Indie dev builds own serving library to train SVG model, finds evolution loses to SGD
cephaloform · x · 2026-10-13
An indie developer training an SVG generation model from scratch shares progress: they wrote their own serving library so the model can call backprop on itself, inspect gradients, and make kernel engineering easy.
They abandoned an evolutionary approach — small runs showed it needs a massive population size, and quantizing n for speed doesn't make the math work out — and went back to SGD.
Related event: Indie Developer Builds Own Inference Library to Train SVG Generation Model(2 posts)→
More from Research
- 15-minute QLoRA finetune turns Qwen3.5-4B into a decision model playing Flappy Bird on 4GB VRAM — danielhanchen · 2026-10-13
- Claude tries to interpret an MNIST neural net from just the weights — Sauers_ · 2026-10-13
- AgentSWE benchmark: no coding agent comes close to delivering the agents it's asked to build — Jiahao Wang · 2026-10-13
- Schema agent hits ~99% RHAE on ARC-AGI-3 public set, full paper and code released — wzenus · 2026-10-13
- World model research: more data won't suffice—squeeze more from it; even intuitive physics is still out of reach — mangahomanga · 2026-10-13
- Harvard & NTT Research: cooperating AI agent populations hit an Allee-effect takeoff threshold — Hidenori8Tanaka · 2026-10-13