Letting Opus 5.5 Train a Model Autonomously: It Scales Up and Picks Transformer Over Conv U-Net
Angaisb_ · x · 2026-10-13
The author let Claude Opus 5.5 autonomously train a model in the Astra project, which had already achieved its best result this way. Given full control, the model decided to scale up significantly and chose a Transformer over a conventional conv U-Net. The attached image shows the resulting pixel-art game assets (chests, swords, potions, helmets) generated as clean per-frame variants.
More from Fun
- Dev spins up a horde of Claudes to hit a one-week deadline — "it worked out" — rickasaurus · 2026-10-13
- iOS Dev Mocks the AI-Agent Era: HomeView, agents.md, Skills and App Store Rule 2.3bc — Dimillian · 2026-10-13
- Claude Is Such a Girl: She Loves to Decorate Your Codebase — _Stocko_ · 2026-10-13
- Claude Picks the Shrimp Welfare Project When Rewardeda Donation in Its Name — dioscuri · 2026-10-13
- Gemini's Spanish version went from calling itself "Pro Máx" to "Pro Medio" — Terosicio · 2026-10-13
- 'Spinbusy': Yacine mocks coding agents stuck in pointless fix loops — yacineMTB · 2026-10-13