AI Briefing — May 25, 2026

🔬 Research

TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation

169 upvotes · Hanyu Guo, Jiedong Yang, Chao Chen, Longfei Xu, Kaikui Liu · Paper

This dataset of 13 million transit records from Chinese cities shows LLMs can plan public transport routes without traditional mapping infrastructure. The trained model can convert GPS coordinates to appropriate stations and generate structurally valid routes purely from data, potentially eliminating the need for complex routing engines in transit apps.

SkillOpt: Executive Strategy for Self-Evolving Agent Skills

97 upvotes · Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou · Paper

The first systematic text-space optimizer for agent skills that treats skills like neural network weights, using a separate optimizer model to make bounded edits based on validation scores. Across 52 evaluation cells, SkillOpt achieved average accuracy gains of +23.5 points on GPT-5.5, with skills transferring effectively between different model scales and execution environments.

Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps

87 upvotes · Yanke Zhou, Yiduo Li, Hanlin Tang, Maohua Li, Kan Liu · Paper

RTPurbo exploits intrinsic sparsity in full-attention LLMs to achieve massive efficiency gains with minimal training. By identifying that only a subset of attention heads need full context processing, it delivers up to 9.36× prefill speedup at 1M context length while preserving near-lossless accuracy with just hundreds of training steps.

🛠 Tools

Lum1104/Understand-Anything — Interactive knowledge graphs for code exploration

3999 stars/day · Repo

Transforms any codebase into an interactive knowledge graph that you can explore, search, and query. With 29K stars already, it integrates with major AI coding tools including Claude Code, Codex, Cursor, and Copilot, making code understanding more visual and accessible than traditional text-based exploration.

colbymchenry/codegraph — Pre-indexed semantic code intelligence

3003 stars/day · Repo

A 100% local solution that pre-indexes codebases to reduce AI coding tool costs by ~35% and tool calls by ~70%. The self-contained binary works across platforms without Node.js, providing semantic code intelligence to Claude Code, Cursor, Codex, and other agents through efficient graph-based indexing.

multica-ai/andrej-karpathy-skills — Claude Code enhancement based on Karpathy insights

2551 stars/day · Repo

A single CLAUDE.md file containing four principles derived from Andrej Karpathy's observations about LLM coding pitfalls. With 153K stars, it addresses common issues like wrong assumptions, overcomplication, and unintended code changes by forcing explicit reasoning and surgical modifications in Claude Code workflows.

📡 Analysis

datasette 1.0a30

May 24 · Post · tags: projects, datasette, annotated-release-notes

The major feature in this Datasette alpha is a customizable "Jump to..." menu accessible via keyboard shortcut, with a new plugin hook allowing extensions to add searchable items. This continues Datasette's evolution toward a more interactive data exploration platform with enhanced discoverability features.

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