It would be very interesting to see how the whole scene is structured, not just the final rendered result. A lot of the work you put in when making Blender scenes is to make sure the scene stays nicely editable, in particular around the controls you are likely to tweak in the future. Kinda like software engineering when you think about it. Which makes me think that the Blender scene generated by a coding agent probably has the same issues as an LLM-generated codebase, e.g maintainability.

I think that beyond marketing this is also a way to collect "AI browsing" trajectories. Not an expert but I think this is the kind of great quality data that can be useful when training LLMs.

And then you have to click "update visualisation" otherwise you're still seeing the data from the previous bird

I hope the media industry will soon realize they are just being used as a very loud relay in a communication strategy feeding the hype of "AI" companies. This is quite sad (but I'll admit reading the article also made me laugh quite a lot).

Good read thanks, quite well documented!

I see your point but I'm not sure there is much more to say on the matter: apart from having fun, learning stuff and taking pride in your work, what other reasons do you see for "doing things the hard way"? You could argue that it makes you more competitive because you'll be better at babysitting a coding agent, but this feels less and less true with every new model. Unless you're seeing something else to take into account here?

> Frankly, I was hoping for a deeper, more interesting conclusion.

Pretty much this. There was a lot of build up in the first paragraphs, only to conclude something that, at least in my view, is remarkably uninteresting.

Great job, very interesting! However I wonder: since you are targetting iOS, why only restrict yourself to generalist frameworks such as llama.cpp and LiteRT-LM? MLX Swift looks like a better fit, and the model catalog on HF's MLX community is pretty huge and diverse.