I would definitely say it’s very much early days for our foreign language initiatives, but we’re very excited about it because there are people in other countries that are now starting to talk about Fireflies and tweet about it. And it’s been very cool because all of a sudden we’ve increased our total addressable market from just English speaking countries to the world and also pushes us to build our product and also all of our AI features to be able to do that. And we’re very much LLM based now as a product, so we don’t need to create manual NLP rules like in the early days, so it can seamlessly work across any of these things. We introduced a product called AskFred, which is like a chat assistant, where instead of you having to review the entire transcript or read the summary or listen to the call, you can just Ask Fred questions about what happened on the meeting. Like what were the decision points? What were the blockers? What’s the timeline for getting the deal closed? So these sort of questions you can ask. And what’s so cool about AskFred is because it’s built on LLMs, if the meeting is transcribed in Spanish, you can ask it questions in Spanish and it’ll give you answers in Spanish. Or you can take an English meeting and let’s say you’re not a native English speaker, you can say, hey, take this transcript and summarize it for me in Spanish or in French or in German. So we’re almost becoming inclusive of every language and language agnostic. So Fireflies can not only transcribe, but it can also translate. And it can also translate the summaries and answer questions in the language that you’re coming from because people will have… We will have customers from global companies where the meeting might have happened in English, but that person is located, let’s say, in another country where they speak another language and they’re more fluent in that native language. So it’ll help them understand the meeting to be able to go through it and have a conversation with Fred in the language of their nationality.

Florian: I’m getting live sold here like, this sounds super useful when you can… Yeah, as you said, there’s all this hidden intelligence and if I could very easily Ask Fred and get information out of all these calls we’re having, I will definitely look at it later on. But you also launched Super Summaries. Tell us more about that and what that does.

SlatorCon London 2024 | £ 980

SlatorCon London 2024 | £ 980

A rich 1-day conference which brings together 140+ industry leaders views and thriving language technologies.

Buy Tickets

Register Now

Krish: Super Summaries really takes the baseline summaries we have and almost ten X’s it. So we’ve always had a lot of feedback from customers about, I wish my summary could be slightly longer, I wish my summaries could be bullet points, I wish my summary could be like this. And after a culmination of listening to all of that feedback, we said, let’s give them the best summaries possible in the market and let’s supercharge it with technology like GPT-4 and let us make the best quality summaries that we can. So that was something that we started doubling down on and today Super Summaries after every meeting will give you a paragraph synopsis of what happened. It will give you a meeting outline with timestamps that you can click on and so it’ll create an outline of the meeting and then you can click on different parts. It will generate shorthand bullet point notes like a human would and then it will wrap it up with action items and next steps. So it’s four components and those summaries are extremely valuable for people to get a full thorough thing. So those that wanted bullet points, those that wanted paragraphs, those that wanted outlines, we do it all after a meeting and we don’t compromise on quality there. And I believe Super Summaries is just the start. We will eventually be building an engine where every meeting, every type of meeting may need to have a different type of summary. So instead of one general summary or general Super Summary, we will have a way where if you have a board meeting, you can have minutes done like you would need for a board, like a professional. If you are having a meeting around your one-on-one or performance reviews, then you can create a meeting around that. You can even create a template of these are the type of things I look for in these type of meetings and you can build that into Fireflies as an extension and you can turn on that extension. So summaries is just one extension on top of Fireflies or one application on top of Fireflies. You’ll be able to build hundreds of applications on top of Fireflies. So that’s something we’re very excited about and this is all possible thanks to LLMs.

Florian: In a lot of these industries, they have very niche type of vocabulary, terminology. Is it possible to upload certain term lists and then the system learns from it and spits it out correctly or?

Krish: We do have that where you can put in certain terms and the system learns from that. You can also, when you edit the transcript, we have the functionality where you could do that. That will also be feedback loop for the system so that it can continue to improve over time. I do believe that the technology, especially getting in if you’re a SaaS customer, it’s able to pick up acronyms, unique words really well and that’s really like the improvements with the training data and where ASR has gotten in the last couple years. I’m just amazed at how it can pick up company names and unique acronyms. So I haven’t even used custom vocabulary, to be honest, in the last like six months because the default engine is rock solid.

Florian: For us it keeps SlatorPod with an E instead of the O. That’s the one thing that in our transcription engine.

Krish: That you might have to feed into it so that it learns since it’s not like your usual word.

Florian: Tell us about your experience in the last six months, right. I mean, we had Whisper, we had ChatGPT, we had kind of this general LLM boom. I mean, just Hugging Face just launched HuggingChat, which now you could see it’s free, basically, and it works okay, it’s not quite as good as OpenAI. But how was it having built a product that is so kind of based on these types of technologies and then all of a sudden you have this huge leap in quality from all angles? Like with Whisper for the ASR and then with ChatGPT for just the general kind of reasoning and summarization. How was this and how did you guys react to it?

Krish: We work closely with OpenAI. I definitely think that they have a great head start in terms of what they’re doing. Other companies are fast following. You have some really great products coming out, and in general, I like the excitement around the LLM space. So this makes me believe what happens in a world where you have several companies that are working on this at the infrastructure level and powering companies over time? Could there be a case for people wanting to use different LLMs for different problems that they’re solving? So that is something that I’ve believed in. I also like that the quality is improving from these LLMs. I think the jump from GPT-3 to GPT-4 is quite noticeable and we’re also seeing the costs start to go down. They reduced the cost of GPT-3.5 Turbo. So those are all really promising signs, and it means that more and more companies are starting to adopt this. I heard that more than 50% of the companies in YC’s recent cohort were all building something around leveraging OpenAIs or GPT or LLMs. What’s interesting is we had access to this technology a year in advance, right, before ChatGPT even came out as a result of working with OpenAI and I believe the technology was there. It’s just that ChatGPT opened up people’s eyes to what is possible and that created a lot of buzz and excitement. So I believe that the technology has been there and it’s improving rapidly, but it’s the use cases and how creative people can get. So for me, it’s almost like the launch of the Apple App Store and then people coming out with fun, silly apps like Flappy Bird back in the day, took off. It was this game that went viral and then you had Angry Birds and all these apps that went viral. And so people start doing the fun stuff and then later you start building enterprise-level use cases. Today, like without an iPhone, if you’re on a construction site, you need to take photos and upload something to the cloud, and you need to do all these things, or you need to press a button and book an Uber ride. All of those things are possible because of the device that enabled it and the marketplace that enabled it. I think we’re in an era right now with LLMs and technology like ChatGPT and all of these generative AI technologies. It’s that on steroids, it’s 20 X, 100 X the potential there. We’re still in very early days. There’s a lot of companies that are, one, trying to be playing the platform play, like competing directly with OpenAI, and then there’s other companies that are fast following and building on top of OpenAI. It just so happens, for us, language is a big part of what we do because we’re doing transcription, we are doing summarization. We’re now doing foreign language translation. So because of all of the things that we’re doing, we felt that it can be one component that we can leverage to help people better understand what happened in their meetings. We have a goal that if it’s an hour long meeting, you should be able to go through it in five minutes and understand what went on and I think these LLMs definitely help us do what we need to do. We had so much interest in this new technology and what we’re doing with these Super Summaries and all of that. Our AskFred, because that’s our own chat module that people can interact with, has been so successful with our customers that we decided to take it a step further. And so we’re running an experiment today where we have built a Chrome extension. If you install that Chrome extension, AskFred will now pop up as a widget on your browser. And when you go to any article on the internet, any YouTube video, you can just press a button and have Fred summarize that for you. So Fireflies not only summarizes your meetings, but any email I’m reading, any article I’m reading, any YouTube video I’m watching, it will be able to summarize that with a press of a button and then I can also AskFred questions about the content that I’m seeing on the page. So this is a way for us to also take advantage of all of this progress in LLMs to make Fireflies not just an assistant for your meetings, but an assistant for all of your workplace, for everything that you are reading and consuming. So we’re curious and excited to see how that… Ever since the launch of that, every article I’m reading, I’m just having Fireflies summarize it into bullet points. Or if I have to have a meeting that I have to catch up on, or I have to go into a presentation. I was at a conference and I wanted to understand this topic before I went into that conference, so I had like eight YouTube videos opened up on multiple tabs. I’m trying to power through all of them, and I can’t sit through all of them and each one was like 30 minutes long. So instead, I just said, Fireflies, can you summarize this YouTube video for me? And I did it and I learned that information in about half the time. I Recruit Talent. Find Jobs

LocJobs is the new language industry talent hub, where candidates connect to new opportunities and employers find the most qualified professionals in the translation and localization industry. I Recruit Talent. Find Jobs

Florian: I definitely need that. Now with this whole boom. Some call it, maybe hype. I’m actually more in the boom camp, I guess. It’s also tough to break through the messaging, right? I mean, there’s like a gazillion, front-end only, API only companies kind of being spun up at the moment. How do you go about kind of go-to-market / marketing in an environment like this? How do you position yourself as somebody that’s been around now for three, four years or even longer and has kind of a very robust product when there’s so many kind of two, three, four-month old companies out there now?

Krish: I do see that a lot of companies are essentially building a wrapper on top of OpenAI and that is their hook or unique differentiation in order to stand out. But to be honest, that isn’t a very defensible moat. I believe the defensible moat is having an actual platform, a unique data set that you are able to provide value on top of the LLMs for. For example, if everyone is training on top of the same data, or doing something creative on top of the same general models, like if everyone’s going to a public data source or just relying on the data that OpenAI has in its model, it’s not very valuable. So that I think, can be easily replaced. Like you have to figure out what is the moat? How do you become a differentiator there? With Fireflies, our unique data with the meeting data, the transcription data, the conversation intelligence data, that’s already been there. So I think we’ve had the luxury of building out the actual platform before these LLMs were available and then LLMs just take what’s there and double it or triple the value on top of the data, right, so you have all this data you’re sitting on. I have six months worth of meetings. I don’t have time to go through all of them. Let’s have LLMs make it faster for me, right. So I think in that way it’s extremely valuable. I do see that a lot of people are going after simple use cases, specifically like customer support. A lot of companies are trying to tackle that and a lot of investments have been made in that space at the early stage where they’re like using LLMs to make customer support or support to answering support tickets better. So you’re going to see people go after the lowest hanging fruit. That’s what my belief is and you’re going to see people that just build quick wrappers and see if something sticks. My analogy here is if you built a cool flashlight app on Apple, and eventually Apple says, now we’ll just build it ourselves and we’ll just have it as a native app and you can just turn on the flashlight, so it doesn’t make a lot of sense for you to do that, right? So you have to create enough value where the default platform just can’t replace you. I think there was a lot of worries for people when ChatGPT announced plugins and what that could do, right? Because if you’re just building a wrapper, you no longer are creating the same value because ChatGPT and OpenAI could easily replace you. So I always believe use technology as a means to an end and you still have to do the fundamentals of SaaS, right? You still have to build a platform that users use and you still have to have a unique angle into the industry you’re selling into. So for us, I think, because we had to build all those fundamentals, the LLMs are just like the magic dust that you sprinkle on top of the product to make the end result even better, rather than building your entire company off just an LLM. And I don’t think you can build, today at least, something and stand out because everyone is doing that. What’s different with you?

Florian: Let’s talk briefly about the kind of the fundraising environment. I believe you guys raised in the past, and now you’re sitting at this intersection of the hottest sector on Earth, AI and kind of B2S SaaS, which had its moment maybe a couple of years ago. But now, if you’re only B2B SaaS, it’s a little tougher to raise going forward than maybe it was 18 months ago. So what’s your take on that? How do these competing forces kind of play off each other?

Krish: We’re fortunate to be in a space where it’s always been hot for various reasons. When we were working on voice technology, voice chatbots was really, really hot at that time. Then when we started working on all the meeting stuff, remote work, people were investing at a rapid pace in anything that can enable remote work. So all of these video and audio products that helped with remote work, that was taking off like a flywheel. And we saw the public markets where the stock were trading at crazy multiples because of that and so people were investing like crazy. And then we are now in a place where we’re now categorized as a leading generative AI company because we use generative AI and because we’re working on language, text and transcription. So I think that the hype was there in different ways. I definitely think this is the biggest one. And even though as an industry overall, fundraising has slowed down, especially late stage fundraising, I think at the early stages, if you are a generative AI company, there is a lot of money being thrown around right now, because they want to know which works. Like which is going to be the next YouTube, which is going to be the next Facebook of this space, and I think a lot of money is being thrown around. We definitely get a lot of inbound interest as a result of that these days and because we’re that intersection of we built a SaaS product, but we’re also like, we triple down on becoming an LLM based company in terms of leveraging every aspect of LLMs in every part of our business. So I think, yeah, it’s a really exciting time. But again, I think that for the dozens of products that are in a market, maybe only a few will eventually weather the storm because it is still early days. You’ll have to wait till the dust settles to see which type of companies settle down and then you’ll see some big funding rounds and things like that happen. I think we’ve seen that with a company like Jasper, which has raised tons of money and has grown rapidly over the past two years. So, TBD, we have to see where this goes. But what I can say is generative AI companies in general are getting a lot of interest. High valuations, I’ve heard of like a company that raised without any revenue at hundreds of millions of dollars in valuation, so there is definitely hype.

Florian: I love that clip from probably ten years ago from Silicon Valley. Like, no, you can’t disclose your revenue, you’re pre revenue. So how do you hire in an environment like this, especially on the AI kind of ML talent side? What do you do to get some of these probably very much in demand engineers?

Krish: I don’t believe in prompt engineering, so I’ll just put that out of the way. I don’t think that is really as hyped up as it’s meant to be. For me, prompt engineering is just basically there’s a gap in the skill sets of what an AI can understand. And once the technology gets much better, the need to fill in that gap with the right words will go away. I’ve already seen that with GPT-4 where I can phrase something less clear, but then get it to generate some magical things for me. So I honestly think prompt engineering is like the phone operators back in the day that redirected your calls. Eventually you’ll have automatic routing, eventually the AI is going to be good enough to understand what you’re trying to say and you can convey it and people will build their skill sets over time. So I’m just going to put that out of the way and that’s my contrarian view and maybe I could be wrong there, but I would say that once the technology is good enough, even your grandmother should be able to converse with it and get the output that it wants out of it. I think it’s a learning curve, just like how people had to learn how to use touch screens on a phone and get used to the user interface. I think we’re learning to talk to machines and one, the machine is going to get more sophisticated, and two, we’ll try to teach the machine and talk to the machine like we would talk to a little kid, it ends up doing much better because it understands everything quite literally in the literal sense, so that’s something I would start with. And in terms of the talent and the space, I also have another contrarian view that now you don’t need these PhDs and those companies that are working on these heavy ML companies. I wonder what happens now that you have LLMs, right, and how that impacts what you do? So it’s hard for me to say right now, but I think there’s a lot of companies that went with, I’m going to pick a very specific industry. I’m going to hire a bunch of machine learning folks, and I’m going to go solve that for particularly that industry. And then you’re going to have an upstart startup come around saying, I’m just going to use OpenAI on the same problem, and I don’t need to hire 10 ML PhD engineers. So OpenAI is really democratizing AI in that sense. Again, you’re going to need incredible talent if you’re Anthropic or one of these other companies that is directly competing with OpenAI. But if you are building on the application layer or the end user layer, the idea that you need to hire tens or 20 ML PhD engineers, that is the wrong idea, I believe, at this point in time, because of what’s happening with OpenAI.

Florian: You’re not the first person that states this view in the past couple of weeks that I’m having conversation. I’m trying to absorb this because I keep thinking like, okay, it’s a hot space, you need to have these ML engineers, you need to hire them. But what you’re saying makes a lot of sense. I mean, you don’t need it if you’re not the fundamental base layer anymore, potentially, right? So this is a big disruptive moment and if you build on the application layer, you need different kind of talent, so very, very interesting.

Krish: Back in the dot-com bubble, if you wanted to spin up a product, there was no AWS, there was no Google Cloud. You needed to have skilled people on infrastructure and DevOps to even just get the product off the ground. You also needed millions of dollars to just get the product running because you had to build out all of your infrastructure, your servers, all of that. And then you have Facebook that was running off a box at the time, and they started scaling. And then you have when Snapchat came around or Instagram came around, they’re just plugging into AWS, and like, two college kids can spin up a product. So I think what happened there where you don’t need these industry veteran infrastructure prodigies, same thing is going to happen when you’re working on these sort of AI related problems. You can at least get started without these ML engineers to get off the ground. And so there’s a lot of veterans or older folks in the industry who probably came into it with the mindset that I need to raise tens of millions of dollars, I need to hire all of these industry leaders in AI and ML in order to stand up a product. Whereas you can have people work on this stuff in a matter of days and weeks. So I saw some really cool projects come out where Google announced Duplex some years ago, where it’ll call the restaurant for you and make reservations for you. Someone with Whisper and the OpenAI LLMs was able to build that in a day, so that’s what I’m saying, right? I’m sure Google had to have lots of PhDs work on that problem, but now you have students hacking it together in a couple of days. That’s how far we are in this market. And it’s crazy how fast it’s evolving and translating over time, like even Auto-GPT, the new thing that’s come out where you can have the AI solve recursive problems. So I think we’re in for a crazy time, and it’s hard for even me to keep up with how fast it’s changing right now.

Florian: That’s the slightly scary part when it comes to the kind of solving cursive problems and having the ability to click on things on the Internet, that’s the part that’s slightly freaking me out. Now, can you disclose anything that’s on the roadmap for you guys in 2023 and maybe ahead in next maybe 18 months?

Krish: I alluded to this a little bit, but one of the biggest things that we’re working on is creating an ecosystem of extensions on top of Fireflies. So we believe that every person has a unique reason for using Fireflies and wants unique data pulled from their meetings and I don’t think a general summary would justify that. And here, by building extensions on top of Fireflies, people can do some really powerful things. So, for example, I’ll give you a scenario where I’m in sales and I want Fireflies to score this call for me, right? Do a lead scoring sort of thing. I can actually build an extension on type of Fireflies with no code that will score the call based on exactly the criteria I want, maybe even get me down to a number, maybe use a couple of heuristics. Was this mentioned on the call? Was that mentioned on the call, et cetera? And then it creates some sort of score. Now, I’ve automatically built a lead scoring app on top of Fireflies. So just like how summary’s an app on top of Fireflies, lead scoring extension is another app on top of Fireflies. And so if you think about it, there’s going to be hundreds of unique apps that you’re going to be able to build on top of Fireflies or extensions. So if you’re in recruiting, you might be trying to screen for certain things, and you can have Fireflies do that for you, right? So I think it’s very, very exciting. This weekend I was playing around with the healthcare use case where I fed in like a sample audio, doctor patient audio, and then I said, please write this in the way that a doctor would write it. Also provide recommendations and suggestions about the diagnosis and it did just that, and it was accurate. So that gets me excited about the potentiality of this and I think Summaries is just utilizing 10% of what the technology can do. There’s so many other things that it can do, so we’re very, very excited about the launch of extensions for users. So Fireflies is no longer going to be just a note-taking platform or note-taking app. It’s going to be a platform with thousands of different apps on top of it.