Newer model shows better visual self-correction, no longer gaslights itself on renders
teortaxesTex · x · 2026-10-17
The author shares iterative progress rendering a "nacre fan" (mimicking aragonite microstructure and thin-film interference) with the latest model — still not perfect, but hopeful.
The more notable observation is behavioral: the new model's vision is markedly better, no longer gaslighting itself into "everything is fine" or hallucinating features — in the example it actively notices an issue with the arms. This improved self-correction tendency matters more than single-image quality.
Related event: DeepSeek V4.1F reportedly beats Opus, rebuilding projects in 5 minutes(4 posts)→
More from Models
- DeepSeek V4.1 Flash talks like a human while coding: "I love this", "Disaster", "Ugh" — teortaxesTex · 2026-10-17
- Theory-of-mind test: Claude spots AI-written text, ChatGPT and Gemini fail — Aizkmusic · 2026-10-17
- Rumor: Core automation to release a model literally called 'AGI' on Nov 10 — imjustnewatai · 2026-10-17
- AISLE found 500+ zero-days across 100+ codebases in one year, ~1% of 2026's open-source CVEs — brhydon · 2026-10-17
- Abliteration releases abliterated GLM-5.3 for offensive cyber, CEO pushes back on regulation — andrewchen · 2026-10-17
- Nemotron 3.5 Bolt 74B quantized hits ~169 tok/s on 16GB VRAM via llama.cpp — yehyakar · 2026-10-17