AI Briefing — March 14, 2026
🔬 Research
Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training
66 upvotes · Fangfu Liu, Diankun Wu, et al. · Paper
A novel approach to spatial intelligence that adapts model parameters in real-time while processing video streams, enabling understanding of 3D spaces through continuous visual observations. Uses test-time training with "fast weights" and 3D spatiotemporal convolution to maintain spatial memory across potentially unbounded video sequences. This tackles the fundamental challenge of how AI systems can build and maintain spatial understanding over time, crucial for robotics and autonomous navigation applications.
Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections
41 upvotes · Łukasz Borchmann, Jordy Van Landeghem, et al. · Paper
Introduces MADQA, a benchmark of 2,250 questions across 800 PDFs to evaluate whether multimodal agents truly reason strategically or just brute-force search through documents. The research reveals that while top agents match human accuracy on document tasks, they succeed on different question types and rely heavily on trial-and-error rather than strategic planning. This exposes a critical gap in agentic reasoning that affects real-world document automation workflows.
IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse
33 upvotes · Yushi Bai, Qian Dong, et al. · Paper
Optimizes sparse attention in large language models by reusing attention indices across layers, reducing computational overhead from O(L²) to O(Lk) more efficiently. Tested on a 30B model, IndexCache removes 75% of indexer computations while achieving up to 1.82x prefill speedup and 1.48x decode speedup. This directly impacts the cost and speed of serving long-context models, making them more practical for production deployment.
🛠 Tools
msitarzewski/agency-agents — Complete AI agency with specialized agent personas
5,745 stars/day · Repo
A collection of battle-tested AI agent personalities for different development tasks, from frontend wizards to Reddit community managers. Each agent comes with specific expertise, communication style, and proven deliverables rather than generic prompt templates. With 41K stars and active development, this represents a mature approach to persona-driven AI assistance that focuses on measurable outcomes rather than general-purpose chatbots.
microsoft/BitNet — Official inference framework for 1-bit LLMs
2,227 stars/day · Repo
Microsoft's production-ready inference engine for 1.58-bit quantized models, achieving 1.37x to 6.17x speedups on ARM and x86 CPUs with 55-82% energy reduction. The framework can run a 100B parameter model on a single CPU at human reading speed (5-7 tokens/second), making large models feasible for local deployment. This addresses the critical challenge of model deployment costs and opens up new possibilities for on-device AI.
obra/superpowers — Agentic skills framework for software development
2,106 stars/day · Repo
A complete workflow system for coding agents that emphasizes proper software engineering practices including TDD, YAGNI, and systematic planning before implementation. Features "subagent-driven development" where agents work through engineering tasks autonomously for hours while following detailed implementation plans. With 82K stars, this represents a mature approach to agent-assisted development that prioritizes quality and maintainability over quick code generation.
📡 Analysis
1M context is now generally available for Opus 4.6 and Sonnet 4.6
March 13, 2026 · Post · tags: anthropic, claude, llm-pricing, long-context
Anthropic removes the long-context pricing premium for their 1M token context window, applying standard pricing across the full window. This contrasts with OpenAI and Google's tiered pricing models that charge more for longer contexts above 200-272K tokens. The move signals aggressive competition in the long-context space and makes complex document analysis more economically viable for developers.
Craig Mod's Custom Accounting Software Built in 5 Days
March 13, 2026 · Post · tags: ai-assisted-programming, vibe-coding
Craig Mod describes building his ideal accounting software in just 5 days using AI assistance, creating a multi-currency system that handles US and Japan tax requirements, ingests any CSV format, and learns from categorization patterns. The software feels "organic and pliable" and can be modified through direct conversation with Claude for batch operations. This exemplifies the emerging pattern of bespoke software creation becoming accessible to non-programmers through AI collaboration.
Shopify CEO Uses AI to Optimize 20-Year-Old Codebase
March 13, 2026 · Post · tags: performance, ai-assisted-programming, coding-agents
Shopify's CEO Tobias Lütke used autoresearch (Andrej Karpathy's automated experimentation system) to achieve 53% performance improvements in Liquid templating engine through 93 commits from 120 automated experiments. The process found micro-optimizations like replacing StringScanner with String#byteindex and caching small integer conversions. This demonstrates how robust test suites unlock agent-driven optimization and enable high-interruption executives to contribute meaningfully to technical codebases again.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net