#307 Human Guardrails in Generative AI with Wendy Gonzalez & Duncan Curtis, CEO & SVP of Gen AI at Sama

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or your team need to adapt and thrive with AI. Whether it's business users looking to get the Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models. You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@datacamp.com.

Welcome to Data Framed. This is Richie. There are now many great AI models that perform well off the shelf. But for some business or scientific use cases, performance is still lacking. Behind every great model is a great data set. And if you want your computer vision model to reliably detect your product, or you want your call center AI to reliably answer support questions that your customers care about, then you're going to have to provide your AI with your business data. Today we're looking at the process of collecting and enriching your data, then feeding it to your AI for better performance. I also want to know about the data annotation workers who are the hidden stars of the AI revolution. I have two guests for you from the data annotation and model evaluation company Sama. Wendy Gonzalez is steering the ship as CEO, having moved into the role from coo. She joined Sama from a background as a product lead at an IoT startup and she's also spent time as a management consultant consultant with stints at EY and Capgemini. In 2023, Wendy made Women Tech Network's 100 Executive Women in Tech to Watch list. Duncan Curtis is Senior VP of Product and Technology. He's got 15 years experience as a Product manager, having previously run Google Play games. There are also products at Zoox and Active. Let's find out how to make great data sets for your AI.

Large language models are really good at getting generic answers right. However, when it comes down to actually doing what we do in our day to day businesses often requires much more specialized knowledge.

So We've moved almost 70,000 people out of poverty since we have started this this company. What we do is we have an impact hiring model where we hire on the basis of impact.

Hi there Wendy and Duncan. Welcome to the show.

Thank you. Glad to be here.

Brilliant. So there are a lot of really great foundation large language models. So why do you need to use your own data rather than just an off the shelf model?

Large language models are really good at getting generic answers right and so they've been trained on a very large corpus of data across the Internet or a large portion of it. However, when it comes down to actually doing what we do in our day to day businesses, it often requires much more specialized knowledge. And so what we find is that while models are really great for things like writing an email, that's a generic email. It might actually really struggle if you wanted to do something within a specific domain like referencing law or referencing, you know, HR policies or even better, what about referencing your own internal documents? So if you wanted to write an email out to your employees, but oh wait, I need to make sure that it actually, you know, follows our policies, that's not going to be as easy to do with a generic model. We do have, we do see people doing something that's called distillation of models. So often when you see large language models come out, you hear like, oh, you know, ChatGPT 4.1 has just landed on Monday. But they also release smaller versions. And the reason they release smaller versions is they take that big model and they eliminate certain parts of it to try to keep, to get that model smaller, but to try to retain as much of that bigger information. And the reason you want it smaller is because you want it to be faster. So when you're actually more real time chat, but you also want it to be cheaper to run and be able to run more on, on cheaper hardware. So that's hope that answers your question.

Okay, so that makes a lot of sense. If you're doing something that's a little bit out of the ordinary, not generic question in some sense, then that's when you're going to want to use your own data or get some domain specific data. Maybe we can dig into some of these ways of making your models faster and cheaper as well. Both those things are quite useful as well. So yeah, we'll get into that later perhaps. All right, so I would love some concrete examples. Do you have any real examples of where a company is trained or they've made made use of data enrichment to improve their own model and make things better?

Yeah, absolutely. So a great example would be we were working with a company in the HR space and so one of the things they wanted to do was train a model to be able to understand what was in job applicants resumes. So they had some matching that they wanted to do where they had, let's call it, I think it was somewhere in the like 180 sort of skills that they wanted to tie to people's resumes. But in those resumes you Call different skills a lot of things. So let's say, let's say executive presentations might be the tag that they wanted to use, but that might appear in someone's resume as anything from presented to the board, regularly present to, you know, my executive team, it might be, speaks at conferences. And so what we would do in that case is enrich a set of resumes with getting those different examples within resumes and aligning them with that taxonomy, that list that they had and then they can use that to train an AI or to basically find new instances of that moving forward.

Okay, so this is one of those high risk use cases. I know certainly under the EU AI act, any kind of HR like AI use cases is considered high risk because the risk of discrimination happening there. So by the way, you've got to be really, really sure that the AI is highly accurate.

And it's not just highly accurate. I completely agree with you. It's also not biased. I mean we've seen things in the past where I think it was, there was some very public ones where it turned out that AI had been trained on what previous good applicants looked like and it was extremely biased to people of color. And so it was quite unfortunate that the way that that model training was designed inherently picked up a set of bias there. And so one of the things that we do not just when we're enriching data, but as we're going through that process is that we also work with our clients to show them the distribution of their data. So you know, in, in something like a self driving car case, it might be as easy as saying, hey, yes, you've got a million vehicles but you only got 200 motorcycles. You really, let's find some more motorcycles within your data or you can collect some more so that you can represent motorcycles. Well now in the HR use case it might be, hi, you don't have a good representation gender, you don't have a good representation of educational background. Are we biasing towards particular schools? Because we're seeing this as, you know, is your training data set up in that way? So we can we help with that as well for our clients?

Yeah, certainly. Even just crunching some numbers on what are the distribution of the outputs. That's going to give you a lot of transparency on what's the model doing. Are you being fair in general? If you're making your own own model or at least modifying your own AI does that. It sounds like it's quite an intensive process. So I'm wondering, is this just something for big businesses or can anyone do this? Maybe. Wendy, can you talk me through which industries or business types are making use of this?

Yeah, so I mean, for a lot of enterprises, they're leveraging these foundation models. It doesn't make sense to build your own. It's, you know, billions of dollars, lots of GPUs, lots of cloud computing. It makes no sense. They're really advanced. You can find a way to apply them to your own use cases and as Duncan said, do things like rag embeddings which allow you to train the model on, you know, your own proprietary data for the right application. So there's a lot that's happening there. But then in some cases your, your business, you know, maybe so core to your business. So some easy examples are self driving cars. It is core to the hardware in your vehicle, the cameras, the sensors and the placement. So a lot of companies are building their own or you know, they might be leveraging, for example, like Nvidia's platform, but they're still ultimately taking the time to build it because it's part of the hardware integration. Some cases, yep, companies are still building it. Or another example might be things like E Commerce. Right. So if you're going online, your product catalog is key. The ability to search for products and to do some of the data enrichment that Duncan had just mentioned, that your products are searchable and you can recommend things appropriately. That's another example to where it's so core to the business of actually selling the product that even if you might be leveraging some models, a lot of companies are choosing to make those investments themselves as well.

I think a good example there would be, say, ThreadUp. So recently we were at a conference called Shop Talk and there was a great presentation by a company called Thredup who basically allows secondhand clothes to be sold. So their product catalog for each item is one. Each item is a unique sku. Yes, they do have their categories, but I think it was when he was at Kendrick Lamar at the super bowl, so ended up with flared jeans and it was like, oh my God, this is amazing. And a new trend again. But they, they're using AI in a really interesting way because they, and their problem is different to what you're mentioning in really large businesses where they have a unique issue where each item is being uploaded in its custom. And so looking at how they can match those and having an efficient system of data enrichment as well as AI to, to help with that, data enrichment as a first pass is really important. Product and user experience straight out of.

The kit okay, so those are quite wildly different use cases. Certainly I can see how a self driving car, you don't want to be sending your video data to chat GBT and saying well should I turn now? Yeah, it's going to be a terrible thing. So you want to build your own solution. But the retail example also. Yeah, I can see how you got loads of different products and you need to have I guess, product descriptions for everything. You'll have photos for everything and AI assistants, there's just managing or generating content. I'd like to know a bit more about what data enrichment involves because I guess in my mind is like the sort of Amazon Turk type situations where you've got a lot of people just staring at photos and writing descriptions of them. How accurate is that?

Oh, data annotation. So data annotation and enrichment, it's really about tagging basically or enhancing unstructured data. Okay, so it could be visual data, it could be text data. It's about adding additional attributes. And the reason why that's important is number one, you either are training a model to perform in some ways on a self driving car is pretty clear. You need to recognize lanes, you need to recognize traffic signs, you need to recognize pedestrians and other vehicles to operate the, to do things like collision avoidance. It also can be used to validate that your model's performing properly. Okay, so you, you need to know what does good look like? How is your model performing against what the expectations are? And so that's really what data data annotation is about. And oftentimes in a number of cases it's not just about capturing the right sort of class or the right descriptions. It can be enriching that data. So I think a good example of this, just to kind of take the parallel from the retail example is think of like Spotify or think of like Apple music, right? So they had to basically build a genome of what is music. And there are a lot of attributes. Right? I mean this is where I suddenly realized when we're talking about all the different descriptions of like you know, K punk, you know, I, I wouldn't even know what K punk was like 10 years, you know, like two years ago probably as an example. But there, there are so many different genres and then you, from the, the song itself, the, you know, the, the AI needs to say hey, where does it fit? And it's, it's literally GM. I mean there may be like 50 or 100 different attributes that can make something be versus a K metal versus a K pop type of song. Right. Product catalogs run very similarly different products, lots of attributes. So when it comes to being able to search, you know, recommend, identify, to allow the AI to really perform as optimal as possible, you need to not only tag it, but oftentimes you need to enrich it.

Okay. Yeah. So it's not just writing descriptions, it's about, I guess, feature engineering stuff on certainly with music. Okay. So you got the people who are in the band and I guess the instruments they're playing and there's probably some kind of chord progressions and tempos and yeah, this could very easily become a very complex modeling situation. Just have subtle differences between genres.

And it's interesting because, like when the reason you want to have this depth of attributes or genome as, as Wendy was pointing out, is because it allows that search relevance or that recommendation, the end use case, to be so much more personalized. Whereas if you just said, you know, if you just started with, oh, it's music and it's Kpop and that's it, and that's all you had about it. It's. You don't. It makes it a lot harder for that next recommendation of a really great song to find its way to a listener or for a really great product that's going to fit a need for someone or for products to be matched near, near each other or shown together. It really powers a lot of that roi. I'm going to say for the user experience where getting good ads is like, or good recommendations is such a better experience than getting just something random or something near what you were looking for.

Yeah. And bottom line, it increases, you know, cart sizes and purchases online. So it's core, core part of it, and that's really where we typically see AI being implemented is if it's going to make a core differentiation in your product or it's going to help you save money.

Oh, absolutely. I have to say a lot of the streaming music websites do this incredibly well, but for many retail websites, absolutely terrible. So the experience of searching something like, well, I know what I want, I can't see it in the results. And so, yeah, I can see this is incredibly important and a lot of shopping websites really need to take note. So hopefully anyone in the audience working on a website. Yep, better search is a good thing. Suppose you're interested in this. I guess to start you need to find yourself a data set. So where does the data come from? If you decide you want to get some data to feed into your own model, what do you do?

Data can come from so many places. So let's say if I was taking a retail example, a lot of that information is going to be already there. So it's the product catalog of all the images and descriptions that your vendors have uploaded. And so you've now got access to that data and you can enrich it from there. If we're talking about something like self driving cars, you need to actually have a fleet of vehicles models, you need to put cameras and lidars on them and drive them around and actually record data. But there's also, that's. So that's for the raw data collection. But I think we should also touch a little bit on what kind of mix of synthetic data can also be part of your AI solution. When we look at how models are being developed, especially even the large language models, we're now seeing synthetic data from the generation before being a large portion of the training data for the next generation. Because as the data needs are exploding in terms of they're running out of the Internet to continue to scrape and so they're generating more, more content, more specific content from. You know, if you take for example with llama 4 out, llama 3 actually produced a large portion of the debt training data for llama 4. And so you can see and if I, if we go to self driving cars for example, being able to generate using top of the line game engines or custom like if you think of, if you think of the movies and video games, the top of the line ones we have today are photorealistic and so they're able to use those technologies to also simulate more hours of driving and see how the vehicle couldn't react there. Same with your product catalog is that you might have, you might be going into a new category that you do not actually have information on. You can actually create synthetic data from, you know, image generation as well as product catalog description. You can actually build a since synthetic data pipeline in order to be able to have new training data, let's say ahead of a new category that you're going to have on your website. And so the funny thing about synthetic data though is that you have to be careful. Like what we were talking about before is taking inherent bias into your model. And so it's really important to have your synthetic data validated. Wendy, do you want to talk a little bit about some of the.

Yeah, I was, I was going to actually just touch on the exact same thing that you can leverage a model, train a model. But that model training that model could carry forward biases from the previous model. So at the end of the Day, you still have to know what good looks like. You still have to know how you evaluate your model. And even if you're leveraging something like synthetic data, is it in context, you know, is it realistic, is it achieving? So you know, number one, you have to say, hey, what is I'm trying to build in terms of my data set? What's missing? What can synthetic data support then? Second, is that synthetic data going to be a, the quality that we expect it to be? So what happens oftentimes, especially in like image generation, you might like have a phone, but maybe it's floating exactly one centimeter above the table. Right. There are these things that, you know, it is not perfect. And at the end of the day, the quality of the data is what matters most. Whether it is captured or synthetic or generated by a model. You still need to know, how do I evaluate it? Is it accurate?

Okay, that makes a lot of sense. So if there are cases where you don't have good real data or is very expensive to collect, then yeah, I think that's going to be incredibly useful. I heard a case a while ago, it was from the US Air Force where they were trying to train their drones to do something and they said, well, we only ever fly drones during the day, so we don't have any data on how well they perform at night. So they had to deep fake nighttime flying in order to generate a data set that, that they could then train on. Yeah, I like that idea. Certainly I can see how if you're generating garbage quality data from a model, feeding that into another model, it's just got worse and worse and worse and that's not going to benefit anyone.

Yeah, it's like a copy of a copy kind of a scenario. And we see it. Synthetic data has good uses or great uses in edge cases. So imagine like nobody wants to real time capture a stroller in the street or an empty stroller or stroller with a baby in it or I heard one actually from a, from an auto company client where they're like, you wouldn't think of this but sometimes hogs like drop out of the sky and get in front of vehicles. That's probably not something you're ever going to catch. It has happened before. It's probably not something you're going to catch, you know, in any regular real life. So it's a great example. Some edge cases where it's like, okay, synthetic data makes total sense.

Okay, that feels like a very Texas scenario somehow. Yeah, I can see how it's gonna be very hard to create that kind of real, that data in real life. But yeah, you might need to deal with it at some point. So it's important to have a data set. So suppose you've got your data, some of it's synthetic, some of it's real. What next? I guess you have to choose a model to enrich.

Yeah, there's some, some really good benchmarks out there. But one of the really nice things about where we're at today is that it's actually very easy to just try a model out. And so being able to pick the right model for your use case can be as simple as actually running through a bunch of tasks or of the type you want and seeing what the performance of those base models are. You know, for example, Claude is extremely good at language, writing emails and writing personalized responses. It, it's a really. It, it far outperforms others, even ignoring the benchmarking. It's certain, even more so than the benchmarking would, would, would show, for example. And so what it really comes down to is test out the existing models for your use case and see how it feels. Because at the end of the day this is a, almost all of these come down to a customer experience for a lot of these. And so being able to get those intangible human elements of that sounds right, that sounds better. That's not messing up in this way. Also set up a. Before you begin, you should also set up how you want to assess these. And so if you're doing a like an image detection model and you wanted to start with something like yolo off the shelf, great. You can also, you can start with your metrics being like, oh, okay, well I care about, do I care more about accurately detecting everything all the time. And I'm okay with detecting some things that are the wrong things. Or is this case actually it's really more. It's okay if I miss some things, but I should only detect the thing that is the problem that I'm looking for. So you can set up those metrics for yourself before you begin your investigation.

Yeah, I, I like to just attack onto that one thing I share often with people are kind of contemplating, well, which model should I use? Is models change all the time. Okay, 4.5 just came out. Right. And there's gonna be something that comes out in another two weeks and probably an open source version of it. Right. Or an open weight version of it, I should say. It'll happen. And the one content though is model valuation. You have to know what are you trying to Achieve with your model, what are the right outputs that I'm expecting, What are the appropriate parameters? And then the, you know, the models are like widgets. They could come and go. And what we're seeing actually a lot more is that companies are looking for flexibility. They know that there may be a next better thing that's coming out. So how do I ensure that, you know, I have the appropriate sort of, you know, forward and backwards compatibility and an ability to, to swap something out if something better comes along, which inevitably will probably already has today. You know, they're coming out like every single day, literally.

That's useful to know that you need to bear in mind that you're probably gonna swap out your model at some point, probably fairly regularly. But there's so many models around, it's relatively easy to try different things. I guess the trick is just make sure you've got something like some measure of how good it is. So if you swap it out, you're not gonna make things worse accidentally. You can have a way of measuring, is this gonna be better on whatever the use cases?

Yes, yes. Especially as, you know, models get more. Because it's all about cost, right. So certain models can be cheaper to run as the technology improves. That's going to be a constant evaluation. Just like people think of like cloud optimization, you know, similar kind of concept is it's always going to be evaluated for the ROI and the cost.

Okay, yeah. And I guess here we're mostly thinking about the inference cost. Is that the primary expense here?

Yeah, that'll be your primary, your primary cost there. Which is, as I was saying earlier, one of the reasons people choose smaller models. And for example, you might need a large, larger model. Like, you know, you might be up in the several hundred billion parameters at the moment, like with your use case, and then something like a deep seat comes along or, you know, two or three generations later, you can actually get away with a 10 billion parameter because it's more specific or it's more, you know, and so you could really drop your inference costs over time, I guess.

If you have more specific data, is that also going to allow you to use a smaller model compared to having general purpose one? Like, do you need a larger general purpose model compared to a smaller, more targeted model to get the same accuracy?

That's a really good question. I think the size of model generally relates to the capability level, the general capabilities you need. So if you just need it to kind of know how to, how to talk and how to think about the world, okay, you can Go smaller. And the other side of that is when you said, is your data really specific? I would look at it more as, is the data you have or can create or you know, as we talked about before, does it really encompass the, the entirety of your problem? So if I took. Took an HR example is that maybe I'm doing that resume, resume piece where I'm just matching keywords to a taxonomy. But do I need my model to do more about understanding the, the maybe the history of someone? Or I needed to understand all of the different educational institutions in the world. Is that really captured within the data set that I've got? Or, or how much am I relying on the LLM or the AI model to have that general understanding there? So I say the more complete your data, your data is for your problem. Another way you can also go generally smaller.

Okay, so this sounds like it's gonna be something a lot of people might fall over because there's very subtle differences in questions. What university did this person go to is very different from? Is this a good university? Was it a good university at this particular time?

At that time? Yeah. I mean, wow. And it's something that you could create or have that data available as part of the models of what the model has access to. Or you could be like, hey, and do you want your model going and checking the Internet every time that it's doing a, you know, a check here? That inference cost can go much higher if you need to check each university, each resume that you're looking for to go figure that out at the same time. Which is just interesting to think about the cost.

Absolutely. Okay, so I guess the last step is how do you combine all your own custom data with the LLM? So I know there are a few techniques for this. I mean, the simplest thing is just use retrieval augmented generation rag. And there are fine tuning options and a bunch of other techniques. So when do you choose each of the different options to combine your data with the LLM or other AI?

Yeah, so I mean with things more like a, like, with simpler models, like say if you're using a vision detection from like YOLO or something like that, it can be very cheap for you to actually add that data and do a retraining or a fine tuning of the model. But when we're getting up into that LLM space is that it's. Fine tuning is probably not the right term as people think about it. More from the sort of like traditional, like CV vision models. It's much more about how do you give the model the Right context and the right access to the right data. So more in that sort of rag kind of approach where you're like, okay, here's my document storage. And even for some use cases, it can, even with the larger context windows we're seeing now, we're actually seeing some clients that are able to incorporate the data that the model needs into the, into the prompt at inference time, depending, once again, depending on how large the data set size is. So you, those are the two common ways that we've been seeing for smaller problems that only require like a few documents worth for context versus ones where you're like, no, I've got millions of images that need to be, or millions of documents that need to be included and, or, or are regularly changing. That's one of the other advantages of a, of a RAG system is not having a fixed, like it stopped at a point in time. You can update the data sources as, as the world evolves and you can keep your model continuing to perform there.

Yeah, we, we see. I mean RAG embeddings is definitely, I think, the predominant, the predominant approach. I, I just, I say that because supervised fine tuning for one, models change all the time, right? So as soon as you choose to fine tune a model, it kind of limits your ability to transition to another model. The second component is that it takes a decent amount of expertise. Right. And a lot of these models are, you know, they're open weight or they're, they're closed. Right. You, you don't really know what's happening in the background to really understand those models and have the expertise to fine tune it. In addition to the fact that it makes it a little bit harder to swap out or you might be losing some sort of ROI in terms of being able to transition to another model. We're seeing that rag and prompts, like Duncan was saying, is really the most predominant way most companies are thinking about it now.

So going back to your threadup example where anytime someone uploads a new clothing item, you don't have to completely retune your model. So you say, okay, I just want to add it to a data store and then that's done.

Yeah, just pull it in. Exactly.

All right, cool. So from that, I guess the other big thing is how much can you just leave these models to just go and do stuff on their own? So when do you need a human involved? So what can you automate and what can't you automate in these situations?

I think automation is a really interesting topic. I think that automation should always be, you should always ask yourself the question, what can I automate? And start with automation as your first place. What you then should be asking yourself is where do I definitely need humans involved here? Whether that's on the data enrichment side, whether that's checking the synthetic data quality, whether it's also in an ongoing manner. So one of the things that we found, you kind of mentioned that how much can you just let them go? And with the rate of change that we see in the world and the data sets for people, is that having an ongoing model evaluation or validation in an ongoing nature is a really great way to be able to continue to keep your model performing at a very high rate. Let's, let's take say Thredup as an example is that if as fashion, maybe their models are performing extremely well now, but maybe they weren't actually that great at the. Sorry, what were the type of jeans again? They were the flare jeans. Flare jeans. Maybe flare jeans actually wasn't something their model was actually performing particularly well at. But it's become a really popular thing because of, because of that super bowl halftime song. And so what you could detect on an ongoing basis is hey, what are the most popular categories people are looking for searching for and are we doing well at that data? And so having a set of humans who can actually on a regular basis plug into how your to your use case and how it's evolving and check that and see that the model is performing and then being able to do something with that response like actually it's not performing really well. Okay, well let's make sure that we go look at some of the data for, for flare genes and actually go and enrich that. And great. Now the model is back up to performing. So we see it more as an ongoing process because once you've found that business value that sort of like, hey, this is actually really valuable to customers, it's, you know, increasing my cart size, it's, you know, lowering card abandonment. What, whatever the business value you're gaining there, you've got to ask yourself for each percentage of performance in that model, what, what's it going to cost me if am I willing to accept 5% less, 10% less, where 1%? Like does each percent mean a million dollars a year in difference or $10 million depending on the scale of the company? And so then you can offset that with keeping, having that human model validation occurring.

Yeah, keeping, I mean that, that's really the key is these AI models, they're experiential as well. Not only do they learn you know, there are new edge cases, new data that comes up like flare jeans or like whatever the, you know, the latest and greatest current trend or event is. But beyond that it learns something from that. You need to make sure it doesn't go off piste. That model evaluation is really important because data is constantly changing and you need to know is it still performing as we expected it to perform? Another example to think of simple one is like the self driving car. You may have a car that works really, really well in you know, Pacific Highway 101 here in California, but you take that, you put into, you know, Rio de Janeiro where you also have coastline and you roads and maybe it doesn't do so well. And plus it doesn't really understand Portuguese or doesn't understand this types of the cars. So there is going to be edge case data that's going to be important. And given that these technologies are global, like that's kind of the trick. So like an agriculture example that is, you know, like wheat in Kenya looks different than wheat in Russia. You can't just assume that the model that's been trained. Right. Knows what the differences are. And that's the thing about these models is that yes, they're, they're trained on the entire corpus of the Internet. But does any believe the Internet is either unbiased or has every single piece of data on it for every application? Right. So that's the other way to think of it.

I'd also mention Wendy, that we think about automation in the same way as well. While we're focusing on that human validation element, we actually, we have automation that's built into our processes as well. So if we were like, if we take that thread up example is that if we were doing data validation and saying, hey, do we have the right category for this new product item? Is that we're not just going to sit there and have someone go, okay, let me look at the entire list of 10,000 options for what category this could be. No, we will introduce it. We have an LLM at the beginning saying hey, I think it's one of these three categories. And that allows us to capture that human insight in the most efficient way possible. And so that's how we really think about sort of human in the loop in automation.

Okay, that certainly makes a lot of sense. If you've got severe consequences when the AI makes a mistake, then you need to, well, you probably need to have both things. So you're going to need to have a software monitoring tool that says, okay, the AI has done something Wrong. And then you're going to have to have humans in as well just to make a decision on what to do next. Verify that there really has been a problem with the AI. I'd like to know a bit about how you go about implementing these things. So suppose you say, okay, we need to have our own custom systems. We're going to use our own data in an AI. First of all, who tends to be in charge of these projects? Is it going to be what, Chief technology officer, chief AI officer, chief data officer, who tends to run this sort of thing?

Yeah, it generally comes from the top down in terms of how the need for the business to say, like, hey, we need to be further ahead in AI. It's a very common thing that we're hearing. I'm sure a lot of the audience are like, my boss or my board is telling me, like, why are you not using more AI? That's generally where it initially comes from. But in terms of those running the projects, it goes down the chain where we do see a much more prevalent set of like a Chief AI officer as a much more common title. Now we're seeing that it then goes down to sort of like your directors of engineering or a director of, of product who may have a particular surface area or part of a product or service that they, they manage. And that's when the project starts becoming more real because you've actually got people who are, you know, partially, if not directly responsible for the P and L of a particular area. And they're like, great, okay, I know what I need to move from a business perspective. I have a target area because I know that, you know, let's say it was thread up, for example. It could be something where you've got, okay, well, I'm in the recommendations team. Okay, well, I know that our recommendations are performing like this and I needed to perform at a different level. Okay, now I can get in and start saying, well, let's try to address this with AI. What kind of model, what kind of data are we going to need? Oh, wait, I need human enrichment. I should probably, you know, reach out to Sama.

And with the fact that make AI easier to adopt, we're seeing also a bit more on the business side. So it's not unusual for a head of customer experience to say, hey, how am I going to integrate, you know, Chatbot, Omnichannel experience. Right. So before Iman or experience, he's reaching out.

Yeah, I think it's very true that stuff like this tends to start right at the top of the business. So yeah, there's always a CE going yes, let's have more AI. And then it has to filter down to all the different departments and yeah, I guess a lot of different departments end up having to contribute in some way. I'd like to talk a little bit about SAMA itself. I know we have B Corps status. So just for the international audience, can you explain what is a B Corp and tell us why did you choose this approach? Wendy, do you want to take this one?

Yeah, yeah. So a B Corp stands for benefit corporation. So Asama, we are a public benefit corporation and it's basically a designation that states that a company has both a profit purpose. Right. But also a social or environmental purpose as well. And so it means that you can do. It's the, the double bottom line or the triple bottom line. B Corp and the B Corp Certification group is basically the best globally known standard for how you about evaluate companies who have purpose and profit. And so we are B Corp but they have a really rigorous evaluation process that is based off of the United Nations Sustainable Development Goals. And in our case we are really focused on workers and impact. So part of our core social mission is to bring people from underserved communities into the digital economy by providing them not only training but full time employment not on the basis of just their work experience or their resume, but on the basis of their impact. And then we provide the upskilling and training.

So actually that sounds pretty wonderful. I do like they have a social mission. So can you describe some of the positive impacts you've had so far?

Yeah, definitely. We're super proud of this. So We've moved almost 70,000 people out of poverty since we have started this, this company. And what we do is we have an impact hiring model where we hire on the basis of impact. So that means household incomes that fall below the World bank poverty standard. Really the key here is that there's so many talented people, lots of just incredibly talented people, but they lack the opportunity, opportunity. And so SAMA means equal in Sanskrit. We are really just trying to level the playing field by providing job opportunities to people who would have huge barriers otherwise. But they're still really talented. So we open the door but our great and talented workforce is the ones who walk through it.

That's very cool. And do you want to describe some of the jobs that these people are doing?

Yes. So we're doing a lot in the data annotation validation space, Data insights. But, but I'll give a couple of examples. So we have people who are Helping do annotation for self driving cars. So that could be, you know, complex sensor fusion, 3D, LiDAR and 2D data to help detect, you know, vulnerable road users or traffic signs. And it sounds fairly simple, but it's not. I would actually challenge anybody to try to run some of these. It requires no depth perception, like lots of, you know, complex taxonomy. In other cases we're doing validation. So an example that Duncan was using earlier is you are building, you know, you, you're, you're, you have a retail application, you have images, but the images that were provided by your, you know, individual, you know, small business, don't include any descriptions. An LLM auto generates the description, but it's written out of context or has weird grammar or it's not exactly accurate, it's chartreuse instead of green or you know, examples like that. And so the human loop might be there to help validate the reasoning, the context, the accuracy. And then they're also doing things really intelligently. So it's not just about, okay, let's annotate this or annotate this edge case. But it could be the evaluation of the data set. So Duncan's example of hey, there are not enough motorcycles or We've noticed that there's a really common failure point with your model. It always fails when it sees, you know, a red, I don't know, vehicle. I'm making that up. But the idea is that they're evaluating, understanding the data, then providing insights back to our clients on how they can improve the accuracy of their models.

Yeah, so I do feel like these are the unsung heroes of AI. So because you've got all these sort of rockstar AI researchers earning a fortune, but then actually in terms of model quality, just getting correctly annotated data in the background is incredibly important. Probably has an equal impact, I think.

Absolutely. I mean it is still true to the state. It really is garbage in, garbage out, like data is king. You have to make sure you're using the right data and that data is valid. That's, that's the entire goal behind the social mission is there's a lot of business value that's being created. Want to create a lot of employment at living wages along the way and let a larger group of people participate and, you know, the, the effects of the digital economy.

All right, wonderful. So actually I have a tricky question because this is the one issue we tend to skirt around a lot when we talk about AI. Since you're running a socially conscious company, I wanted to ask you about It, a lot of AI, particularly generative AI is incredibly energy intensive. So there's this trade off between do we go all in on AI and just kind of burn the world or how do you deal with that? Do you want to talk me through how you think about these energy problems at sama?

Yeah, it's really interesting. We, we all, we also are actually part of the science, space, climate action. So we have a, we have a sustainability objective as well and we're carbon neutral in all of our North American offices and we are working towards that with our locations in East Africa as well and, and tracking and measuring that too. It is really interesting. As you can see, the largest tech companies, right, the hyperscalers are literally doing things like buying Three Mile island and building data centers and billions and we got some pretty interesting data points that you know, even the, the average price of, you know, of utilities and energy has actually affected the average person in Ireland because there is so much energy that's being used because so many tech companies are based in that area. So I think it's an under discussed topic that is something that should, deserves, deserves more because at the end of the day the more sort of commoditize and lower the cost the AI is going to, you know, it takes to be, to be built. Right. So you know, the GPUs are going to get cheaper, they're going to get smarter and yes, they'll use less energy. But what happens on the flip side is that adoption goes up was the first thing Amazon says. They said we're going to, you know, be building more in terms of cloud computing and we need more energy. Right. Because a lot more people end up leveraging AI. So I don't know that I have a good, like hey, what do, what do we do to manage this challenge? I think it is something though that needs to be really raised because at the end of the day a lot of people believe, and I believe too that you know, economies are going to be driven by how, how intelligent people are at AI, right. How these countries are at AI. If you don't have a policy review on how you're going to be leveraging it, it will and is contributing to the, you know, climate crisis. We already have.

Yeah, I kind of agree with you on this. There's probably no easy solutions at all. But I do like the idea that legislators need to start thinking about this and I guess any companies that are using this also need to think about what they're doing.

Yeah, and some are, you know, I think some, some, some have really clearly stated kind of Net zero approaches. And so I think, you know, it's, especially in this day and age, it's, I think we'd be hard pressed to see regulation being, at least here in the US A, a top priority. But I do think the, you know, the notion of both standards and talking about it is going to be important because then each company can choose to take their own, you know, actions towards this. So here, here's a challenge, I should say, with not having any policy or regulation is we saw the leaps forward in green technology take place when California set some really important emission standards. Right. And then all of a sudden everybody's motivated to think really intelligently about green technology. I think there is a role for policy to play in this. Whether that will happen in the next few years here, probably, maybe not, maybe not here in the US But I think it's a very important consideration. And we're seeing some companies lead the way. They're really challenging to get Net zero and challenging their suppliers, like us, to get Net zero.

Yeah, I like the idea that just one part of the world needs to come up with some regulations and then it provides inspiration for the other parts of the, the world. But yeah, maybe this feels like a whole separate episode.

It probably is.

So, yeah, the audience has to watch this space. We'll see what we can do. All right, so just to wrap up, do you have any final advice for companies who are interested in doing data enrichment?

One of the things we haven't discussed, but I, I think we've been discussing components of it, is this notion of responsible AI. And responsible AI isn't the same thing as like ethical AI. What it means is it's really, have you built the AI in a responsible way? And it's typically around four different pillars. It's around data governance. So what data are you using? Where'd you get it? Whether you have the right to use it or not. All that's coming into question, I think, with everything going on right now. But still, do you have the right data? Do you have the right sort of privacy and security? Do you have a way in which you can evaluate your model? How do you evaluate it? Right. Some high risk applications require a different level of scrutiny, maybe some human insight, et cetera. So lending, systems, safety, applications, you know, et cetera, defense, you know, cybersecurity, all those kinds of things. And so at the end of the day, developers are the ones who build this. And so I'm definitely an advocate of making sure that those pillars of responsible AI which are really just development best practices are really communicated because at the end of the day it's the developers who's picking that model, who's building that model that needs to be aware of these practices so that you can really sort of measure twice, cut once, if that makes sense, have a plan, have a way to evaluate it, be able to do so knowing that you're going to get to the right, perform an outcome. If you leverage some of these best practices out there, you can make sure that you know your, your bot or whatever it is you're building, you know, gets red teamed. Right. And doesn't say or do something inappropriate. That there are ways in which there you can put some appropriate best practices into your development. Much like you would for security. Right. You don't want to leak your customers data, so you're going to build an architecture with security in mind. I think of responsible AI development practices is the same thing.

I'd add one more thing I love that Wendy, is especially for potentially some of your audience, depending on how far along they are in their AI journey, is that when you're thinking about this is do some research or talk to some experts like ourselves, that doesn't have to just be a plug for us, but really there's really good resources out there. There are people in the industry, people like ourselves who can help, who have seen this, like we know what good looks like. We've been there, we've done that for extremely large names named companies and we've seen what helps them drive success is that there's help out there. You don't have to figure it all out by yourself.

Yeah, certainly both great ideas there. I like the idea of responsible AI. You don't want to expose yourself to business risks by doing something stupid. And again, same with your punt, Duncan. It's like, yeah, you don't want to do something stupid like talk to an expert, have a think about what you're doing before you just dive straight in there. Which is a shame. I always hate planning, I always want to dive in. But yeah, it's much more sensible to think first and then ask for help if you need it. So finally, I always want recommendations for people to follow. So I'd love to know about whose work you admire at the moment. Who are you listening to?

This has nothing to do with AI, but I've been listening to Michelle Obama a lot. So nothing to do with AI, but that's the first thing, first thing that came to mind. Go ahead.

Okay.

I actually thought it was going to be a KPOP recommendation.

Yeah. I mean, I haven't been following individuals as much, but I've been following trends. And some of the things that I find really interesting is the introduction of, let's call it more standard engineering practices as it applies to LLM. So we saw this with Deep SEQ as it came out where they took an approach of different innovative, not even necessarily super innovative ways, but instead of making the goal, how do we make the biggest model, how do we make the most efficient model? And seeing different ways to tackle that problem, that's been super fascinating to me as I nerd out in that space.

And I have a serious answer. Actually, Michelle is the first one that came to mind. But Andre's Karpathy, he does a great set of series. I think anybody who is really trying to understand more about AI, the technical concepts of AI, he's I would subscribe to his YouTube channel. It's excellent.

Yeah, both Michelle Obama and Andre Carpathi definitely great speakers, although on very different topics. And yeah, there are a lot of amazing foundation model researchers speaking about their work as well. And I do like the idea of making things more efficient. Saving money, me. But on that note, I think we have to stop now. So.

Yeah.

Thank you both.

Thanks, Richie.