Vulpes

[Vulpes @VigorFox]

Added a one-page flowchart for GALT.

Idea: turn LLM post-training from single-loss tuning into constraint-graph learning with task, safety, and memory as responsibility channels.

Paper/code/results:
https://t.co/egWjAAZJth
#LLM #AIAlignment

[Vulpes @VigorFox]

@MaxForAI 一眼假，因为我真有一台 mac studio，后面根本就不是这种接口布局

[Vulpes @VigorFox]

Safety may not just constrain intelligence; it may help organize it.

[Vulpes @VigorFox]

GALT is surfacing a surprising pattern: memory and safety are distinct, but memory seems to grow on top of safety boundaries. Training the memory path alone helped less than getting the safety path to route cleanly first. Safety may be scaffolding, not just suppression.

[Vulpes @VigorFox]

@cem_yuksel @AnkaHeChen @karpathy @ylecun This was a 2-week sprint from idea to release. I wanted to get the core direction out quickly while the physics→ML connection was fresh. More experiments and ablations are coming in v2.

[Vulpes @VigorFox]

@cem_yuksel @AnkaHeChen
GALT is out: Graph-Parallel Augmented-Lagrangian Training with Responsibility-Separated Channels.
I took the core local solve technique from your AVBD paper and applied it to neural network training.
Repo + full paper: https://t.co/egWjAAZJth

[Vulpes @VigorFox]

Thanks to @cem_yuksel, Chris Giles, Elie Diaz, and @AnkaHeChen for the AVBD foundation and Physical AI direction.
Tagging some deep learning leaders: @karpathy @ylecun
Brutal feedback welcome on the physics-to-ML transfer and scaling ideas.
Repo: https://t.co/egWjAAZJth

[Vulpes @VigorFox]

@47htx @Locz123a @Trump_Fact_News That’s the exact event related to the video because I lived nearby at that time.

[Vulpes @VigorFox]

Big refactor landed in PhysX_AVBD.
All joints (Spherical / Revolute / Prismatic / D6 ) now unified into one clean D6 solver.
→ ~400 lines of duplicate code deleted → Much better stability & convergence
Ordinary joints = AVBD’s real battlefield.
Repo: https://t.co/NVacjxXwJc

[Vulpes @VigorFox]

Excited to share my latest project: PhysX_AVBD! Forked NVIDIA's PhysX SDK and integrated an experimental position-level Augmented Vertex Block Descent (AVBD) solver—99.9% coded by AI via @GitHubCopilot powered by @AnthropicAI's Claude models.
Repo: https://t.co/NVacjxXwJc

[Vulpes @VigorFox]

Building on the PhysX_AVBD project: This AI-powered (via Claude in Copilot) position-level solver tackles high-mass-ratio issues in rigid/soft-body sims. Demo GIF shows the potential! @oegerikus @dickson_tsai @AlexAbreuDev @minchen_li_ @SimAI4Science @algoryx