AI Briefing — March 11, 2026
🔬 Research
Geometry-Guided Reinforcement Learning for Multi-view Consistent 3D Scene Editing
97 upvotes · Jiyuan Wang, Chunyu Lin, Lei Sun, Zhi Cao, Yuyang Yin · Paper
RL3DEdit tackles the notorious problem of 3D editing consistency by flipping the script: instead of generating perfect 3D content from scratch, it uses reinforcement learning to verify and reward consistency across multiple viewpoints. The system leverages VGGT foundation model priors to provide reward signals that guide 2D editing tools toward maintaining geometric coherence, achieving 74% better results than current methods while remaining computationally efficient.
Lost in Stories: Consistency Bugs in Long Story Generation by LLMs
74 upvotes · Junjie Li, Xinrui Guo, Yuhao Wu, Roy Ka-Wei Lee, Hongzhi Li · Paper
ConStory-Bench exposes a critical weakness in LLMs generating long narratives: they frequently contradict their own established facts and character details. The benchmark includes 2,000 prompts across five error categories and reveals that consistency failures peak in the middle of stories, correlate with higher token entropy, and cluster around factual and temporal inconsistencies—crucial insights for anyone building narrative AI systems.
LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory
43 upvotes · Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, Ming-Hsuan Yang · Paper
LoGeR breaks the quadratic attention bottleneck plaguing 3D reconstruction from long video sequences by introducing a hybrid memory architecture combining Test-Time Training for global consistency and Sliding Window Attention for local precision. Trained on 128-frame sequences but generalizing to thousands of frames, it achieves over 74% error reduction on KITTI and handles sequences up to 19,000 frames without post-processing optimization.
🛠 Tools
openclaw/openclaw — Personal AI assistant that runs locally
9,080 stars/day · Repo
OpenClaw positions itself as the anti-cloud AI assistant, running entirely on your devices while connecting to every chat platform you already use (WhatsApp, Discord, Slack, and 15+ others). With 301K stars and active development, it represents a significant push toward personal, privacy-first AI assistance that feels "local, fast, and always-on" rather than dependent on external services.
msitarzewski/agency-agents — Specialized AI agent personalities for production workflows
6,223 stars/day · Repo
Agency-agents delivers battle-tested AI specialist personas ranging from frontend developers to Reddit community managers, each with distinct personalities, proven workflows, and measurable deliverables. Born from Reddit discussions and months of iteration, it's gained 27K stars by focusing on production-ready agents rather than generic prompt templates—think assembling your dream team of AI specialists who never sleep.
666ghj/MiroFish — Swarm intelligence engine for predictive simulation
4,504 stars/day · Repo
MiroFish creates parallel digital worlds populated by thousands of AI agents with individual personalities and memories, designed to predict real-world outcomes by simulating social interactions. Users upload seed data (news, policies, financial signals) and watch autonomous agents evolve scenarios, offering a "God's eye view" for decision-making through digital sandbox prediction—a fascinating approach to collective intelligence forecasting.
📡 Analysis
AI should help us produce better code
March 11, 2026 · Post · tags: coding-agents, ai-assisted-programming, generative-ai
Simon argues that shipping worse code with AI agents is a choice, not an inevitability. He advocates using asynchronous coding agents for time-consuming but conceptually simple refactoring tasks—API cleanup, nomenclature fixes, file splitting—that teams often postpone due to resource constraints. The dramatically reduced cost of these improvements enables a "zero tolerance attitude to minor code smells," potentially eliminating a major source of technical debt.
Perhaps not Boring Technology after all
March 9, 2026 · Post · tags: ai, generative-ai, llms, ai-assisted-programming, boring-technology
Contrary to expectations that LLMs would push developers toward well-documented mainstream technologies, Simon observes that modern coding agents handle novel or proprietary tools surprisingly well by consuming documentation and existing code patterns in-context. This challenges the assumption that AI tools would enforce a "Choose Boring Technology" approach, though he notes models still show clear preferences (shadcn/ui, GitHub Actions) when making their own recommendations.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net