Utilizing Tech - Season 7: AI Data Infrastructure Presented by Solidigm - 3x10: Democratizing Data Infrastructure for ML with Melisa Tokmak of Scale AI

Episode Date: November 9, 2021

Data is the most important component of AI implementation, but most companies neglect data infrastructure and focus too much on the ML models. In this episode of the Utilizing AI podcast, Melisa Tokmak of Scale AI joins Frederic Van Haren and Stephen Foskett to discuss the democratization of data infrastructure to support machine learning projects. Enterprises often don't have a good understanding of their data, and this can undermine the success of an AI project, and this must be addressed before the project can proceed. Companies also must consider the quality of their data, beginning with a definition of the metrics that will properly assess the data foundation for their ML models. Three Questions Frederic: Will we ever see a Hollywood-style “artificial mind” like Mr. Data or other characters? Stephen: How big can ML models get? Will today's hundred-billion parameter model look small tomorrow or have we reached the limit? Alexandrine Royer: What do you think is one of the biggest ethical challenges that comes with AI that often goes under-discussed and should be more present in conversations surrounding the deployment of AI models? Guests and Hosts Melisa Tokmak, GM at Scale AI. Connect with Melisa on LinkedIn or on Twitter @MelisaTokmak.  Frederic Van Haren, Founder at HighFens Inc., Consultancy & Services. Connect with Frederic on Highfens.com or on Twitter at @FredericVHaren. Stephen Foskett, Publisher of Gestalt IT and Organizer of Tech Field Day. Find Stephen’s writing at GestaltIT.com and on Twitter at @SFoskett.       Date: 11/09/2021 Tags: @MelisaTokmak, @scale_AI, @SFoskett, @FredericVHaren  

I'm Stephen Foskett. I'm Frederick Van Haren. And this is the Utilizing AI podcast. Welcome to another episode of Utilizing AI, the podcast about enterprise applications for machine learning, deep learning, and other artificial intelligence topics. One of the things that's come up again and again on our episodes is the democratization of AI technology. Essentially, there's only a few companies in the world that are big enough to own everything, to build everything, and to really get benefit from machine learning. And a lot of that comes down to data science. Frederick, I know you've been involved in this space as well. Talk to us maybe a little bit about the link between data and machine learning.

Right, so AI in general, it's all about data, right? It's not about formulas and mathematics like we used to do. You need the data in order to apply mathematics and to build models. And in order to create models that are accurate, provide the necessary performance, and reduce the development time, you need good data, good quality data. A lot of times people say you need a lot of data from a quantitative

perspective. The reality is you better have less data, but qualitative than quantitative. And I think that is a very good topic for this episode. Yeah, absolutely. And it seems as well that sometimes the data science aspect and the management and organization and structuring and collection and use of the data to train the ML model is overlooked because people are so focused on machine learning and GPUs and basically the doing that they don't think about the groundwork. And that's why we wanted to invite Melissa Tokmak from Scale AI to join us here to talk a little bit more about the data science and democratizing data infrastructure. Welcome to the podcast, Melissa. Hi, thank you so much for having me.

I'll maybe introduce myself a little bit. I'm Melissa, and I work at Scale AI. I've been here for about two years, and I'm the GM for our Document AI and Models as a Service products. And in my time here, I really have worked across the board in different products from data labeling, machine learning, augmented data labeling to the models itself. And before joining scale, I used to work at Facebook in new product monetization area. And before then, I was at Stanford University studying computer science.

So today, I think we're going to talk a lot about how data is really the most important component when we're looking to implement AI into different industries in our daily lives, right? And I actually got really inspired about Scale's story that really matched a little bit my story before I joined that specifically focuses on this problem. Right. So I remember when we were talking with our founder, who's himself a machine learning engineer as well. In college, he used to try to build models. Right. Which is like all of us. Maybe if you went to school and dabbled a little bit in computer science and machine learning,

you have tried to do, but guess what? You kind of stop early on because you're like, okay, well, I'm not really getting the results I'm hoping for. Like why? Because you as a student don't have data. Well, you're not lonely.

A lot of the different industries and companies, enterprise companies are going through that problem themselves. They might have lots of raw data, but what they don't have is the high quality training data, right? To be able to not only train these models, but actually get the results that you're hoping for. And that's how the story of this company started. And as AI took on, we really see this problem being very prevalent across the board, right? When we look at various numbers, right? Like recently, I was looking into this for the industries we support, but about 6% of industries only adopted AI, right? What does that mean? That's crazy. Because if you talk to anyone in Silicon Valley or around us, right?

AI, AI, AI, machine learning is like going off the top of mind. But when you look at the world, various industries, enterprises, like these companies that are actually holding life together from financial services to logistics to right like autonomous vehicles and even beyond like big tech companies that are not the five we'll maybe touch upon a little bit it's very new for them when it comes to AI adoption they might be dabbling in various solutions that give them rule-based approach, right? You write code and you put some rules and try to make decisions based on that rules, but really implementing machine learning, operationalizing, what I call operationalizing AI, is very new to these companies? And that's really about two big reasons, right? One is exactly what Frederick was talking about, the data. Having high quality data and then feeding that into it, really investing in the ML to get the results.

Not everybody can do that. And we'll touch upon that. And really, the second part, which hopefully we'll talk about today, is really patience as well to get the results what you're looking for. Because machine learning is really all about the iteration, that constantly the iteration you do on the data. And a lot of the industries might be kind of losing that patience early on and see, OK, do we really need this now or is it ready?

So today we'll talk a lot about how to tackle those problems if you are a company and what can you do if you can't build huge ML teams yourselves to be able to get some ROI on your business problems. Yeah, I totally agree. I think a lot of enterprises have difficulties with the concept of data. And what I mean by that is that they don't really have a good understanding what qualitative data is, and so they're looking for answers. So in your opinion, do you think a lot of enterprises understand

that their problem is data? Because they might be pointing at petabytes and petabytes of data, but most of the data could be irrelevant to the problem they're trying to solve. So how do you as an organization help those enterprises first understand what they have and kind of provide a metric maybe, I don't know, and how you help them kind of improve not only the concept, but also the quality of their data? Yeah, yeah. No, that's actually a great question. But you're right. I think we're in the early days of that education. However, one thing we have on our side, right,

like the tailwind here is the fact that everyone knows that ML train has like taken off and you have to jump on, right? If you don't jump on, then like the future is pretty grim, right? So that understanding, just knowing that even we say things like, right, like data is the new code, made it through all industries and these leaders. So they know that they have to pay attention to this. However, when it comes to the data part, right, like looking into, okay, how you have been doing things and how can we really optimize that for you? What does high quality data look like? How can you even define that and get it at the end is still an early on for these industries and enterprises to be able to understand and invest. So what we do in that situation,

I think these companies, a few things to understand is very important. First of all, like having ML models or deploying ML models is really 50% done. Recently, we hosted Andrew Ng, who's really like a prevalent, obviously name in machine learning. That's what he says. And a lot of people are starting to understand that more. It's not like you have some data deploy something

and go ahead and you're now down, done machine learning magically will solve all your problem is not the case, right? Like you're really in the beginning of that journey. So one thing we have done, and I'm seeing a lot more in the industry for that is being that technical partner to the companies you're working with, right? Really talking about, okay, where is the data situation right now? Like, do you have the raw data? What does the raw data look? What have you tried? Did you have the training data? What are the models looking for? We really try to be that technical partner to understand and provide that. And I think one of the, I guess, like with you, you were asking, you know, what are the metrics that can it be? It's very difficult to identify

a metric early on, right? Because what matters at the end is how is the model improving on that data? So you have to look at the end result, right? Like it's not just I provided data for you and you look at some samples, it looks great. Sure, you can continue getting that. But the actual metric is, did that help my model? Did that help the result I'm getting the model? That's quite difficult to understand in the industry today because machine learning engineers themselves have to write these right like applications to be even able to understand and visualize the data and see the results of the

model and make any changes and I think a lot of the machine learning engineers obviously you guys know the best like don't want to spend time on that right so what we do on top of it is like really enable these engineers and teams with the tools for them to put their models in, for them to have the data and see as the data quality improves, how would the model perform better? That's quite important to provide like these tools to the machine learning engineers so that they don't have to waste their time, right? Like focusing on that, they can only focus on, okay, what does my data not cover? So how can I solve that with the data? And what do I need to do in the model architecture as well separately? But we

try to provide those tools as well to these teams. So it's easy for them to see the end metrics, which can be very different according to what problem you're taking on. The second one that is more indicative, maybe it's a directional metric early on is the quality, right? So I think a lot of the time we call it just quality, quality, but quality can mean so many different things according to the problems that you're trying to solve. So in that approach, we really take like defining those metrics, right? According to what the data needs to get and being that partner that if something doesn't really matter to that problem, you know, making that clear so that we're not measuring the wrong things. For example,

not every data has to be pixel perfect, but it might be you need to understand a little bit more about what are the critical errors and what are non-critical errors to really focus on resolving the critical part to see the biggest impact. Then you can optimize, right, like your data even more if you're not getting the full results you're looking for. But you have to optimize for the big jumps in the results that you're hoping for. And that also helps, to be honest, I mean, we're all working. And when we are prioritizing our work, we're looking for quick wins, right? Is this going to work? How can I get the indication, right, to invest in this further? And this is even more

prevalent for enterprises where machine learning is a new investment, right? And they may not immediately decide, okay, I want to put like millions of dollars, billions of dollars, like the big tech companies on it, but we have to be there showing them, okay, you know what, this works. How can we help you see this works, but we need, right, we need to work on more iteration on it to be able to get them that trust and the trust to invest more before continuing. Right. I mean, definitely AI is an iterative process. I know I always think about the best data you can collect is kind of a chicken and the egg problem, meaning that the best data you can collect is the data you get out of your model, but you need that data to build your model, right? So there's a little bit of a problem there.

Now, many more enterprises are entering the AI market, so I would presume that the average AI slash ML knowledge across the enterprises is actually going down as opposed to going up. Is that a correct statement? I mean, considering there is a lot of pressure to do AI and ML without really having the right background. And then maybe a follow-up question there is, how do you help? Is it just tools or is it tools and services or is it mostly services that's it no no that's a great question so like what we see is um you have to meet people where they are right like in the industry right now where you see is that there's a bucket who have been investing in machine learning for a very long time right they have their machine

learning teams and uh they know to expect. They know it's not going to take a, you know, like immediately one iteration to take results. These are really highly educated enterprises in machine learning who have been investing in it for a long time. And then the second group, I think you'll see these companies who have tried various options, may not have done very well. So they're skeptical, right? And we'll get to that group a bit more, but they know that they need this, right? They are growing, their volumes are increasing, there are highly unstructured, but repetitive data that they need to solve with AI. They know they have to make it work, but some of the things they have tried didn't work in the past, so they're skeptical. Then the third group is really

about people who know they have to jump on the ML train, but have not done much, and they have lots of convincing to do in their companies, so they want to show some results. All the while, they also can't build the ML teams immediately, right? Like there's not that justification. Probably it's even better for them not to build it, but find partners to be able to solve this problem in terms of cost and value and like peace of mind into the value of their businesses. So what we see right on the first group, it might be, it's a lot more about efficiency and increasing the results immediately for them. And the solutions really focus on providing that high quality data, right?

Fast and like super, super high quality and the tools. So like in that group, like what scale approach is that we have a full product suite. And one of them is obviously just the tooling to help to see for these ML teams to see the results. But the biggest area there when they know they have the raw data and they need high quality training data is what we call machine learning augmented data labeling. And what all it means, right, you're marrying machine learning with actually the humans, right, like the operational side that what we do. So in that case, the models, the base models we have already built in all these different data types that are ranging

from video, you know, images to text documents and 3D, like, or any other sensors data, right? Our base models already can label that data all by itself. And we utilize the humans to make sure that it's actually on par with quality, super high quality to review. And if anything that needs to be fixed, to be fixed, right? But as we're doing that, those base models are also getting better to produce higher quality output in the first go, right? So in that first bucket, it really focuses on their need is getting better data because they have the models, right? And equipping their teams with the tools to improve the models further. The second group is a bit different, right? And in that group, I think what I call, sometimes they're the victim of what we say, the AI marketing noise, because everything they have tried and went in, every website you go in says AI and machine

learning, right? Even in the most basic rule-based systems will say that, so they have tried it a little bit and they got burned, right? That group, you really have to focus on, okay, how can we help you? Like you might have some models, right? We want to support and like provide the output, the ROI you need, but especially the areas you cannot afford to build those themselves, we provide turnkey solutions. So a good example for that is, you know, if they need the data for their models, go ahead and do machine learning augmented data labeling and give them the high quality data. But second, in the areas they can't build it, whether it is right document processing models or like, you know, content moderation models, we give them turnkey solutions so that they can see

the results immediately. And the third group is really about that area who don't know where to start, right? Like we really provide even more technical partnership there, even the conversation, right? Like they don't need to work with us. Let's see where you are, right? And from where you are, here are the solutions that we can support you with, right? Like from our product suite. And that tends to be really the solutions that are fully turnkey, right? Especially in like, whether it's in financial services and logistics

and like e-commerce, it can be like the turnkey solutions we provide to build their catalog so they can reach to their customers faster and at a high quality output or like taking on all of their processing with the long tail and the majority of the documents so that they, their people don't have to do the work and like do the tedious work,

but we provide models to do this, deliver this in seconds at a high quality. So it's like, you can, as you can see, it's more of a spectrum in these buckets according to what the companies need and how you can offer various solutions in your product suite to solve that. So I would definitely say a majority of the solutions we offer are just directly enterprise solutions and products, but for the groups that who want to learn more on how they can partner and how they should even think about machine learning, we do go above and beyond to be that machine learning partner for them. And I think that's quite important to find those partners for yourself. Yeah. And given that, one of the things I've been eager to talk to you about is you've got a lot of

experience in the real world with companies that are trying to do this. And one of the things that we talked about previously was the challenges in and making a machine learning project successful in the real world. And we talked about the fact that very, very few projects actually ever make it from conception to production. And so many of them fail. I think that some of them fail for purely technical reasons, but I think that a lot of them fail for a lack of communication between the various constituencies that are involved in the whole process of building a successful ML model. Given your experience with these companies, maybe you can make our audience feel a little bit better.

How often is it that companies are actually ready to go forward and have good quality data and kind of know what they need to do? They've got a good ML model and they're able to put it in production versus where they kind of fall flat at one point along that path. Yeah, I mean, that's a great question. I would say definitely majority of all companies today are in the second bucket, right? Like they maybe have dabbled on implementing AI a little bit in here and there, but it didn't work or it worked,

but like it's kind of siloed solutions and have no idea on how to implement that for the rest of the company. There are definitely, I think, the people at companies who have invested in this problem so much, like will be in the first bucket, right? They have figured out their infrastructure, how to apply it. Like a lot of the big tech companies have been investing in machine learning for many, many years. And they have full on not only teams to operationalize this, but also

teams to figure out the partners to work with, like the infrastructure to work with, research teams, like to be even 20 years ahead, right? Like what's going to be the next thing we can apply in 20, like 50 years. So, but a lot of the companies, especially I would say, and this is a term I'd like to use, like the industries, the enterprises that are keeping our lives together are more early in this journey. So that can be right, like we go and like you, you purchase something and you want to pick from like these offerings. So like commerce companies, or it is, you know, financial services companies that can change from your own, you know, like financial services industry, whether it's, you know, you're applying for a loan or a mortgage, or, you know, you want to expense your bills,

or as a company, you want to do your accounting, like across the board, or even logistics, like how do you even get the things at our door when they're shipped, right? These are industries that are ripe for this transformation, but very, very early because many times the data that you need to use to automate some of the processes with AI is not even digitized, right? You're starting from there. Like how do I get this data, then understand the business problem, and then put AI to solve that. So in that manner, I think what is the best path forward is that, at least like something that I have seen work great in my engagements with various customers, is that you are the domain expert for me in the business problems that you are trying to solve.

But what I can bring, right, the infrastructure and the processes and the output you need to get there. So that like tight collaboration is extremely important for us to be able to get to the results. And on top of it, it doesn't end there, right? We call this an ML full loop. So as you are implementing these solutions, how do you see and review results to do the next stage of iteration, right? So I think a lot of the companies may be listening to this or leaders like, no, like you're not the only single dot person who may not be seeing like immediate results or you're like

afraid to invest more. Like there are a lot of companies there. I think the best thing to do is like figure out a problem that's high priority enough that you, your team will put in the time to figure it out, but try it out there, right? High priority enough, there is urgency. And if it did work, it will get you quite a bit of return on investment. Then if you see and like make it work with either your own team that you're investing in machine learning or with your partners, if you don't want to do that

internally, at the moment you can produce those results, then you figure out how to expand in your organization. Because in that expansion, right, you need to be what we call AI ready, right? Like is that infrastructure ready? Many times in multiple industries we see, let alone replicating a solution we made it work in other areas of the company, they may not be even sharing data internally within themselves, right? So how do you make that work as a whole if we're even seeing these kind of problems? So there's lots to fix there in the long tail of these companies in various industries. But I don't want anyone to think that they have already missed the boat. There are a lot of companies who are in this boat.

And I meant it when I shared only 6% companies like in our world right now have implemented AI in one way or another in some business problem, there's lots of opportunity to be able to get the benefits from this. Yeah, what I see a lot is that solutions are being replicated in verticals. For example, in pharmaceuticals, you will see that one organization is more savvy on AI

and their processes and their tools are copied over completely as a solution and that those become reference systems across the board. So I think maybe another way to say it is that AI solutions are somewhat contagious reference systems across the board, right? So I think maybe another way to say it is that AI solutions are somewhat contagious within certain verticals. I think an interesting topic is that when you don't have enough data, the ability to create or generate synthetic data, is that something you can talk a little bit about? Because I think a lot of organizations who think, hey, I might not're seeing this company is doing, like doing well in something.

What are we not doing? So it's really contagious in that way. And I think to cut down the time to be able to implement these things, like genuinely look for partners that are offering these solutions for you, like in that industry directly, instead of trying to reinvent the wheel internally,

because that's only going to take time for you to catch up, right? So I would really think about that. And in terms of data, if you don't have enough data, there's actually two paths to go about it. One of them is you have talked about synthetic data, right? How can you get synthetic data and make it work? Actually, like synthetic data gives a huge promise.

And it's actually something we're ourselves investing in to provide people with synthetic data, but at the same time it really depends on your industry, right like, what do you need, like what is your business problem and can it be solved with synthetic training data. requires complex problems in real world where there's a lot of very like different variables and you may not be really catching the edge cases you need to catch and if you don't do this right the consequences are worse uh than uh you know like the limited it may not be the right solution for you right so in that case you want to mix up right you want to mix up like you want to use some synthetic data and you want to see how you can get more data yourself, like collect data. This is why, right, lots of, we work with, one of the biggest industries we work with

is autonomous vehicles industry, right? All of them will make an effort, really collect that data themselves and use synthetic training data as a way to like combine and maybe increase efficiencies along the way if they don't have the data for a specific scenario, right? Or like a specific world where it is hard to collect the data. So like that's quite important. The second way you can do if you're more in the industries like, right, so like financial services or like logistics or like, you know, like in this case, even

going from there, some processes that you can increase like efficiency, whether it's, you're looking at IDs or trying to understand more information when you see something new, instead of synthetic data, you can use that as a combination, but you can also look at solutions where the models are already working, right? So like you don't need to use your data. instead of synthetic data, you can use that as a combination, but you can also look at solutions where the models are already working, right? So like, you don't need to use your data, you may be like, you may not have it. So we do that quite often, especially in the document processing realm. It's like people may not, if you're a newer company, you're growing, or you have a new product or whatever it is, you might have some data and will make use of that data, but you can

directly come in and those data types, like we have already built all these base models that are trained on millions and millions of different types of data already that can start giving you results from the get-go. And then in the meantime, when you're getting those results, what you can do with the new data you collect, right, that given we are both machine learning plus also the human in the loop like operations side, we can start making sense of that new data, like label that to improve your own models only. So you're making use of the data you're collecting, but you don't wait until you collect that data, you already can have a turnkey solution to get results. So I would really suggest it's both ways, like synthetic data has great promise. And second, look at those solutions and partners that have those turnkey solutions and can utilize your coming data. And that should be really combined depending on your use case. Because as I said,

like in, you know, if your business problem is like a life and death issue, you don't want to rely on only synthetic data and only see that your models are obviously not performing in those difficult edge case situations. It's all about the edge case, right? When it comes to AI. And I always think of this as, you know, like the thing that we do, like it's, think about it, like it's a child, right? Like when you do, when you're teaching things to a child, you give them lots of data, you put them in situations where they can collect that data and provide it with the framework to make the judgment, right, when they see something new afterwards. It's exactly the same, right,

with machine learning and you can give lots of data, right, like and really high quality data and you're providing it with the framework to make the decisions and in this case you can support that data with synthetic data as well, which in this case for the child would be maybe the scenarios you tell them in words right you paint out the picture for them to act on but it only goes so far depending on what the business problem is um yeah but it's a very exciting area yeah and i think um a lot of organizations are scared of starting from from scratch right they want to do something specific like maybe in speech words they want to use speech recognition but only want to do like keyword spotting right so they they can start with

a base model and then use transfer learning to kind of build that out now when you talk about creating or providing base models and and sharing data is is are you talking about data when you talk about the models or are you effectively you talking about data when you talk about the models or are you effectively just talking about models without the data that you then kind of provide a customer with transfer learning or any other augmentation? Yeah, it's definitely the second. So like, as I was mentioning,

if you have your own ML teams and models, it can be only the data you provide, but not everyone right now there in our world, the majority is in a place where they don't have these things. So they need the models, right? So what we do is like start with the base models that we have already built and trained on so many different types of data with like both supervised learning and transfer learning. And even like, you know, like to pinpoint, to do active learning on top to make sure that it's performing on specific various data types. But what you can do is you can either provide those base models ready for use. However, you have to

acknowledge that there's so much more you can do there with increasing the quality, right? At that point, we can also do what we call fine-tuned models. So those base models exist, but we'll take, if you have data, your sample data, and really annotate it ourselves to improve those base models just for you and for your use case, right? And those are a lot higher in terms of quality and the problems that they can solve for you without you even having to have a machine learning team, right? So you're getting really specialized, like fine-tuned models with, and it's okay if you have like small amount of data because this still can be done given we already have the

base models, then we improve them for you and you can immediately use it via an API actually. This is why I think we've been talking a lot about democratizing AI. You have to make it so that for a lot of these companies, get in the race and actually use AI to improve the problems, their solutions and the problems they're trying to solve, you have to make it very easy for them to use, right? So like, it should be as easy as you take the API and send the data and receive back to use, right? So like it should be as easy as you take the API and send the data and receive back the response, right?

And that it's readily available for you because if you go the other route of trying to do everything yourself, it is going to take years, right? And especially some of these companies tend to be very big, right? So it's like even trying to go through the internal hurdles to have that investment, it's like even trying to go through the internal hurdles to have that investment

is like more than, you know, three, five years, which definitely you don't want to wait around when everyone is investing in ML. So we really believe in that democratizing access to machine learning infrastructure and machine learning. So you get the solution today and start seeing the results today, right? Like for you to get in the race. We're actually seeing that, you know, one great example of this year, like this past year was in financial services. So the VC industry is booming, right? A lot of financial services companies are coming up, startups are coming up. They have amazing capital and they have been building all the solutions really ML-based across the board, right? Like we're seeing it with companies like Brex, right? Really, they have built all the solutions they have like based on

that technology because they started really like, you know, quite early enough where machine learning was a thing. But now the rest of the financial services industry that might have incumbents trying to catch up, right? So what do you do? You have to keep finding partners to start giving you results today to basically have that merged environment in financial services with these like startups

with lots of capital going so fast. They already believe in ML. They already have partners like scale, right? And the rest, like you have to figure out a way to convince your institution, look, like I have tried it in this one problem, it worked really well, we got to go invest more so that we can even, you know, compete for the attention of consumers who already started expecting everything instantly. No one wants to wait 10 days anymore if you're applying for a loan or if you're opening an account, right?

Like everything right now is based on the demand that's instant, high quality. So how do you get there? As yeah, like you have to figure out and find these partners like scale and start investing in solutions and expand from there to become an AI ready and AI implementing institution. Yeah. And I think that this really, it all kind of comes together with this and that, you know, we've got a lot of

companies who really want to implement ML, but they are, you know, they don't even know where to start. They're not thinking of all the questions. And frankly, if all they're focused on is the ML aspect, they're really missing something if they don't have the data to back that up. So I really appreciate this discussion. You know, again, it's one of those things for our audience, you know, you have to think about this, you have to think about the data that's supporting your ML model, just not just the ML model itself. So the time has come in our podcast now where we shift gears. And we are starting back in here with our tradition of asking our guests three unexpected questions. So a note, our guest has not been prepped for these questions. These are a surprise.

And we're going to get some off the cuff and hopefully some fun answers. This season, we're also going to be spicing things up a little bit by bringing in a question from a previous guest. So let's get started here. You've mentioned the word data a few times. Frederick, do you have a question involving data? Indeed, I do. Will we ever see a Hollywood style artificial mind like mr data or other characters um like in real life uh let's see i think well there's a lot of speculation about this right there's a lot of people who are also afraid what would that mean i think we're quite a long way uh from there like genuinely some of the problems we're working with enterprises right now is how do you even effectively read unstructured data, right? We're there. I think

it's a possibility in our future, but I would definitely say that the ML world and the solutions we're focusing on are not there today and won't be most likely for the next foreseeable timeline. But I personally would love to actually see something like that and make sure that we can really have these characters and solutions take on most of the tedious, repetitive work and provide some insights, right? It's just like the models we're running right now that give us insights. We can still make the decisions, but you want all the information and the insights possible. Yeah. So we're still working on lowercase d data

before we have uppercase D, Mr. Data. So yeah, correct. All right, next question. You've talked a lot about data. You've talked a lot about size of models. How big do you think ML models are going to get? We've got a hundred billion parameter model right now. How big is he going to get it?

Will that look small or have we reached some kind of limit? Yeah, yeah. I actually think this is actually a great question. I think, you know, GPT-3 is one of the like biggest models that have been really built. And recently when we were hearing from Sam, Sam Altman, who talked a lot about how GPT-4 not being higher in size, but really focusing on the specific quality, right? Like off the model. So I think increasing the model size only gets you so far. And already, I think hitting these numbers with, you know, like solutions like GPT-3, we're seeing, right,

like people are starting to focus, okay, increasing it further than this is not really helping, but really thinking about the quality and the parameters themselves will be the next to go. So I genuinely think that will be the next way for these extremely tech forward companies to explore because a billion parameters is already there, right? So how, like, I don't think people are investing in going further than that. All right, next and final question

comes from a previous podcast guest. This following question is brought to us by Alexandrine Royet, a student fellow at the Leverhulme Center for the Future of Intelligence. Take it away, Alexandrine. Hi, this is Alexandrine Royet, a PhD candidate at the University of Cambridge and a student fellow at the Leverhulme Center for the Future of Intelligence.

My question for you is, what do you think is one of the biggest ethical challenges that is that so AI is happening whether we want it or not. It is already permeating our lives across the board. It will only continue, right? And it's not only here in different industries but in the world with different countries and different solutions really adopting it. What we are not talking enough about knowing this is happening and it's continued to happen. How do we deal with bias, right? Because it's real. We face that every day of our lives, even without AI, right? Like the people I'm interacting with, right? They might be biased according to what I look like, my gender,

my opinions, whatever it is, it's happening today in our world without AI, and it is going to happen with AI. So how can you accept that AI is happening, but come up with various solutions and like guidance to the people who are implementing AI to solve that problem? And what I believe actually, it really goes back to the people who are implementing AI to solve that problem. And what I believe, actually, it really goes back to the data, right? It really goes back to the data of, you know, do you actually have enough data across the areas that you want to be conscious of? And when you discover more, when you discover your model might be biased according to like a specific group, so how are you handling that to improve that bias?

Again, it comes to with humans, right? We give them data. We get them introduced to different types of people and cultures to really remove that. And when you spot something, you have a correction. You talk to the person, educate the person. It's the same. So I would love to see a lot more about knowing AI is happening. How does data come in with this implication of bias? And how do we make sure

that the people who are creating this take the necessary precautions to have that data diversity and the continuous iteration on the data diversity as we move forward. Well, thank you very much for that answer, that thoughtful answer to Alexandrine's question. And we're looking forward to what your question might be for a future guest. If listeners want to be part of this, you can join us. Just send an email to host at utilizing-ai.com and we'll record your question for a future guest. So, Melissa, thank you very much for joining us today. Where can people connect with you and follow your thoughts on enterprise AI and data and other topics? Yeah, that's great. So, one way I share

quite a bit on my Twitter account, that's at Melissa TalkMac. But more importantly, there are lots of resources. Recently, we have done a conference, virtual AI conference from scale called TransformX. It hosted amazing folks from, you know, Fei-Fei Li to Andrew Eng, Sam Altman, Eric Schmidt, across the board from different industries, people who are making these decisions day to day to invest in AI or not, how to invest in it. And you can see quite a variety from different industries on the companies to just go to scale.com and on the event side, you can watch all the sessions now, according to your interest. That would be a great place to follow us on our thoughts.

Great. Thank you. And Frederick, what's new with you? Yeah. So I'm currently helping enterprises with data management and designing large GPUs clusters for AI. You can find me on LinkedIn and on Twitter as Frederick V. Heron. And you can find me on Twitter at S Foskett on LinkedIn and other sites as well. And right now, one of the things that I'm pretty excited about is we just had our Cloud Field Day event. So if you go to techfieldday.com, you can see some interesting technical presentations about the future of enterprise cloud computing.

Check that out as well. So thank you for joining us for the Utilizing AI podcast. If you enjoyed this discussion, please do subscribe in your favorite podcast application. Maybe give us a rating review while you're there. Also, please do share this discussion with your friends. This podcast is brought to you by gestaltit.com, your home for IT coverage from across the enterprise. For show notes and more episodes, go to utilizing-ai.com, or you can find us on Twitter at utilizing underscore AI. Thanks for joining us and we'll see you next week.