People misinterpret the Ramanujan and Einstein examples and assume this just means AGI has to be a super genius otherwise it doesn't count. To me this doesn't seem to be the point; they're illustrations of categories more than thresholds. Demis previously described creativity as falling into three buckets:
Interpolation, i.e. averaging many data points and recombining them, e.g. an image model producing a new photo that would have not been in a dataset before. Extrapolation, i.e. going beyond the convex hull of the training data to produce something experts recognise as genuinely new - we have a lot of examples of this with existing systems, like Move 37 and all the recent maths examples. And thirdly, invention: actually coming up with the game of Go in the first place. It's the third type that existing systems, at least as currently scaffolded/used, appear to lack.
There may be a difference between solving existing conjectures or problems, and actually coming up with the theory in the first place. Language models produce all sorts of new outputs, but the outputs differ qualitatively in kind even if they're novel to varying degrees. Does the new output live inside the conceptual space defined by the training data (interpolation), push outside it along existing dimensions (extrapolation), or reframe the space itself by proposing a new conceptual structure (invention)?
Ramanujan illustrates the depth of intuition and innovation that the third 'true creativity/invention' category represents. There may be recombination and extrapolation as part of that process, but at least so far they don't seem to be sufficient. Of course you can argue that most humans don't do this - when was the last time you invented a new abstraction?
My uneducated view is that this third type of creativity doesn't have to lead to crazy new inventions - obviously inventing the theory of relativity is both more impressive and more *valuable* (and thus depends also on a social aspect and utility), but I think individuals do this third type of creativity in many smaller/micro ways too. For example a teenager on TikTok warping a meme format into something the format didn't previously allow/cater for. This kind of type-3 operation should be observable at small scales, frequently, with low stakes too.
I think it's very reasonable to argue that this is *just* interpolation/extrapolation, but I personally don't think this is all there is. If it was then it should be feasible for e.g. Talkie (the model trained up to 1930s data) to be prompted/scaffolded to create a new abstraction entirely. I haven't seen this, except by teaching it basic coding through fine-tuning - but that's not the same thing, since you're showing it the new abstractions (coding) to start with!
I'm not sure if this is because (a) language models lack the cognitive operation entirely, given the architecture; or (b) it has the operation but lacks the appropriate scaffolding, memory, and process. Maybe it's (b), though I lean towards (a): language models are better at paradigm exploitation than paradigm generation. Of course practically speaking, you don't strictly need this third type of creativity for the technology to be transformative and revolutionize fields; but it's a difference that is still worth highlighting and accounting for, since it also implies certain ceilings determined by the shape of the existing conceptual space.
