DeepSeek V4.1 gets smarter but discards its best ideas; Opus struggles to rescue the mess
teortaxesTex · x · 2026-10-19
User teortaxesTex ran a multi-model relay: Opus 5.5 wrote a design spec, DeepSeek V4.1 Flash executed it, then Opus tried to fix the mess — and "almost had a stroke." DeepSeek built tons of modules it just dropped; V4.1-1017 is "massively smarter, often tests cooler things, but mainly discards them," e.g. an amazing skylab.py it abandoned. Screenshots show Opus's rescue: per-commit render comparisons, diagnosing another agent's uncommitted edits that dimmed lighting and let grass clip through the skateboard. "Lots of wasted potential."
More from coding & agent
- Dev suggests Claude Code compact context before notifying, not after — tonygwu · 2026-10-19
- Air Force recruit open-sources RaiGolmi, a VM workspace for running Claude Code agents unattended for months — BoshBoyBinton · 2026-10-19
- Consultant open-sources agent-first slide editor where Claude Code edits decks from comments — beh_zod · 2026-10-19
- Chat is the wrong shape for AI brainstorming: a dev experiments with tree-shaped idea maps — S_B_B_B_K · 2026-10-19
- Addy Osmani: writing code isn't the bottleneck anymore — build agent self-verification loops — bibryam · 2026-10-19
- Local models are now good enough for first-round code review, dev finds while paying frontier API prices — draginol · 2026-10-19