Idea amplification is the really exciting potential of generative AI
When I was 18 years old I discovered I had an unusual ability. One day I found myself humming a tune and realized I couldn’t place it. So I decided to sing it into my computer’s sound recorder program, including the lyrics. When I played it back, it was obvious I had just made up the whole thing, melody, lyrics, and all. Not only that, but I kind of liked it. So soon enough when I found myself humming another unplaceable song, I did the same. It started to happen so often, it became a kind of hobby. Over the next few years, I accumulated hundreds of original improvised lyrics and melodies.
Unfortunately, beyond this unusual hobby, I don’t have any well-developed musical skills or talents. I can barely keep to a beat, read musical notation, or play any instrument well. And anyway, I was dedicating most of my energy back then to becoming a computer scientist and AI researcher, so I didn’t feel I had time to really dedicate to developing this questionable “talent.” As a result, while my song collection accumulated, there was nowhere for it to go.
But you have to wonder, do I have some kind of latent musical talent that I simply lack the tools to express? Are there tiny seedlings of musical grandeur lurking in the deepest reaches of my mind, never to be planted? Maybe there are creative seedlings in your mind too — tiny sparks of inspiration utterly exclusive to you, yet beyond your ability to ignite with your current skill set. The sadness for us all is that these sparks from beyond our domains of expertise have nowhere to go — locked inside, never to see what might have been.
Until now, that is.
Something fundamental is changing. Suddenly, there is a technology—generative AI—that promises to seamlessly fill in the gaps between inspiration and expressive ability. There is still a tiny (yet profound) place where human inspiration must provide the spark, but the fuel for the fire doesn’t need to come from your own mind and body—AI will do most of it for you. Consider OpenAI’s DALL-E or Stability AI’s Stable Diffusion. The spark is the written prompt, through which an idea travels from your mind into the computer, but the beautiful, complex image that emerges is all DALL-E or Stable Diffusion. They take your creative inspiration and amplifyit into real creations.
In effect, it brings millions of spectators on the sidelines of the creative enterprise right into the fold and makes them participants instead. That is the promise of generative AI: We will be liberated to see our creative visions realized without the need for skill or talent beyond the initial spark of inspiration.
This promise is not exclusive to any single domain, but let’s imagine what it means more concretely when it comes to music, using my improvised vocal tracks as an example. In principle, generative AI could someday soon take even an off-key or off-beat recording of someone singing a tune in their bathroom and turn it into a polished, full-fledged musical production with backing instruments, and perfect pitch and timing! You could be a rock star and experience the joy of hearing your spark of inspiration fully realized.
Of course, the AI training behind such a capability fundamentally depends on the available data. We’ve seen that the best models in this area thrive on access to massive volumes of data. The consumption of such data to train AI models is controversial. But in the long run, while companies controlling highly valuable potential training data (for example, record labels) will understandably hesitate to share their intellectual property for AI training purposes, they will also face the possibility that a competitor takes the leap before they do, meaning they could miss a valuable opportunity. Individual artists will also rightfully demand a say (and a stake) in such decisions. Excruciating deliberations on whether to share data (as well as confrontations on whether it should have been shared) can be expected in coming years.
This idea of amplification of initial creative sparks through generative AI reminds me of the earliest days of my career when I started to see students working on projects that I had conceived. It hit me that there are countless programmers in the world just as skilled as I am or better. But I still possess a set of unique ideas. And if I have the privilege of working with programmers who bring those ideas to life, that is a form of idea amplification. If I can focus on the ideas and not the programming itself, many more of my ideas can come to life. The amazing thing about the new creative AI is that the privilege of experiencing idea amplification is coming to everyone. You can provide the spark and the AI can make it into reality.
Stepping Stone Creators
Several years ago I wrote a book (with coauthor Joel Lehman) called Why Greatness Cannot Be Planned. One of its central themes is that the real key to groundbreaking discovery and innovation is not setting an objective, but rather identifying stepping stones that might be interesting. Often we don’t actually know where interesting stepping stones might lead, but we follow them because we understand that they open up entirely new vistas of possibility. In the book, we make the case that one of the greatest unsung talents of human beings is our ability to understand where there might be interesting potential even though we don’t know its precise destination.
Generative AI is a monumental new stepping stone. It will lead to new companies, new applications, new professions, and even more new technologies. But it’s also an obvious stepping stone at this point. Everyone is already talking about it. Often, for individuals, the greatest successes come from identifying an important stepping stone before anyone else realizes it’s there. With the entire world recognizing a new paradigm emerging, it’s a little late for that now. But that doesn’t change the fact that collectively we are going to be seeing a lot of interesting innovations jumping off this stepping stone, some of which will be unpredictable.
Even so, creative AI is no normal stepping stone. In fact, it does something radically novel with respect to steppings stones in general that few technologies have done before: it can instantly transform innumerable former dead ends into viable stepping stones. It is this ability—to turn the dormant into the vibrant—that distinguishes it so uniquely.
Consider all the ideas trapped inside our heads that are seeds and sparks for something interesting, but that cannot escape because of our lack of individual skills and talents (or inability to convince others to work with us). Until now, just like my songs, these are all just dead ends, almost certainly leading nowhere and therefore uninteresting.
But if AI can now take one of these sparks and truly realize a full vision around it, then these former embers are suddenly live stepping stones! In effect, an incomprehensibly vast collection of stepping stones has suddenly sprung into existence all at once. There are few precedents for such an event, and interestingly the people who recognize these stepping stones will tend to be the individuals within whom they lay dormant. The potential innovation to be unleashed is indeed unprecedented.
In this way, generative AI is not just a new stepping stone in and of itself, but a new stepping stone creator.
The Fall of Talent
Even as we celebrate this new creative empowerment, it comes to us bittersweet. All these amplifications and stepping stones—music, art, coding, writing, whatever it may be—are human skills and talents. For many of us, they play a central role in our identity, our uniqueness, and our self-conception. I look at my 8-year-old son reveling in the joy of drawing cartoon characters with his own unique style and personality, and I think of how it used to be that those drawings might be the seeds of future dreaming, of becoming someone uniquely capable and expressive. But now, I feel less sure. What would it mean to develop a unique artistic style and talent if a computer can output anything in any style anyone wants instantly? In fact, the computer could even rip off my son’s own style within seconds after seeing his drawings.
What does talent or skill mean anymore in such a world? Talent and skill are falling to computers, and we haven’t seen that before. There are profound professional implications, such as which careers will remain viable, but also psychological and even philosophical fractures that will need to be addressed. What will make us feel special or valuable in the era of creative AI?
Self-expression is so essential to joy and humanity that the crumbling of our uniqueness is inevitably a crumbling of our self-conception. Of course, my son can continue deriving pleasure from drawing pictures, and it is no doubt healthy for his development. But a time will come when the meaning of such self-motivated personal development confronts a novel reality, for him and for us all. I don’t think this issue should be swept aside, even as we derive new pleasures from our interactions with creative machines.
The Last Remaining Talent
Even as talent and skill are broadly challenged by creative AI, there does remain one vestige of human talent that seems safe for some time to come: the ethereal spark of inspiration itself.
After all, someone still has to come up with the idea for the image, song, story, program, or invention. Wherever such ideas come from, AI doesn’t seem to be quite there yet. Though they can arbitrarily juxtapose different concepts to give a veneer of creativity, LLMs and other large models still struggle with genuine originality. The originality they do achieve, to the extent it’s truly inspirational, still tends to come from the human using the AI.
For example, if you really push an LLM to give you something genuinely original (i.e. not just a superficial pastiche) completely on its own, often what you get back is actually something that was already invented before. In a sense, the ability to conceive the spark itself is gradually becoming the last talent standing.
So we can imagine, perhaps hopefully, a future where we are all sort of creative gurus sending sparks into AI models that then explode into realizations and amplifications of our reveries.
But before we embrace that vision (to the extent it provides any comfort from the fall of talent), there is one important dose of sobriety. It is possible that the spark of genuine, paradigm-shifting creativity is still vanishingly rare among humans anyway. If that is so, then connecting people’s ideas to powerful idea amplifiers will mostly just amplify mediocre ideas, leading to a slew of derivative creative products. In other words, maybe those liberated stepping stones weren’t that interesting after all. Of course, the tiny minority who have incredible vision will indeed be empowered to the benefit of us all, but the overriding flood of derivative mediocrity, though splendidly packaged, would be much less exciting than it seems. That world, which may or may not be coming, might be a let down. True sparks fly only as often as they do, and that has not changed.
The Fall of the Spark
As we attempt to connect the dots even farther toward the horizon, there is an almost unthinkable possibility that must at least be considered: The spark of inspiration itself, the nascent seed of ideation, the source of a billion future stepping stones, could one day fall to AI as well.
While “where do ideas come from” is tough to answer formally, I see no reason to believe that our ideas are anything other than algorithmic. In other words, while the process of idea generation is elusive, it is ultimately a computation. Most likely it involves in part a synthesis of vast amounts of experience, which is basically just data. And synthesizing monumental data is exactly what large models do. So generating that spark is not beyond their purview.
I don’t think we’re there yet, or even close. One problem is that the data we’re currently feeding these models is not necessarily capturing the right information to discern the origination of ideas. An important factor in originating impactful ideas (discussed often in my work) is an intuition about what’s interesting. Yet much of the reasoning behind why ideas are interesting (including how to evaluate their novelty) is left unsaid and implicit, which means that even the entire internet of data may not be sufficient to cover this critical bridge toward the pinnacle of creativity. Even given such a viable notion of interestingness, to get to new ideas from where we are today there is also an algorithmic process needed to proliferate interesting stepping stones in a systematic, divergent way (which relates to my own work in novelty search, quality diversity algorithms, and evolution through large models). So we may have significant research to do before we have machines that are truly inspired. But that doesn’t mean it’s impossible in principle, just that it may take more work, and (to give an example) tools for extracting human intuitions like reinforcement learning through human feedback (RLHF) are waiting in the wings.
If that does happen, then even the spark can fall, and the world will become an even more confusing place to be human.
That said, we should embrace that even if computers could someday match the ethereal quality of our inspirational sparks, it would still be true that your own life experience itself will still be unique, and so your own sparks, whatever they may be, will still remain uniquely you.
For that reason, it’s important to remember that even as AI ascends, though we don’t know how far or how long it will take, humans will remain a unique source of stepping stones for each other. (The importance of continuing to elevate and experience our uniqueness as humans is partly why I recently decided to start a new kind of online service called Maven, which is a serendipity network that helps people connect on ideas and reveries they would not otherwise be able to discover online.) It will be increasingly important for us to highlight the opportunities that remain uniquely human as AI becomes a growing participant in our creative lives.