Originality in the age of AI
Dear XYZ,
If you’re the kind of person who believes that every idea that will ever live already exists and originality does not or no longer exists, then you can stop reading the essay here.
Otherwise, if you, like me, believe that people and culture at large are continually creating new ideas that are distinct from prior ones, then a pertinent question in the age of AI becomes the relationship between AI (specifically LLMs) and originality.
My main explanatory model today will be the idea of dimensional spaces. Anyone with a basic understanding of geometry, the difference between 2D and 3D, or better yet, the concept of embeddings (if you have learnt machine learning), should be able to grasp the gist of my argument.
I think novel ideas fall into two categories, which we can differentiate using spatial dimensions. I don’t have good names for these yet, so I will be relying on the model to refer to them.
The first type of new idea is about filling in existing known space. If you have an idea represented by the 2D coordinate (1,2), then you know that there can be a coordinate (54,376), because that is already captured by the axial logic of the 2D coordinate system. We can call this intradimensional originality. In other words, we’re extrapolating and interpolating between known ideas to get to new spaces. Let’s use the example of a horse and a man. In reality, a combination between those doesn’t exist, but our minds can conjure an interpolation: a centaur. Extrapolation could be maximizing the quality of some attribute. For instance making the best chair (the chair-est chair, as Platonists might say), or perhaps making a car go faster.
The second type of new idea creates a new dimensional axis. Its existence expands the possibilities of thought. It’s akin to going from a circle in 2D space to a sphere or cylinder in 3D space. 3D forms cannot be fully represented in the prior 2D space. Our 2D coordinate (1,2) becomes (1,2,x) and an entirely new axis becomes available for us to explore. A 2D plane becomes a 3D space. We can call this neodimensional originality. Examples of this may include discovering that light behaves both as a particle and a wave or a carpenter having screws as opposed to a saw and glue — it expands the way we see and the possibilities we have at our disposal.
We can understand the most recent AI achievements using these notions of originality, the most salient of which being the AI-fueled disruptions in Mathematics. AI will continue finding proofs and solutions like Navier-Stokes. However, I don’t see it inventing Lean, new mathematical questions, or novel perspectives to these problems. It extrapolates and interpolates, it does not create new dimensions. And perhaps we can argue that this is fundamental to the training regime of LLMs, they are trained on existing corpuses. This is also a point of contention in the Navier-Stokes debacle, where OpenAI cannot confirm that their training data did not include transcripts with Buckmaster and Alpoge.
I’m sure that there are ways in which this distinction can be disproven and as with anything, things are likely gray at the edges. I think it is possible if we give LLMs increasingly vague and open-ended instructions that begin to model human-like ontology. However, I would argue that this would be a terrible idea in the first place.
Beyond the question of originality, we should also interrogate, like Terence Tao has, the underlying value system, beliefs, and culture of our practices and communities of practice. Why we continue making and inventing things and how that relates to what we understand to be “original” are possibly even more important meta-questions we should consider.
xoxo,
K