Building a Deep Tech Startup in NLP with Nasrin Mostafazadeh - #539

[Music] all right everyone i am here with nasreen mostafade nasreen is the co-founder of vernique naserine welcome back to the twin wall ai podcast oh thanks so much for for having me again sam uh for those that recognize the name nasarin is a long time friend of the show and this is your third interview maybe um the last time we caught up was in january of 2020 before so much in the world changed and we were talking about trends and natural language processing and today we'll of course be talking a little bit about that but we'll be focusing on some big changes in your world since uh we last spoke and starting with the company that you founded so maybe uh catch us up a little bit on on big changes in your world since the the last time absolutely well it's so interesting that the last time that i talked with you was january 2020 it was precisely third of january 2020 and the reason i remember this is that you may recall it was literally the day after us basically went on the brink of having a war with iran so we were all extremely stressed that was the night i literally didn't sleep we were all like checking our phones trying to see whether or not there would be a literally a var and it was just just a very bizarre time so that's kind of how 2020 started for for a bunch of us not even knowing where it's headed so even 3rd of january was a very very weird year um so anyways since then that that very event actually played a major role in how the rest of my uh life went on um from that point on so basically um diverse so many other things that happened like on 8th of january there was this plane that was down with like 176 uh people on it a lot of them iranians some of whom i literally went to school with you know who could have been literally me so after so much thinking i always you know i've been working in the startup scene for the last five years or so and i truly believe in wanting to make impact through the power of startups and through being able to laser focus at a particular fundamental problem uh in real world so anyways i always wanted to to do this and this whole series of events in around like january and february 2020 made me kind of think how blessed i am as an individual and the kind of opportunities that i have and i'm not necessarily leveraging uh with all the people that didn't even get to live to want to fulfill their dreams like it felt like it's time so i should just just do it and go for it um so anyways the process of starting learning was so many you know years of in progress basically but it was always about waiting for the right time for the right time um and then in february 2020 i was like that's it i'm gonna start it and hilariously the first day basically official first day of vernick is march 1st 2020 not knowing what we are really up for so we had all these omid my co-founder and ivy had old and in new york city exactly yeah we had all these grand plans of starting this deep tech ai company in new york city with like you know a bunch of plans as to when we would go fundraising maybe we'll hire our team and of course mid march 2020 the uh you know shutdowns started and then we in april so we realized that okay we are in the middle of a global pandemic um so the rest is kind of history um it's the timing was impeccable to say the least um so we basically kind of put everything on hold for a couple of months just to see where the world is going like it wasn't necessarily clear that anyone wanted to uh to even think about a brand new ai company um while everyone was basically licking their wounds um so yeah we just waited a couple of months and then uh around like late 2020 we were like okay let's just go for it so we started uh fundraising and in about a month or so we could like close our round and then we started hiring and now we are based in new york city we are in flat iron uh specifically it's like the dream neighborhood we're working in a physical office so this is our space and it's been just a really really great uh time to to want to do this honestly despite all the hurdles how hard it was to go through the pandemic to start a pandemic company basically i think it has helped us build all sorts of muscles that we never thought we can and hence we are really in a very um good shape in terms of all the other hurdles that we have to overcome moving forward that's awesome that's awesome congratulations on that um so the company is as yet stealth and not much is known publicly about what you're up to but you've promised that you're going to share a little bit of uh pull back the kimono a bit so to speak um uh what's the company up to what are the the challenges that you're hoping to take on absolutely so um our mission is to enable anyone to make data informed decisions be for their personal matters or for their businesses without having any kind of technical background and what that means is that we are basically innovating in the human machine interfaces area where we want to replace hard-to-grasp things such as programming languages with other intuitive modalities of interaction including of course human natural language which you know given our backgrounds is abundantly clear but we are not just you know biased thinking that natural language is the answer to anything and everything in terms of intuitiveness we are thinking uh actively and working actively towards other like modalities of interaction that would be really easy to use and intuitive in a given context so uh basically this is like two degrees how much we usually share publicly but i love talking with you so i would love to start sharing a little bit more in terms of our vision for um for vernique and what kind of a technology we are building so very deep tech company meaning that we are really um we are overcoming various like scientific and engineering challenges that would ultimately enable us to build such a platform that i was talking for data informed decision making so what does data inform decision making means it means that there's all sorts of data out there they can come in any shape or form they can basically have any format they could be any small any large they could basically be anything and everything and when you want to put a interface on top of it and enable anyone to use it like let's say get analytical insights about it or like just ask any kind of questions that they can they have that can help them uh better their lives or their better their businesses that means that you have to have interoperability and that means that you have to be domain general we are not trying to build like something that will be just working in one sector we don't want to build something that is going to be just in a particular domain operational we want to build a truly domain general ai platform that can have interoperability across all sorts of data so it's a pretty grand mission in terms of the technology that it needs and we are actively working on all sorts of problems that in the academic sense has been uh you know an issue for the past couple of years as well so we are working on um basically been building first and foremost intuitive and easy to use interfaces including natural language building a domain general natural language interface in and of itself is very very challenging but mainly we are trying to make it controllable we are trying to make sure that we assign provenance as to where the data source comes from and what kind of data we are basically putting in the loop on top of it we are trying to make sure that these basically interfaces that we put are controllable meaning that we can have interaction with them and we can teach them or help them forget something that they've learned that was wrong you're also trying to make sure that these models are actually able to get all sorts of feedback and make better decisions down the road for the sake of the user so basically they should be able to not just base their decisions that they help the user make on one source but like just bring all the bits and pieces together and basically do reasoning but so many other characters is that this this kind of a fundamental ai platform that you're building should have including the fact that it should be amenable to data privacy issues and so there's so many other things that we're actively working on but in terms of our lines of ai research we are of course working on conversational agents we are building basically dialogue systems that have these features that i was just outlining we are also working on um like all sorts of learning phenomena that are not data-hungry that actually can help us go from a domain to another with like the least amount of time and the least amount of supervision of course but we also have problems on the data side itself so building a platform that has interoperability across the board requires like innovations in distributed systems requires innovations in building like runtime runtime engines that can actually digest all sorts of data and understand it so uh it's not just ai research that is a problem that we are tackling but also you know other kind of computer science uh issues that we have to grapple with on a day-to-day basis and i will add that um we are very mindful of having it like design elements and design driven thinking to be the front and center of what we are building so that we don't kind of like repeat the mistakes that have been made in general in the tech scene of like bunch of technical people getting together thinking they have an intuitive solution and then turns out it's really not usable so that's that's our front and center we are really striving to build something that works for the masses as opposed to something that is just um you know for the sake of pushing the boundaries of ai how big is the team so far we were a total of eight people we had like six or so interns also that loved us um so vr is very small we're hiring across the board we used to be just hiring on for like research team and the engineering team but now we're also hiring on the business side so we are hoping to grow to 12 or so more people as soon as possible we wanted it yesterday but hiring is the toughest one of the toughest things yeah in my experience yeah i asked because the you've outlined a pretty broad set of uh areas not just that you want to build product in but that you need to do fundamental research in and that sounds like a ton for eight people oh my god how do you how do you um before we get to kind of how you take that on uh maybe another way of coming out that is like how do you think about the mvp for the product like when i think of you know the simplest thing that you know kind of the simplest version of the the vision you outlined i think of something like uh you know natural language query generator but those you know exist they're not without challenges like how do you think about the the mvp and how do you contrast with something like natural language query generation good question so i will just first and foremost say something as a uh in like in parentheses we call mvp legume vernique so legume is this swedish term i don't know if you're familiar with it that means just the right amount i just personally like we take issues over the night we take issues on um like minimum viable product often not being really minimum and not being really viable so it's very improving so anyways the question is was just the right amount for us right which is what we've been dealing with for the past like year now doing like r d until uh we get to a point that we're comfortable with taking our product out so it's a very co i would say complex question what the lago basically should be for our domain uh on the business side what we're doing is that we're very practical we are taking this technology one domain at a time we are like right now focusing on a particular domain for example that we are hoping to come out with in the next couple of months and in that particular domain then you have this uh you know flexibility of narrowing down the data that you're training on narrowing down basically the the capabilities as well right so for example right now we we know that we want our system to be instantaneous and it is so we have like a threshold that we know okay if it's responding after one second it's like definitely a downer but if it is uh like we can just keep pushing the boundary so there are such parameters that we are kind of tuning uh but anyways so in terms of the taking it one domain at a time uh we are doing that but we have this rule in-house that we are like if it is um basically um diverging more than 20 percent from the general domain general platform we are just working too much on one particular domain so basically we are making our legume or mvp to be an 80 of the same kind of code base same kind of technology that we would have and spending 20 of time in a particular domain without it generalizing out of it so it's it's been it's not easy i would say and i think i will have a more clear answer to that when we are out with our product i know what i said is a little vague as to what the mvp should be but there's so much to talk about in terms of how to curate the exact packaging of such a technology so that it's still meaningful you can get enough uh signal back from the market and from the actual users and then keep iterating and is the the example of some kind of natural language query is that directionally um you know accurate accurate enough to give us some to provide some concrete grounding for the conversation you're going domain at a time so i'm envisioning you pick a domain who knows what that domain is let's say contracts you know legal contracts right and so you know some lawyers want to do discovery and it's hard for them to find what they want and so you're essentially giving them a box to type in and you're doing smarter things with what they type in to get them to you know to turn that into a query that you can run against whatever systems they're trying to search against as an example but that's you know that's the unstructured text example you know there's the structure text example you're working with uh you know power companies and you're trying to help them manage their grids better or something and you give them a box to you know tell me the you know aggregate power across whatever whatever six regions uh i guess i'm you know the picture that is coming to mind is maybe uh you know ai or natural language interfaces for like business intelligence types of interests or problems yes that's that's very close that we can talk for hours and hours right about the state of the field like what are the relevant technologies how they're failing and why they're failing and why what you're doing is is really different but yes at the end of the day it's it resembles that so basically let me give you an example this is not what we are doing right now but one of the reasons we wanted to do what we are doing is that even as technologists it's so hard to make decisions based off of your data because uh it just requires so many bells and whistles so for example you just asked me right before this like what did you eat for breakfast right and i told you well i don't need breakfast so as an individual i do intermittent fasting i have this app that i'm recording um the number of hours that i'm not eating a day i've been just doing it habitually in the past couple of years then i have this other app where i'm recording what i'm eating every day and then just because i have some like underlying conditions that i have to be careful with whatever x y and z that i'm like putting in in my body basically and then i have this other after i'm like uh basically recording my weight i put i hop on a way to scale every morning personally so there's this app that is capturing the data from my health on a day-to-day basis on my weight on a day-to-day basis so anyways as an individual like myself although i'm a technical person it's so hard for me to get the answer to a question like i don't know what uh is there any correlation between my weight loss and the number of hours that i'm intermittent fasting right yeah and why is it hard because it requires me figuring out how to download the data from all of those apps and if you even can if you even can exactly and then maybe i don't know going to a night like opening up nope like a python notebook and trying to remember what is the correlation function of some sort and then trying to call it and then coming up with an answer right it just takes so much time and i am a technical person so imagine what the world looks like for other individuals and small businesses who have all these kinds of data and they are making wrong decisions on a day-to-day basis because of their lack of access to such easy to use interfaces so anyways we want to very much in line with what you said you're not like you know basically doing any of these things that i mentioned right now but in the long run for the company that's division we want it to be that we have this one interface that we can put on top of anything and everything and it can smartly navigate its way to find the right sources and then come back with the with their results but it definitely is basically like a natural language querying interface but as i mentioned we believe tremendous if you believe in the tremendous value that other modalities of interaction uh bring into the scene and that's something that has been definitely neglected but on top of it you know i've i've worked on natural language understanding basically my whole like life like not of an adult life like whatever 15 years right they're just no one knows how to build a domain general language interface like that we've had tremendous progress in the field in the past couple of years which is why likes of myself are motivated to want to finally take of you know this kind of technology to the market so that we can get you know clear signal as to the flaws right and the problems and the solutions and feed it back into the basically academic ai world we could go down the rabbit hole of uh wearables and personal health tech and all that kind of stuff and maybe that's a fourth conversation but for now i think that the thread that i want to pull a little bit on is the the state of ai research in the domains relevant to what you're trying to build because a lot of what you're i think your contention is is hey the you know the idea of a search box that you know lets you find things out right that's something we've been pursuing for a while but we've been failing and some of that is execution and um you know or different execution different vision um but some of it is that the research or the technology just fundamentally isn't there and so maybe a next place to explore is you know where you see the gaps in the technology and how you're using you know that assessment to kind of prioritize the way you're approaching research at veronique absolutely so i think this is very much actually in continuation of the conversation we had in january 2020 when you know i was reviewing basically the state of the field um if you you may recall so my net was that we've come a very long way in natural language understanding uh and national language processing in general in the past like now six years or so it's been tremendous how much progress we've made not on just down the stream tasks but also on real-world products basically that have been impacted by these kind of technologies but i characterized uh like the flaws in a few ways uh that i think some of which are now kind of being um addressed by the recent developments so like just to backtrack a little bit so it's kind of interesting so i started my personal work in natural language understanding because i was back in time this is back in high school i used to work in robotics then i switched to national language understanding and comments as reasoning in particular because i came across this motivating example that i saw in a random book that was like uh for an ai system for a machine to understand natural language it requires to put together lots of bits and pieces about the world that it is grounded in and hence it's the so-called ai complete problem so the motivating example was that uh the monkey ate the banana because it was hungry and is that it referring to the you know monkey or the banana basically was the the the riddle for the ai system so i literally literally started my life in a in sorry natural language understanding for for that very same problem for tackling that very same problem which i found fascinating and then fast forward these kind of problems were really not the focus for the field for a very long time until in 2015 and 16 or so that these um you know at the time it was like biolus teams that were suddenly starting to to to leave their mark on some comments it's reasoning and natural language understanding benchmarks so much so that it was touted as the solution but then of course with the transformers in 2017 or so everything changed for the better so much so that it just kept kept going right they had all these other models and for me personally always i was on the camp of questioning whether or not any of this is real progress as someone that cared about common sense reasoning which is sort of historically been this line of thinking about reasoning which supposedly conflicts with like um you know uh machine learning and stochastic uh and like look into to the data et cetera um so anyways for me it was very natural to want to be biased against the progress so much so that in 2016 basically the work that i personally did was on building at the story closed test benchmark which was basically trying to push these systems to showcase whether or not they have any sort of thomas's reasoning by continuing this like a short story in uh basically finishing the story right in a right way that is like reasonable and logical um so anyways we uh when i did that work the intention was to assess whether or not these models have common sense and i wasn't that uh convinced that they do i wasn't that convinced that they have any reasoning capabilities and then the transformers came of course gpt one did a great job on story close test and then we were like oh is it just picking up on the intricacies of the data is it just uh basically learning the biases that exist in the data data set and not doing true natural language understanding and then turned out it maybe was to a degree but we even changed the test set to another version of story close test that was kind of de-biased and then gpt-1 was basically the only model that could sustain its performance and anyways still i'm a skeptical let's say and i'm like okay i want to see deeper language understanding and deeper common sense reasoning and then gpt um three came out basically this this uh in 2020 that was uh doing zero shot performance on story class test meaning using none of the training data at all and it was getting to 80 something percent performance which if you had told me in grad school i would not have thought is possible right in 2020. so these this is tremendous progress right i think it's really hard for us to try to um sweep it under a rug that these models are not showing any fundamental language understanding and all they're doing is pattern recognition because one can even argue what it is that we do as understanding human beings and how much of it is recognizing patterns and the priors we have built on the world throughout our lifetime um so anyways uh i think this is tremendous progress and i'll add one other note that since this to me the progress of the field of natural language understanding has been so intertwined with my personal line of research as well so in 2020 i also did this really interesting work that i personally believed in with my great colleagues at elemental cognition called glucose that was basically about building uh these board models while you're reading a story so story close test was all about read four sentences predict the ending but now let's go way beyond that that's just not just predicting but also come up with this deep like a word model of the you know a person's kind of set of states and like events and their causal chains and like draw a coherent picture of the narrative that you're building let's make this as the new uh basically um benchmark for evaluating whether or not a system is showing any sort of deep understanding so anyways this was we you know came up with this work and basically did all of it which you know we can delve deeper dive deeper into uh but it when it came out was it like two or three months after uh the gpt3 work and it's fascinating how gpt3 was doing better than gpg2 on if even in this particular basically task of ours and how far it has gone in showing that it has some sort of a board model so i personally think that whoever kind of claims that these models do not have any word model don't have any kind of a human-like cognition don't have any deep understanding of language it's it's really incorrect to say the least of course these models are fundamentally flawed no question right we talked about this in the last session that these models are extremely biased they are uh really easy to get tricked which makes them in my opinion brittle you know how like it was the the thing to call symbolic models back in time brittle i think these models are also quite brutal right it's easy to for them to get sidetracked and make really stupid mistakes although they work say well like often um so these and you know of course these models are also really uh not controllable which as i mentioned is something that we are working actively at vernique on so these are all flaws that they have no question but i think saying that these models do not have a world model do not have any understanding is is really not correct so that was one of the contentions of the stochastic parrots paper wasn't it yes and i i so it's it's we can talk for hours and hours about this as well so many things i i personally think that look um these these models are really great pattern recognizers and one can argue that recognizing patterns and then trying to stitch them together is not real understanding but i would refute that and i would say that look at the end of the day for me as a researcher all i could do throughout the past six years or so was to think about ways of evaluating the and like nlu systems for deeper understanding and these models are proving consistently that they're making progress towards doing what constitutes us having understanding so we can of course argue what is a good benchmark i think benchmarking is one of the problems we've had for years and years we are making progress but for me one of the reasons that why i want to work in startups is to you know build something in real world that actually works for end users in the messy noisy real-world environments as opposed to our lab settings so that's definitely a problem we have that we have really narrow inherently narrow and biased benchmarks but setting that aside i think that um you know this is like if you're kind of theoretical right like people there are people who don't believe in distributional semantics being the uh like expression kind of meaning that you can represent and believe that formal semantics and formal kind of meaning representation is the way to go which i would argue against i think having a way of representing meaning distributionally sort of representing a word by the um context in which it just is uh often occurring at is is a viable representation of meaning so i think as at the end of the day if you have the right benchmarks for evaluating representation of meaning we have the right benchmarks for evaluating common sense reasoning and if these models pass past them that that tells you something and this is the take i had when gpt3 were of work came out so many people were asking my opinion as to how i think about this because you know i was one of the proponents of let's push these systems for deeper understanding and like you know they're not they're really lacking it uh but the truth is we've made progress i would say that we need to move the uh basically um the the bar you should you know raise the bar of course as to pushing these models to go further and further but they've definitely come very far already so i don't think that these models are parrots really it depends on the definition of course of of a parrot but if it means that it's really just repeating um without having any kind of an understanding i think that's that's not the case so going back to kind of the the state of of research broadly and the the gaps that need to be overcome for you to accomplish what you're trying to do at vernique um it sounds like you know the first one of those is or the first you know area you know of exploration is just understanding in general in order for you to do what you're trying to do you need a system to be able to understand to some degree or another or according to some you know metrics the what the user is trying to express in whatever box they're typing in or whatever interface they're using um and it sounds like you're saying that models that you would lean on for that understanding are broken in lots of ways and that's part of what you need to research is how to fix the ways that they're broken you know bias and transparency explainability however you want to put that um but you are you have seen enough evidence of enough understanding for you to do what you're trying to do for them to be promising yes absolutely and honestly not even just talking about the particular technology that we're building at vernier but generally for the field i think that first and foremost i hope that so many people will work on so many other directions of ai so that we really have diversity of thought we have diversity of thinking we have other models that may get to flourish if anything a deep learning trend has taught us that by a couple of people not giving up on what they believed in they could prove us all wrong right and then like all the amazing progress that we're seeing is is the fruits of that basically but anyways on our end in particular for the sake of natural language understanding yes i think these models are to have tremendous flaws uh but they've shown enough evidence of being um basically foundational to say the least so i know that this is also something pretty um controversial right now like you know when stanford started calling these this whole like transfer learning and pre-training and um fine-tuning paradigm foundation models so many people started raising eyebrows as to okay fun foundation should be reliable foundations should be something that you can poke in like just just change things which i completely agree but setting the kind of arguing over the um terminology aside i think that these models have shown enough evidence that they can give us like lexical and board knowledge which i think are very foundational for for building natural language understanding and dialogue systems like you know i come from the school of thought of people who have spent their lifetime trying to build semantic parsers which is which has to do with these formal like semantics of representing each and every word in a way that like conforms to an ontology and then how you would go about like representing a verb in conjunction with something else and like connecting the dots and everything so that you can represent the meaning and then on top of it building a dialogue system that recognizes intents and like tracks and does planning and everything um so i think that what these models are doing uh and you know you they are even very promising in doing things in a multilingual sense but anyways these what these models are doing is that they are enabling us to really let go of some of these pipeline like uh things that we had in nlp and just not starting from scratch basically starting from a model that comes at some sort of a lexical onboard knowledge and turns out common says knowledge baked in um in ish you know i they have flaws as you mentioned and we've talked about the fact that they're not controllable the fact that they're not transparent and the fact that you can't let them like teach them interactively and let them forget about things that they're doing wrongly these are all problems that need to be fixed which may require its own paradigm shift it could be architectural it could be on the data side but i think yes to answer your question in one final sentence these models have shown enough evidence that they could be used as foundations so that we don't start from scratch yeah yeah maybe in kind of keeping conscious of time uh maybe we can switch gears from talking about kind of the the research broadly to the the research that you're pursuing and um you know i think we can infer from there what you think is missing in the the research broadly so it sounds like some controllability of language models is one of the areas what are some of the other areas of research that you're digging into um so yes controllability is one absolutely um vr so our line of thinking is that we believe in um i think i've seen people use this terminology so i'll try to repeat it a retrieval augmented generation so we want to make sure that we don't build uh you know like wishy-washy language models at the end of the day just you don't know where their information is coming from we want our models to be able to retrieve from existing data sources as i mentioned it's all about data and from decision making we want to you as a user to know where your information is coming from or if we are like actively telling you what uh basically uh we think we want you to know the source of it so we want to basically work on ways of retrieving what is out there with the resource and the provenance intact this is another thing that we are pursuing and in general generalization right we want to build a domain general platform how do you make it so that the models that you have are truly generalizable right this is a this has been an ongoing line of work for uh in deep learning for for many years but it's you know of course not solved i think last two years or so has been phenomenal in terms of how uh how much more flexible and generalizable these models are but it's still not you know there's a very very long way to go the other thing is a data right these models are extremely data-hungry i love how there has been a lot of progress on kind of zero shot and in context learning kind of paradigm uh paradigms but at the end of the day still whenever you go to a new domain and you see this first of the first hand when you work in a real world setting like in startups these models are truly data-hungry right how can you make it so that these models are more sample efficient and they don't really need that much training data for adapting to a new domain and this is for us of utmost priority because we literally want to make a domain general model so how do we just go from a domain to another on the technology side without spending a lot of time just collecting data and then tuning models learning like new intricacies of that that existing um domain and i will add that uh i i see a lot of value and someone like myself having spent time thinking about problems that like a pure deep learning person would not have had and the value that it kind of brings to building a real world product so like we need to work on how to build conversational agents right it's not just about natural language understanding like one utterance at the time but it has a lot to do with word modeling really and kind of building a belief system right like all the skill kind of dialogue systems had this framework called bdi that was really cool so belief desire intention so when you have a full-fledged conversation with someone you basically think about uh building models basically and you think about them about what the other party knows what they don't know what you know how you can help them reach to a certain goal and you plan right so the existing systems out there don't really do any planning right these are all kinds of things that need to to be worked on how do you basically um build a sustainable world model that includes the sets of beliefs desires and intentions and you dynamically basically monitor write what's happening and another thing is this is also very important is that we need to have memory right systems at the end of the day controllability comes from them being able to store what they've learned from the interactions separate from their actual like prior knowledge separate from their other kind of conditions so that they can make an informed basically decision and these are also all kinds of things that are not in these so-called like foundation models right that need to be worked on and so you're given the the again kind of the breadth of all of these things that you need to figure out in order to to build a product how do you how do you scope that down like you know each of those could be a you know a four-year phd effort right or or many right how do you scope that down and connect that to uh you know your your mission as a startup founder to get a product out the door that meets a need yes so we are not of course i outlined our years of planning right outline the kind of problems that we have to work on it doesn't mean that we're working on it at this moment or we need to work on it immediately um so that just to answer you on the business side how we are uh practical but to be honest with you so the last like year and a half now has been truly the most rewarding part of my entire life literally in terms of just seeing firsthand how far you can go then you have no other choice but to innovate but to just keep working at what you need to and um the truth is although you're very small and which is the nature of a lot of startups anyways there's so much you can do if you really uh you know gets scrappy and like uh know over to spend your time and effort so we have this thing in at vernee that i don't know if you know what i'm gonna just mention so there used to be this meme going on that was the sketching of spiderman that was done in 10 seconds versus the sketching of it in 10 minutes by an actual artist um so even if you are the master of your skill being art you can always produce something in 10 seconds and you should be always able to to do it and then there is a 10-minute version of course that will be your best work so that's the person we take at vernee for anything and we say this to all of our employees and it's we even said it to all of our interns that look you have you don't have 10 minutes you have 10 seconds you have to have a first version of this go go for it you've done it on the design side as well like our we have an exceptional designer and then she started we were like you have two weeks you have to come up with a full fledge like ui ux and just go for it and she really did it and this is a test we do even when we are doing interviews so anyways we are not working on everything and anything at once a lot of what i outline or on our in our kind of a longer term planning uh but i think we've gotten really good at uh trying to to do a 10 second version of everything just to to know that we we have a something in place you know if you were advising you know someone who was thinking about starting a deep tech type of startup what are kind of the general principles that you would suggest or or put forth for translating um you know going from this kind of gap in the research to to product like what are the general ideas there to me honestly so there there were so many reasons why i wanted to embed myself in this startup scene as opposed to like other maybe industry research labs or even academia and i think the main driver is and should be the fact that you want to build an ai system that actually works so i think for anyone who is working on a fundamental like a research at a fundamental research setting they can see what kind of thing they're really passionate about and see how far it has gone in like you know lab settings and see if it makes sense for it to want to to be basically in an actual product so i don't know if i have necessarily principles but i can just talk a little bit more about um my own frustration as to the progress that i couldn't make outside of this this kind of startup world that may resonate with someone out there that they might see it themselves as well i would say that so i spent about like a year and a half or so back when i was in grad school in industry research labs and to me at the end of the day publishing for the sake of publishing was not really satisfactory so as much as we are all seeing the fruits of academic like scientific research right then sharing which we will you know do at very nick as well as part of our i think duty right as researchers to contribute back like having it as the the main thing that drives you the main thing that you have to report on the main thing that you care about is very counterproductive to me right and i feel like okay i'm spending all of this time of mine on uh basically this this line of work uh where is it going right what kind of value am i bringing to the broader world so that's like one of the reasons i felt like personally the the kind of research i was doing for the sake of publishing is not really a good like bar to have for my life i want to do it for the sake of making your progress but if you had something worthy of saying you should say it i think it's our you know duty as i mentioned to to contribute back so that's an argument maybe for having a product to focus your research and i think maybe put differently from principles i'm trying to get at the kernel of like you have so much that you could possibly research like just how do you prioritize how do you lead or focus a research team that could go in lots of different directions like you can't be bell labs and just do a little bit of everything and kind of have it raw i guess that's the the example that you were just talking about you know with a product in mind like how do you prioritize what you're going to spend your time on so i think actually so we should talk in a few months when we have the actual product and then i would love to be more specific about how we did that because i think that answers a lot of these questions so the truth is honestly having a product helps with narrowing down the focus so much we are we are truly laser focused right now on a particular domain which drives a lot of our decisions and gives us a lot of insights as to what to prioritize and what not to so just talking about the issues that i was outlining that we have with the existing you know models which kind of applies to the internal in-house models that we have as well like lack of control lack of transparency like not being able to for them to work fast enough right because the user has a certain level of expectation as to how quickly they should respond these completely change when you go to a new domain so that just dictates our um basically priorities it's like v it's so funny about like a year and a half ago or so when ahmed and i we were sitting down to to just outline these features that we want our ai platform to have we literally wrote down the list like it should be instantaneous in terms of response time it should have controllability it should be i don't know first start being like single turn and then multi-turn and then it should be domain general from the get-go blah blah and we literally on the notion page which is the software we use for knowledge management we have a priority assigned to them which is like hi media blah and then it changes per domain so if i could share that with you that notion page you would see that we are literally doing it like the features that even we care about in terms of the uh the the technology itself is very much dependent on the domain and then we keep going back and forth on them depending on what you're focusing on for a particular basic recorder got it got it got it so i so summary there is you know if you're very focused on kind of product and features then that will tell you what you need to figure out to deliver those and in order to tell me more detail you'd have to talk about the specific features and research uh and that's going to come soon because i think actually you would really like it for for us we had this dilemma of what our net our first kind of sector would be first domain would be and probably will be even coupled when we come out with it but regardless uh the the thing that is the closest to our hearts is the one that has a lot to do with some of the things that we were just discussing and it really applies to everyone's day-to-day decision making and it's really exciting so that's another thing there's like this kind of hidden decision like a feature right when you're prioritizing this gigantic space in this gigantic space which has to do with your personal passion it really is right then you're especially so people call this product market fit but i think in our world it's called technology market fit that like founders like us you have a particular technology that could be fitted into any market but in any product and now you have this dilemma of even thinking about the technology being fitted into market not just taking the product being fit into markets it's really even a larger search space for a founder to want to to navigate um but yeah i think that personal passion and personal care is something that will play a role and has played a role in you know hopefully a couple months from now when we chat i'll happily spill all the beans awesome awesome well nasreen it was wonderful catching up with you as always thanks so much for taking the time and looking forward to next time absolutely pleasure was mine yeah looking forward to that awesome thank you thanks bye you

Channel: The TWIML AI Podcast with Sam Charrington

Views: 406

Keywords: TWiML & AI, Podcast, Tech, Technology, ML, AI, Machine Learning, Artificial Intelligence, Sam Charrington, data, science, computer science, deep learning, nasrin mostafazadeh, ai rewind, startup, verneek, human computer interfaces, product development, ai research, productizing, natural language processing, natural language understanding, stochastic parrots, GLUCOSE, deep tech

Id: FEVIGcL57Xc

Length: 53min 16sec (3196 seconds)

Published: Wed Nov 24 2021