Pod Insights Latent Space: The AI Engineer Podcast From Astrophysics to AI: Building the future AI Data Stack — with Sarah Nagy of Seek.ai About this episode If Text is the Universal Interface https://scale.com/blog/text-universal-interface, then Text to SQL is perhaps the killer B2B business usecase for Generative AI. You may have seen incredible demos from Perplexity AI http://preplexity.ai/sql, OSS Insights https://ossinsight.io/explore/, and CensusGPT https://twitter.com/0interestrates/status/1633992774394544128 where the barrier of learning SQL and schemas goes away and you can intuitively converse with your data in natural language.
But in the multi-billion dollar data engineering industry, Seek.ai https://www.seek.ai/ has emerged as the forerunner in building a conversational engine and knowledge base that truly democratizes data insights.
We’re proud to present our first remote interview with Sarah Nagy https://www.linkedin.com/in/sarah-nagy to learn how AI can help you “seek what matters”!
Timestamps
* 00:00: Intro to Sarah
* 03:40: Seek.ai origin
* 05:45: Data driven vs Data backfit
* 09:15: How Enterprises adopt AI
* 12:55: Patents and IP Law
* 14:05: The Semantic Layer
* 16:35: Interfaces - Dashboards vs Chat?
* 21:05: LLM performance and selection
* 26:05: LLMOps and LangChain
* 30:55: Lightning round
Show notes
* Sarah Nagy Linkedin https://www.linkedin.com/in/sarah-nagy
* Seek.ai https://www.seek.ai/
* Sarah on the dbt podcast https://roundup.getdbt.com/p/ep-40-what-can-generative-ai-do-for
Lightning Rounds
* Favorite AI Product: Stable Diffusion https://stability.ai/blog/stable-diffusion-public-release
* Favorite AI Community: Eleuther https://www.eleuther.ai/
* One year prediction: Things will move fast!
* Request for Startup: Scheduling/Emails (shoutout Ipso.ai https://ipso.ai/ from our hackathon!)
* Takeaway: Automate everything!
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe https://www.latent.space/subscribe?utm_medium=podcast&utm_campaign=CTA_2
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:08 Alessio Fanelli? Hey everyone, welcome to the Late in Space Podcast. This is Alessio, Partner and CTO in Residence at Decibel Partners. I'm joined by my co-host swyx, Brother and Editor of L Space Tyres.
0:19 Speaker 01 Today with a special guest, Sarah Neji from Seek AI. Welcome, Sarah.
0:23 Speaker 02 Yeah, thank you so much for having me.
0:25 Speaker 01 Awesome. I like to introduce guests on your behalf so that you don't have to always introduce yourself and then also you can get a chance to correct me. So you were an Astrophysics major at UCLA and then you're a Master's in Finance at Princeton. You spent something like, it looks like 10 years in quantitative trading, which is fun because I also was briefly a quant in a hedge fund. And you were most recently a Citadel before you started Seek AI. Anything in your bio that people should know about you that people don't find on LinkedIn?
0:53 Speaker 02 Sure, there's a lot of hobbies I had throughout the years, things I did for fun. So I mean, you know, that's not on my LinkedIn. I actually played classical piano for over 10 years and actually was a DJ for a little while, just DJing around Williamsburg and Lower East Side. Just had random hobbies here and there. I also used to do improv comedy. Maybe I'll talk about that later, like why it's relevant to ChatGPT. Because there's actually some similarities. Besides that, I mean, you know, pretty much everything's on my LinkedIn. So what you said, you know, that's pretty much my background. Like you mentioned, started out doing Astrophysics.
1:35 Speaker 02 I was working at UCLA in Caltech, doing a lot of research using data from the Hubble Space Telescope. But like you mentioned, you know, I saw a lot of my colleagues going into quantitative finance. And so that's kind of what brought me out here and that was my background before starting Seek.
1:49 Speaker 01 Yeah, that's super cool. Actually, I don't mind going into improv. What got you into, this is a little bit off topic, but what got you into improv? And I assume ChatGPT is kind of like a yes and in terms of how it agrees with every question that you ask it.
2:04 Speaker 02 Yeah, I actually, I was kind of a theater kid in high school. So, you know, I had actually been doing musicals since I was in like middle school. And I just was really, really bad at improv in high school. Like I auditioned for comedy sports like two or three times and just never made it on the team. And then later, as an adult, I started taking improv classes. And after working hard enough at it, I ended up becoming a lot better. So that's kind of why I chose to do it. But yeah, I think it's actually really helpful to use that to describe ChatGPT because what I realized, you know, working with these models for
2:47 Speaker 02 actually, I I can almost call it several years now. In in improv, if you have to pretend to be a character, um, for example, a doctor, when you're on stage, you can just talk about you know, things that sound believable to the audience, but they may not necessarily have to be factual. And so I I've found that talking with these large language models, they can say things that sound believable. But, you know, they're not necessarily true. Uh, so you know, I I I just think it's kind of an interesting analogy that I noticed.
3:24 Speaker 01 It it is, it is. So then that's an interesting lead in to using GPT three and large language models to be a source of truth of data.
3:36 Speaker 01 Well, uh, maybe we should, we should just set the context. What is SEEK and how do you explain it today?
3:40 Speaker 02 Yeah, so SEEK AI is a natural language interface that anyone in a business can use to ask questions about the data within the business and get the answers that they need much faster than it would take talking to the data team. So to tell you a little bit about why I started SEEK, it really kind of arose from this pain point that I just kept encountering pretty much everywhere I was working. Which was, you know, I want to focus on projects that could really help the business. What really excited me about so-called big data was being able to just unearth all of these insights.
4:20 Speaker 02 And, and you know, especially in the quantitative finance world, it gets really exciting when you can put together these novel trading strategies. You know, and you're the first one to discover them and, you know, they can make the company a lot of money. That's just an example of the type of value that excited me about becoming a data scientist and a quant. But what I kind of kept seeing was that my less technical colleagues really didn't have the right tools to be able to answer their own questions about the data. And so they would constantly kind of come up to me throughout the day and ask me to help them find insights from the data.
5:02 Speaker 02 You know, and what that actually looked like was very kind of mundane, you know, I would get a question like, "Hey Sarah, I'm talking to this customer tomorrow. Can you just pull this data really quick so I can show this to the customer when I meet with them?" And I would just have to write a lot of very kind of manual code that was very simplistic and kind of just give them the answer. And that would take time away from the projects that I wanted to be working on. And since I started SEEK, which is almost a year and a half ago now, I've learned a lot more about the pain point from the business user side as well. You know, and it, it is very tedious having to just wait for the data team to get back to you when you need some data.
5:45 Speaker 02 You know, that's a little bit about SEEK and the problem that we're solving.
5:48 Alessio Fanelli? I think like one of the things about organizations, it's data-driven versus just using data, right? And I think in one of your interviews before, you talked about how sometimes organizations kind of use the numbers to justify the qualitative decisions that they made before versus going at it the other way. How do you think about that in the context of SEEK? So you're going from you, Sarah, being in a way the gatekeeper, right? To the data when these people come to you versus now you're letting everybody query, which also means everybody can kind of look for the slice of data that they need to justify what they want to prove. How do you think about the balance and like how data teams should think about this
6:30 Alessio Fanelli? as the access to it becomes more widespread versus them being the central place where?
6:35 Speaker 02 That's a really good question. The kind of challenge you're describing is how do you actually educate the business users on how to actually utilize the insights once they get them? I mean, there is a lot of education involved here. To your point, you know, I think the example you're talking about, I remember one time I was talking with a sell-side research analyst. And they were showing me this data, you know, that was pointing towards a certain prediction about a KPI of a stock. And I was trying to learn more about how did you actually get to this prediction? And they were actually telling me they tried a lot of different hyperparameters, basically.
7:20 Speaker 02 And they chose the set of hyperparameters that validated their thesis, which was a very qualitative thesis. And I mean, I just was so surprised, you know, I'd never actually seen anyone do that before. And that is the totally wrong way to use data. You know, it's not about like trying to figure out how do I manipulate the data to support what I already think is true? You know, that's just using data to prove a point. When you really should be using data to help you uncover the truth and make a better decision. So when I saw this, it really changed my perspective about
8:01 Speaker 02 just how people may be misusing data and insights even today. So I mean, as a startup founder that goes on podcasts like this, you know, and does some public speaking, I can definitely help kind of talk about what I think is the right way to use data. But I also think it's the job of the leadership of every company to learn more about what does it really mean to be data-driven? And you know, educate business users on how to properly use the insights that they're getting from the data. But you know, that that's a really good point that you're making. So I'm I'm really glad you brought it up.
8:42 Alessio Fanelli? Yeah, I think it's the usual dilemma in the organization. It's like, give everybody the power to query the data, then it's how did you actually come up to this conclusion? You know, is this actually a actual statistical thing, or did you just put together a bunch of queries that made it look like what you wanted? Do you see companies using SEEK as kind of like, yeah, replacing the data analysts in a way and then still funneling all the reporting through the data teams? Or are you seeing, for example, yeah, the the sales team doing all the reporting on their own and then looping in their data team? Once there's maybe some complicated queries or just to fact check it, how do you see SEEK playing the organization?
9:22 Speaker 02 Well, I think this is also, it goes to a bigger question of how is AI going to start getting adopted in organizations? And I think, you know, in general, it's going to be the most manual, time-consuming, soul-sucking tasks, for lack of a better word, that, you know, people just don't want to be doing, that I think AI is really going to help with. You know, so when you talk about what is SEEK going to be replacing? It's not replacing data analysts or, you know, data people. It's just replacing the stuff that they don't want to be doing, which is taking time away from the real value that they can be adding to the business. And instead, just automating the stuff that is very ad hoc,
10:07 Speaker 02 that feels so repetitive, you're like, this is so repetitive, you know, I should be able to automate this with a Python script. Then you go and try to automate it and you find, oh, this is actually way too complicated to automate. So you keep doing it yourself. You know, that's that's the kind of stuff that AI can be really helpful with. And you know, that that's kind of how SEEK. You know, that that's the role we aim to play in the business.
10:34 Speaker 01 I definitely recommend everyone go to seek.ai and check out the both the video and the screenshots that you have. I think it gives a really good picture. I think one of those things where, uh, you know, using natural language, people get to be very lazy and imprecise about what they define. And and you have a, I I wanna say like a novel solution to to do that. Maybe you want to explain like how you can get people to use natural language and be precise about data.
10:58 Speaker 02 Yeah, so what you're touching on is a very interesting challenge in our space. Where we're we are using AI to automate the querying of data to be able to get answers. You know, from the natural language interface. And so it's a very precise kind of problem to solve, right? Like, you can't just generate code that, you know, kinda to go back to my example about improv. Like, you know, it it sounds right, you know, you look at it and you're like, that looks right. But then you run it and it doesn't run or it gives bad data. That's a big challenge in our space. Very early on when I was building Seek, I was just thinking about, you know, how do we get around that challenge?
11:43 Speaker 02 And so if you go on our website and watch the video, it talks a little bit about the Seek workflow. And kinda just how it works, how it enables collaboration between the business users and the data team. Part of that is putting guardrails around some of this, you know, code that's being generated. But I'll also say Seek is doing a lot more than just generating code with generative AI. It's a very challenging space to be in, just, you know, the data space. Why is that? It's because data is really complicated. You know, companies spend tens of millions, maybe hundreds of millions a year on the modern data stack. You know, just putting data into the data warehouse.
12:27 Speaker 02 And extracting ROI from it is very challenging. You know, and that's why we have data teams. It's literally fleets of PhDs and master's degrees, you know, very smart people querying the data. Doing any sort of automation in this space, it it requires a lot more than just like, oh, GPT three can write code, you know, let's. Let's go into Bubble and build an app and, you know, and call the API. It's a lot more high barriers to entry than that, I would say.
12:58 Speaker 01 Yeah, I actually just noticed it's patent pending. You you apply for a patent. That's that's so cool. Is that your first patent?
13:03 Speaker 02 Yeah, thank you. It's my first patent, but others at SEEK have published patents before.
13:10 Speaker 01 Yeah, yeah, I I did a patent application once and it it was more extensive than I had had hoped. And there's sort of like two stages to the patent process. And I think we never got through to the to the ultimate one. But it it was it was such a long process. It's interesting how to defend IP.
13:26 Speaker 02 Yeah, and that's the really interesting thing about just being in a startup. You know, you learn about all these different areas, like I've learned so much about IP, law, trademark law. You learn a lot about legal in general, how to incorporate a company. And I'm I'm a lifelong learner, so it's something I've really enjoyed. It is just getting the opportunity to to be become more well rounded, learning about just all these different aspects of the business.
13:53 Speaker 01 Yeah, yeah, absolutely. So so basically, you know, so something that I I liked quite a bit is how you essentially have like preset definitions. We talked a little bit before the show about how the data world also has this concept of the semantic layer. How do you think about like helping people ask the right question? Like, I I think in in your in your interface, like you you essentially have ways to suggest questions and help people get to where they want to go.
14:20 Speaker 02 You know, the semantic layer is a really important kind of topic of conversation right now for people working with data. You need somewhere to organize all of your SQL code that you're you're using to get different metrics or, you know, do do different things for the business. And without DBT, there really isn't a good place to kind of organize all of that. So that's one big kind of utility of that to me. And then where it gets into the semantic layer is, you know, how do you map that to natural language metrics? Where that gets really important is, once you have all these natural language mappings
15:00 Speaker 02 between metrics and, you know, the typically SQL or Python code that that's being used for the data, that's when the code can just get really well organized and you have all these really nice building blocks that you can use to calculate more metrics. You know, I think in the long run, I think it's what a lot of people in this space are expecting to happen is that, you know, semantic layer is going to get more and more fine-grained and just more mature within businesses. And then, you know, that'll enable business users to have better access to these natural language kind of metrics.
15:39 Alessio Fanelli? And do you help at all new organizations coming to the platform into the same vertical to maybe kickstart some of this understanding? Is there anything in the product that says, or that you plan on building that says, hey, usually people in your industry, like, these are the main questions that they ask and like, this is a good way to formulate the question, kind of helping with the prompt engineering in a way?
16:02 Speaker 02 Yeah, absolutely. No, it does depend on the the business, but, you know, in general, our goal is to make it as easy as possible to ask questions and get the data you need faster. You know, that's our mission. So if you're asking a question and it's already been asked before in your business, there should be a really easy way to just, you know, fetch that question and be able to see the data. So that sort of retrieval is just essential for any business.
16:34 Alessio Fanelli? Say we're 10 years in the future, you're the hottest company in data, everybody's using SEEK. What do you see the future of dashboards versus this Haddock interface? You know, should data teams now focus on just building more comprehensive dashboards for specific reporting and like leave all the Haddock work to SEEK? I'm curious to to hear how you think about it.
16:57 Speaker 02 Yeah, I mean, I think it's a really good. Kind of vision, the last part of what you said, leaving the ad hoc work to Seek. I mean, unless the ad hoc work requires a lot of kind of new idea generation. If it's really just ad hoc work that needs to be done, but you're not really generating any new ideas. And you're using a precious resource, which is the data team and their time. That would be better allocated to some deeper research. Then yes, that would be a perfect example for Seek. What I think is going to be the hardest thing to automate is things that require new ideas. I think this is true for AI in general.
17:43 Speaker 02 When we think about artwork generation, for example, or poetry generation, or whatever it is. You know, the AI isn't, it's not really built to do something new. It's predicting the most likely next word. Or image generation, to be perfectly honest, I know less about. I know a lot more about, you know, text generation, but assuming that's kind of the same. You know, you're generating a most likely next image. Like, it's meant to be the most likely, so it's meant to be kind of like the average. Like, this is what I would expect on average as an output from the thing that I request.
18:26 Speaker 02 And so because of that, you know, we don't really see, I don't think, kind of AI-generated stuff in general. That doesn't remind us of something that a human already did, you know what I mean? Uh, it's always kind of reminiscent of something, you know, and that's also because of the training data. We're training it on lots of data generated by people. So what does that mean for new ideas? It just means, in my opinion, that these models aren't the best suited for idea generation. And that's something that I think people are gonna excel at for a very long time.
19:05 Speaker 02 And when that comes to building dashboards, or doing research, or doing deeper data analysis. Or it is that is a longer-term project. Those are the things that are gonna be the toughest to automate. And I think that's why people go into this field. You know, just speaking for myself, kinda what I said earlier, I wanted to become a quant. And and then, you know, when I was a quant, I transitioned to becoming a data scientist. And part of what excited me about data science was the ability to just analyze just tons of data.
19:46 Speaker 02 You know, when I kinda made that pivot in my career, it was when big data was kind of a thing. Now I feel like we just take it for granted, like, I don't really hear people talk about big data anymore. It's all just big data at this point. Back then, this was around, you know, twenty fifteen or so. I just thought big data was so cool. It's so much more data than I could ever just look at on my own. And the ability to write code that could extract value that I could put into place in the financial markets and just kind of be smarter than the other traders, you know, and be able to win in the financial markets. That was just so exciting to me, and I think a lot of data people I talk with have their own variation of that.
20:30 Speaker 02 You know, they they wanna be just doing research, you know, finding really just cool insights in the data that they can show to the leaders of the business and say, hey, you know, here's what I found. Is there any way that we can deploy these findings to generate more revenue? Save money in a really big way? Like, that's kind of the dream, I think, for a lot of people that go into this career path.
20:59 Speaker 01 Yeah, it uh it can be as well. I'm curious about how you pick models. Do you use fine-tune models? Do you just use OpenAI RAW? Probably not. Do you want to specialize them eventually? Like, how do you think about what models you use under the hood?
21:14 Speaker 02 So Seek's philosophy is just use whatever produces the best results. So as a result of that, we use a combination of third-party software and for that is in-house. It also depends on the customer. We have some customers that tell us they really just want to work with, you know, certain vendors and if we can provide that as part of our platform. So we do have some partnerships with third-party vendors. But long story short, it's really to me all about just what works the best. At the end of the day, the performance is the key.
21:53 Speaker 01 Well, how do you measure performance then?
21:55 Speaker 02 Well, it's customer happiness, you know, we're not just limiting ourselves to something quantitative. You know, we're still a startup at the end of the day, like we're almost a year and a half old. And we're, we're still kind of doing, you know, the Y combinator thing where we're building product. We're talking to customers. We're trying to learn about what is it that's going to make them happy. And so what we're not doing is we're not just like in a black box, trying to just crush a benchmark. We're building a product that people want to use and want to buy. And so having a great user experience is, you know, really important to us.
22:37 Speaker 01 Yeah, I think we can all agree that is the ultimate north star. Would you ever sort of build your own models in-house instead of doing third-party stuff? Like I, I don't know what the calculus is around that, you know, like you do have to have, I guess, research talent in-house. You, you need to have a training budget.
22:53 Speaker 02 Yeah. And again, it just goes to what's going to work the best for your business. We actually don't just use one model anyway. Building a AI workflow for people to work with data requires more than one model. And so when I say, you know, we work with multiple types of models, and that's part of the reason is we have our own proprietary architecture that consists of more than just one model. And so because of that, you know, that's when you can benefit especially from using in-house models. Could be forked from GitHub or you're using Hugging Face, you know, whatever it is.
23:35 Speaker 02 But also if there are APIs, then you can use those as well.
23:39 Alessio Fanelli? Do you have any maybe mental models on like how to think about the user happiness and performance? So, you know, previously if you had to ask the data analyst, you would get a response in hours, maybe days. Now you can get a response pretty quickly. How do you balance like getting getting the response back as quickly as possible versus like? Maybe taking a little longer but giving a better answer like? Have you seen anything in terms of where the threshold is for the the user? Would they rather wait five minutes for like a pretty good answer? Like would they rather wait an hour for like a more in-depth report?
24:18 Speaker 02 Well, and this kind of gets to the high barriers to entry of our space. It's in general not a great practice to provide a software that just is is risking giving bad data or bad answers to business users. So when I was building Seek, you know, I I just made a decision. This product needs to be trusted to deliver accurate answers. Part of why I made that decision is when I was a a data scientist, we actually had self-serve tools at some of the places where I worked. And I would give those tools to, you know, the business users that I worked with
25:00 Speaker 02 in the hope that they would just use those tools instead of asking me to do manual work for them. And what I found is that they actually kind of misused the tools, not not on purpose. But, you know, it just kind of gets back to what I was saying as well about just working with data being complicated. You really need some training to be able to work with data effectively. And using a self-serve tool for the most part, it's really hard to get around that. And so it's really easy to make mistakes. And the stakes are so high using these self-serve tools. If you make a mistake and you don't know it, you could make a decision.
25:43 Speaker 02 And the decision actually could be a bad decision because it's based on bad data. My philosophy is build something that provides accurate results and don't cut corners around that.
25:57 Alessio Fanelli? No, that makes that makes a lot of sense. The last few years, I think all the VCs were writing about ML Ops, kind of the hottest space. I think now the new term is LLM Ops, I guess. How do you think about what's new and what's old? You know, obviously, you still have a lot of the same problems when you were building on your own models. Now you have a lot of new stuff, like how do you prompt them? How do you figure out how the same prompt works across different models? Do you have any insight into, you know, what are staying to build Zeke? And maybe what other founders out there should think about when they start to build on these foundation models?
26:32 Speaker 02 I've come across really cool projects in this so-called LLM Ops space. I think Langchain is a pretty cool idea, and it just makes a lot of sense to me. You know, when I was starting to build with some of these models, I would have to just, you know, it's like, great. You know, you have an API call that you can make, but what about all of the handling that goes around that? Say, you know, say that you're just using a sentence completion and, you know, you ask, "Write me a sentence about the American Revolution," or whatever it is.
27:12 Speaker 02 You know, it'll it'll write a paragraph for you, for example, but then it might do a new line. And then it might say, "Write me a paragraph about about America," or something. Then it might just write another paragraph. So it sometimes it just picks up on these patterns if that if the example I'm describing makes sense. But you might want to just cut it off, you know, at that new line before it repeats itself. So like, that's an example of something that I used to have to do in the early days of building Zeke. I used to just have to write code to handle all that kind of stuff. But now there are great projects like Langchain. I mean, I don't know if it solves that exact example, but that's the, you know,
27:57 Speaker 02 that's the vision, I think, for that project is just building this toolkit that people can use to more easily have building blocks to, you know, work with these models and and build stuff quicker. I think that's a really cool project and. You know, I I think when it comes to L L M ops, that is some of the most novel kind of projects that I've seen.
28:22 Alessio Fanelli? Um hum. Yeah, you know, we're recording this two days after the chat G P T A P I was released and the the new turbo models are out. I think there was the first day a lot of excitement and then the second day, a lot of people saying that some of the prompts they were using before were performing much worse on the on the new A P I. What have you seen there at S C I C? Like, how are you thinking about building infrastructure that helps you do this kind of tracking and similar for your customers like? As the model change, i- is there any way for them to see how that impacts it? Or is that not a problem that you think is top of mind for people today?
28:58 Speaker 02 Yeah, we definitely get a lot of customer feedback to provide analytics into how productive is C I C helping them be? You know, what is the R O I on C I C and provide analytics around that? Which is kind of funny because we can actually integrate those analytics with C I C. So it's kind of this meta project for us. We definitely are, you know, tracking these kinds of metrics for customers. I guess to your other question about are we building things ourselves to, you know, kind of handle these model inputs and outputs?
29:37 Speaker 02 Again, we are a startup and so we're not really trying to reinvent the wheel outside of our core competency, which is the natural language product for our customers. So anything that falls outside of that, we work with a lot of different vendors and try to just build on top of what what's already out there as much as we can within certain constraints versus just build everything ourselves.
30:05 Speaker 01 Yeah, just a quick follow up on, I think the the primary interface being chat and the effectiveness of chat as we've seen with chat G P T. And you guys having a Slack bot, do you think there's a lot more potential for? Let's just let's just call it conversing with your data, like, you know, like actually literally conversing with your data. What needs to happen in order for that to be a reality?
30:28 Speaker 02 Well, I think what you're talking about is C I C.
30:32 Speaker 02 You know, it like we are a, you know, when I say a natural language interface, there is a conversational component of that. That's a really important component versus just a pure like search bar. So that that's already underway, you know, that's kind of already the product vision. So um.
30:53 Speaker 01 And you've been doing it for uh one and a half years. Shall we jump into the lightning round stuff?
30:59 Alessio Fanelli? Yeah, we only got five short questions, Sarah. Uh you can give us a a quick answer. Uh so the first one is, what is your favorite AI product that is not Seek, obviously?
31:12 Speaker 02 I really like Stable Diffusion.
31:15 Speaker 01 Do you have any favorite uh text prompts that you like to throw in there?
31:18 Speaker 02 My background on Zoom, if you, I think you may have met with me on Zoom. At some point, it's a spaceship. So I, the prompt was AI spaceship office. So it's in, it's a office as well. It even has a a gaming chair.
31:33 Speaker 01 No tricks like 8K render, Octane, uh Dreamweaver.
31:37 Speaker 02 No, it was a very simple prompt, but it did a it did a great job just on its own.
31:43 Speaker 01 Yeah, yes, uh only getting better. I I I've heard very good things about Stable Diffusion 3. How about favorite AI people and communities that you like to follow or shout out?
31:52 Speaker 02 That's a good question. I was actually on the Aluether AI Discord. I I've been on that Discord for a long time, you know, just lurking. Like, I don't know if I've ever posted anything. But but I've actually been following that project for a really long time. I thought it was really cool how it started. It seemed like kind of a grassroots kind of project. So I think that's a really cool project. Like I said, Langchain, you know, that that seems like just very well thought out project. GPT Index, I think that's a pretty cool project. Uh so you know, tho- those are a few that I can think of.
32:31 Alessio Fanelli? You know, we almost had, I think now, half of our guests mentioned in Luter AI, so that's a one.
32:36 Speaker 01 Yeah, we're gonna have something on. They're they're official foundation now, so it's it's becoming more official than just a bunch of people on Discord.
32:45 Alessio Fanelli? Sarah, a year from now, what do you think people will be the most surprised by in AI?
32:51 Speaker 02 I mean, I think it's gonna move really fast for the next year. There are just a lot of people building in this space right now. So everything we're seeing right now is just the beginning. So a year from now, I think there's just gonna be some really cool projects that go viral that we have no idea today what they're gonna be. And that's what I'm really excited about. So I I think that's that's kind of a cop out, but I mean, that's I I guess all I'm just saying is people will be surprised by new projects, but do I know what they're gonna be? I mean, hopefully, Seek will be one of them with with the evolution of our product.
33:30 Speaker 01 Yeah, it's hard to predict the future. Requests for startups. What's an AI thing you would pay for if someone built it? Personal or work?
33:37 Speaker 02 Well, I can think of something right now. It's just kind of a boring answer. You know, definitely whoever's listening, please send me an email if you have something that you want me to try. But I haven't really found anything I'm not happy with when it comes to things like, you know, scheduling meetings, managing my calendar, automating emails to to people. Like, just those basic productivity things. I haven't really found anything that I really love using. So, you know, when somebody gets that right, that's just gonna make me really happy, because it it's gonna make my life a lot easier.
34:14 Speaker 01 Yeah, we had a hackathon submission actually for that. I I need to go look up the form for the name, because the name is takes me.
34:21 Alessio Fanelli? Ipsso.ai, I P S O dot A I.
34:24 Speaker 01 Yeah, it's a it's a co-founder pair. They've been working on it for a while, and it's pretty impressive. They they handle multiple calendars of multiple people as well.
34:31 Speaker 02 Cool, I'll check that out. Ipsso, I P S O dot A I.
34:36 Alessio Fanelli? That's it. Shout out the latent space hackathon. That's what we make them for. Exactly, exactly. Sarah, last question before we let you go back to building a great company. If there's one thing that you want everyone to take away about AI and its impact, like, what would that be?
34:52 Speaker 02 Well, I would say that in general, something I've learned about just people and humanity when it comes to AI is like, just look at all the content that's being created about AI. A lot of it is very fearful and dystopian. I would just say there are definitely pros and cons of AI. You know, there's a lot of cons, a lot of things that can go wrong. But there's also a lot of research being done to protect against all of the downsides. Versus, there's so much upside as well. There's so much that we all do today that we don't even know is just so, you know, mundane, taking up so much time
35:36 Speaker 02 and taking time away from the things that, you know, we could be doing that bring us so much more joy. And I think there's just a lot of upside to automating things we may not even realize can be automated today. And so I would just say, you know, there are definitely pros and cons, but personally, I see a lot more upside than downside. And I think we're already doing a really great job just protecting against the downside. So I think if anybody's, you know, wants something to take away, I would just say, it's always good to just learn more about new technologies in general. And, you know, the, the pros and cons, you know, of what we can expect in the future.
36:17 Speaker 02 But personally, I'm very excited about all of these innovations in AI. And I'm just so excited to be working on SEEK because I just can't think of a better project to be working on.
36:30 Speaker 02 So that's my two cents.
36:32 Alessio Fanelli? Well, Sarah, thank you so much for coming on Seek.ai. Everybody should go check it out. Anything else you want to shout out, Sarah? Like your Twitter, like where should people find you?
36:42 Speaker 02 Yeah, and by the way, thank you guys so much for having me on. This was really fun conversation. So you can check out Seek AI. It's just S E E K dot AI. Seek dot AI. Yeah, you can also follow me on Twitter. It's just at Sarah with an H R Nagy. And as in November A G Y. Uh, so that's, that's where you can find me. You know, you can also find me on LinkedIn. Uh, you can also email me, just Sarah with an H at Seek dot AI.
37:15 Alessio Fanelli? Awesome. Thank you so much, Sarah, and thank you, everyone, for listening.
37:20 Speaker 01 Thank you.
37:20 Speaker 02 This is great. Thank you, guys.
Previous 97% Cheaper, Faster, Better, Correct AI — with Varun Mohan of Codeium Next Emergency Pod: ChatGPT's App Store Moment (w/ OpenAI's Logan Kilpatrick, LindyAI's Florent Crivello and Nader Dabit) Page updated November 13, 2026 · transcript and summary are machine-generated