~ 3D modeling with ChatGPT - Solidified ephemerality

I have asked ChatGPT to generate 3D models. ChatGPT can not generate 3D models directly but 3D models can generated via intermediary OpenSCAD scripts: OpenSCAD provides a scripting language to describe objects which can be combined to form 3D models. ChatGPT understands the syntax of this scripting language and generates perfectly cromulent scripts. I have asked two versions of ChatGPT to generate a 3D model of a house, a cat, a stick figure, a chair and a tree. The results are interesting…

The models immediately make the difference between ChatGPT 3.5 Turbo and ChatGPT 4.0 clear: 4.0 generates much better models with, at least, recognizable elements: a chair has four legs, a cat has a head and a tail. It is impressive that reasonable 3D models are generated but there is still room for improvement: proportions are not respected and elements are not always connected. Anyway, if the 3D-models can be seen as a way to visualize code quality, then 4.0 is a clear improvement and it makes me curious about future ChatGPT versions. It also made me reflect on a couple of aspects of LLMs in general.

Fig: a black box generating 3D models.

• ephemerality The response of a LLM to a prompt is ephemeral: the same prompt causes a different response depending on context, randomness and the position of heavenly bodies - or so it seems. Traditional software systems follow a strict set of clear rules and provide deterministic, predictable and reliable results. The inverse is true for LLMs which takes some getting used to. As a user, an LLM is effectively a vantablack box - there is no way to know why a certain response was given instead of another.
• Updates LLMs services - and SaaS in general - have an additional feature which makes them even more unpredictable: updates to systems can happen without notice. After a recent unannounced update, for example, ChatGPT 4 started to produce gibberish. This adds another layer to the already uncontrollable and ephemeral nature of responses to LLM prompts.

To counter the ephemeral quality of prompt responses, I have 3D printed the generated 3D models. Some pictures can be found below. I find that these physical, tangible, immutable objects provide a comforting counterbalance to the digital, ephemeral nature of LLM responses. Additionally, it highlights the absurdity of the generated models.

There are other ways to solidify ephemerality: crochet patterns, juggling patterns, guitar tablature, music notation all have some kind of structured text representation which LLMs can generate and which can have a physical representation. I would encourage people to bring prompt responses to the physical world: it really makes the - current - limitations of LLMs very clear.

UGent and Code