AI Briefing — March 15, 2026

🔬 Research

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

69 upvotes · Liu et al. · Paper

A new approach that enables AI models to continuously build spatial understanding from video streams without needing massive context windows. The method uses test-time training with "fast weights" that adapt in real-time, plus 3D spatiotemporal convolutions to maintain geometric consistency across frames. This tackles the challenge of long-horizon spatial intelligence where traditional models struggle with unbounded video sequences.

Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections

49 upvotes · Borchmann et al. · Paper

Researchers created MADQA, a benchmark with 2,250 human questions across 800 PDFs, to test whether AI agents actually reason strategically or just brute-force search. The findings are sobering: while top agents match human accuracy, they succeed on different questions and rely on trial-and-error rather than strategic planning. There's still a 20% gap to optimal performance, with agents often stuck in unproductive loops.

IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse

36 upvotes · Bai et al. · Paper

A technique that speeds up sparse attention in long-context models by reusing attention indices across layers instead of recomputing them. Since consecutive layers select very similar top-k tokens, IndexCache eliminates 75% of indexer computations while maintaining quality. Tests on a 30B parameter model show 1.82× prefill speedup and 1.48× decode speedup—crucial for making long-context AI workflows more affordable.

🛠 Tools

msitarzewski/agency-agents — AI specialists with personality and process

4,280/day · Repo

A curated collection of 44K+ starred AI agent templates, each with distinct personalities and specialized skills ranging from frontend development to Reddit community management. Born from Reddit discussions, these aren't generic prompts but production-ready workflows with specific deliverables and success metrics. The agents work with Claude Code and other tools, offering a plug-and-play approach to assembling specialized AI teams.

lightpanda-io/browser — Headless browser optimized for AI automation

2,069/day · Repo

A Zig-based headless browser built specifically for AI agents and automation, claiming 11× faster execution and 9× lower memory usage than Chrome. It supports JavaScript execution and Web APIs while maintaining compatibility with Playwright, Puppeteer, and Chrome DevTools Protocol. With instant startup times and ultra-low resource consumption, it addresses the performance bottlenecks that make web automation expensive at scale.

volcengine/OpenViking — Context database for AI agents

1,610/day · Repo

An open-source context database that unifies memory, resources, and skills management for AI agents through a file system paradigm. Designed specifically for agentic systems, it enables hierarchical context delivery and self-evolving agent capabilities. This addresses the growing need for persistent, structured context management as AI agents become more complex and long-running.

📡 Analysis

Quoting Jannis Leidel

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

Jazzband, a collaborative Python project maintenance organization, is shutting down due to GitHub's "slopocalypse"—the flood of AI-generated spam PRs and issues. With only 1 in 10 AI-generated PRs meeting project standards and confirmation rates dropping below 5%, their open-membership model became untenable. This signals how AI-generated content is fundamentally changing the economics and safety of open-source collaboration.

1M context is now generally available for Opus 4.6 and Sonnet 4.6

March 13 · Post · tags: ai, generative-ai, llms, anthropic, claude, llm-pricing, long-context

Anthropic has made 1M context windows generally available for Claude Opus 4.6 and Sonnet 4.6 with no long-context pricing premium—standard rates apply across the full window. This contrasts sharply with OpenAI and Google, who charge extra for tokens above 200K-272K thresholds. The move could significantly reduce costs for long-context applications and pressure competitors to adjust their pricing models.

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