Color me skeptical. LLMs seem to make writing code faster, so of course that means that people can iterate on ideas faster, but I have yet to see actual creative output from an LLM that wasn't coached into it or random juxtaposition.
Perhaps I don’t understand relativity and its precursors well enough to see how it was incremental, but either way i’ve yet to see anything remotely like relativity, or even anything like more traditional examples of creativity like picasso or van gogh, come out of an llm
Einstein contributed an argument (relativity) in 1905 for why this should be the case, which lead him to make important inferences such as mass energy equivalence. i.e. he took existing maths (derived from experiments) and was able to extrapolate it to other cases.
General relativity 1915 is similar. Starting from general covariance (the equations should look the same in all frames, even if the numbers/coefficients change) and his equivalence principle (acceleration indistinguishable from gravity) there are only so many places in the equations where a simple formula for gravity can fit (and there were sound reasons to expect the formula is simple). Einstein (and others) developed several of the possibilities before he settled on the famous field equations.
I say all this because once you realize each of these ideas was built up incrementally by dozens of people over the course of years/decades and was deeply rooted in experimental observations, then it's harder to believe that it is beyond the capacity of AI. In fact, AI is contributing to cutting edge understanding of gravity https://bigthink.com/starts-with-a-bang/ai-first-breakthroug...
In the first case you are directing the creativity, in the second you can ask for non-random all you want but the output will be definitionally random based on the training weights and the temperature setting of the prompt.
... provided you keep eyes closed when examining the output.
To those gulled by the AI coding hype, I suggest try a trivial code task with e.g. immediate feedback in a domain you can swiftly verify e.g. HTML or SVG. Your red pill.
Example: https://chatgpt.com/share/6ac0d871-6314-83eb-990a-41baed9642...
But your example feels like a bad illustration of your point. LLMs operate on words, and can’t read your mind - if you had given me (a human) this task with these prompts i too would have failed because i can’t divine your intent even slightly.
It didn't need to. It operated on my prompt words just fine. It provided precisely the starting shape I wanted.
> - if you had given me (a human) this task with these prompts i too would have failed because i can’t divine your intent even slightly.
Then I prompted "Show as html svg". Clear?
It was clear to the bot. Bot showed HTML SVG proving it understood perfectly.
The problem is, it garbaged the content. Imagine the equivalent buried in 1000s of lines of code.