In this episode of the Innovators Inside Podcast, Wendy Gonzalez, CEO of Sama, shares a practical look at what it really takes to build trustworthy AI in the real world. For founders, operators, and innovation leaders, this conversation goes beyond the hype and focuses on the systems, standards, and leadership decisions required to move AI from experiment to production.

Wendy brings a rare perspective to the conversation. As the leader of Sama, a company that helps enterprises improve AI through data annotation, validation, and human-in-the-loop systems, she sits close to the real operational challenges behind AI deployment. The episode covers everything from data quality and edge cases to responsible AI, regulation, language bias, and trust.

Here are the five key takeaways from their conversation:

1. Great AI starts with great data

One of the clearest lessons from the episode is that AI performance depends heavily on the quality of the data behind it. Wendy explains that “dirty data” can show up in many forms, including missing information, inconsistent labeling, weak attribution, and incomplete context. When that happens, model outputs become less accurate, less useful, and less trustworthy.

This matters whether the use case is e-commerce search, recommendation engines, autonomous vehicles, or generative AI. If the training data is flawed, the model will struggle in production. Founders and innovation teams often focus on model selection, tools, and interfaces, but the underlying data layer is usually where success or failure begins.

For operators, this is a reminder that AI strategy is not just about choosing the right model. It is about building a strong data foundation. Better inputs produce better outputs. That sounds simple, but it is still one of the biggest competitive advantages in AI.

2. Human-in-the-loop is still essential for production AI

A major theme of the episode is that human-in-the-loop systems are not a temporary workaround. They are a core part of building reliable AI. Wendy explains that while models have become far more capable, there are still many situations where human judgment is necessary to annotate edge cases, validate outputs, and improve model performance over time.

This is especially important in higher-stakes applications. In areas like autonomous driving, financial decision-making, product search, and enterprise workflows, teams cannot afford to blindly trust model outputs. Human oversight helps companies catch errors, improve quality, and maintain trust with customers.

The broader lesson here is that AI is not simply replacing people. In many cases, it is creating a new operating model where machines and humans work together. The strongest AI systems are often not fully automated systems. They are carefully designed systems where human review improves reliability, especially when the cost of failure is high.

3. Edge cases and quality standards define whether AI works in the real world

Wendy makes an important distinction between AI that looks good in a demo and AI that performs in production. The difference often comes down to edge cases and clear quality standards.

A model may perform well under common conditions but fail when it encounters unusual scenarios, local differences, language nuance, or unexpected user behavior. Wendy uses examples like autonomous vehicles navigating new road environments and e-commerce platforms handling complex taxonomies. These are the real-world moments where AI systems are tested.

Her advice is simple but powerful: define what “good” looks like before deployment. If a team cannot clearly explain the desired outcome, it will be difficult to measure model quality, improve accuracy, or know when the system is production-ready.

This is one of the most practical takeaways for innovation leaders. AI evaluation cannot be vague. Teams need clear benchmarks, feedback loops, and review processes. Trustworthy AI requires ongoing calibration, not a one-time launch.

4.  Responsible AI is not just a policy issue. It is a business issue

Another strong takeaway from the episode is that responsible AI is not only about ethics or compliance. It is also about product quality, customer trust, and adoption.

Wendy frames trust as a business outcome. If an AI product performs poorly, gives wrong answers, or behaves unpredictably, users lose confidence quickly. Once that trust is damaged, adoption becomes much harder. This is true for chatbots, recommendation systems, financial tools, and any customer-facing AI experience.

She also highlights the growing tension between innovation speed and regulation. Different governments are taking different approaches, and AI policy is evolving unevenly across markets. But even without perfect regulation, companies still need internal standards for responsible development. That includes knowing where the data came from, documenting how models are trained, setting evaluation criteria, and testing whether systems operate within intended boundaries.

For founders and operators, the implication is clear. Responsible AI should not be treated as a legal checkbox. It should be treated as part of good product and operational design.

5. Leadership in AI requires humility, clarity, and continuous learning

One of the most memorable parts of the episode has nothing to do with models or data pipelines. It is Wendy’s leadership advice around “firing yourself” as a CEO each year. The idea is to step back, ask what the company needs from its leader in the next stage of growth, and then assess whether you are showing up in the right way.

That mindset connects directly to the broader AI conversation. In a fast-changing environment, leaders cannot assume that old playbooks will keep working. They need humility, adaptability, and the willingness to question their assumptions. They also need to stay intellectually engaged rather than blindly trusting tools just because they are powerful or popular.

The episode makes the case that critical thinking is becoming even more important in the age of AI. Leaders need to ask better questions, define clearer goals, and resist the urge to hand over judgment too quickly. AI can improve productivity, but it does not remove the need for strong thinking. In many ways, it raises the standard for it.

This episode of the Innovators Inside Podcast offers a grounded and highly relevant conversation about what it takes to build AI that actually works. Wendy Gonzalez brings operational depth to a space that is often dominated by hype. Her perspective is especially valuable for leaders trying to move from AI experimentation to AI execution.

The biggest lesson is this: trustworthy AI does not happen by accident. It is built through strong data, human oversight, clear evaluation, responsible practices, and thoughtful leadership.

For anyone working on AI products, enterprise innovation, digital transformation, or machine learning operations, this episode is worth paying attention to right now.

Have a question for a future guest? Email us at innovators@alchemistaccelerator.com to get in touch! 

Timestamps

🤖 What it takes to get AI into production 00:00

💡 Surprising leadership advice and firing yourself as CEO 01:41

🚀 Wendy Gonzalez’s path into AI and Sama 07:16

🧹 What dirty data means in the real world 12:15

🏗️ How Sama helps companies train and validate AI 15:29

🌍 How AI work creates economic opportunity 18:16

🔍 What is really happening with AI right now 22:04

⚖️ Should AI access be democratized? 25:19

🗣️ Why language and cultural context matter in AI 28:30

🏛️ Regulation, trust, and responsible AI 36:20

📈 How enterprises measure AI quality and risk 42:27

💸 The cost of failure and defining what good looks like 49:34

🧠 How to decide whether AI output is trustworthy 52:08

📚 Why AI education matters for everyone 56:07

👋 Final thoughts and where to follow Wendy 59:00

Full Transcript

00:00:54:11 - 00:01:27:24
Layne Fawns
What does it really take to get AI into production and keep it there? Today's guest, Wendy Gonzalez, is the CEO of Sama, a company trusted by more than a quarter of the fortune 50 to deliver the data and human in the loop systems behind production grade AI. With over 20 years of experience scaling technology organizations and a mission driven approach to ethical AI, Wendy brings a rare perspective on how high performance teams, high quality data, and social impact come together in the real world.

00:01:28:01 - 00:01:31:09
Layne Fawns
Wendy, thank you so much for joining us on Innovators Inside.

00:01:31:12 - 00:01:33:15
Wendy Gonzalez
There to be here. Thank you.

00:01:33:18 - 00:01:41:01
Ian Bergman
Welcome. We are really excited. Can't wait can't wait to dive in with you. And what do we got to start?

00:01:41:03 - 00:02:05:23
Layne Fawns
Yeah. So I want to start off with, something that's just a little bit more kind of, like, work oriented, in, like, the sort of, like, value sector. What is the most surprising piece of advice you've ever received? And, Wendy, I'd like to start with you. This stumped me. I thought about this for about ten minutes before we hopped on here.

00:02:05:25 - 00:02:24:03
Wendy Gonzalez
The most surprising piece of advice, I think. Yeah, surprising. All the surprising was, when I first, became CEO. So there's always a first. I was like, okay, I had a little bit of doubt. I'm like, I don't really know what I'm doing. I better start talking to people. I better start figuring out what this whole gig is about.

00:02:24:05 - 00:02:45:08
Wendy Gonzalez
So I am probably at like, I don't know, 30 different conversations with CEOs. I'm just like, hey, help me understand, how do you manage your board? How do you do this? How do you do that? You know, give me some advice. And one of the most surprising pieces of advice was from, you know, a multi-time CEO that said, you know, what I do every year is, I find myself.

00:02:45:10 - 00:03:08:08
Wendy Gonzalez
And I was like, what do you mean, find yourself? So, well, I think about what the what the company needs in a CEO the next year, I fire myself mentally. I pull up a list of what do I think the company really needs as a CEO. And I set my objectives and I sort of measure both my performance, what I need to do, who and how I need to show up on that benchmark.

00:03:08:11 - 00:03:28:26
Wendy Gonzalez
And I thought it was pretty amazing advice to not just, think about, okay, you know, I've gone here, I've earned it. This is what the company company needs is to really take a step back and say, all right, you know, where's the company going? What does a company really need? How can I measure up? How can I, you know, supplement my my gap's things of that nature?

00:03:28:28 - 00:03:30:06
Wendy Gonzalez
Yeah. It interesting.

00:03:30:09 - 00:03:30:25
Ian Bergman
I really.

00:03:30:27 - 00:03:33:17
Layne Fawns
Love that. Yeah.

00:03:33:19 - 00:03:37:29
Ian Bergman
Although, although lay you're not allowed to look at me and be like. Ian, time for you to fire yourself.

00:03:38:02 - 00:03:48:11
Layne Fawns
No, no, no, I was actually thinking I was like, man, I should start firing myself too. So it's it's I don't know. That's, very profound. I love that. Well, and.

00:03:48:11 - 00:04:15:20
Ian Bergman
It's and it's, it's it's interesting in a changing world, like, it's it's funny, like you got the annual cycle, right? And yes, you have fiscal year cycles and we all work on years and all of that. But like, it's really interesting to think about what, what feels like in an, in a, in a world where the, the, the change around you in a year feels like it's accelerating and more every year, like it's almost more important to take that moment and take stock of the change in context.

00:04:15:22 - 00:04:24:00
Layne Fawns
Well, and I think, too, it takes a lot of humility as a CEO to be able to do that. And I think humility as a leadership quality is incredibly important.

00:04:24:05 - 00:04:37:27
Wendy Gonzalez
It's good advice. Yeah, because it's not it's exactly it's not about you. It's about what the company needs. And it's kind of interesting. And I sort of liken it to budgets. There's a point in time where, you know, you run a budget and b everybody looks that well, this is what you spent last year. So this is this year.

00:04:37:27 - 00:04:52:02
Wendy Gonzalez
And it's like you're you're asking the wrong question. Right. Yeah. It's like don't take and repeat. You know rinse and repeat. Take a look at what has changed. And that goes beyond just, strategy. It goes to competencies. Right. And skills.

00:04:52:05 - 00:04:58:27
Ian Bergman
Yeah. I've got some finance managers from early in my career. I'd like you to talk to, you know.

00:04:59:00 - 00:05:11:24
Layne Fawns
Yeah, I like the, the days of, like, like, if we, if we come in under budget than our budget gets slashed next year. So we going to try to figure out how we're going to spend the money and it's, you know. Yeah. Yeah.

00:05:11:27 - 00:05:19:19
Ian Bergman
Eileen, you said you thought about this for ten minutes, but you need to answer this question. Surprising advice you ever got.

00:05:19:22 - 00:05:45:08
Layne Fawns
So I think that stepping into a management position for the first time, I had kind of this notion that I needed to understand how to do every job, and then that I had to know best practice for it. And so I think, like the most surprising advice was, actually, it came from you in where I was.

00:05:45:10 - 00:06:11:01
Layne Fawns
Yeah. But it basically it was like, I would hope that the next person that you interview teaches you something that you don't know. And, and I think it comes, it comes back to that humility where it's like, yes, I think like management isn't about knowing everything. It's about knowing what you don't know and finding the right folks to help fill in those gaps and supplement that.

00:06:11:03 - 00:06:26:14
Layne Fawns
And a willingness to learn from the people around you. And, you know, and I've always believed that you're never done learning and that life is a growing process. And I hope I'm not the same person next year that I am this year and things like that. But I think just the way that it was framed, it was like, oh, that makes sense.

00:06:26:14 - 00:06:27:15
Layne Fawns
So yeah.

00:06:27:15 - 00:06:48:09
Ian Bergman
I love well, I'm glad, but you know, and okay, so I meant you guys, I'm so sorry for this. I'm going to do like the, the most artificial segue, but it's so good right. Because, you know, aren't we in a world where all of a sudden the tools that we use can hypothetically teach us things or collaborate with us, with us?

00:06:48:11 - 00:06:58:00
Ian Bergman
And doesn't that sort of change the nature of our relationship with these tools and systems, etc.? So let's let's build on that thread later. I want to pull that.

00:06:58:03 - 00:07:01:05
Layne Fawns
Yeah. Well, let's build on that thread right now.

00:07:01:07 - 00:07:16:12
Ian Bergman
Oh. That thread right now. I mean, aren't we in the world? Okay. So Wendy, like, I, I want to pull on that thread, but I want to do it. I want to start with an introduction for you. I want to start with like how did you come to be sitting on innovators inside? Tell us about your background.

00:07:16:15 - 00:07:28:28
Ian Bergman
What brought you to build some, what you've learned along the way? Like, you know, why are you here to talk with us about the changes that AI is bringing to the world and how we deal with it.

00:07:29:01 - 00:07:57:20
Wendy Gonzalez
You know? So, well, as I go all the way back, I, I'm a reformed accountant, so I actually went to, started in auditing, realized I absolutely despised it. And, leaned on my, information systems, part of my degree to focus on technology. So I did, consulting for, over a decade, working with large companies, figure out how to leverage technology to kind of transform businesses.

00:07:57:22 - 00:08:20:06
Wendy Gonzalez
Just sort of realize I wanted to I wanted to, to own the technology. So I went to some public companies, you know, took technology leadership roles. And I was like, well, that's that's not enough. I want to build a company. So then I went into startups. I co-founded, a SaaS startup. All of this is really pretty, pretty great kind of leveraging technology in all, in my entire career, it's been primarily driven by data.

00:08:20:09 - 00:08:37:21
Wendy Gonzalez
So how do you bridge data to make intelligent systems? It was big data. It was internet of things. How do you leverage telemetry data, etc.? And then, you know, one day, as we were kind of working through getting to, you know, 10 million plus, I was like, okay, what am I what am I doing here?

00:08:37:24 - 00:09:04:23
Wendy Gonzalez
We are building intelligent systems to make cigars not dry up in humidor. We are building, you know, intelligent systems to make sure that the golf course gets, you know, stays green. You know, during the hottest months of the summer but uses, you know, water efficiently. And I was kind of feeling like, there's probably something more. This is how we answer the point where, we're trying to, show are we had little kids at the time, and like, hey, here's how you get involved in community.

00:09:04:23 - 00:09:23:20
Wendy Gonzalez
And like, you walk past a piece of garbage, you know, just step over it. You pick it up and you throw it in the trash, you know, all those kinds of things. So, it kind of came to a bit of a conclusion that, moving into a space where where I could kind of, you know, kind of connect interests and, sort of values with my work became really interesting to me.

00:09:23:20 - 00:09:46:07
Wendy Gonzalez
And then, as I was looking around, I had, my father passed away unexpectedly. So that was really sad. But when it kind of made me realize is, you know, life's short, like. Yeah, okay. I can't, you know, co-founded this, you know, startup it, like, seems to be getting on its own feet. So I ended up getting, connected with, the founder of Salma.

00:09:46:09 - 00:10:15:01
Wendy Gonzalez
Leila. And, I was like, okay. Wow. What an interesting concept. You know, there's this belief that talent is distributed equally, but opportunity is not. And that resonated with me greatly. Oh, my. My, my husband, my partner, he, you know, he grew up in the projects. First person, you know, in his family to go to college, like, we, we, you know, child of immigrants, like we really kind of found our way into, I wouldn't say overwhelming success otherwise.

00:10:15:01 - 00:10:42:02
Wendy Gonzalez
It probably retired. But certainly we felt really fortunate, you know, and or like, hey, work has been incredible. Well, you know, it is a culmination of a lot of hard work, but it's also access. So this idea that, providing access and opportunity is really what is behind sama. Sama means equal in Sanskrit. So the way that I look at it is, you know, all these super talented people in, you know, underserved areas, they just like access, right?

00:10:42:02 - 00:11:06:13
Wendy Gonzalez
So, you know, we don't make up talent and brilliance, right? You just got to open the door for somebody be able to walk through. So it resonated with me. A lot that financial independence and, you know, business is can be a force for social good. In other words, instead of, you know, worried quite so much about donations, if you can create financial independence and technology ecosystems, people can create their own financial independence.

00:11:06:13 - 00:11:27:14
Wendy Gonzalez
And, you know, kind of break the poverty cycle. So long would you say? I thought it was just a really brilliant the straightforward way to think about connecting, people with work. So I thought that was fantastic. As I, started to look into it, I was like, okay, wait a second. I've got a background in technology, I've got a background in services.

00:11:27:14 - 00:11:52:11
Wendy Gonzalez
Like I work with enterprises. Wow. And this business can really leverage, kind of human and human judgment to, improve data. So I thought, wow, what a great way to to do this while, you know, creating meaningful living wage jobs, you know, parental benefits, you know, you know, medical safety nets, basically to, have people not only, take on employment, but the idea was jobs beget jobs.

00:11:52:11 - 00:12:15:20
Wendy Gonzalez
So if you have no access and you don't have a, you know, CV or a resume, but then you start to work, your ability to get hired and kind of permanently move out of the poverty cycle is much greater. So I love the concept so long story. That's what interested me in it so much that as I dug into the company, the core focus was really about how do we leverage, you know, all this great human judgment to make data better.

00:12:15:22 - 00:12:34:05
Wendy Gonzalez
And dirty data is like a massive problem. So, immediately I started focusing on leveraging, you know, all of this great talent and, we built a technology system to be able to, you know, capture human insight for training data. So, okay, all I write, all I, you know, lens for.

00:12:34:08 - 00:12:35:18
Ian Bergman
To be tagged and trained.

00:12:35:21 - 00:12:43:06
Wendy Gonzalez
All this to be tagged and trained. Right. So, you know, AI is like, they're like humans, right? Data is like experience, right?

00:12:43:08 - 00:12:43:27
Layne Fawns
Yeah.

00:12:44:00 - 00:13:08:27
Wendy Gonzalez
The more, data that is ingested, you know, the the better, more accurate models can be. So, that's really how I got into this. And I thought this combination of, you know, leveraging a technology platform, thinking about where human judgment is necessary to really help improve and train AI. Yeah. Was really fascinating. So super long winded answer, but that's all I got in the first.

00:13:09:00 - 00:13:16:08
Layne Fawns
Sorry. And just really quickly for the for the audience, what do you mean by dirty data. Like what? What does that look like to you?

00:13:16:10 - 00:13:39:06
Wendy Gonzalez
So dirty data can take like a ton of different, you know, forms. So, imagine this. You're trying to buy something on like eBay or marketplace, right? There are a bunch of sellers. They have data. You're looking for something gray. Gray. But like half of the, you know, product SKUs and items on there are spelled dry, right?

00:13:39:06 - 00:14:01:13
Wendy Gonzalez
So it's a kind of an example of dirty data, right? You know, things don't match. Maybe they're missing attribution. You're not able to search for something. That's an example of you need to kind of clean up and enrich data so that it is, searchable and trainable. Yeah. Or or it could be, you know, missing, you know, I mean, basically, by definition, we structure unstructured data.

00:14:01:15 - 00:14:24:12
Wendy Gonzalez
So imagine, an autonomous vehicle, how you train an autonomous vehicle to recognize, different cars, you know, a Vespa, maybe its first time, you know, the car is driven. You know, there's no no, data for for mopeds. What about, vulnerable road users? So people on the side of the road, what if it's a small person, a big person, a person of color?

00:14:24:15 - 00:14:35:02
Wendy Gonzalez
You know, like, how do you recognize all that? So how much data you ingest, the quality of that data and how it gets properly attributed? Is really critical to making models function.

00:14:35:04 - 00:15:00:11
Ian Bergman
And so, you know, it's really interesting. I think I maybe late in my career was first exposed to the notion of data labeling, actually through, vehicle vision, right in kind of the early 20 tens. And, at the time, one of what I learned was that it was an extraordinarily human intensive process. Right? Literally flash image up in front of person.

00:15:00:13 - 00:15:05:22
Ian Bergman
Right. You may or may not have attempted to automatically label some things and be like, okay, this.

00:15:05:24 - 00:15:07:16
Wendy Gonzalez
Captions and stuff like that. Yeah.

00:15:07:16 - 00:15:29:20
Ian Bergman
And then details. Yeah. So and it was a, you know, it wasn't as a quote unquote expensive human process, but, but to your point, necessary. So tell me, like, how does how does summer help people that want to implement fine tuned models. What do you what do you actually do. And then we're going to connect that right back to your story around, feeding entrepreneurial ecosystems.

00:15:29:22 - 00:15:50:28
Wendy Gonzalez
Yeah, absolutely. So at the end of the day, all and models need to be trained on data. Sometimes companies use synthetic data, right. Some they're all different methods in which you can, you can train, but the, the real what, what Salma does is we do a couple of things. One is we can do straight up data annotation.

00:15:50:28 - 00:16:15:27
Wendy Gonzalez
So in that example of complex edge cases, there's a lot models of advanced dramatically, but edge cases are really prominent. So I'll use a simple example back back when Tesla's when the when the autopilot was first being released, they worked super well on highway 101. Sure. Because that's where all the road, you know, data was, right, was driving up and down the California coast and through the highways.

00:16:15:29 - 00:16:22:26
Wendy Gonzalez
But trying to take that same vehicle and have it work in a nine lane roundabout in China. Right. Not going to work so well. Right.

00:16:22:26 - 00:16:23:26
Layne Fawns
So yeah.

00:16:23:29 - 00:16:44:22
Wendy Gonzalez
So you have to be able to capture and annotate those edge cases and use scenarios that the model doesn't recognize. Right. So it knows how to how to learn from that. In that case the the nine lane roundabout, we also do something called data validation. So models have gotten a lot smarter. So this is really about, hey, that this model has anticipated or has identified what the objects or what the classifications are, but are they correct.

00:16:44:24 - 00:17:03:09
Wendy Gonzalez
All right. So then the human then goes and kind of does an audit if you will, and provides oversight. And then of course, the same thing can happen for a generic AI. And LMS. So how do you evaluate that the model's responses, like the prompt and response pairs and things like that, are actually accurate for context, tone and sentiment.

00:17:03:11 - 00:17:24:03
Wendy Gonzalez
So that's effectively what we do. We can process large volumes of data through our platform. And something that I think is important to note, you touched on this end, is that like gone are the days of, you know, is there a dog or cat in this image? Right? Right. I mean, things have advanced so much further, so much further.

00:17:24:03 - 00:17:50:16
Wendy Gonzalez
So a lot of times, you know, for, for basic applications, you can leverage, you know, models, right? You can leverage models, you can, but the more complex the more stories saved your outcome based. Right. Or customer experience based then. Yes. You you want to make sure that you not only have captured all of the appropriate data and edge cases, but likely have a, like human in the loop effectively to validate that the model is actually given the right answer.

00:17:50:19 - 00:18:16:03
Ian Bergman
And how do you connect this work to the sort of the social responsibility piece of your mission that you talked about, like, you know, from the very beginning, you have talent distributed globally. Access can be and and resources could be hyper local. How does the work of some, create more opportunity in the world, not just for your customers?

00:18:16:06 - 00:18:38:07
Wendy Gonzalez
So we have an impact sourcing model, which means that you don't you don't have to have, you know, a, specific set of experience you can qualify to test out your skills, basically, based off of, both household income. We focus on gender diversity, right? So we make sure we have at least 50% women in our workforce.

00:18:38:10 - 00:19:03:18
Wendy Gonzalez
And we pay living wages. That's really the key. So living wages are not, minimum wages. There was it take to support, somebody in a, you know, with, rent, you know, healthy foods, 10% savings, etc.. And, the key is that this is an employment model that includes benefits so that means maternity leave, paternity leave, mental health, you know, vision, dental, you know, pension plans.

00:19:03:20 - 00:19:24:18
Wendy Gonzalez
And the reason for creating that safety net. And the big distinction is that we're hiring people in, we can train them and we can incentivize people on quality because they get a salary. Yeah. And the crowd model. Right. And crowdsourcing model, you only get paid if you complete the task quote unquote, correctly. And so what happens quite a bit is number one, these are contractors not employees.

00:19:24:18 - 00:19:47:21
Wendy Gonzalez
So they aren't able to get training. So the description of a task you get within one screen, right. And okay, this is how we need to get the answer. And there's nothing. Exactly. There's something called crowd bias. And I've I've seen this happen before where, you know, you can get paid, you know, whatever X amount per task if you are a botany specialist.

00:19:47:28 - 00:20:09:01
Wendy Gonzalez
And so you answer a bunch of questions, the system doesn't really know whether you're a botany specialist or or a botanist or not. And what you're kind of, you know, qualifications are and you just keep clicking on the answer until you get it right. Right. So that can be a little bit of a challenge. One of the models that is, is, different for us is that we can train, we can employ.

00:20:09:03 - 00:20:29:28
Wendy Gonzalez
And what happens is that because we're able to train, we're able to really focus on the right taxonomies and like quality evaluation methods that our clients want. We have the same people who are looking at the data on a regular basis. So when something changes or looks different, we can catch those edge cases. So we think there's a lot of value towards being able to have a trained and retained workforce.

00:20:30:00 - 00:20:52:09
Ian Bergman
Well it's amazing because it there are implications as well beyond the value to your clients of having a trained and retained workforce. And so do you have any, stories, anecdotal or data you as a data person on, sort of what the second order impacts of your trained workforce might be?

00:20:52:12 - 00:21:14:26
Wendy Gonzalez
Yeah. Well, I mean, absolutely, like, you know, imagine, like, search relevance or a product, product catalog as an example. There are some these large, you know, retailers and companies out there that have super complex taxonomies like, you may not realize it, but there's like 20 different ways you can search for granola bar, right? Is it breakfast?

00:21:14:26 - 00:21:37:29
Wendy Gonzalez
Is it a snack? Is it organic? Is it you know, healthy? Is it related to outdoors and camping? Right. So, some of these tax on these can be actually quite complex. And so, they are specific to that company. So that trained, you know, resource that understands the company's business rules can be, incredibly valuable to getting the quality right and being able to, to validate the data very quickly.

00:21:38:02 - 00:22:04:23
Ian Bergman
Yeah. Amazing. Okay. So we started before recording with, you know, interest in AI writ large and the impact on society, the impacts on all of us. And there's ethics and moral and political stuff. There's all kinds of fun stuff around this. And, you know, fundamentally, you are enablers of, you know, effective quality AI at scale, right? That is something that you enable.

00:22:04:25 - 00:22:25:00
Ian Bergman
So lots of people out there have opinions on AI. Oh it's evil. Oh it's amazing. Oh, it's going to put us all out of jobs. The data centers are taking water. Oh, it's going to, you know, give a new lease on life for the creatives in us, whatever it is. What's going on with AI?

00:22:25:03 - 00:22:45:03
Wendy Gonzalez
What is going on with AI? Well, yeah, I mean, it's an incredible. And I'm sure you see this every, every day too. Like it's incredible. Productivity booster. I mean, I'm leveraging it myself on a regular basis, to help, you know, distill information to aggregate data. But, you know, that's just on a, on a personal basis.

00:22:45:03 - 00:23:05:11
Wendy Gonzalez
The key is, is how can you get the AI model to perform effectively and accurately, and how can you really do that? It's at scale. So, you know, it's kind of interesting is that, I'll maybe use an analogy. Right. I was reading something actually the other day. People don't really know how to read maps anymore.

00:23:05:13 - 00:23:25:09
Wendy Gonzalez
They are really not very good at navigating, understanding models and describing how to get to places. Well, that's, you know, all of our, like, Google Maps and GPS systems, right? I mean, and have you ever been where like the, the, satellite isn't working well, and you're like, driving yourself in a circle for a while when you actually, if you just look up.

00:23:25:12 - 00:23:33:01
Ian Bergman
You ever been under, you ever been under ground. It's called what, lower Thacher or something in Chicago. Like it's famous for this. Like, because it's.

00:23:33:01 - 00:23:51:12
Wendy Gonzalez
Like spinning around and you're like, oh, why did it take 28 minutes to get there when it's actually just down the streets? Because your brain's on like autopilot, like just trusting that the air is, is working. Right. And I, I think the, the what I find what's interesting about the, what's going on with AI is that, yeah, there's all these incredible like, amazing advancements.

00:23:51:15 - 00:24:07:06
Wendy Gonzalez
But what is really the key to deploying AI successfully is going to be is well, first of all, do you know what you want it to do? And can you articulate that? Do you know what good looks like. And then once you can describe what what good looks like, how are you going to find ways to validate that it actually works.

00:24:07:08 - 00:24:22:22
Wendy Gonzalez
So it's kind of I use that example of the sort of blind, you know, usage of GPS is because I have had it before. I will admit, it's kind of autopilot, like following the GPS you've lived in that city a million years and you're like, oh, wow, it's just taken me down the wrong route because it got confused.

00:24:22:24 - 00:24:39:25
Wendy Gonzalez
And you're, you know, you just added 30 minutes to your trip. Well, I think that's really the the interesting aspect of this, if you put it a question into ChatGPT or Gemini, why are you just taking that the are you are you assuming that the answer's correct? Right. Like, oh, I'm going to take it for what it's worth.

00:24:39:25 - 00:24:54:03
Wendy Gonzalez
And so I think that's going to be really the the challenges is as we adopt AI and as we adopt and build AI, can you articulate what you want? What do you want the AI to do? And can you articulate what good looks like?

00:24:54:06 - 00:25:19:28
Layne Fawns
Okay, so articulation is key. But then there comes that that educational component. Like for example, like my parents didn't know for the longest time what catfishing was like with the internet, right? But that was something that like, we grew up just automatically knowing like if someone random message you on Facebook and they claim to be a famous person, it's really probably not that person, right?

00:25:19:28 - 00:25:53:26
Layne Fawns
So I think so we're talking about almost like a standardized educational component around how to use AI, how to engage with it ethically. So then this leads me to another question, which is then, much like the internet, do we I say like, do we as a group here? But do we believe in democratized access to AI as something that's almost going to end up being a basic human right, the way many people believe that the internet should be?

00:25:53:29 - 00:26:04:19
Layne Fawns
And if we believe that, what do we see the effect on society being in terms of our ability to continue to think autonomously and do things for ourselves?

00:26:04:22 - 00:26:37:00
Wendy Gonzalez
That is a that is a heavy that is, that is a heavy, heavy question, which I really, really appreciate. That's, that's, that's probably like, oh my gosh. Well, it's so interesting. Right. And all of these different I see this with like policy. Every country is building their own AI, policies in their own AI standards. And it's pretty interesting because the companies that are building the biggest models in the world are all based in North America.

00:26:37:02 - 00:26:56:22
Wendy Gonzalez
Right. And so it's kind of interesting about this is that, you know, these LMS are the are the building blocks, right, that many other that consumers use and that many other companies are building their AI on. And these LMS, they've been trained on all publicly available data primarily right. The internet.

00:26:56:24 - 00:26:57:22
Ian Bergman
Primarily.

00:26:57:25 - 00:27:09:01
Wendy Gonzalez
Primarily does, does anybody believe the internet is not biased? And then the internet is 100% truthful, right. So what's really, when you say.

00:27:09:03 - 00:27:12:04
Ian Bergman
We all believe everything that's been written, what are you talking about? Yeah.

00:27:12:06 - 00:27:13:14
Layne Fawns
Yeah.

00:27:13:16 - 00:27:48:26
Wendy Gonzalez
Exactly. So I think was kind of fascinating. About what you just said, Lane is, is like, democratized. Well, gosh, like, how do we how do we do that? Right. Some I think countries are trying to establish that through through policy, like the EU AI Responsibility Act, saying, hey, if there's AI that's being built in Europe, it's got to come, you know, with with data that, you know, it's been built off of data that, you know, data that is, traceable and that they've got the right to use, that there are model evaluation standards, that there are things such as high risk systems.

00:27:48:26 - 00:28:14:29
Wendy Gonzalez
So there are I think there are some protections, in different, countries that are trying to be leveraged to, at least sort of, if not democratize, like, make safe if you will, or make, make responsible. But I think the inherent challenge of that is, is that all of the core model builders are, you know, they're they're building models off of this publicly available data.

00:28:14:29 - 00:28:30:25
Wendy Gonzalez
And so even if you have all these different approaches to sort of either make it safe or provide access, they're being built by a handful of companies around the world, and some of them are doing a really great job. But we did one thing as an example. So we have a lot of our workforce is based in East Africa.

00:28:30:27 - 00:28:51:29
Wendy Gonzalez
And, you know, everybody it's interesting. I think, on a per, sort of capita basis, I believe it's India and Kenya that actually have higher uses of LMS than even the US. The US may have the highest number of paid, subscriptions, but like the, you know, a lot of these emerging, countries are using it greatly.

00:28:52:01 - 00:29:15:19
Wendy Gonzalez
So we did, we did a research project where we took Swahili. That's the language, words, and put them into with English, into ChatGPT. And do you think it could find anything? Well, no, it could, it would like not even close. So that's kind of the example of, you know, you can take this based AI technology, but if it hasn't been contextualized right, it's not going to work effectively.

00:29:15:19 - 00:29:19:15
Wendy Gonzalez
So how is that, democratized?

00:29:19:17 - 00:29:47:21
Ian Bergman
Well, and actually, that is super interesting. I want to come back in a second to the tension between safety and democratization. And I like that is some rich fodder. But, to your point, this language translation, I think, is something we don't often think about. Like, because we can go to any land, we can go to Google Translate as a, as a different model system and we can say, hey, like take this sentence and make it French, or take this sentence and make it Tagalog or vice versa.

00:29:47:23 - 00:30:17:19
Ian Bergman
But that is very different than training data sourced from French or from Swahili. And I wonder if you can maybe help us understand, unpack that a little bit, understand why it is so, meaningful to ask the question, hey, it does. Has the model been trained on and have the context of all of the publicly available Swahili data for instance?

00:30:17:21 - 00:30:35:20
Wendy Gonzalez
Yeah. Well, I mean, in that case, especially with some of these, and it was popular, the right word, but maybe, you know, less captured, languages is that, when you're, you know, you're training off of a limited data set. So how much will will these, these model know, know based off of what is available on the internet.

00:30:35:20 - 00:30:59:07
Wendy Gonzalez
So a lot of there's some pretty interesting groups out there like we've worked with this group just trying to get improve Southeast Asian like, you know, languages or help help these large frontier model builders improve, their understanding of, you know, Malay, you know, Lao and, like, you name it. And so what they're doing is they're encouraging people to get data sets out there so that these models can be trained.

00:30:59:10 - 00:31:21:15
Wendy Gonzalez
So that's one way to do it, is increase the data sets. The other is, is have local people help do that training. So as an example, what we did in our, in our study is we, we said you know, create your own prompts. Right? So we had to define what the prompts were, to really be able to help do the evaluation of, of the model itself.

00:31:21:20 - 00:31:36:12
Wendy Gonzalez
So having that local context is, really key because you, you know, you can you can find a translation dictionary that's online, but that's not going to give you any level of context. And language is really, really interesting because language changes. Right? There are new.

00:31:36:12 - 00:31:36:29
Layne Fawns
Yes.

00:31:37:06 - 00:31:47:01
Wendy Gonzalez
Catfish wasn't a word, you know, whatever. Four years ago. Right. So and I mean, like, I've got teenagers, so I mean, every day I'm kind of like, what did you say?

00:31:47:07 - 00:32:13:06
Ian Bergman
Cause I mean, I yeah, my. Yeah, I've got young children that have new words all the time, but but it's a new. So I had, a really interesting moment. I want to say it wasn't that long ago, maybe 18 months, maybe 24 months ago, which, you know, is, is an is an eternity. I guess. But I had a really interesting moment where I was talking with a native Japanese language speaker about the state of the art LMS and foundation models at the time.

00:32:13:06 - 00:32:49:26
Ian Bergman
Right. And so it was only two years ago or whatever it was, but it was still much earlier than today. But what she told me was that even speaking in Japanese, these models spoke like they were speaking English. And, you know, it's something really interesting because if you think about the difference in the way communication and thought happens between a high, extraordinarily extreme, high context place like Japan or the polar opposite, low context, direct place like the US, you're actually you're not just speaking differently, you're actually conveying very different representation of intelligence and thought.

00:32:49:28 - 00:33:15:00
Ian Bergman
And anyway, I mean, that really jumped out at me. And I think when it kind of builds on what you're saying, like, you know, the fact that the model could theoretically communicate perfectly in Japanese did not mean that it represented the sort of the thought patterns, the practice and the language of Japanese. And it felt inherently foreign to her.

00:33:15:02 - 00:33:30:06
Ian Bergman
And I wonder how many of us, you know, in North America where, you know, despite incredible work coming out of China. Yes. The biggest and most used foundation models are all created. Don't even realize how much is missing from those models.

00:33:30:09 - 00:33:56:00
Wendy Gonzalez
Yeah, I mean, completely that that just that one, you know, research study on Twi. Lee was, was, really interesting because it wasn't like complicated things. It was like, you know, how, you know, how is your day? Where is the restaurant? You know, top five common most foods. And it was really interesting to see that. Yeah, it's really difficult to make a one size fits all.

00:33:56:03 - 00:34:10:00
Ian Bergman
So who tackles this? Who actually cares about tackling this? Like, this is always the problem is like, you know, where's the resources, the money, the the impetus, the customer demand. Like, you know who who yeah. Who tackles this.

00:34:10:02 - 00:34:47:01
Wendy Gonzalez
Yeah. It's really it's really interesting given that it's all, you know, private companies, right, that are building these models and where people are paying for them. Who's your audience in the subscriptions? And so, yeah, it's not soup. You know researcher I think researchers I think are trying to do, you know, the, the right things. But at the end of the day, I think it's fairly complex, which is where some of this policy and, policy and regulations come into play because, how else do you sort of have, risk factors, right?

00:34:47:01 - 00:35:08:18
Wendy Gonzalez
Because if the only people who are kind of, I won't say funding, but like the focus is going to be on getting additional subscribers and applications, but you're going to focus there. So how else are they going to fit in other applications. And these lenses, frontier models are so, vast and expensive will happen. I mean, this is my my belief is that, you know, they'll continue to consolidate.

00:35:08:19 - 00:35:26:23
Wendy Gonzalez
It's kind of like cloud computing. You don't have 50 large cloud computing companies. You basically have three. Right. And that's probably what will happen in this space is a cost. So much money and so much energy. And to get these models, to that level, you won't. They're only like 40 or so in the world. That number is going to continue to drop.

00:35:26:25 - 00:35:48:17
Wendy Gonzalez
So yeah, it is interesting. And it is challenging. And will we start to see some more, you know, kind of country specific models. Right. Like you're I'm sure like astrologers, French out of France. And then you've, you know, DeepMind out of China, then, I don't know, it's really it's a really, really, interesting space, but I.

00:35:48:18 - 00:35:56:09
Wendy Gonzalez
Yeah. Is it going to be, inclusiveness and maybe we'll maybe I'll describe it. Yeah. I'm not so sure about that.

00:35:56:12 - 00:36:20:28
Layne Fawns
I know Ian's got another question here, but I just want to follow up on something that you kind of touched on. And that's the tension between government regulation and corporations that are trying to develop something that's, you know, for profit and for a specific customer demand. And that tension is the fact that this technology evolves a lot faster than regulation can put things through.

00:36:21:00 - 00:36:46:11
Layne Fawns
And so what are your thoughts on like, do you foresee the integrity of AI then? This is not to be negative, but the integrity of some of these looms and some of this, you know, software not diminishing, but finding workarounds quicker than regulation can keep up with them. Or like what does that look like to you.

00:36:46:14 - 00:36:50:21
Wendy Gonzalez
Yeah. Or or kind of lack of policy and regulation. Right.

00:36:50:23 - 00:36:57:17
Layne Fawns
And the lack thereof as well. Right. Like, well, if we can't keep up then why try. I don't know if there's.

00:36:57:24 - 00:37:23:22
Wendy Gonzalez
Yeah. I mean it's kind of interesting, right. Because I mean certainly in the US like the an under, you know, kind of current current administration, it's sort of like even the word responsible is kind of pulled out of most of the, the, the language. There's not really a policy and there's definitely not regulation. What's interesting though is that they're different states doing things like Colorado as well as California, which took, kind of a carrot approach instead of a stick approach.

00:37:23:24 - 00:37:40:04
Wendy Gonzalez
So there's this pretty interesting, legislation that happened in California where instead of saying, like, hey, we're going to regulate you like Europe because I mean, there are arguments, you know, absolutely for, well, I mean, since when does the government know more about technology? Right? So there's a challenge right around regulation.

00:37:40:04 - 00:37:41:23
Ian Bergman
That's a pretty real argument.

00:37:41:25 - 00:38:01:05
Wendy Gonzalez
It's a very real argument. So you talk about, you know, the European approach in a moment and why I think, you know, kind of why they did and what the impacts are there. But this California legislation, which is really interesting, basically says, hey, we want people to build AI responsibly. So responsible is different than than ethical.

00:38:01:05 - 00:38:21:28
Wendy Gonzalez
It includes ethical. But it's about how do you build trustworthy AI. So Lane, you you mentioned trust. Well, everybody should care about trust right. And private company should care about trust when they're building their models. Because trust means that customers will adopt their software because they'll use it. They believe in it. I think there's going to be a good outcome and value out of the software.

00:38:21:28 - 00:38:45:24
Wendy Gonzalez
So trust is really about customer adoption. At the end of the day. To build trust responsibly. AI frameworks are around, you know, where'd you get the data? Did you, you know, did you get it? Lastly. Right. As opposed to, you know, incorporated, etc., do you have an audit built trail of it? So how did you train your, your model to, to a certain extent.

00:38:45:24 - 00:39:01:08
Wendy Gonzalez
And did you, have a method of evaluation? So do you know what good looks like? Right. Just kind of like the measure twice cut once it. If you can't articulate what you want the AI to do, how do you know it's performing well or not? Well, how do you read team. How do you ensure that is operating within policy.

00:39:01:10 - 00:39:25:08
Wendy Gonzalez
Right. And then even ongoing way to do that evaluation. So those practices are really important. And what this legislation does, it said, hey, for for developers who kind of get themselves certified in responsible practices. And under that framework you'll get indemnity. So if you build something bad happens with your software, you'll be protected. And so I thought that was a really interesting way.

00:39:25:08 - 00:39:50:05
Wendy Gonzalez
And then, you know, companies can sort of sign up to be certified. They have responsible. They've been built responsibly, which I thought was a pretty clever right. So it's not necessarily regulating you, but incentivizing companies to take on best practices. In Europe. The same pillars exist too, but it's actually regulated. And I think their approach is more, hey, we we, you know, like, you know, like rotten tomatoes, right?

00:39:50:05 - 00:40:10:01
Wendy Gonzalez
You know, the movie rating system, it's like, yeah, Certified Fresh, you know? So like if you pass our, you know, you developed it in the EU and you developed it under this framework, it's like a certified fresh kind of accreditation where I think part of the view is that then Europeans, because data privacy is like a really big deal and there's GDPR and all that kind of stuff.

00:40:10:04 - 00:40:34:11
Wendy Gonzalez
Can feel better about using this. But then the, the other component of it is around high risk systems. So for example, like if you're doing something like financial lending, right, that's pretty challenging if you have not trained the data properly, like I'm sure you this is this is obviously all but you know, the the old horror story of the, you know, the Amazon recruitment engine, right?

00:40:34:12 - 00:41:03:08
Wendy Gonzalez
So they built a AI model that was recruiting for software engineers. They inferred, based off of all the candidates, that women aren't qualified because they only represented like 5% of the population. So that's that's a training data, you know, example where, what what level of, you know, in higher risk systems like lending, like, you know, lending, like, you know, insurance things of that nature.

00:41:03:10 - 00:41:14:07
Wendy Gonzalez
Would you want people to get excluded or be treated unfairly because the models and trained properly? So yeah, I went like super wide there, but then, but but it's, it's important.

00:41:14:07 - 00:41:38:06
Ian Bergman
Right. Because like, I mean, look, this stuff, this stuff, this is the conversation not just in Silicon Valley and not just in technology ecosystems globally, but in business globally. Right. And like, you know, yeah, we're touching on classic things like, you know, governance around free market incentives versus sort of prescribed regulation. And, you know, we could wax eloquent on strong opinions there.

00:41:38:08 - 00:42:05:20
Ian Bergman
But it is it is important because I touches all of this. So I've got kind of, we this might be a bit of a tough question. Feel free to, you know, deflector adapt or just dive right in. Do you get pulled into these conversations with your clients, with your clients who are not just sort of thinking about, okay, like, how do I make sure that I can run my business in France as well as in Canada, as well as in Australia, as well as in Argentina or wherever?

00:42:05:27 - 00:42:27:26
Ian Bergman
But clients that are thinking about, what position do I want to stake out? What claim do I want to make? What position do I want to stake out for our company in terms of how I should and could be used, is do you get involved in that, or is that sort of outside of the core of what you do?

00:42:27:28 - 00:42:50:04
Wendy Gonzalez
So when we get involved, I mean, usually the we get involved after the company have said, hey, we're going to build some enterprise AI application. So we're we're not usually involved in the hey, how could I make a difference in our product? They're kind of already there saying, you know, hey, I want to build, you know, an anti, you know, collision feature in my car.

00:42:50:04 - 00:43:14:10
Wendy Gonzalez
I want to, you know, have this like, amazing recommendation engine in my e-commerce platform so that I can provide substitutes, you know, if we're out of stock of X, Z. So they've already kind of said, hey, we think that there's a really good, you know, all right, there is a model that can add value to our, to our product, where we start to get engaged in what we see a lot is really around quality and accuracy.

00:43:14:10 - 00:43:36:04
Wendy Gonzalez
So how do we know when the model is working properly? And what's really what's interesting is that, you know, for for all the different enterprises we work with, you know, they they are leveraging a in a way that is going to either be safety related, it's going to generate more revenue, it's going to generate outcomes. It could generate a really amazing customer experience or terrible customer experience.

00:43:36:07 - 00:43:58:20
Wendy Gonzalez
So for example, like you know, when you're when you're using like a virtual sitting room or you know, like filters and stuff, right? It wouldn't be a very cool experience. Like you're if your bunny ears were out floating, you know, off on the side and not on top of your head. Right? So there's usually some sort of a, desire to have a, a high kind of quality threshold.

00:43:58:23 - 00:44:21:20
Wendy Gonzalez
But what we see that's really interesting is that, to be able to do it and do it at scale, you have to have an ROI, if you will, like what is the cost of failure and, and is that worth the investment in to basically training your model to the level of, of kind of accuracy and completeness?

00:44:21:22 - 00:44:38:07
Wendy Gonzalez
It's kind of getting a little bit into sort of the nerdy part of things. But that's that's really what we see being, a pretty interesting question as well. How much is good enough? Yeah. It's like, well, how bad is it if it fails? Is it going to cost you, you know, money? Are you going to have lower revenues?

00:44:38:12 - 00:44:42:14
Wendy Gonzalez
Is it a safety application. So safety application people care deeply.

00:44:42:15 - 00:45:08:26
Ian Bergman
What's what's the answer like that you're getting from clients. And I realize it ranges. But like let me just if I frame this right there's often a tension between, you know, innovation to drive an outcome that you're going for and risk management. Right. There's a natural tension there where your clients landing on this are they are they trying to design for the five nines edge cases, or are they trying to step back and say YOLO?

00:45:09:00 - 00:45:13:00
Ian Bergman
Let's learn. Let's see. Let's, you know, throw something out there?

00:45:13:02 - 00:45:43:04
Wendy Gonzalez
I would say it's we're we tend to work with larger enterprises. So they're pretty well thought out. Right. So, they are I mean, I use some pretty obvious examples, like self-driving cars. They're going to be thinking about it pretty deeply. Yes. Surgical robots. Yeah. They're they're going to be thinking pretty, pretty deeply about what this looks like because, of the implications, at play.

00:45:43:06 - 00:46:10:02
Wendy Gonzalez
But, we, we see very few, enterprises who are willing to large, large enterprises who are willing to sort of risk their reputation, by not having a thoughtful approach. So it's kind of it's interesting. I think there's something that's pretty admirable, that, that Waymo has done. Right. So one of the things they've done as a public safety reports, like 100,000, miles driven autonomous.

00:46:10:02 - 00:46:10:26
Ian Bergman


00:46:10:28 - 00:46:35:19
Wendy Gonzalez
Yep. And I think it's a very smart thing to do because basically what they're, they're showcasing the casing is, you know, not only does this I work but look at this. Compared to our, you know, sort of accidents per, you know, property damage, you know, accidents etc. and compared to human drivers. So I think one of the things is interesting is in that case, there's sort of a recognition that the expectation of that vehicle is actually probably higher than a human driver.

00:46:35:21 - 00:46:41:09
Ian Bergman
Yeah, right. I mean, it's, I mean I it's probably a thousand higher than a human driver in terms of what you're.

00:46:41:11 - 00:47:10:11
Wendy Gonzalez
Walking into the car. If you're like, this is like, you know, my crazy neighbor Joe's driving habits, like, you wouldn't do that, right? Yeah. You, you, definitely have a higher expectation. So I think at the enterprise, level, some of these comes they're really thinking thoughtfully and saying, hey, we know that the expectation is higher. So if you have a, like a chat bot experience, for example, and the chat bot says something incorrectly, the threshold for tolerance and the customer experience on that, it's trust lane.

00:47:10:13 - 00:47:30:14
Wendy Gonzalez
I mean, I keep going, like you said, trust. You're not going to trust that AI. So yeah, I mean, I think that there going to be a push to say, hey, how do we, you know, reduce costs, how can we deliver this AI more efficiently? The, the, the companies that are serious about this as a differentiator are looking at it pretty hard and saying, you know what?

00:47:30:14 - 00:47:36:12
Wendy Gonzalez
We recognize there's a cost to making this product effective. The product is not effective. We're going to lose trust.

00:47:36:15 - 00:47:38:19
Ian Bergman
Yeah. We're not I think that's.

00:47:38:21 - 00:47:39:12
Layne Fawns
I think.

00:47:39:15 - 00:48:00:16
Ian Bergman
I loved it I love to hear that. And it's it's reassuring. Right. Because I think a lot of people are just looking at industry writ large in any sector and saying, oh, people are forging ahead with new technologies, new ideas. And so it's reassuring to hear that. But I will say I have a certain admiration for for companies that take calculated risks with AI.

00:48:00:18 - 00:48:25:04
Ian Bergman
And let's be very clear, like your risk threshold is wildly different in an autonomous aircraft versus, you know, a chat bot. But I, I've talked about it on the show before, but I, you know, there's this wonderful example from Air Canada from what actually might be 18 plus months ago, two years ago now. Right. Where, you know, Air Canada, they threw out.

00:48:25:06 - 00:48:46:03
Ian Bergman
Oh. I'm back. Chat bot. It promised a fair that was incorrect. To a customer. Air Canada refused to honor it, and then they got slapped in the courts for it. I'm summarizing. This is effectively what happened. Right. And I think rightly so, because an agent of the company had promised a fair and, you know, they didn't honor it.

00:48:46:06 - 00:49:10:06
Ian Bergman
There's a lot of when you see examples like that or you see me like the MIT headline from last fall that said, like 90 plus percent of AI projects are failing. You know, there's a there's a rush to critique. But I will say I have a certain admiration for the, the calculated risk takers that are like, actually, we're going to design this to the quality threshold we think we need and then throw it out there and then learn.

00:49:10:08 - 00:49:34:06
Ian Bergman
And I wonder if you're seeing in the industry, like, I'm just wondering, if you're seeing any changes, perhaps in the industry in how companies are approaching risk thresholds or is it they just, you know, are they just saying, hey, like, we have to bring AI and we know we do or we're not going to be competitive? You know, what's what's going on there?

00:49:34:08 - 00:49:53:29
Wendy Gonzalez
Well, I think it's, I think it's like, two, two fold. So when I, when I mentioned the, the sort of cost of failure or the cost of quality, that is a that is a sort of a bit of a calculated risk. Right. So you know how you want. I'll, I'll use a benign example. You're an e-commerce platform.

00:49:54:01 - 00:50:18:00
Wendy Gonzalez
You, you know, lots of these now. E-commerce platform for, for food. We'll say, you know, oh, if you bought that, you know, French bread, then, you know, here's some butter, right? Like they do recommendations and things like that. Is it, a just terrible thing if it says, you know, instead of butter, like Alfredo sauce, you know, right.

00:50:18:01 - 00:50:34:17
Wendy Gonzalez
Probably. Like, is it. So will you go through, will you spend X amount of money like so. So there are some calculated decisions that say what is the what is the sort of cost of failure in the ROI. And in that case it would be a when you can do pairings, you increase your cart size. So there's revenue there.

00:50:34:17 - 00:50:59:07
Wendy Gonzalez
Now if you don't get it right every time, that's probably okay. Right. And so there are going to be I think we see those kinds of, decisions, you know, being, being made, I think where things tend to go quite I don't see quite wrong or can be challenges when, when the definition of quality or what, like what I mentioned before, what good looks like isn't well defined because then.

00:50:59:07 - 00:51:18:09
Wendy Gonzalez
Yeah, you you're really lost, right? Like, how do you know if it's working well or not working well? Can you even identify what those quality thresholds are if you can't articulate what you want it to do. And so that's the notion of edge cases and y kind of quality calibration. And you know you don't just sort of set it and forget it.

00:51:18:11 - 00:51:29:01
Wendy Gonzalez
You can data evolves and like, which developer in their life ever knew every single outcome that they wanted? They didn't. That's why you do it iteratively, right? Yeah.

00:51:29:03 - 00:51:38:06
Ian Bergman
I love I mean, I think we just keep coming back to like, in all things. Right. Like measure thrice cut once, you know. Do. Yeah I love it.

00:51:38:09 - 00:52:08:06
Layne Fawns
Yeah. Absolutely. Okay. We've we've talked about trust in a couple different capacities. So trust in terms of, say for example like your chat bot example. And then we've talked about trust with AI in things like insurance or, or finances. So what about the trust that you can't see. So for example, it's even just like, like we talked about very early in the conversation, a conversation with ChatGPT.

00:52:08:09 - 00:52:24:18
Layne Fawns
And you asked a question and it gives you an answer. What is your advice to people in general with regard to engaging things and being able to measure whether or not something is trustworthy?

00:52:24:20 - 00:52:47:04
Wendy Gonzalez
Very good. Very good question. I think I probably start, start with you know, this sounds really basic critical thinking before you just start jumping in doing like oh okay. I'm going to trust what that, you know, with what the data says. And I'm not blindly, you know, kind of operate on is just take a step back.

00:52:47:04 - 00:53:11:22
Wendy Gonzalez
Look at the answer. Does that look right. You know, each what we find is that each you know, if you if you ask, certain models to do a calculation like I always double check the calculation, right. Like I did, this actually work. Was this correct? I've seen things are wildly, wildly off about companies calculations. So do a sanity check.

00:53:11:22 - 00:53:35:15
Wendy Gonzalez
Don't just sort of read it and trust it light blindly. I think the, second thing is, is certainly is that, you know, the benchmarks are helpful sometimes. Right? So different models are good at different things. Yeah. You know, we all kind of have our favorites. I think, you know, anthropic when it comes to writing and content creation is pretty solid, right.

00:53:35:18 - 00:53:55:02
Wendy Gonzalez
Is it the best, you know, engineer? You know, software, code assistant? Well, I think that there probably some others like first that might be a little bit. So, you know, there's, there's different, I think benchmarks can be very helpful, but it absolutely starts with, you know, did you set your prompt correctly? What is the answer you are expecting?

00:53:55:02 - 00:54:08:15
Wendy Gonzalez
And and take a moment to think about it before you just take it for for what it's worth, I don't like, like it's the same thing, but yeah, you got to actually think about it critically. Models are not good at doing everything.

00:54:08:18 - 00:54:37:23
Ian Bergman
You know? I mean, friend, a friend of the podcast, Vivian Ming, came on recently. She's got a book, on this, on this topic. We'll put in the show notes. But, you know, she kind of made a point really similar or basically said, but if I, if I am to paraphrase it is that if you accept the model's answer or the text answer every time you become dumber as a person, if and you're going to get things wrong if you accept it is the beginning of a back and forth and you ask the question, is this correct?

00:54:37:23 - 00:54:57:26
Ian Bergman
Both you and your tech are going to get smarter every time you know, if you accept the GPS navigation into the Thames, you're going to drive into the Thames. If you ask, is this the right route? You're going to reinforce, you know, decision making and knowledge. I think that's actually kind of important, because I think a lot of us sometimes default to, oh man tech.

00:54:57:27 - 00:55:08:05
Ian Bergman
You know, it gives me an answer. Let's just accept it. And that way actually lies quite a lot of danger. Right? Bad training data leads to bad outcomes. It's. But there's other danger as well.

00:55:08:07 - 00:55:31:20
Layne Fawns
Well, because my argument has always been that that goes against human nature, like human nature, we we want profit. We seek the easiest course. Right? That's that's why innovation happened in the first place. The industrial revolution, the agricultural revolution, all of these things, right, is we found a way to make things more efficient and easier for ourselves. We can't help it.

00:55:31:23 - 00:55:59:07
Layne Fawns
And and that harder. And however, if I personally just kind of believe if we're going to survive the age of AI, we actually have to learn how to inherently go against those instincts and be mindful of all of this moving forward. And I think there has to be an educational component to this in terms of, you know, in schools and things like that.

00:55:59:07 - 00:56:07:14
Layne Fawns
Is, is there has to be a way that you approach all of these systems so that you don't become dumber as a human.

00:56:07:16 - 00:56:27:07
Wendy Gonzalez
100%. I mean, I don't know if you guys have seen Wall-E. There's some truth to that. That movie, like there's some planet will be floating, you know, in the galaxy on, on on these little hovercrafts. Right around like that. How to walk. Right. So, but it's an incredible movie. But, yeah. No, I mean, I think it's I think it's true.

00:56:27:07 - 00:56:43:17
Wendy Gonzalez
It's interesting. Was just on a family trip, and, you know, we driven out of this, like, cul de sac, like, the same time ten times. But, you know, we follow the GPS, so, like, we, you know, has the time. We did the wrong way. Right. It's like, well, wait a second. Like we've been here before, we actually know that we need to go.

00:56:43:17 - 00:57:07:23
Wendy Gonzalez
Right. So like you have to sort of have that question and, yeah, trust your instincts. And what comes to education? I think that's, exactly right. There is, everything from how do you build AI. Right. So the education around responsibly AI practices, which are again, there they are best practices. It's the measure of the cut ones is have a way to do the evaluation.

00:57:08:00 - 00:57:30:25
Wendy Gonzalez
Know where your data is coming from. Actually articulate what data. Articulate what you want the outcomes to be. Things that all of a sudden very self-evident but are not a standard. Right. So, that is one which is the first place to start is the developers who are building these AI applications and models like, you know, let's, let's communicate and make all of the best practices, including things like red teaming, which is okay.

00:57:30:25 - 00:57:48:02
Wendy Gonzalez
Is the model operating within the parameters you want it to and having, you know, tools like I think GitHub and, and Google, they both have some developer tools out there that allow you to kind of leverage this. So if you're building it, you can have these different checkpoints, you know, in place kind of like standards if you will.

00:57:48:04 - 00:58:09:18
Wendy Gonzalez
Yeah. And then for individuals, I mean, absolutely. It's like I hope everybody can understand this by now. But my goodness, you know, if you don't have a licensed, you know, software to these, a license to some of these arms, you know, you're giving your data away. Do you want your data to train the model? Are you aware of that?

00:58:09:18 - 00:58:35:22
Wendy Gonzalez
Do you understand the parameters and these sort of. This disclaimers of how to use this data? So yeah, I agree one 100%. I don't you know, we shouldn't bury our heads in the sand in terms of, training about this and informing people how to use the AI for, for their own benefit. Right. Yeah. Completely. Completely. Agree.

00:58:35:24 - 00:59:00:11
Ian Bergman
Well, this is incredible, Wendy. We've we've covered a lot of ground here. Actually. But I just want to say thank you for joining us on Innovators inside your Business is that, you know, a very fascinating place. Like, sort of. I mean, fundamentally, you're sort of sitting between the reality that exists in the world, the real data, and how it's used and ingested effectively into these AI systems all around us.

00:59:00:11 - 00:59:20:22
Ian Bergman
It's got to be got to be wild. Some of what you see day to day and really appreciate you coming here on innovators inside to share your stories, your thoughts, etc.. For audience members who want to follow what you're doing, your work, where should they find you? You on LinkedIn? Should they head to the company's website?

00:59:20:24 - 00:59:22:02
Ian Bergman
Where should they get in touch?

00:59:22:04 - 00:59:40:19
Wendy Gonzalez
Company's website. And so sarma.com or LinkedIn, happy to connect. And, yeah, this has been fun. It's not often you can, range from, you know, policy AI democratization to some in your blog to wali all in one conversation. So, good stuff. Thank you.

00:59:40:22 - 00:59:41:24
Ian Bergman
Glad we got only.

00:59:41:24 - 00:59:42:29
Layne Fawns
On innovators inside.

00:59:42:29 - 00:59:47:15
Ian Bergman
Only on innovator. That's right. Thanks so much, Wendy. Really appreciate it.

00:59:47:17 - 00:59:48:01
Wendy Gonzalez
Thank you.