For a technology as powerful and exciting as generative artificial intelligence (AI), I’m often surprised (and sometimes embarrassed) at the ways in which people use it! Far, far too often within today’s market, I see people and organizations use generative AI in ways that are, simply put, boring.
What do I mean when I say these generative AI use cases are boring? I mean that most of the time, these are use cases simply using generative AI to do things we’ve already had a handle on for decades. These include using generative AI to perform basic tasks like text classification and content retrieval. Frankly, these aren’t problems that are worth using generative AI on – in fact, doing so is probably costing organizations. Using generative AI to do these tasks is like using a super space laser as a hole puncher.
Let me preface this article by saying that I’m no clairvoyant – if I had all the right answers and could predict what the best generative AI use cases will be five, 10, or even 20 years down the line, I would do so – and be a very rich man. But alas, I can only dream. What I can do, however, is use my decades of data science and AI expertise to analyze today’s generative AI use cases and give people and organizations a quick, easy set of frameworks and tools that will help them sort the gems from the sediment when it comes to generative AI use cases.
Let's Not Retread Our Steps
It’s human nature to use new technologies in familiar ways – people have been doing that since the dawn of time, and I doubt 2024 will be the year we stop. And just because I’m calling certain generative AI use cases “boring” doesn’t mean these applications aren’t useful, it just means that you don’t need – and often don’t want – generative AI doing those tasks. Generative AI is only one type of machine learning under the broad umbrella of machine learning. Other types of models, like discriminative (“traditional”) machine learning models, have been in mainstream use for much longer for good reason – they’re great at what they do while being cost- and resource-effective.
Let’s think about it this way: discriminative machine learning models are superb at taking some inputs and generating simple outputs. You have customer data and want to predict churn? Done. You have inventory data and want to predict how much will be in stock next week? No problem. And not everything even needs machine learning in the first place: looking for an answer to a question? Google it. So, let’s save those simpler types of tasks for machine learning models, search engines, or other, simpler tools that can do them quickly and efficiently. We don’t need a tool of ChatGPT’s power and scale to help us set our alarm for 6:30 tomorrow morning.
But what are we currently doing? Throwing this amazing technology on those simple, well-understood and well-served problems. This is ok if you do this for personal entertainment, but it’s not the best idea to apply a powerful technology like generative AI on enterprise use cases that can be solved more efficiently otherwise. We can generate anything with generative AI, and all we can come up with is to use it as an expensive (and sometimes hallucination-prone) search engine? Please, let’s save generative AI for tasks that have complex inputs and outputs.
Now, I don’t want to seem like I’m ranting. There’s always a learning curve when transformative new technologies arrive on the scene, and the generative AI use case landscape is far from a complete disaster. But I know we can do better. And I know that organizations looking to reinvent themselves or make their mark on the world usually don’t have the luxury to undergo trial-and-error in a competitive landscape. So how can we better identify effective ways to utilize generative AI? Enter the impact/feasibility matrix.
The Impact-Feasibility Matrix: Your New Best Friend
When looking to implement generative AI, businesses and individuals should always be asking themselves: What can we pursue that’s both feasible and impactful? This may seem self-evident – who would pursue the opposite? – but anyone who has seen the inner workings of an organization knows that it’s quite easy to lose track of priorities and get tunnel vision, which robs us of crucial perspective and can cost us dearly.
I’ve populated the matrix below with examples of generative AI use cases ranging from low impact/low feasibility to high impact/high feasibility. Let’s call the latter quadrant the “golden zone.” What makes these use cases so special? They’re both highly feasible and provide immediate, significant business acceleration. These use cases aren’t theoretical, either – some organizations are already leveraging them, though many, many more could (and should) be. Let’s break down each axis.
Business Impact: It’s All About Acceleration
When we talk about impact, we should always be thinking about acceleration. Code generation, for instance, is a great example of generative AI’s ability to take complex inputs and produce complex outputs. Writing code, even for master coders, can be a complex and time-consuming task. Great code is also invaluable to organizations and their solutions. With generative AI, an expert coder could prompt a model with something like, “Generate some code that sorts this array here in descending order” and receive a working output almost instantaneously. Sure, they could write that code themselves, but generative AI accelerates their job multiple times over. This is a capability that, until now, was only a dream.
Conversational analytics is also another “golden zone” use case because of its accelerative potential. Within business intelligence (BI) and other advanced organizational data/AI tools, you could ask a model something like: “What have our sales been for each of our product groups, and are there any trends?” To do this in code would take a person, team, or department a long time. But with generative AI, you can have insights in a couple of seconds – without even needing any data analytics skills.
Code generation, image generation, conversational analytics, generative design – these are certainly not from boring applications. Much the opposite – this is what generative AI was meant for! Doing things we have never been able to do before with (relative) ease. They’re also achievable and will become cheaper and easier to use as the technology and the people that use it get better with time.
Feasibility: Costs Come First
“Impact” is usually a straightforward concept, but “feasibility” can be much more elusive. Obviously, there’s the technical aspect of feasibility – what can current models do and not do, how easy is it for our organization implement, deploy, and maintain these models, and so forth. But the financial side of feasibility, too often overlooked or underestimated, is my main concern.
In today’s market, there are a lot of false impressions out there. That’s because the most popular generative AI tools, like OpenAI’s ChatGPT and Google’s Bard for instance, are essentially being subsidized by massive organizations. The ease of use and cheap sticker prices of these tools (and similar ones) belie their true costs. I assure you, if these tools’ true costs – including the financial, infrastructural, environmental, human, and beyond – were reflected in their sticker price, we’d be paying much, much more than, say, $20 a month.
This is a great example of how many costs can be “hidden” to consumers. For a real-world demonstration of “hidden,” costs, let’s examine single-use plastic bags. The financial cost of these bags to consumers is often $0 since stores give them out for free. But the “costs” of plastic extend beyond the world of dollars and cents. Plastic is produced from oil, which has to be drilled, refined, and transported; bags end up in landfills (and often in the environment); and since we don’t reuse them, we use an estimated 500 billion single-use plastic bags worldwide every year. If we account for the toll all these bags took in 2019, the World Wide Fund for Nature (WWF) estimates the lifetime cost to society, the environment, and the economy of plastic produced in 2019 alone would be $3.7 trillion. By 2040, it could grow to $7.1 trillion. So when we get “free” bags for our groceries or our takeout, we aren’t getting something for free – we just can’t see the costs in our bank account.
The concept of costs – especially “hidden” costs – must be at the forefront of any generative AI strategy. And unfortunately, in my experience, this isn’t how most people are thinking. Want your own secure, proprietary ChatGPT? You’re going to need massive HPC infrastructure, data experts, time, a whole bunch of testing, and more. If you do all this just to reduce a simple 30-second task to a simple 15-second task, you’re going to end up in the red. This is why I get so frustrated when I see people and organizations using generative AI for tasks that simply aren’t worth it.
And lastly on the topic of feasibility, I also believe organizations must start quantifying their generative AI projects to a much greater extent. Just because you completed a project doesn’t mean it was a success. So you implemented your generative AI model for content retrieval – does it perform the task faster, cheaper, and with greater accuracy than a traditional discriminative model? If not, then you essentially wasted your time, your effort, and your money. And if you can’t even measure its impact, then we have a bigger problem afoot. Success is measured in data. Too often right now, people aren’t doing enough quantifying because they aren’t keeping costs and feasibility at the forefront of their strategy.
I don’t mean to sound like I’m a generative AI pessimist. Much the opposite – I think this technology is going to be one of the defining technological catalysts of our era and beyond, our version of the World Wide Web or smartphone. It’s exciting to watch this development unfold because I know how much good it can bring us!
New technologies bring a learning curve, that’s all. As someone passionate about machine learning and technology in general, I want us to succeed as quickly as possible so we can bring about the next generation of innovations that will reshape our world for the better. I believe thinking in terms of the combination of both impact and feasibility for generative AI can help people and organizations accelerate their operations.
Obviously, the use cases on this matrix are going to look very different in a year, three years, five years, etc. What isn’t feasible at all now may be trivial in a couple years. This is also true from industry to industry – what may be feasible and impactful in aerospace may be the complete opposite in retail and healthcare, and so on. But while the use cases and industries can vary, there will always be tools and frameworks that can guide our thinking so we can maximize our efforts and resources.
Like I said, I’m no clairvoyant. But I know we can do better! So please, if you’re reading this: promise me you’ll start using generative AI, your metaphorical super space laser, for exciting, high-impact things like harvesting asteroid diamonds – the humble $10 hole puncher can handle the papers on your desk. I guarantee your team and your balance sheet will thank you.
To learn more about Altair’s data analytics and AI capabilities, visit https://altair.com/altair-rapidminer.