Opus 5.5 Turns Papers Into 3Blue1Brown-Style Videos; RRSI Paper Tops alphaXiv
deedydas · x · 2026-09-24
deedydas demoed generating a full 3Blue1Brown-style educational video from any research paper with Claude Opus 5.5 — an 8-minute summary — claiming "the 90th-percentile educational YouTuber is fully automated," while noting 3Blue1Brown is still better and video beats dense academic prose.
The paper demoed is RRSI (Regularized Recursive Self-Improvement of Agent Harnesses) by a Google-led team, #1 trending on alphaXiv. Key points:
- An LLM agent's capability depends heavily on its harness (prompts, control flow, tooling, memory, context management); recent methods automate harness editing for agent-level recursive self-improvement (RSI)
- Such recursive evolution overfits training tasks: in-distribution gains shrink or vanish out-of-distribution
- RRSI adds regularization: a proposer with a temporally annealed edit budget favoring unexplored trajectories, and a selector with a critic (screens benchmark-specific proposals) and pruner (removes tiny, costly, or obsolete changes)
- Across 8 benchmarks spanning coding, agentic workspace, and engineering design, gains reach 14.1 points in-distribution and up to 4.7 points on five OOD benchmarks
Related event: Opus 5.5 Turns Research Papers into 3Blue1Brown-Style Videos(3 posts)→
More from Multimodal
- World Labs Previews Chisel in Atlas Beta: Sketch a World and Bring It to Life — theworldlabs · 2026-09-25
- A sub-$20 LoRA makes Qwen-Image 2.1 rotate transparent objects with a prompt — ben_burtenshaw · 2026-09-25
- Lingbot World v2 runs at 60 FPS, hinting world models could reshape game dev — bingxu_ · 2026-09-25
- Image-video feedback loop yields bizarre organic shape-shifting dance moves — pixlpa · 2026-09-25
- Creator makes a silly 'music video' with Luma AI, purely for the fun of it — mrjonfinger · 2026-09-25
- HuggingChat's ML Intern trains a LoRA from one prompt in ~75 minutes — Gradio · 2026-09-25