AI Briefing — 2026-03-16

🔬 Research

Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training

76 upvotes · Fangfu Liu et al. · Paper

A novel approach that maintains spatial understanding through potentially unbounded video streams using test-time training. The system adapts "fast weights" during inference to capture and organize spatial evidence over long sequences, achieving state-of-the-art performance on video spatial benchmarks. This matters for robotics and AR/VR applications where continuous spatial awareness from streaming video is crucial.

IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse

45 upvotes · Yushi Bai et al. · Paper

Optimizes sparse attention by reusing index selections across transformer layers, reducing 75% of indexer computations with negligible quality loss. Testing on a 30B parameter model shows 1.82x prefill and 1.48x decode speedup over standard sparse attention. Critical optimization for long-context applications as LLMs increasingly handle extended conversations and document processing.

LMEB: Long-horizon Memory Embedding Benchmark

32 upvotes · Xinping Zhao et al. · Paper

Introduces a comprehensive benchmark for evaluating embedding models on memory retrieval tasks across 22 datasets and 193 zero-shot tasks. Results show larger models don't always perform better and traditional passage retrieval performance doesn't generalize to long-horizon memory tasks. Essential for developers building memory-augmented systems and choosing appropriate embedding models.

🛠 Tools

666ghj/MiroFish — Multi-agent prediction engine

2782 stars/day · Repo

A swarm intelligence engine that creates parallel digital worlds populated by autonomous agents to simulate and predict outcomes. Users upload seed data and describe prediction needs in natural language, then watch thousands of AI agents with independent personalities interact and evolve in digital sandboxes. Targets strategic planning, risk assessment, and scenario modeling with a "God's eye view" interface for injecting variables.

volcengine/OpenViking — Context database for AI agents

1870 stars/day · Repo

A specialized database that unifies memory, resources, and skills management for AI agents through a file system paradigm. Addresses the fragmentation problem where agent context is scattered across different storage systems, enabling hierarchical context delivery and self-evolution. Built by ByteDance's VolcEngine team to solve context management challenges in long-running agent applications.

obra/superpowers — Agentic development framework

1867 stars/day · Repo

A complete workflow system for coding agents that emphasizes proper software engineering practices over quick code generation. Guides agents through specification gathering, design review, implementation planning, and test-driven development using composable skills. Enables autonomous development sessions lasting hours while maintaining code quality and following established engineering principles.

📡 Analysis

What is agentic engineering?

14th March 2026 · Post · tags: coding-agents, agentic-engineering, ai, llms

Simon defines agentic engineering as developing software with coding agents assistance, emphasizing that human engineers still determine what to build rather than just how. Code execution capability distinguishes true coding agents from simple code generators, enabling iterative development toward working solutions. The practice requires providing proper tools, specifications, and verification processes to maximize agent effectiveness.

Quoting Jannis Leidel

14th March 2026 · Post · tags: ai-ethics, open-source, github

GitHub's "slopocalypse" of AI-generated spam has made open-source collaboration models untenable, with only 1 in 10 AI-generated PRs meeting project standards. Jazzband's open membership model designed for accidental mistakes can't survive in an environment where curl shut down bug bounties due to sub-5% confirmation rates. Highlights the urgent need for better AI-human collaboration mechanisms in open source.

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