A skeptical take on the disruptive potential of generative AI
Most people think generative AI will be disruptive. This article argues the opposite – generative AI and the underlying technology of foundational machine learning models will be incremental.
Over the past two years, AI has blown up. The fear of AI ranges from replacing all jobs to becoming a Skynet-like superintelligent entity that results in the demise of humanity. Like the crypto buzz from a few years ago, AI seems to be the next technology everyone’s talking about, and every funding announcement seems to be about. But something about AI feels a lot more real and useful. It genuinely seems like the generative AI wave will be a disruptive technology that changes how we work and live.
Having worked in the AI space for the past five years at Grammarly, I’m here to argue the opposite. I’m here to share why I think generative AI, an incredible technological advancement, is not disruptive but rather incremental. I’m claiming the way it will impact society will be through small percentage improvements on a long tail of industries rather than unlocking new ways of living and working.
What makes a technology disruptive in the first place?
At its core, disruptive technology makes people do something in the world radically differently than they do today. Obviously, that’s a retroactive definition, but there are many leading indicators we can use to gauge how disruptive a technology will be. A great leading indicator for what could be a disruptive technology is if it makes it 100x faster or easier to do something foundational people are doing today.
For example, cars were 100x faster than horses, and planes were 100x faster than cars. Both of these were disruptive technologies. However, what ultimately made them disruptive is unique. Cars allowed people to live farther away from city centers and quickly get around to anywhere connected by a road. Infrastructure investment in roads and the federal highway system quickly took advantage of this in the US, allowing cars to become the primary tool for daily transportation across the country. Today, the average American spends an hour a day in a car.
Said differently, cars allow us to live in one place and work somewhere else. It’s a similar story for airplanes and commercial air travel, which allowed us to get from one point to another in less than 24 hours. Both of those capabilities didn’t exist for humanity before those technologies. Some other popular examples of disruptive technologies were the printing press, which allowed us to print books 100x times faster than before, and the internet, which allowed us to communicate 100x times faster.
Another characteristic of disruptive technologies is that they create significant market value for their industries and have several other multi-billion dollar industries adjacent to or dependent on them. For example, the automobile industry is massive, but equally massive are the rental, residential real estate, and restaurant industries in the US, which depend on consumers needing cars to get around and live away from work. Cars allowed US residents to spread out and purchase their own homes in the suburbs, while Michelin, a well-known tire company, even started its Michelin Guide to encourage people to use their cars more to travel to famous restaurants on road trips.
Similarly, for planes, we have the entire tourism industry, including rental cars and hospitality, built on top of planes. And it’s very fresh on everyone’s minds after the COVID pandemic what happens when air travel suddenly stops – the dependent industries suffer too.
So, we have three (non-exhaustive) characteristics of disruptive technologies:
They make something people want to do 100x faster/easier.
They generate significant market value.
They spawn other dependent industries also of significant market value.
What do we call non-disruptive technologies?
I call it incremental. The vast majority of technologies fall into this bucket since human progress by nature is incremental, with once-in-a-century innovations that are truly disruptive.
Think of the yearly iPhone updates that Apple makes, which, rightfully so, seem like small changes. Surely, today’s iPhone is leaps and bounds ahead of the first iPhone released, but they still resemble each other. And we got to today’s through years and years of iteration and small improvements, feedback, technological advancements, experimentation, and so on.
But the first iPhone introduced back in 2007 seemed, at the time, revolutionary and disruptive in that it brought on the true worldwide adoption of touch-based smartphones that could browse the internet similar to a computer.
Most technologies make percentage point improvements to a process, capability, or functionality and allow us to do something slightly better and faster than before. Of course, this can compound over time, and we may find ourselves slowly doing things differently. But that’s quite different from disruptive technology, which creates a new status quo for society seemingly overnight.
What bucket does generative AI fall into?
First, some quick background. When companies have a problem they think can be solved with machine learning (ML), they need to follow a pretty extensive process to develop an ML model specifically for the task, domain, or application. For example, if you want to do sentiment analysis on some type of text, you’d need to collect a lot of sentences, label their sentiment, and then train a specific ML model on that data. If you also wanted to modify the text to make it more positive, you’d usually have to train a new model with new training data that models the task you’re trying to accomplish.
But collecting the data, training a model, and deploying it to be used within your product isn’t trivial and usually requires very specific and expensive expertise. This is why, for most businesses, ML is either outsourced or avoided entirely. At the end of the day, lots of businesses can’t justify the upfront cost of building an ML team just to solve what, for them, is usually an optimization problem.
However, things we frequently call “generative AI” are tools like large language models (LLM) used in ChatGPT or large vision models used in DALL-E or Stable Diffusion. Those fall into a different category called foundational models and work quite differently.
Foundational models are inordinately complex ML models trained on a mind-boggling amount of data and function like general-purpose models that can do almost everything well. For example, the LLM that powers ChatGPT, GPT-4, can do all kinds of complex tasks with text, such as sentiment analysis, grammatical improvements, text generation, text modification, code generation, etc. Similarly, tools like DALL-E can do image generation, image editing, conversion, and more.
So, how does this change how someone can use ML in their business? Whereas previously, to use ML, you needed to train a custom model, which requires significant time and expertise, now you can just use an off-the-shelf foundational model and get pretty good results just by telling the model what you want it to do in human speak (aka “prompt engineering”).
This is an essential point. These foundational models allow businesses that previously couldn’t solve complex problems with machine learning to use machine learning overnight. And what types of problems are we talking about? Most likely, optimization and efficiency problems, such as:
Optimizing shipping routes for companies operating fleets with international and domestic shipping. (bearing.ai)
Improving crop yield and reducing waste in agriculture. (taranis.com)
Helping large automotive industries with EV battery generation. (aionics.io)
Allowing companies to model and engineer proteins and improve their synthesis. (cradle.bio)
These are all examples of optimization problems – shipping routes, crop yield, battery manufacturing, and protein engineering. The technology of foundational ML models makes it easier for businesses to solve complex optimization problems without in-house ML expertise.
This trend will continue across the long-tail of businesses and industries as well, as famous AI researcher and investor Andrew Ng talks about, over the next few years as foundational models become better and easier to use via lightweight UIs such as ChatGPT and other startups that will inevitably crop up.
So, in conclusion, the impact of “generative AI,” or really, foundational machine learning models, is not disruptive. There’s nothing vastly new that these models will necessarily allow us to do that will change how we live our lives. But that’s not to say I’m discounting the significant world impact that AI will have. Giving all businesses the ability to make a 10% improvement to their complex optimization problems is not, by any means, insignificant. It is definitely an incremental and not disruptive change, but a 10% incremental improvement across almost every industry will have an incredible impact on the productivity of our workers, countries, and, eventually, people.
Too often, the incentives are aligned that the people talking about the technology are the most incentivized to hype it up, and this creates a situation where its usefulness and disruptive power are blown way out of proportion and create unnecessary hype, which I believe does a disservice to the actual productive impact this technology actually has on businesses and people.
I hope this article adds more nuance to your thinking the next time you read about how generative AI will “disrupt” the world and change everything we know and are accustomed to today.