@ga0_lab I forgot to say how you go from isosurfaces to gcode: After you have an isosurface, you can then generate a compute a distance field across the surface, where all the edges are zero-distance. This allows you to extract perimeters (contours).
@ga0_lab You can model this in voxel space easy enough, and then extract a series of isosurfaces from the field. Each isosurface descibes a truly 3d "slice" along which the head will move. However, this is limited by the voxel resolution.
@ga0_lab 2) You cannot print where it will block future parts of the print.
You can model (1) as field propagation. You can model (2) by after each propagation step, checking reachability.
@ga0_lab Sure! You can model printing as putting down plastic on other plastic - constrained by the shape of the geometry you wish to make:
1) You can only place plastic where there is something to place plastic on (print bed or other parts of the print.
@spandrell4 Because that is a 4k parameter model. Even a low end local model is 8b parameters, and a SOTA model is >1T parameter. So the LLM in the video is 6-9 orders of magnitude too small to be useful.
What if AI agents could build interactive widgets right in the chat? If AI agents were 100x faster and 100x better, all software would be disposable. Even today it works very well.
@trq212 Turns out that if you build your chat interface around this things work out very well too - then theres no need to tell the agent to generate it as a separate doc, it can just generate diagrams inline with it's response. Examples: https://t.co/Pph5SfGiAm
Imagine when software costs ~$0 so there's no point in an app store. Or when hardware is so cheap that you can talk to your microwave about philosophy. Or when the keyboard is deprecated as AI is your interface, and how that will change office space design...
@emollick I tend to model the future as if intelligence was free and trivially available, as if each human had an army of human experts available at all times.
@emollick So imagine the end state. Imagine when software costs ~$0 so there's no point in an app store. Or when hardware is so cheap that you can talk to your microwave about philosophy. Or when the keyboard is deprecated as AI is your interface (which will change office space design)...
When working with AI, Sometimes you want an autonomous agent, sometimes a pair coder. Sometimes you care about backwards compatibility, sometimes not. Lots of the problems people seem to have with coding agents is the general purpose behaviour specified in the system prompt.
@halvarflake Most coding agents (claude, codex etc.) are designed for interactive coding. They include a stop tool and limits on the number of iterations. A simple solution here is to make yourself an agent/harness that doesn't have a stop tool. Then the agent cannot stop.
@owickstrom Embrace throwaway code. If your aim is to write "good maintainable code" - YMMV (it can be done in some cases). If your aim is to solve a problem, coding agents are wonderful. If an AI can produce code on demand that solves your problem, there is no need for maintaining code.
People say that AI sucks at architecture, but when was the last time you asked an AI if it wanted to refactor it's code? In my experience, coding agents know the codebases they output are messy, but the user never gives them opportunity to tidy it up.