AI Briefing — May 23, 2026
🔬 Research
TransitLM: A Large-Scale Dataset and Benchmark for Map-Free Transit Route Generation
166 upvotes · Hanyu Guo et al. · Paper
TransitLM introduces a massive dataset of 13 million transit route planning records from Chinese cities, enabling LLMs to generate valid transit routes without traditional map infrastructure. The trained models can implicitly ground GPS coordinates to appropriate stations and produce structurally sound routes with high accuracy. This represents a significant shift toward end-to-end, data-driven route planning that could eliminate the need for complex routing engines and structured map dependencies in public transit systems.
Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?
156 upvotes · Caixin Kang et al. · Paper
This research exposes a critical flaw in multimodal LLMs: 51% of correct personality assessments aren't grounded in actual behavioral evidence, with models often relying on superficial pattern matching rather than genuine understanding. The new MM-OCEAN benchmark reveals a "Prejudice Gap" where models get the right Big Five scores for the wrong reasons. This finding has major implications for AI systems deployed in human-facing roles where personality assessment matters, highlighting the difference between appearing accurate and truly understanding human behavior.
DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards
126 upvotes · Kaiyi Zhang et al. · Paper
DelTA tackles a fundamental problem in RLVR: how to properly assign credit to individual tokens when training on response-level rewards. The method addresses the issue where standard approaches can be dominated by high-frequency formatting tokens, diluting the learning signal from truly discriminative content. On mathematical benchmarks, DelTA achieves 3.26 and 2.62 average point improvements over baselines on Qwen3 models, demonstrating that better token-level credit assignment significantly improves reasoning capabilities.
🛠 Tools
colbymchenry/codegraph — Pre-indexed code knowledge graph for AI coding assistants
3,684/day · Repo
CodeGraph creates semantic code intelligence that makes AI coding assistants 35% cheaper and 70% more efficient by reducing tool calls. It works across Claude Code, Cursor, Codex, OpenCode, and Hermes Agent, providing a pre-indexed knowledge graph that eliminates the need for repetitive code analysis. The tool runs entirely locally with no Node.js dependencies, offering a bundled runtime that auto-configures with popular coding assistants to dramatically reduce token usage.
anthropics/claude-plugins-official — Official Claude Code plugin marketplace
2,549/day · Repo
Anthropic's official plugin directory for Claude Code provides both internal plugins developed by the team and curated third-party extensions from partners. The marketplace includes MCP server configurations, slash commands, and agent definitions that extend Claude Code's capabilities. With over 25k stars, this represents Anthropic's commitment to building an ecosystem around Claude Code, though users are warned to carefully vet plugins since Anthropic cannot guarantee their safety or functionality.
Lum1104/Understand-Anything — Interactive code knowledge graphs
1,393/day · Repo
Understand-Anything transforms any codebase into an interactive knowledge graph that can be explored, searched, and queried. The tool emphasizes educational value over visual appeal with its "graphs that teach > graphs that impress" philosophy. It integrates with multiple AI coding assistants including Claude Code, Cursor, and Copilot, making complex codebases more navigable by creating visual relationships between code components and enabling natural language queries about code structure and functionality.
📡 Analysis
The memory shortage is causing a repricing of consumer electronics
May 22, 2026 · Post · tags: memory, ai, ai-ethics
The AI boom is creating a memory crisis that will make consumer electronics significantly more expensive. Memory manufacturers are shifting 20% of wafer capacity to HBM for AI data centers (up from 2%), and HBM consumes 3x more wafer capacity per gigabyte than consumer RAM. With only three major memory manufacturers learning to under-provision rather than over-provision, this constraint will particularly hurt sub-$100 smartphones in markets like Africa and South Asia for several years.
Datasette Agent
May 21, 2026 · Post · tags: projects, sqlite, ai, datasette, generative-ai, llms
Simon Willison launched Datasette Agent, bringing conversational AI to database exploration after three years of developing the LLM Python library. The tool generates SQL queries from natural language and can create charts through plugins, with a live demo running on Gemini 3.1 Flash-Lite. The extensible plugin system already includes chart generation, image creation, and code execution capabilities, with Claude Code and OpenAI Codex proving excellent at writing new plugins.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net