The Original Tech Aunties
Welcome to the Tech Aunties podcast, where we're bringing you industry context and vision from myself, Angelia McFarland, and Gina Rosenthal. On each podcast, we will share our marketing and technology industry experiences along with the team. Listen to us as we explain the past so you can have context to understand and create your own version of the future. So let's get into it.
Hey, Gina.
Hey, Angelia. This is exciting. This is episode three.
Yeah, and this is one of our tech episodes, so you're going to be leading this one and you're going to help us understand ChatGPT. What is it? Why is it important? What should we be doing with it? Tell us about ChatGPT. I know you are very passionate about the fact that it is not AI — but it is a tool.
It's an AI tool, though. We can say that. Yes, it's an AI tool. Well, ChatGPT is a large language model. Basically, you give it a query — there's a little text box at the beginning. And actually, OpenAI was pretty brilliant with that. They were like, how can we get people to help us see more use cases? We should just make access free. And it took off. That was pretty smart. So a little chat box makes it seem like you're just talking to the AI and the AI is talking back to you like a person. And it tells you stuff that is amazing and you don't have to do all the research yourself. So it's this generative AI tool. Basically what it's designed to do is — the data scientists write an algorithm to take all the words in your query and they assign weights to those words. So then the AI knows to go out into the data set — the huge data lake that they scraped off the World Wide Web — and come back with the answer it's pretty sure matches what you asked for. That's pretty much what happens.
So ChatGPT — earlier you said GPT was generative, but what does the full acronym stand for?
Generative Pre-trained Transformer. I think that's actually pretty important, because if you think about this as a transformer — like a transformer in an electrical system — you're putting in a question, you're putting in some words, and it transforms the answer into something that makes sense to you as a human. That's why it seems like there's a human on the other side of the AI talking back to you. It is a very descriptive name once you get under it. It generates a response from pre-trained data and transforms that data into something you've asked it to do.
Right. I think the pre-trained piece is essentially about training — it knows where all the data is. Some of the pre-training is about putting the weights on the words in the prompt you've placed in the text box. So it knows, for example, that if you say the word "email," it knows what that means and will bring back results in that context.
So I think one of the things I would say to people is — remember, this is not some fancy, magical, mystical AI like in movies. This is literally something that really smart people with a lot of really hard math have put their time into: number one, making the algorithms to go find the correct data, and number two, training the data to spit it out in the format you're looking for. Stop there for a second, because there's a show that came out on Paramount called Rabbit Hole. In it, a character speaks into a computer and asks it to identify, cross-match, and index specific people across all data on the World Wide Web and government databases — and on television, this happens in literally less than a minute. I was watching it with my husband and he said, "I think it can happen that fast." I said, "There is no way that can identify and pinpoint that quickly the information he was asking for, and then index and compare it to other databases." So I say all that to explain — when I describe the entire World Wide Web as the data set, how long did that actually take?
First of all — that was fiction. It was a cool TV show.
Okay, okay. Just making sure. But you know what — I do think your husband is partially right. There are real systems that can bring those results back. Maybe it takes a little longer than on TV. And there is actually a company called Palantir that creates software for law enforcement and the military. It was designed to go through all the data already available to law enforcement — driver's licenses, tax information, where you live — and combine it with social media data, license plate capture data from parking garages, all of that. So now you can say: look for this name or this picture. It trains its neural networks to go look at a whole bunch of untagged data — just everyday things happening — and it can find you. And today, you still need humans to sort through what comes back. But it's not fiction that it exists.
So what you've captured for ChatGPT is more than just a big pile of data. You also have a model that has been trained by data scientists so that it knows what to spit out. The AI is a large language model — in ChatGPT's case, it goes into the data lake and looks for patterns in the data. The chat box in front of it gives you a natural human way of interacting with all of that data. And data scientists use deep learning and machine learning algorithms — which is just a fancy word for math I don't really understand — to train it to pull a sentence out, write a story, create the resume, whatever you've asked it to do.
It's powered by extremely powerful computers. I know you probably saw this when you were working at your last employer — I saw it at mine — how powerful the computers have become and how fast the chips are these days to make these processes work. And I think facial recognition on your iPhone is a good example of the same kind of algorithm. When I say "find pictures of me," it leaves some pictures of me out. But it also finds pictures of my daughter. She can open my iPhone about 90% of the time. It freaked me out the first time it happened. That's why I don't use any biometrics on my phone.
That is both fascinating and alarming. So I still feel that ChatGPT is something that is valuable — but we cannot use these things without knowledge and care. Not fearful. But careful.
Exactly. Because if you think about what ChatGPT and other large language models can do, there are some pretty cool applications. We both came out of the corporate world and now have our own startups doing product marketing. Running a launch is hard work — it's tedious, and you have to get every little thing right. If you've got a mature product and two or three groundbreaking new features, you have to go in and replace old stuff with new. That takes forever. Imagine if you had your own version of a language model that knew your business — trained on what data sheets look like, what blog posts look like, what all the different elements of a launch look like. You could say, "Run my launch 3.0," and it would gather all the documents for you. All you'd have to do is have the right people review them. My daughter has also written a blog post about how she uses it as an aid — she's on the spectrum and has a really hard time with text. It stresses her out because the language feels so different every time and she doesn't want to upset anyone. She's tried ChatGPT to generate emails that should be quick and easy for anyone else but for her feel overwhelming. So there are a lot of really good uses — but there are things you have to think about.
And number one is that data. Where did the data come from?
There are some issues with the ChatGPT data set. One of them is — since it was scraped from the World Wide Web, and especially in our industry, the tech industry has been predominantly male. If all the writing out there is predominantly male, all of the things scraped from the web carry that viewpoint. There are biases that people are blind to that will get encapsulated in that data set.
I think we should be factual and allow people to come to their own conclusions. Because if not, they won't understand. The World Wide Web is about 30 years old, but in common usage it's probably been the last 20 to 25 years — and it's been male-dominated for most of that time. That doesn't mean men are bad people. We all have biases. Humans have biases — that is how communities work. For example, if anytime a paper referred to an engineer as "he," then when you ask the model what gender engineers are, the answer is male, because that's what it found. I saw something on Instagram — a prompt was given: "Many non-human cultures have developed complex rituals surrounding death and the afterlife," and the example given was Native American cultures. Native Americans are not non-humans.
Exactly. When you were reading that, I was like, where is she going with this? Because I'm thinking — who are non-humans? Animals? Birds? And that's one of the things to be careful of. If you don't think about that because it doesn't impact you, then the problem is that you'll write something based on the prompt you were given and continue to share it. It goes back into the corpus and it just strengthens that bias within the body of work. So it makes everything that comes out of it questionable. The data set is a little flawed for ChatGPT. The thing is not to decide not to use it or to be afraid of it. It's more important than ever to edit and really read what comes out of it with a critical eye.
I felt where you were going with that earlier is — the impact on society when it comes to cleaning that data. Someone had to clean it. There are biases in the data. There are nefarious things in the data — how to build a bomb, pornography, violence, all of those things are on the web. So all of those things were in the data. And to deliver a tool on top of a data set that can be used globally, all of that stuff had to be sanitized. And it was someone's job to do that.
Yeah, that absolutely happened. The bias is harder to weed out — and that's probably, maybe always will be, on us to fight back and say, "Come on, that's not right." To have a viable product to sell, they had to look at the data set and sanitize it. OpenAI hired people in Africa to do that. And they did not give them the mental health care they needed — and that's coming out more and more. And the pay was not what it should have been. That's why they did it in Africa.
So if you have any kind of ESG — Environmental, Social, and Governance — process in your organization, there are things to think about when adopting this technology. If you bring your best and brightest to do a job that consists of sanitizing horrific data out of a data set — data that isn't even going to be used in your country, a product whose end users aren't you — and you don't adequately compensate them or provide mental health care because you just want to crunch through and get it done... what happens when your best and brightest are traumatized by that? So there's that piece to think about.
And I don't want us to beat up OpenAI. Because most technology companies have done something similar — always in countries with a lower GDP and a large workforce. In Africa especially, the concentration of young people is larger than anywhere else on the planet. So there's also the question: why aren't we doing better by them? You've got all these young people trying to get into the world of tech. Why are we giving them jobs that are potentially harmful to them? If you're going to use a service, you've got to think about how the service was created — how the data set was created, and how the people who cleansed that data were treated. And there's also this: you've got so many people rushing to do AI and dropping everything because they can see generative AI as being able to solve problems that are hard for humans. But you've got to think: are you using the right thing? Is this something somebody just slapped together real quick to get in on the curve?
Going back to the ESG angle — it takes a lot, a lot, a lot of energy to run these models and especially to train them. You have to have lots of computers, lots of storage, lots of cooling. The impact on the environment is significant. I think that's something the data scientists and architects are working on right now — looking at new technology so they can run these large algorithms with less environmental impact. But that's something to think about.
So if you're listening to this podcast and saying, "I was going to try ChatGPT, but now I feel conflicted because of all the harm it has possibly done" — I don't think that's the message. What would you say to that person, Gina?
Any AI model is the workload that's coming out of all the digital transformation we were forced to market about in the last ten years. Back when I first came into the industry, the jump was from mainframes to three-tier applications and Linux at the same time. This is the jump we're making now — away from that three-tier application model into containerized workloads that can make sense of what's coming back from these large AI models. No — I think we have lots and lots of smart people that can get on with it and say, "This is great, we're going to figure out how to make these run more efficiently. We're going to figure out what steps we have to take so that we aren't using all of the encapsulated bias." There's going to be bias. We're humans. That's just going to happen. The goal is — instead of encapsulating it and accepting it, how do we make sure that when we see it, we find a way to work around it or repair it? It's all hands on deck. It's not being afraid. And it's definitely not letting the technical CEOs say "we're so afraid of this, everybody needs to stop right now." The cat is out of the bag — and they put the cat out of the bag. What has to happen is: get educated about how these things work and try to figure out the workarounds.
To shorten Gina's passionate answer — yes, use ChatGPT. But be aware of the areas where we still need to work to make it a better technology: a more socially conscious technology, a more equitable technology, and a more environmentally friendly technology. That's going to take time and people — people who care about it and not just care about getting a product out the door to make a big profit. This will make a big profit because we got a product out the door that is great for everybody. That's what I think.
Thank you for joining us today on the Tech Aunties podcast. If you have a topic you would like us to cover, please connect with us on LinkedIn and Instagram. You can also find this episode and others at Tech Aunties dot com. Until next time.