AI Briefing — March 10, 2026

🔬 Research

Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

48 upvotes · Yuanyuan Gao et al. · Paper

The first fully automated pipeline for creating large-scale 3D spatial intelligence datasets from raw videos without human annotation. Holi-Spatial-4M contains 12K optimized 3D Gaussian Splatting scenes with 1.2M spatial QA pairs, addressing the scalability bottleneck in current spatial understanding benchmarks. This could significantly reduce the cost and time needed to train models that understand 3D environments, making spatial AI more accessible for robotics and AR/VR applications.

Believe Your Model: Distribution-Guided Confidence Calibration

34 upvotes · Xizhong Yang et al. · Paper

DistriVoting improves test-time scaling by using Gaussian Mixture Models to separate confident from uncertain model responses during multi-candidate generation. The method decomposes confidence distributions into positive and negative components, then applies filtering to reduce overlap between them. Tested across 16 models and 5 benchmarks, it outperforms existing approaches, potentially making reasoning models more reliable in production environments.

Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model

30 upvotes · Dongwon Kim et al. · Paper

CompACT compresses observations into just 8 tokens instead of hundreds, making world model planning orders of magnitude faster for real-time control. This addresses a key bottleneck preventing world models from being used for decision-time planning in production systems. The dramatic reduction in computational cost while maintaining planning performance could enable real-world deployment of learned simulators for robotics and autonomous systems.

🛠 Tools

openclaw/openclaw — Personal AI assistant running on your devices

9,164 stars/day · Repo

A self-hosted AI assistant that integrates with 20+ messaging platforms (WhatsApp, Telegram, Slack, Discord) and supports voice interaction on mobile platforms. With 294K stars, OpenClaw positions itself as a local, privacy-focused alternative to cloud-based assistants. The TypeScript implementation and extensive platform support make it appealing for developers wanting full control over their AI assistant's data and behavior.

msitarzewski/agency-agents — Specialized AI agent personalities collection

4,415 stars/day · Repo

A curated collection of specialized AI agent prompts designed for specific roles (frontend developers, Reddit community managers, reality checkers). Each agent has distinct personality traits, workflows, and deliverable-focused approaches rather than generic templates. With 21K stars and production-ready workflows, it represents the growing trend toward role-specific AI assistants with measurable outcomes and success metrics.

666ghj/MiroFish — Swarm intelligence prediction engine

2,294 stars/day · Repo

A multi-agent system that creates digital parallel worlds from real-world seed data (news, policies, financial signals) to predict future outcomes. Thousands of autonomous agents with individual personalities and memory interact within these simulated environments. The system allows "god mode" variable injection to test different scenarios, positioning itself as a digital sandbox for decision-making and strategic planning.

📡 Analysis

Production query plans without production data

March 9, 2026 · Post · tags: databases, postgresql, sql, sqlite

PostgreSQL 18's new pg_restore_relation_stats() functions let developers copy production database statistics to development environments without the actual data. This enables realistic query plan testing with tiny statistics files (under 1MB) instead of copying hundreds of GB of production data. SQLite already offers similar functionality through writable sqlite_stat1 tables, making query optimization testing more accessible across database platforms.

Perhaps not Boring Technology after all

March 9, 2026 · Post · tags: ai, coding-agents, boring-technology

Modern coding agents with sufficient context length can effectively work with brand-new tools by consuming documentation and existing code patterns, contrary to expectations that LLMs would bias developers toward "boring" mainstream technologies. Agents successfully adapt to private or cutting-edge tools by iterating and testing their output, though a separate study found Claude Code still shows strong preferences for specific technical stacks like GitHub Actions and Stripe.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net