Feb 3
Thinking Outside the GPT Box
A [Human] Chat about GPT-3 with Michael Schrage
With all of the buzz surrounding the nascent Generative Pre-trained Transformer 3 (GPT-3) technology — from talk about displacing huge sectors of workers to student plagiarism running rampant — the MIT IDE wanted to take some of the sting out of the conversations and offer a high-level perspective about the capabilities, as well as the limitations, of this emerging AI.
GPT-3, and ChatGPT are state-of-the-art, language processing AI models developed by OpenAI. ChatGPT, which launched in November 2022, is specifically designed for conversational tasks, whereas, GPT-3 is a more general-purpose model that can be used for a wide range of language-related tasks. It is based on a neural network machine learning model trained with Internet data to generate text. The tools are free to download and require small amounts of human text commands to generate large volumes of relevant and sophisticated machine-generated text.
These abilities — along with image-making tools like Dall-E — have led to lots of speculation about what’s next as the technology matures. Clearly, GPT-3, like many AI technologies, calls for a re-examination of the concepts of work, human creativity and automation, and their intersection points. IDE Content Director, Paula Klein, asked IDE Visiting Scholar, Michael Schrage — who studies human-machine interactions — for some fresh insights on the latest wizardry. His comments follow.
MIT IDE: As someone who has studied human/AI interactions and the potential impact on business, what are your thoughts about Chat GPT in its present iterations, and what’s the potential?
MDS: ChatGPT and Dall-E are clever instantiations of what is called generative AI, or generative ML — “pre-trained” models capable of algorithmically generating original and contextually relevant content in all manner of media — text, images, sound, software, etc. They can go beyond clever, and when appropriately prompted, are impressive, too. At least, I’m impressed! Generative AI can learn, but it also can learn how to learn so we will increasingly see “mashups” of words, imagery, code, animations and artifacts that people — amateurs and professional alike — will tune and edit to their satisfaction. Just as social media influencers work really hard to curate selfies and digital identities on TikTok and Instagram, my bet is that
we’ll see more of what I call, “interactive iterative intimacy,” between individuals and their generative platforms.
IDE: Will anyone learn to write anymore!?
MDS: Sure, people will still learn how to write — just as people with calculators and computers still do math, but it will be more of an iterative relationship. Generative algorithms and their human collaborators will team up and edit each other. To add some humor, but also some perspective, think of it like this: “Do people train their dogs, or do their dogs train them? Are cats really our pets, or do we reorient ourselves around our feline’s whims?” I see generative algorithms evolving — I pick that word deliberately — to both effectively and affectively inform how we create. This creative evolution will prove as true for our personal lives as our professional practices.
IDE: The threat of job elimination by technologies like ChatGPT — without replacement employment — is a legitimate concern for workers and employers alike. What are the labor implications, as well as the upside for businesses such as publishing, marketing, PR and consulting?
MDS: This is always the question that gets to me most because I’m much more comfortable anticipating how technologies diffuse, adapt and scale around certain use cases than how they impact jobs. The work of MIT economists like Daron Acemoglu is remarkably helpful in analyzing the impact of technology and automation on the tasks that make up jobs as we now know them.
I expect the top decile and quintile of clever knowledge workers who rely on — apologies to Robert Reich, “symbol manipulation” — will take advantage of generative to differentiate themselves, their creativity and productivity. We’ll no doubt see generative templates for median and typical knowledge workers, and sure, a good chunk of what we now think of as creative will be generatively automated sooner rather than later. From where I sit, the past quarter century has made clear that it’s getting harder for people with so-called average or typical skills to command a premium in the market.
So, yes, the composition of and expectations around jobs and their deliverables will continue to shift to smarter machines.
But remember, the same AI that can do more of what you already do can also give you actionable insight and advice into how you can make more money from working better with it. One of the clear takeaways from the book, Recommendation Engines, that I wrote, is that smarter technology increasingly can give you advice worth following and profiting from. Increasingly, the best recommendation engines will be generative; they will coach, tutor and train you, if you are so inclined. That’s where I believe the market and these technologies have to go.
IDE: From an investment perspective is ChatGPT a niche market, or do you see it growing and expanding to include other forms of human/AI collaboration apps?
MDS: If you look at Microsoft’s investments in Open AI, Google’s efforts to make search more chat worthy, and where billions in venture funding are going, it’s clear that expectations are grand and growing. Generative will increasingly live in the cloud, and the growth of multiple-cloud climates will prove economically hospitable to further generative innovation. Mobile devices will help pre-train those models, as well as datasets, that develop more bespoke generative recommendations and creations.
Bluntly, we’re already cyborgs; generative will enable ever-smarter and better-structured “choice architectures” and options for both our productive and consumptive lives.
IDE: There are also concerns about plagiarism, protecting IP, and the notion of blunting human creativity and expression. What scenario do you envision when we no longer have to labor over essays and snappy product ads — among other efforts?
MDS: Some of these assumptions must be challenged and reexamined. A lot of us — myself included — like the option of being intensely involved in creating something versus simply being able to swipe left (or right). This is why I referenced choice architecture. What are the choices we make by default versus by design? This line of thinking directly relates to the brilliant Nobel Prize-winning behavioral economics and experimental economics research of people like Richard Thaler, Daniel Kahneman, Michael Kremer, Vern Smith and Esther Duflo. How should people come to create, generate, edit, modify and experience choice? When is it better to automate choice versus augmenting human decision-making with generative algorithms and data? These used to be rhetorical questions and research themes, now they’re the organizing principles for innovation initiatives worldwide.
IDE: Did you anticipate some of these trends in previous work on AI and individuals?
MDS: I’ve written about “selvesware” — technologies that amplify and enhance elements, aspects and attributes of one’s self — as an essential innovation trajectory that we’re all on. Technology is transforming both the style and substance of agency; that feeling of control that we have over our actions and their consequences. Generative is both accelerating and intensifying that transformation. To take it further, consider what happens when exercising that kind of agency doesn’t feel like labor or work at all? That’s a fresh way of thinking about new technologies instead of trying to fit them into old models. That’s where the exciting innovation lies. Where will we take it?
