AI Briefing — 2026-05-22

🔬 Research

Video2GUI: Synthesizing Large-Scale Interaction Trajectories for Generalized GUI Agent Pretraining

142 upvotes · Weimin Xiong, Shuhao Gu, Bowen Ye, Zihao Yue, Lei Li · Paper

This work introduces Video2GUI, an automated pipeline that extracts GUI interaction data from unlabeled internet videos to train better AI agents for computer interfaces. The researchers processed 500 million video metadata entries to create WildGUI, a dataset with 12 million interaction trajectories across 1,500+ applications and websites. Pre-training popular vision-language models like Qwen2.5-VL (4.5M downloads) and Mimo-VL on this data delivered 5-20% improvements on GUI benchmarks, demonstrating that web-scale video mining can solve the data scarcity problem for GUI agents.

Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation

124 upvotes · Zhifei Xie, Kaiyu Pang, Haobin Zhang, Deheng Ye, Xiaobin Hu · Paper

Mega-ASR tackles the "acoustic robustness bottleneck" where speech recognition models fail catastrophically in noisy, real-world conditions by hallucinating or omitting words. The system combines a 2-million-sample dataset covering 54 compound acoustic scenarios with progressive training techniques. Results show dramatic improvements on challenging benchmarks: 45.69% vs 54.01% word error rate on VOiCES and over 30% relative improvement on compositional acoustic scenarios, establishing a new paradigm for robust speech recognition without relying on clean audio assumptions.

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

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

TransitLM proposes that public transit route planning can be learned entirely from data, bypassing traditional map infrastructure and routing engines entirely. The researchers compiled 13+ million transit records from four Chinese cities covering 120,845 stations and trained LLMs that can generate valid routes and implicitly ground GPS coordinates to appropriate stations. This end-to-end approach could revolutionize transit apps by eliminating dependency on complex mapping systems, with both dataset and benchmark code publicly available on HuggingFace and GitHub.

🛠 Tools

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

4,294 stars/day · Repo

CodeGraph builds semantic code intelligence graphs that make AI coding assistants 35% cheaper and 70% more efficient by reducing redundant tool calls. The system works with Claude Code, Cursor, Codex, and other agents, running entirely locally with no Node.js dependencies required. With 15,270 total stars and cross-platform support, it represents a practical solution to the token cost and context management problems that plague current AI coding workflows.

multica-ai/andrej-karpathy-skills — Karpathy-inspired guidelines for Claude Code

2,614 stars/day · Repo

This single CLAUDE.md file implements four principles to address Andrej Karpathy's observations about LLM coding pitfalls: think before coding, simplicity first, surgical changes, and goal-driven execution. The guidelines directly tackle issues like silent wrong assumptions, code overcomplication, and orthogonal edits that don't serve the task. With 144,741 total stars, it's become a de facto standard for improving AI coding behavior across the developer community.

Imbad0202/academic-research-skills — Complete academic research pipeline for Claude Code

2,579 stars/day · Repo

This comprehensive plugin suite covers the full academic research pipeline from literature review to publication, emphasizing human-in-the-loop workflows over full automation. The system handles grunt work like reference hunting, citation formatting, and consistency checking while leaving creative and analytical tasks to researchers. With 18,714 total stars and recent updates, it addresses quality over speed, including style calibration that learns from past work to maintain authentic voice.

📡 Analysis

FTC Settles "Active Listening" AI Marketing Fraud for Nearly $1 Million

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

The FTC confirmed what many suspected: Cox Media Group's "Active Listening" service was pure marketing fiction that didn't actually use microphone data, but instead resold marked-up email lists from data brokers. This settlement validates that the microphone ad-targeting conspiracy theory was largely overblown marketing speak, while establishing that even hiding opt-ins in terms of service wouldn't constitute adequate consent for such invasive practices. The case provides crucial regulatory clarity on voice data collection claims in an era of AI-powered advertising.

Datasette Agent Launches with Extensible AI Assistant Framework

May 21, 2026 · Post · tags: projects, sqlite, ai, datasette, generative-ai, llms, llm, uv, datasette-agent

Simon Willison's new Datasette Agent combines his LLM library with Datasette to create a conversational interface for querying databases, complete with chart generation and code execution capabilities. The system runs on Gemini Flash-Lite for cost efficiency and supports local models, with a plugin architecture that's already spawned three official extensions. This represents a significant step toward making SQL databases accessible through natural language while maintaining the flexibility and extensibility that characterizes the broader Datasette ecosystem.

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