tl;dr: Generative models pulls AI out from behind the scenes and make the power of the technology tangible. This is the first time in history that non-techy humans and AI can interact directly with one another, which may propel curiosity and adoption for generations. It also happens to raise fundamental questions about how "human" creativity really is.
The recent breakthroughs in generative AI have left the whole world dumbfounded. It seems as if everyone and their dog have made or seen a piece of AI art in the last six months. Images like the ones shown in this article. Advances in generative pre-trained transformers and, more recently, diffusion architectures, have truly catapulted us into a new era of creative expression. Neat, right? AI art and literature are, indeed, brilliant in themselves, but I'd argue that the societal implications may stretch far beyond. Let's dive into it.
Who really cares about AI?
Outside of the tech echo chamber, the answer is no one. Sure, the average person cares about relevant search results, autonomous vehicles and auto-correct on their phones. But how it happens is not really important. AI or 5 billion if/else statements - It's all the same. If anything, AI may be a bit scary to most people. Damn Skynet.
As data scientists, we often find ourselves evangelising the gospel of AI, but, let's face it, it often falls on deaf ears outside of our labs and conferences. Or it did until recently. With the emergence of creative AI, the power of the technology has been made tangible and apparent. Creative AI is immediately impressive and not just another "think of what this could lead to" act. This changes the value proposition, and, consequently, the societal narrative of artificial intelligence. Effectively, it draws AI out from behind the scenes, into the limelight, as a stand-alone wonder - and not just another cog in the machine.
"I could do that", is a justified response to many AI models. A dog breed classifier, for example, is not likely to be more accurate at its task than a human would be. The fact that the model is factors more (cost-)efficient, doesn't really make it any more magical to a consumer. Creative AI, on the other hand, is as close to magic as it gets, and definitively shows how the capabilities of AI stretch far beyond (average) human ability.
Hi robot, hello human
Text-to-image, image-to-music, X-to-Y. We can think of these conversions as approximations of idiomatic dialogue. To humans, a "blue world" means something sad, a "burning heart" means passion, and "angry, orange man" could mean Donald Trump. Meaning, subtext and idioms are intuitive to us, but harder to learn for language models. In part, this is due to emotive interpretation. In the image below, do you see a single bird on a branch in the rain, or do you see a sad bird?
With the recent advancements, we may be significantly closer to closing the idiomatic gap in interactions with AI systems. That is, systems that understand what we mean, not what we say.
As AI systems learn to decode us, we also learn to decode them. Prompt engineering - The ability to identify which inputs to generative AI systems yield which results - is already a sought-after skill. This job posting for an AI Prompt Engineer, for example, requires "1+ Year of Experience Prompt Engineering for Large Language Models". Figuring out how to tame generative models will be essential in a future with creative AI.
But creativity is for humans!
A curious thing about our relationship with automatons and artificial intelligence is that we tend to move the goalpost for what it means to be "intelligent". While we have (begrudgingly) ceded territory after territory, one area remains a shining bastion of humanism in the eyes of the public: Creativity. The new capabilities of creative AI seem to bulldoze this notion.. Or does it?
DALL-E 2 from OpenAI, possibly the best text-to-image model around, is trained on hundreds of millions of captioned images, sourced from publicly available and licensed datasets. DALL-E 2 has learned the relationship between these images and their captions remarkably well. So well, in fact, that it can extrapolate across thematics, styles and motives to create completely new visuals. Arguably, however, is it not just regurgitating and combining human creativity? One answer is: "Yes, but so are we".
Our brains have been pre-trained by millions of years of evolution, and are then fine-tuned during our lifetime. Just like DALL-E, we regurgitate and combine what we have seen before. So, is creativity inherently human? Art is said to be a reflection of ourselves, and this may hold true for AI as well.
Top comments (0)