AI Briefing — May 24, 2026

🔬 Research

DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards

191 upvotes · Kaiyi Zhang, Wei Wu, Yankai Lin · Paper

DelTA tackles a fundamental problem in RLVR: how response-level rewards translate into token-level learning updates. The method reweights token contributions during training to focus on discriminative tokens rather than high-frequency formatting patterns. Testing on mathematical reasoning shows 3.26 and 2.62 average point improvements over baselines on Qwen3 models, with generalization to code generation tasks.

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

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

A 13 million record dataset from four Chinese cities enables LLMs to learn transit routing without traditional map infrastructure. Models trained on this data can generate valid routes and implicitly map GPS coordinates to stations without explicit geographic knowledge. This approach could eliminate dependencies on complex routing engines for public transit applications.

Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality?

160 upvotes · Caixin Kang, Tianyu Yan, Sitong Gong, Mingfang Zhang, Liangyang Ouyang · Paper

MM-OCEAN dataset reveals a "Prejudice Gap" in multimodal LLMs: 51% of correct personality assessments aren't grounded in behavioral evidence. The benchmark tests whether models truly perceive personality through observable behavior or rely on superficial cues. Only 0-33.5% of models achieve holistic grounding across 27 tested MLLMs, exposing the gap between accurate scoring and sound reasoning.

🛠 Tools

multica-ai/andrej-karpathy-skills — A single CLAUDE.md file to improve Claude Code behavior

3,507 stars/day · Repo

A distillation of Andrej Karpathy's observations on LLM coding pitfalls into four actionable principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The 150K+ starred repo provides a single markdown file that directly addresses common issues like wrong assumptions, overcomplication, and uncontrolled code changes. Each principle maps to specific LLM failure modes with concrete guidance.

colbymchenry/codegraph — Pre-indexed code knowledge graph for AI coding tools

2,456 stars/day · Repo

A semantic code intelligence layer that reduces AI coding costs by ~35% and tool calls by ~70% through pre-indexing. Works with Claude Code, Cursor, Codex, and other agents by providing structured code understanding without requiring external API calls. Self-contained with bundled Node.js runtime and supports automatic configuration for major coding assistants.

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

2,299 stars/day · Repo

Transforms codebases into interactive knowledge graphs that can be explored, searched, and queried. Integrates with Claude Code, Cursor, Copilot, and other AI tools to provide visual code understanding. The 23K+ starred tool emphasizes "graphs that teach > graphs that impress" with practical exploration features rather than just visualization.

📡 Analysis

Memory shortage reprices consumer electronics

May 22, 2026 · Post · tags: memory, ai, ai-ethics

AI's hunger for HBM memory is constraining consumer device production. Memory wafer allocation for HBM jumped from 2% to 20% by end 2026, with each HBM gigabyte consuming 3x the wafer capacity of consumer RAM. Only three major memory manufacturers remain, and they've learned to under-provision capacity. This is already impacting sub-$100 smartphones critical to developing markets.

FTC settles "active listening" marketing fraud case

May 22, 2026 · Post · tags: privacy, microphone-ads-conspiracy

Cox Media Group paid nearly $1M for falsely claiming their "Active Listening" service used smart device microphones for ad targeting. The service actually just resold email lists from data brokers at markup. The FTC clarified that mandatory app terms of service don't constitute valid consent for voice data collection, even if the technology worked as advertised.

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