Did Gen AI Kill NLP? With Meri Nova, Technical Founder at Break into Data

Richie Cotton: Hi, Meri. Thank you for joining me on the show.

Meri Nova: Hi there, thank you for having me.

Richie Cotton: So just to begin with, can you tell me what is natural language processing?

Meri Nova: Sure, so natural language processing is one of the fields in AI that solely focuses on helping and enabling computers understand, interpret, process, and even generate these days. human language.

Richie Cotton: Pretty broad. Do you want to tell me, like, uh, some ideas of things that you can do with natural language processing or with text data in general?

Meri Nova: you can parse text, you can understand the semantic meaning behind it, can even generate these days, as I said, so there are many use cases that NLP has. It's a topic.

Richie Cotton: and so in a business context how might you make use of natural language processing?

Meri Nova: I think the first thing that comes to my mind is having chatbots on your website that answers customers questions and that also Big one is the content moderation in things like Discord or different online communities. You can flag harmful content and you can also process documents and store them and retrieve information from them.

Richie Cotton: So I like the idea of just, chatbots and processing documents, like two very common business cases. It seems like these things have been around for quite a while now. I mean, we've had chatbots for at least a decade. So how has the advent of sort of powerful, alternative AI changed the field?

Meri Nova: That's a big question, I think, because nowadays we are living through it, and every day we see these new models coming up, especially within LLMs, and not that many people realize this, but I feel like this is one of the most exciting times. in the history of AI just because we have seen so much of transformation in the field and not only in NLP specifically, but in every other field, which makes it even more exciting.

 But generative AI specifically made us finally Able to create large amounts of text data, which we couldn't before with regular learning and NLP specifically Models because before it used to be only about parsing, understanding or just classifying a piece of text. And these days you can be a lot more creative because nowadays you can generate lots of data, which is super exciting.

Richie Cotton: Okay. Yeah. So I think certainly like classifying text is fairly standard use case has been around a long time, but now having these sort of more creative conversations going to improve the quality of your chatbots and things like that. So since you've got all these sort of wealth of like traditional NLP techniques, I mean, you have the new generative techniques, when might you want to use one or the other?

Like when do you actually need these new powerful tools?

Meri Nova: Yeah, this is a great question because I feel like these days everyone is just obsessed with LLMs and they want to patch it on every single business problem that you have, but I feel like it's Generally, a very wrong approach because these models are quite expensive and if you do not have a justified business problem that you want to tackle, I think you need to still stick to the traditional models which can help you with just clear You and simple tasks where you can see patterns and like, for instance, if you take spam detection in the emails, you don't have to use LLMs to say, Hey, this is spam or not when you only have to say yes or no, basically.

So I feel like people are just jumping too fast on the LLMs when there's so many great options that you should be using beforehand.

Richie Cotton: Okay, yeah, certainly I can see how just trying to use an LLM for a spam filter, that might be a bit overkill. So do you have a heuristic for when you might want to choose one technique or the other?

Meri Nova: I generally try to stick to um, approach when if I need something creative, which is interesting, and I think it's controversial because we originally wanted AI to take on the mundane tasks, but nowadays we see a lot of it just involving creative tasks like Generating articles and just actually yesterday I saw Google's app that's called notebook LM and they generate podcasts like these based on the text.

And I feel like it's very creative, but at the same time, many people are sad that we are basically giving all these models all the fun. But this is the heuristic. So whenever you have, let's say, very tasks with clear patterns, like classification or detection, then you might want to stick to traditional models.

Richie Cotton: Okay, yeah, so that podcast one is particularly interesting. I have had to play around with notebook LLM, and it seems to be a recurring theme on this podcast that guests come on and tell me that I'm about to be replaced by AI. It's very worrying. I have to say, you know, I think my job's kind of secure.

Sometimes I stutter, sometimes I ask a stupid question, and that way, you know, it's like a proper human artist product. Yeah, not just AI perfection. So if you need creativity, then that's the case when you need to use more gerontology techniques, something a bit more powerful. So one of the grumbles I keep hearing is that actually, once you start putting things in production, generative AI gets quite expensive. So does cost factor into your consideration for like what tools or techniques you want to use for NLP?

Meri Nova: It's interesting because if you want to have your own custom LLM built, it would probably cost you around a couple million dollars to say the least. And that's just training part. But when it comes to inferencing, it's another whole cost that you need to cover. So if you have a Novel ideas that you want to tackle and you feel like that can justify your business model, then sure, go ahead and build your own model.

But many people cannot afford that. And again, if you're trying to build a feature that simple models can tackle, then I don't think the cost can justify itself. So yeah, I think it depends on the use case.

Richie Cotton: Okay. So in general, you probably want to start with the cheapest, actually, I'm not sure whether this is the right round. Do you start with the cheapest tool and then see if it works? Or do you start with. the tool that's going to give you the solution fastest and then see how expensive it is. is there a order to which you would pick things?

Meri Nova: I think it depends not on the cost of the tool itself, but rather how much does running that specific approach would cost you because running again, the LLM would mean a lot more computational load and also a lot of data that you would want to store for your pre training and fine tuning purposes.

But with a simple models, like even not just simple, but like even previous neural networks that we worked with, like BERT they just differ in the order of like thousands of like magnitude in terms of like the cost. And yeah, I think it's worth having those first.

Richie Cotton: related to this I suppose like how well the model performs is going to be an important decision. So how do you go about evaluating the performance of I guess let's start with the traditional NLP models. what constitutes success? is there an accuracy measure or what would you say constitutes good performance?

Meri Nova: I think these days people focus a lot more on evaluating the LLM systems which is part of NLP, of course, but they focus a lot more on evaluation these days just because these are non deterministic systems that we have never dealt with before, especially in the production environment, and it can cause all sorts of failures, including security related ones, too.

Just yesterday. I watched a video where a guy basically went on live and tried to have a simple SQL like injection attack, but this time it was through the chat bot. So what he did was he just typed in the question trying to get the schema of the database of that company. So his end goal was to get the revenue for that month, I believe, right?

And LLM basically gave out all the answers to him even though there were security measures basic security measures that had to do with LLMs, but still we're trying to figure out how to make them more Not only deterministic, but also safe to use. And I think it's still an open ended question.

And there are many benchmarks and evaluation modules that different companies offer. For instance LangChain, since the course that I've been making with DataCamp is with LangChain. They have their own frameworks that anyone can use on Lang Smith. And there are other frameworks as well, of course, which has to do with more like rag systems.

I think there, the one that I was teaching in the course is from Ragas Framework, and it has I believe there were four different metrics, the main ones, which was, which had to do with faithfulness and precision and recalls. they sort of. are relevant to the traditional metrics, but of course they have to take into consideration of the stochastic behavior of the LLMs.

Richie Cotton: Okay. Yeah. So that seems like a very interesting thing. And maybe we'll get back to like Jane and Langsmith and things like that in a moment. But you mentioned the idea that generally these LLMs are probabilistic, so you're going to get a different answer every time. How does that affect how you even go about like saying, well, does this work or not?

how do you measure the performance when it changes every time?

Meri Nova: I think there, there's also benchmarks for LLMs performance without having to look at The task specific metrics. So definitely there is on faithfulness and just the number like of hallucinations that LLM can make those. There are benchmarks that evaluate purely the LLMs performance.

Richie Cotton: Okay. All right. So there are some tools out there might be slightly more complicated than just the sort of traditional thing where you've got a single number as, as, as an answer. So you mentioned some of the tools you've been working with when you were teaching how to work with in your course. So do you want to talk me through, like, what's the current tech stack for working with LLMs, and what are the different tools you use?

Meri Nova: I think everyone has their own tech stack when it comes to LLMs, just because also the number of models, the open source models that we have in Hugging Face is around hundreds and hundreds of thousands of them, and it becomes a task of just like understanding which one that works best for your use case.

And everyone has their own way of deploying those models into production. But again, it depends on the database that you're using. And our case, it's vector store that you're choosing their many third party vector stores that you can use. And that's number one. And also the retrieval systems that you're putting in place.

And in the course, we're going over sparse and dense embeddings that dictate the type of retrieval that you'll be using. And the evaluation frameworks, of course. And where would you like to host your model to if it's not hugging phase, then Where and how you would like to monitor your models and where your logs will go and so on and so forth.

But my favorite ones, I actually I think made a post a couple months ago about my favorite tech stack, but a big fan of what Weights and Biases does. I think they have a great environment and great support. Other than that there are many options like on Hugging Face. Hugging Face is one of the most available and the just popular ways to deploy your models.

Richie Cotton: Okay, cool. So we've got some sort of vector store to store all your data for the fact retrieval, and then you mentioned you need somewhere To host your model. So there's something like Hugging Face. You mentioned Weights and Bias as well, which I presume it's also like a model tracking platform. And then there was some sort of evaluation framework as well. So lots to go there. Maybe we'll start with like the vector stores. Did you want to tell us a bit more about what are these vector stores and. How do they fit into the whole AI system?

Meri Nova: To give you context, the GPT 2 models that we have, I believe, were trained on Forty something terabytes of text data, which is quite a lot, and you need a way to store that unstructured data in one place and be able to retrieve them as well. So there are great vector stores that originated in San Francisco, which is interesting.

I met a couple of the founders and founding engineers of Chroma DB. One of the conferences in San Francisco, and it felt like at one point everyone is surging to like open and start a vector database just because the demand for it is huge. So vector database is basically data storage for Okay. All that text that you have, and it allows you to create these from text which is that hiring embeddings are that highly numerical values in a vector.

So like one token, let's say it's a one vector and it can have hundreds and hundreds of dimensions. And you need to be able to store them in one place. And also you can have a way to understand the relationship and the semantics underneath this text. So it's not only just storing, it also stores it in a way where it understands that let's say gender is a thing, or the transportation is a thing and car is related to transportation and so on and so forth.

 that's the gist of it.

Richie Cotton: Okay, so the beta storage is just a place to store chunks of text, and then you've got an embedded model that's gonna say, tell you what the meaning of the text is. You can find similar text, I guess, the idea.

Okay, just continue on the subject. Do you want to talk us through, like some of the use cases for this?

Like, when do you actually care about having A vector store as well as having the LLM.

Meri Nova: Pretty much all the time. I mean so vector stores are, let's say if you're building a system that has your own custom build LLM. So in order for you to have your own model, you need to have a vector store so that you can store All the embeddings that you have created and be able to retrieve from them.

So in our case, when we created the rag, we used a vector store to basically transform the document corpus that you have into the embeddings and store them in the vector store so that next time a customer, let's say Visits your website and asks a question, you can convert their query into embeddings and then use a similarity search so that you can find similar words or documents that have to do with this query so that you can respond to the user's question.

And that's the case with RAG at least, you need a vector store. But in other cases, you need it to store your training data and fine tuning data as well. But if you're just using a third party API, you definitely do not need a vector store. You just need to have good prompts and work with them.

Richie Cotton: Okay so really it's any kind of use case where you have to have some kind of fact retrieval, I guess the big use case is about reducing hallucinations, right? So it's like if you want to work with a document or you want to have a chatbot or whatever, then it's going to involve this retrieval augmented generation at some point.

Meri Nova: Yes, not only reducing the hallucinations, but also having a way to integrate your own proprietary data into your LLM system so that you do not need a human to be able to answer customers questions.

Richie Cotton: I like that. Yeah. So avoiding humans, I guess a bit of a double edged sword there. It's like, if, if that's your job to answer questions, then being replaced by the AI is kind of hard, but also if you're answering the same questions over and over again, then maybe better done by a bot rather than making you go crazy, answering the same questions over and So actually I think this leads nicely into talking about the organizational side of like working with text. So the point where you start talking about, well, maybe AI is replacing humans in some place, that sort of suggests that some of your processes are going to change. So once you start working with these newer generative AI tools, if you are an AI engineer or a natural language processing scientist, then talk me through how your workflow is going to change now.

Meri Nova: I think the title of AI engineer actually became a lot more popular since the advent of LLMs, just because before it used to be ML engineer or NLP engineer or computer vision engineer. But these days, I guess when people say AI engineer, it's very generic actually. I, I think I would put even ML engineer title underneath the AI engineer.

There's a huge debate on this, that's why , I'm quite specific about this. But the workflow I think it depends again, on the domain. And how are you implementing LLMs? But if you've never worked with LLMs and you would like to start working with them, it depends on choosing the type of path that you're taking.

For instance, again, training and having and hosting your own model is a completely different workflow than, let's say, finding and searching for an open source model and for your specific tasks. That you wanted to, perform. And the third one is just, again, finding the cheapest third party API and integrating it into your system.

Those are three different things. And I believe the skill set does also depend on that because the first one demands a lot more of the scientific understanding of deep learning concepts and how to process large amounts of data and those other two I think more related to engineering side and deployment side and system design side, as well as just having great prompts.

So depends on the approach you take.

Richie Cotton: Okay. Yeah. So that's interesting that although AI engineer is the very kind of, it's becoming a very popular job title. It's actually lots of different things. Involved in that. So messing about with PyTorch to create your own models or fine tune models, that's a very different proposition to just calling APIs. And say, just calling, I told you to be that as a, as a job skill. So, but if you're calling APIs and pulling in uh, or making use of other models, then that's again, a different skill. So in general you mentioned that AI engineer is kind of or at least part of it is the same thing as a more traditional machine learning engineer. How do you see all these roles changing at the moment?

Meri Nova: I think that's the million dollar question. Just because I write for aspiring AI engineers and machine learning engineers, and that's one of the top questions they ask. It's what kind of skills that we need to prepare and especially those who are graduating. or master's or bachelor's degree and are trying to get their first jobs.

I think a lot of it has to do with the strategy that you choose. But the one thing that does not change is if you have a specific domain expertise. And then sort of adapting your skills to that domain. That would definitely give you a competitive edge in any scenario. But if you are looking at the skills as an engineer themselves, I would say just understanding these concepts are super important because if you're trying Deploy again, nondeterministic systems and you do not know how they behave and how are they created and what are their dependencies and the outputs look like, then I think it will be really hard for you to even understand the it.

The opportunities that it has. And yeah, so I think it's just about staying open minded and learning every single day, just because there are so many tools that are coming out. And for my audience, I usually just tell them to stick to the fundamentals, which is in our case, especially these days for deep learning is linear algebra and also of course, statistics and probability.

And once you have these. Fundamentals in check, it will be very easy for you to adapt to any other tool because now you know what are they comprised of and you can even come up with your own tools, which is just a great place to be in, in any case.

Richie Cotton: that's interesting that you recommend learning a lot of the fundamentals then, so linear algebra, statistics, things like that. I think a lot of people that jump straight into, well, I want to do something with a tool and kind of skip over the, some of the fundamentals. So it's interesting that You are like advocating for learning these sort of lower level things. Are there any particular techniques within linear algebra or statistics that you think are particularly important for people interested in machine learning or, or AI?

Meri Nova: yeah, actually I had this debate online with another person, another creator who disagreed with me and we sort of had this argument. And his opinion was that you need to build and you do not really need to understand these underlying concepts and I agree you can build things without understanding them, especially these days.

But my approach is mainly on Just getting on back to the first principles thinking just because I think that people give you a lot more freedom in your creative approach as an engineer as well. And I believe in our world of like following the trends and just especially, especially these days when you have these transformers and large language models that can do almost like everything.

I feel like people forget that. These architectures have also been created quite recently, and it does not mean that it's the only path forward. In fact, there are many architectures that have been invented alongside neural networks in the past, but have been forgotten. And I feel like if there was a lot more creativity when it comes to building architectures and building models we could even arrive to AGI, which is also a hot topic these days in Silicon Valley.

a lot faster or even cheaper and we can even maybe have more control over these models. And yeah, I think I'm mostly I'm excited about these things, but when it comes to statistics and linear algebra, I would say have an intuitive understanding of how matrix multiplication work and how vector transformations work and how optimizations work as well as calculus understanding derivatives and basically how the rate of change happens in these models.

And In statistics, it's causal statistics, but statistics is mainly for traditional machine learning, however with neural networks, it's mainly linear algebra, and people think it's very complicated, but I believe it's just the notation that scares many people away, but when you look underneath the notations, it's simply just multiplying a number.

with B and adding in Y and and just doing it a million times over and over. And I think that's it.

Richie Cotton: I love the idea that, well, yeah, linear algebra is just almost like adding numbers. And it's cool. Dead easy, really. Yeah. So Yeah, can certainly see how understanding that actually everything with a neural network is, basically just matrices underneath and some optimization routines.

So having a basic understanding of those is going to be quite helpful. so you mentioned that people aren't very creative in their architectures and some of the sort of traditional machine learning techniques are getting ignored or forgotten about. What do you think people are missing out on?

Like, what are the most undervalued machine learning techniques at the moment then? 

Meri Nova: so I have a co founder, and his name is Kostya, and we are both building BreakintoData, which is our community for machine learning engineers and, data professionals in general. And when I met him for the first time, he influenced a lot of my thinking and a lot of my just thinking around machine learning in general.

 the first time I met him, he was telling me a lot about Fourier transforms and how you can process audio without text and how you can combine and mix these algorithms and formulas together and just come up with a completely different approach. And I really liked his way of thinking because like before meeting him and, and just understanding his way of thinking, it was mostly just learning.

Pre existing frameworks and just strictly following those to get results. And I finally understood that you do not have to be locked in these conventions to solve even the same problems. for me, we started working and just researching about, Symbolic AI and how you can integrate symbolic AI with knowledge graphs and sort of create this entity that can understand things not in a way of statistics where you have large corpus of data, but rather have it behave just like a human brain does, which is just accumulating knowledge over time and having these Understanding the relationship between things and that's why knowledge graphs there too.

So I think there's just. a lot of it comes from math these days, but even nowadays you can just use Python and nothing else but Python and start with just like simple algorithms and the way he taught me to do this is just Come up with a recommender system that just like links your relatives to one another and just understands the relationships between them and, and so on and so forth.

So like having these like small challenges will make you creative and also will make you understand the practical implications and opportunities of what you can create.

Richie Cotton: Yeah. So lots of very cool stuff. I do like the idea of having wait, was that a recommender system for your relatives? It's like, Oh, I don't recommend Auntie Nelly. Uh, Yeah, but 

uh, 

Meri Nova: yeah, it's just important. We have there's hundreds and hundreds of relatives and it's a lot of data to go through.

Richie Cotton: Big family, big data. Nice.

Yeah.

I, I do that idea of having just a simple project that's just personal to you though, just to practice things on. And that just seemed like a great idea for working with knowledge graphs actually. But yeah also. Interesting that you mentioned Fourier transforms. I think it's like physicist's favorite.

And I guess anyone involved in like music recording, they're going to be very familiar with Fourier transforms. But yeah that is an interesting idea that we've taken over, like just using generative AI for working with audio, but actually there's a ton of techniques that are around for decades. So, it also seems like a lot of the tools are making it easier to work with AI uh, and data in general. So, Is that changing who can build things or who can make use of this? Like, is there, do you think the population of like who is working with AI or data is changing?

Meri Nova: Yes, was also thinking about Langchain a lot. And when I was thinking about what kind of course to create with Datacamp, I wanted to do it with React, just because back then it was very popular. And I mean, it still is, but Langchain specifically, especially in the more experienced crowd of machine learning engineers and just Python developers in general, they do not like it.

And it's funny to see because whenever I tell them, I'm making a course and they're like, what are you making a course on? I'm like, Lang chain, but he's like, why out of all the things you choose Lang chain? But I think there is a place for Lang chain to exist. And I think people do not appreciate it enough, especially for those who do not have engineering background.

And LinkedIn itself sort of brought to the field of what's possible with LLMs and how to digest LLMs better and people don't appreciate this and they think their documentation is terrible and things like that, which might be true because they're just changing lot of versions. So, but other than that, I think they are definitely leaders in implementing LLMs and showing us the way, sort of like paving the way to the production environment with LLMs.

And that's interesting because people that do not have experience can greatly benefit From exploring Lankchain's methods and different modules they have, but people that do have experience and especially engineering experience, they can tinker around even further and have this flexibility of, like, coming up with anything that they want, their own unique evaluator, let's say, or their own unique way of retrieving things or combining it, having a hybrid retriever or things like that.

And just also being creative with the prompts. And yeah, I think that's just two different levels, and I definitely urge everyone to focus more on the engineering side so that they do not rely on libraries like LangChain. But LangChain is definitely a source of inspiration for many of the new tools that are coming now to the field.

Richie Cotton: Interesting. So I hadn't realized there was such controversy around Langchain. So you're saying it's a good beginner tool, but if you're an engineer and you're comfortable tinkering around with APIs, I guess directly, then maybe you don't need it. All right. And you mentioned there are some other new tools that have been inspired by Langchain.

What are these other tools?

Meri Nova: There's other retrieval based tools like Llama Index, for instance. They're also from San Francisco and I've been to one of their talks. they talk a lot about agentic AI and how they are making it decentralized. And they're sort of coming up with their novel ways of implementing AI agents. So that's exciting.

 And I think there are a lot more that for instance, what else? DSPY is also becoming very popular for prompt engineers because they're trying to make it more like a programming, even though it's human language, but I like the way they come up with the systems and frameworks that help us, implement LLMs.

So, there are many more, and I haven't even explored that many, but these are, like, come to my mind now.

Richie Cotton: So you also mentioned knowledge graphs as being an important component of of any AI system. So you mentioned the idea of having like looking at the relationships between your relatives. Beyond this, can you talk me through why would you want a knowledge graph and how would it fit into your AI?

Meri Nova: So in the course that we made with the data camp. our last chapter is on the graph rack. And when I started developing the course, I think about that time Microsoft released their graph rack as well. And people sort of realized that graphs can make rack systems a lot more efficient. And the way they do it is basically the retrieval system that we have with the rack these days has to parse through lots of data to find the most relevant documents or text to put into the output, but graphs, what they do is they give you the higher level a relationship between the main entities that you have and that significantly reduces your. search space and computational load, therefore, and would also allow your outputs be more relevant to the document corpus in general. So, graphs in general they are considered to elevate the RAG systems, and even though we do not see a lot of them in the production just yet, just because it's a relatively new model.

Way of implementing RAG, but I think there's a lot of opportunity there too, because simply because it. Gives us a lot more understanding of the document corpus, which is harder to get with the simple retrieval systems.

Richie Cotton: so yeah, I think like the standard RAG approach, you're looking for things with a cosine similarity or something. It's back to simple linear algebra to find what's the, the best text. So having something more sophisticated might give you better results. so that sounds very cool is using all this sort of relationships for invisible text to like find what's the best text to include in your response. Have you seen any success stories with it yet? Is it too soon or are there any people like building cool things using this graph reg technique?

Meri Nova: I, especially for Graphbrack, I haven't seen as much blogs from BigDeck or any startups yet. I've seen lots of tutorials on this. I feel like this is relatively new approach and people were kind of skeptical about this just because it's it adds another layer of complexity when it comes to dealing with your database.

And not that many people are ready to, to get to that. But other than that, outside of RAG, Knowledge Graphs have been popular since a long time now, decades. And one of the popular implementations of Knowledge Graph is Google. Also having, I think they had a couple hundred millions of, nodes and facts stored in their database around just historical facts on musicians, writers, and so on and so forth.

 but other than that, I haven't really seen any implementations in the production, but I'm excited to I feel like they will come out soon.

Richie Cotton: All right. So it's a, an open space for anyone wanting to be first to build something. Cool. So some good opportunities there. guess if you want to build your own search engine as well. All right. Takeover would be the next Google. All right. So just to wrap up, what are you most excited about in the world of NLP and AI?

Meri Nova: I am honestly excited since the beginning of my machine learning career. I've been excited about automation and specifically. And I feel like there's a lot of opportunity when it comes to AI agents. And AGI in general, just because now that we understood how to communicate with computers with the English language, which is super exciting.

It can do things for us just because the things that we need to manage are in English as well. And before we didn't have the capacity. for the computer to navigate these spaces. So I think this is very exciting. just to see how the field is evolving, especially in San Francisco.

And I go to lots of hackathons, and they always come up with the novel ideas of how to Automate your browser related tasks or have your own assistance, AI based assistance that manage everything for you respond to your emails, schedule your calendar and make decisions even for you just based on the facts that they know about you and just also having it as a companion.

I think it's very underestimated. Just because I am a huge advocate for CPTSD most of my time is just having this, like my prefrontal cortex trying to understand what's the next priority and making the decision is really hard, especially if I do not have anyone to analyze the situation.

So I just take a ChagGPT and I'm like, okay, here's the facts that I have right now, help me digest this and make the best decision. And we'll just go through one by one. And sometimes I'm just like, wow, this is crazy. It simplifies my life. And I know that. I can use this tool anytime I want, and it's just wonderful.

And I feel like people are still not using it enough. There are so many use cases that we have for this amazing uh, technology that we're still discovering. And I feel like this is a nice era for us to experiment. And that's my favorite place to be. It's just the constant experimentation and finding things.

how to use this technology to make our lives better?

Richie Cotton: That's actually a very cool use case. I like the idea of yeah, a particular view of ADHD is just having something to just help you make decisions and sort of take away Having brain fog and just things go around your head. You just got someone to chat to. It's like a more advanced rubber duck, I guess. Just someone to give you a response. Do you have any more use cases like that?

Just for AI and day to day life.

Meri Nova: Oh, interesting. I mean, I use AI a lot in terms of especially content, just because I'm a content creator. And I feel like before AI, if I was, I was not a content creator before AI, by the way, which I to this day, I'm like, Oh my gosh, how did these people come up with like ideas? And how did they edit their text and things like that?

I don't know. But these days, it just makes my workflow so much more enjoyable, and sometimes I can just like, have this creative dual with, with AI where I tell a joke and then I ask it to tell a joke that's similar to this joke. And then I just have this loop of like, funny jokes and that's the way I create content, especially for memes.

But yeah, I think it's great mirror of your own imagination. And I am really grateful again to, to be alive during this technological advancements.

Richie Cotton: That's a, that's a good way to wrap up there. Really grateful to be alive in this amazing time. It truly is an amazing time. All right. Thank you so much, Mary, for your time and your thoughts.

Meri Nova: Thank you so much, Richie, for having me.