In my mind there are 2 raisons d'être for tools and notations: (1) for non-programmers, (2) for programmers and hard-core, control-oriented Engineers who build tools for (1). TLDraw Computer is of type (1) in my current opinion.

I like to peel off the cover and see how things work inside. From this tinkerer perspective, I find this video enlightening: youtube.com/watch?v=Kn1De5uwrlY

... my gut says that this would be trivial to partially mock up in 0D or in Maikel van de Lisdonk's work. I even have an LLM 0D component lying around in my (public github) repo ...

Yes, that's what I was thinking as well. Although I dont have an llm-node in code-flow-canvas itself, I can call the openai api's and transform the inputs and output and stitch together multiple calls. But not (yet) with the same easy UX that tldraw has

Insightful video about how the prompts are constructed!

Something that I am wondering: which LLM is used, is it openai? Lu Wilson can you share something about this?

hey no it's Flash! you need a fast model like Flash or Haiku to be able to do so many calls without it being really slow. 4o mini isn't good or fast enough really

Thanks for sharing that! Is that Gemini 2.0 Flash?

no we aren't allowed to switch to 2.0 yet so just a 1.5 version

(an experimental version)

2.0 seems better and still fast so will switch to that once it out more widely

Do you use imagen3 for image generation with an expiremental version of 1.5?

yes currently using imagen 3 which isn't so good for us because it doesn't do image inputs

but we've been using native image output on Gemini 2 and that's been mind blowingly crazy

being able to prompt an image model like it's a language model is really weird cos you can give it instructions, and get it to modify existing images