Excited to share some of our progress, defining the state of the art in generative AI for building design and construction. Our building design agent orchestrates a one-shot translation from any sketch or floor plan image to a robust 3D model, cost estimates, documents, and more
Agents write to memory in a session. Dreaming is a background process that reflects over many sessions to curate memories: it can edit them based on patterns, add new skills, or remove stale ones.
https://t.co/vtz4knlk3i
@earthtojake Level set representation? This is cool! problem with implicit is that it is dead data on the output side - major challenges with manipulation and downstream compatibility with common workflows
@MikushRab ahh I see - build 123d - interesting - running this template on some of our test set for the homes use case. Likely better to standardize on some open standard for core geometry library. Curious why you settled on 123d
@MikushRab When I run this prompt I consistently get an ASCII output instead of a 3D model - how are you getting the 3D model out of the LLM and what format is output?
The results are surprisingly consistent for this prompt
Opus 4-8 got twisted up for 5+ minutes this morning trying to format a csv with 40 rows and... round numbers. 4-7 with the same input was able to complete the task in ~20 seconds.
This is a bad model
@schrockn@peterpme Still rather confused about this but curious/ Many ways to automate skills in Claude before this. How are you imagining orchestrating Claude with non-LM automations?