**The Challenges and Opportunities of Building with AI**

As I reflect on my experiences working with AI, I've come to realize that it's not just about developing clever algorithms and models, but also about understanding the human side of innovation. When I was asked to write down notes after a coaching session, I realized that this task was not about retaining information in memory, but rather being present and mindful of what was discussed. This experience taught me the value of being more present than writing notes.

However, when I've worked on building AI-powered solutions, I've encountered challenges that are both fascinating and frustrating. For example, during my time at Amazon, I worked on a project where we wanted to predict the right product to add to an accessory based on customer searches. What seemed like a simple problem turned out to be much more complex than initially thought.

Another challenge I faced was making the leap from solving a specific problem too quickly. When I was working at MatterMark as head of machine learning, we were tasked with creating a vector database that could perform semantic search across a vast amount of content. This project was exciting, but also proved to be more complicated than anticipated. Looking back, I realize that starting with defined problems and finding the people who need them solved is crucial. Once we understood the needs of our customers, we were able to automate those things and make significant progress.

However, I've also come to understand that automating away humans from a problem can be problematic. While AI can be incredibly powerful, it's not always perfect, and there are times when it may not produce the desired results. For instance, AI-written SEO content is probably good enough for many uses, but it lacks the nuance and creativity of human-generated content.

**Lessons Learned**

As I continue to work with AI, I've learned several key lessons that I want to share. First and foremost, it's essential to start by defining problems and finding people who need them solved. This approach allows us to understand the needs of our customers and develop solutions that truly meet their requirements.

Another crucial lesson is that starting with a specific problem can sometimes limit the potential applications of an AI-powered solution. By focusing too much on the "what" rather than the "why," we may overlook opportunities for generalization and scalability. As I've reflected on my experiences, I realize that it's essential to take a step back and consider the broader implications of our work.

Finally, I've come to understand that AI is not a magic bullet that can solve all problems overnight. While it has the potential to automate many tasks, it's also important to recognize its limitations and the need for human oversight and review. By acknowledging these limitations, we can develop more effective strategies for working with AI and leveraging its power to augment our capabilities.

**The Future of AI**

As I look to the future, I'm excited about the potential of AI to continue advancing and improving our lives. While there are still challenges to be overcome, I believe that with careful consideration and a willingness to learn from our mistakes, we can harness the power of AI to achieve great things.

In particular, I think that finding people who have specific problems or needs is crucial. By working together with customers and stakeholders, we can identify areas where AI can make a real impact and develop solutions that truly meet their requirements. As I've reflected on my experiences, I realize that the key to success lies in being open-minded, adaptable, and willing to learn from others.

**Conclusion**

In conclusion, building with AI is a complex and multifaceted challenge that requires careful consideration of both technical and human factors. While there are challenges to be overcome, I believe that with the right approach and mindset, we can harness the power of AI to achieve great things. By defining problems, finding people who need them solved, and acknowledging the limitations of AI, we can develop more effective strategies for working with this technology and leveraging its potential to drive innovation and progress.

"WEBVTTKind: captionsLanguage: enSamir it's great to have you here I'm really excited to talk to you you are the CEO and founder of Heyday you're also a participant in our deep gram startup program Heyday is doing some pretty rocking stuff I mean let me just break down the short blurb about what you all are doing and then you can get into it a little bit more because I would love to hear your story and how you created it but for those who do not know Heyday is an AI thought partner that is for knowledge workers and you're primarily working right now with executive coaches so that being said that was a bit of a mouthful I want to know though before we learn more about Heyday and how you came up with the idea what's your story yeah uh I've got kind of I mean probably most s of Founders have like a weird story so but like typically weird story uh my dad was a diplomat I was born in Bangladesh so I moved around a lot growing up um so like lived in the US when I was super young but then moved to uh Saudi Arabia back to Bangladesh to England to Iran Bangladesh again and then my dad very unexpectedly died when I was 14 and my mom decided to move the whole family to uh New York city so uh weirdly i' consider you know have mov moving around a lot in my life like New York is like the place that I consider home uh so um I don't know how I have like an very clean American accent CU I learned English all over the place at some point I had a British accent I don't know what happened to it it comes out when I get drunk uh so that's like the start of that I actually studied like electrical engineering Wireless signal processing with a side like um side in machine learning and like at the time like this is like 14 years ago so like B like you know what machine learning was back then you know kind of started getting into neural networks started my career off at Amazon was doing some like very early machine learning stuff with them like doing recommendations on accessories uh I I was working on the retail side for a little bit where like you know what HDMI cable should you buy with this TV that that was like the life fulfilling work amazing work it's crazy though so every TV would have like a set of recommended products before we used any machine learning for it it would be a human picking out for every TV like there's like a team of humans picking that stuff out so that's that's how weird it was back back then and and like I had to like I was like a new employee and I had to like convince like multiple levels of people that like why are we why are we doing this with a team this is is like uh so uh and then I eventually worked at AWS for a little bit but honestly got realized that I really didn't like working at big companies like multiple times I learned a lot from ad bu but multiple times like bureaucracy getting the way of doing things even at a company like Amazon was just like you know what screw it I just sort to work at smaller companies so uh I was up in Seattle I um worked at another company doing like Consulting work with uh products like you know the Nike fuel band Disney's like Magic Band there was like a bunch of stuff it was around like using small ml algorithms on top in like in devices so like to prct steps and stuff like that super interesting work but then realized like I loved actually being part of the product development process like being a client like being like a sole like development shop it was like Nike or whoever would just tell us we want to build this this is what needs to happen and like so I moved down to work at a small startup uh and I've been in the Bay Area for like nine years now uh started another company before Heyday that I shut down so I'm technically a second time founder even though it feels like it's just as hard anyway maybe a little bit easier but uh yeah um primarily my machine learning work has been in text um and natural language processing and so I've been basically building that kind of stuff for for years you know basically my whole career uh I did do some like you know signal processing based prediction algorithms but that was like like not that much uh but yeah I've been working on like search and natural language processing and text for a long time yeah so you recognized the love for developing products and being on the product side of things and understanding what the user would potentially want and that got you into the startup game how long ago did you start Heyday and what was the motivation so he was two years ago and that was like right after like within a couple months of shutting down a previous company cuz you know I think there's a starup Founders are crazy psychotic people I'd say and I'm one of them uh instead of taking a break just being like well let's I got a another thing why don't we just jump into it um so that's the that's how Heyday started uh and hati started with a lot of the like a mix of like the pre the first company jour we'd started as like hey what if you could have like a Google like search for all your own stuff so it would integrate with sounds very familiar these days yeah yeah and like well I I feel like we have a lot of learnings from that so well I can share that it could either just be execution failure or actually learning that people think it's cool but then won't pay for it uh is uh we integrated with like all the things like Google Google Docs Gmail Dropbox box at the at the time slack and then it would be one place for you to search for everything and it's we got to like 20,000 users in a premium way but very few of them were like actually converting to paying users couldn't figure out the business model and like uh we we still had like an incredible search engine we basically built our own at the time built our own Vector DB because that's there was no Vector DB so uh it would have been nicer to not have to spend like six months building our own Vector database uh and just use pine con or what the hell exists now so many things you've got quadrant you've got weeva you or we could have made that into the business also no one would have wanted that at that point so we were too early for a lot of things I would say um but also I think we learned that like we were building a really General product without a lot of um direct like magical workflows and honestly even that we didn't quite carry over into learning like applying that learning directly with Heyday with Heyday one of the few one of the first things was we got to be like in like we got to be solving a real use case we like I I'm still I'm very passionate about like leveraging information knowledge I'm I'm slightly ADHD uh and also care about like automating away like anything that humans don't need to work on I'm very passionate about like building tools that help like allow people to focus on the things that they're either excited about creating that they're they have like a unique human potential to create and so you know all the things that I've worked on are kind of related to that um probably why I'm excited about machine learning and AI in general but uh so with Heyday we wanted to do the same kind of thing but ALS but be even more AI forward not just like search um so could we figure out ways to be like assistive with without people prompting us you know prompt is a word that's uh overloaded these days but like you know uh I don't mean just like prompt in like the open AI type sense or whatever language model sense I just mean like when I the ideal AI for me like consense when I need something and just like provides it um and we weren't doing that people would have to basically change behavior and come to journal to do that so from the get-go hey Dave you know we we started by hey could we serve like alternate answers to things alongside Google search results so it would be like oh hey like you're looking for some uh way to you know uh like tips on fundraising like creating a pitch deck and turns out you read like three articles about that before or you even took notes about it and it was like years ago and like our goal would be like surface that right next to the Google search result and be like oh this is like stuff from your memory from your previous research um and but like we all and then we that started doing better it was paid from the get-go we had like four or 500 paying users but uh we then got stuck with another thing where it's like it's still too General there weren't enough people like there were people who were enjoying it but no one was like oh my God this has changed my complete yeah this is this has changed my life and so our hypothesis was we really need to like build so we wanted to focus on a like single Persona and we we kind of looked into who was already getting value from Heyday there's like a cluster of like journalists marketers investors startup Founders and executive coaches for some reason so we like kind of like dug in and evaluated across like a different bunch of different you know how How likely are we to help them how much do they care about this problem how good are they as like early startup customers like can they journalists are like on the opposite end of the spectrum like uh because they don't really make buying decisions themselves so terrible at buying tools uh So based on a lot of that we we went with executive coaches and have been working with them for like the last six months and launched a product that's more F Well the product's been launched for long but like the features that are focused around coaches we launched that like a month or I guess two months ago now and um that's been going really really well but we're still uh in the phase of figuring out how to get it in front of more coaches so can you tell me more about this idea around magical workflows that you said you learned back at journal and you didn't quite put it into production or understand or internalize that learning when you started Heyday but it feels like now you have yeah it's actually it's just it's one of those things that I'm like how have I been a startup found her for like 5 six years but didn't do this where we spent like three basically 3 to four months uh you know we're we're building AI ml products so like we were like we set up a process we like okay we're going to work with 10 to 15 coaches who signed an NDA with us and our goal is to like learn what we can help them with and then be a human uh service that does the work for them while while we're doing that our job is to automate the work that we're doing using AI so instead of a human doing it eventually it becomes like the AI starts doing it and then we have like we have like a tuning process so we basically like learned on our own work um and learned in the process what are the problems that are most uh kind of impactful to solve for coaches uh it it also lines up to the you know the the types of problems we we're solving are like very general in some ways so in Weir in weird ways it's going to be like easy to go to the next Persona we think like a lot of Consultants have the same problem coaches are basically like a small like small subset of the Consultants that we could go after soon after but this idea or at least our hypothesis is going Persona at a time we can like talk to real people and find out what their problems are and then balance this like we want to be a in the long run a thought partner an AI thought partner for everyone but but in the short term how can we build things that can help more people in the long run but in focused on this this group of people so it's it's still like a concrete problem that we're helping someone solve but with an eye to well let's not do like really coach specific things like that only coaches need but let's find the like General problems that they need help with that we can automate away so talk to me about how the product has evolved from the beginning and then becoming coaches you said that in the beginning it was something like oh yeah you've actually researched this before in the past and you're Googling it so you probably want these relevant links or this relevant Google doc that you have something like that how did it evolve since then yeah so a lot of the like the product surface area in the beginning was simply we're indexing all your public browsing um so it's browser extension that indexes all the readings that you're doing videos that you're watching um it also um integrates with your email your Google calendar slack notion all that stuff places where important content exists and then the main way is to like surface that while you're doing your normal browsing so when you're searching in Google show like show up with other helpful search results when you're looking at a new like a an article can we like show you other potentially related articles maybe we know information about the author like you've read other things by that author can we show you that can we show you like uh hey have you have has someone mentioned this to you in some other channel like in slack or Gmail so we can show like this was shared to You By Demetrios two weeks ago and you just forgot about so that was like the initial kind of state of the product but um over time we evolved that to you know do more with AI so we allowed people to like organize things automatically into topics so they could tell us like hey I'm interested in machine learning as I'm reading new things about machine learning I can kind of build like an Ever growing knowledge Base by saying every day you we suggest like hey here's like the here's like five things you read about machine learning yesterday do you want to like save them away into your machine learning topic and so it makes it easier for you to like keep your topics growing but now now we wanted to use that backend to serve coaches and the way we've been helping them is two two big ways one how do we make it so that they can have they can focus on their their superpower which is uh having insightful conversations with their clients and not do all the other stuff around it so they they usually have to take a lot of notes they have to like prepare ahead of the session and then they have to review their notes to like be like well what were what was that client working on and they usually have like 10 to 15 clients it's hard to like keep all that in in memory so they're they usually like juggling a bunch of stuff they like frequently recommend like documents or articles or things for the client to read um and track like action items and things like that for that client and so we were we just built an AI that can both understand all the conversations that someone's having with their client because we have an email integration or slack or like the documents that are being being shared between that client and we added a zoom like integration uh which as part of that pipeline that's where we use deep gram where we um transcribe the audio from recordings of a client session and then we extract the most meaningful insights from it with the context of the client overall so previous Rec sessions like all the notes that the cl the coach may have taken the like um email threads between so like we understand like hey here are the things that are important to this client so like you know if you have like a 20-minute conversation about vacation to start the conversation like that's not the most important thing to pull out as an Insight right so that's been like super well received by coaches because they now can go use Heyday to instead not take notes during sessions be really present with their client and then not spend only like 5 minutes correcting the auto generated notes from Heyday and sending it off to their clients but then it becomes like part of their system that they get prompted to review right before their next session by Heyday it's like hay is like oh hey here's like the stuff from the last session here's like General themes this is probably all you need to like review to be preps for your next session so that's been super helpful and then the other piece is coaches are like generating lots of useful insights they're having these thought-provoking conversations but they also have to they're like solo entrepreneurs they have to grow their brand and their business to get more clients or they want to turn that all those insightful things that they were talking about into like writing even if they don't want to grow their brand maybe they want to create like a framework and share it with their customers and so we've also built an AI like content writing assistant which on this is like very general I would say to lots of people because uh we help essentially be like an outliner plus first draft writer using only the content either you've written yourself or you've you know from your own conversations over email or Zoom or whatever or you know if you're if you want to like Source like information from other folks if you've been doing a lot of research about llms but you haven't written much you can also because of the browser extension we automatically know everything you've read about llms you can say no actually I want to use all the readings that I've been doing on my research about LMS too and so it's like not just chat GPT style being drawn out it's our goal is to help people share original authentic insights but for folks who are not like natural writers help them get a first draft out but only from the content that is actually their own yeah so it goes much deeper and it's not just that surface level I can see that and I know just in the times that I've had coaches Interac with me and I've had coaching sessions right one of the biggest things that you do is you write down a lot of these learnings me as the person that is not the coach the coach tells me to go write this stuff down and then they come back the week or two weeks later and they say so all right were you able to do that show me what what you learned from it but that is exactly like what you're saying they have to keep it in memory and they have to know review their notes like oh yeah we talked about this and maybe they're not picking up everything because they're being more present than writing notes and so I see the value in this I love it one last question for you before we jump what have been some challenges when building with AI I mean I've been working on this for forever across many products and so since Amazon days so when you have like very specific problems right like you're predicting in all you want to do is predict the right product to add to the accessory and like it's very simple that's that's like easy AI stuff you know back then it was a little harder but even then it was easy to beat humans at that the next big thing I'd worked on was I I was working at a company called matter Mark as head of machine learning and um we we were a private company database and people wanted to search for you know companies by like descriptions our customers were like VCS and salespeople so if someone searched like hey like I want to find all the U ride sharing companies in in s in the Bay Area or in California or or or all the gig economy companies in yeah in the Bay Area we wanted to be able to like at search time do kind of what is done now the the semantic search we you know we we had a vector database that was like the inspiration for like wait there's a better way to do search we can like do embeddings we can do like similarity and we could do this across like a lot of content so that was like a directed problem we were solving we're helping salespeople and VCS descriptively find their the companies that they were looking for I made the leap to make it super General too quickly and I think I've struggled for years based on that and so when you see things like rewind which is really cool but like again I think they're just like e maybe they're going to execute much better than we did but my problem with it is it's like a it's a hammer looking for a nail even though it's like it's cool that you can index all this stuff but what am I going to use it for like very few companies figure that out like Google Google's probably like one of the only ones that's like all it is is a search box for the entire world and like people will figure out how to use it uh a lot of times I think that like starting with like defined problems and then finding who the people who really need it solved building something for them and then realizing oh it could be generalized to a bunch of other things like let's do it again let's like roll forward and figure out like what are the actual needs um and then automating those things and the other thing is like don't I I think like uh anytime you try to automate away humans from something it's it's going to seem really cool at first but generally speaking it's going to be like 80% good enough and then uh and that's fine for a lot of things right like I might say like all the SEO content that's getting that gets put out was all garbage in the first place and so AI written SEO content is probably good enough but like yeah you don't really want like chat GPT spat out like content about how to build startups in AI right like it should me and you be replaced by chat gbt just like splatting out stuff no not yet at least yeah words of wisdom Samir I appreciate you coming on here and getting to talk to us about what you are doing what you are building and I'm going to take a lot of this to heart because I feel like it resonates a ton with me right now this idea especially that you just mentioned on finding people who have that problem first defining problem and then finding people who resonate with that problem is huge for me so you could you could go the other way too like you could easily just go talk to people and then find the problem by talking to them like like I I guess there's like a highlevel thing like you can't just like show up with like nothing like but you know we we showed up with hey we're we're going to be an AI knowledge base for everything for you what do you what would you use that to like automate that was like a pretty general question and uh we learned a lot of cool stuff from that that we had no idea about well this has been great dude thank you so much for coming on here of course man thank you for having me\n"