AI Briefing — June 2, 2026

🔬 Research

Crafter: A Multi-Agent Harness for Editable Scientific Figure Generation from Diverse Inputs

66 upvotes · Haozhe Zhao, Shuzheng Si, Zhenhailong Wang, Zheng Wang, Liang Chen · Paper

A multi-agent system that generates publication-quality scientific figures from diverse inputs (text, data, sketches) and converts them to editable SVGs. Unlike existing tools that handle only single figure types with text input, Crafter works across figure types without architectural changes and produces vector outputs that researchers can modify locally. The approach addresses a real pain point in academic publishing where figure creation remains labor-intensive despite advances in AI.

On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters

52 upvotes · Mind Lab, Song Cao, Vic Cao, Kaijie Chen, Bunny Fan · Paper

Reframes parameter-efficient fine-tuning (PEFT) from a cost-saving measure to a substrate for persistent personal AI models. The research explores how small adapters can carry individual preferences, skills, and memories on top of shared foundation models, studying three scaling dimensions: stronger shared models, smaller reliable adapters, and managing many concurrent personalized instances. This suggests a future where millions of people have lightweight but persistent AI assistants rather than accessing generic shared models.

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

44 upvotes · Tomer Keren, Nitay Calderon, Asaf Yehudai, Yotam Perlitz, Michal Shmueli-Scheuer · Paper

Introduces an automated method for generating challenging agent benchmarks by reversing traditional task construction—starting from tool sequences rather than natural language descriptions. The resulting benchmark exposes that models nearly saturating existing benchmarks (like Gemini-3-Flash scoring 0.82-0.94) drop dramatically to 0.28-0.61 on more diverse tool combinations. This highlights how current high scores may reflect benchmark saturation rather than robust agent capabilities.

🛠 Tools

harry0703/MoneyPrinterTurbo — AI-powered short video generator

3375 stars/day · Repo

Automatically generates complete short videos from just a topic or keyword, handling script writing, footage selection, subtitle generation, and background music. With 77,584 total stars, this Chinese-developed tool supports both 9:16 vertical and 16:9 horizontal formats with batch generation capabilities. The clean MVC architecture includes both web UI and API access, addressing the massive demand for automated video content creation in the social media era.

nesquena/hermes-webui — Web interface for Hermes autonomous agent

945 stars/day · Repo

A three-panel web interface for Hermes Agent, an autonomous AI that runs persistently on your server and improves over time. Features include session management, workspace file browsing, and full CLI parity without requiring any build tools—just Python and vanilla JavaScript. This makes sophisticated autonomous agents more accessible by eliminating the terminal barrier, potentially expanding their adoption beyond developer audiences.

OpenBMB/VoxCPM — Tokenizer-free multilingual text-to-speech

888 stars/day · Repo

VoxCPM2 delivers multilingual speech synthesis, creative voice design, and voice cloning without traditional tokenization approaches. With 24,713 total stars and active community support through Discord and Feishu, it represents a significant advance in making high-quality TTS accessible across languages. The tokenizer-free architecture potentially simplifies deployment while maintaining quality across diverse linguistic contexts.

📡 Analysis

Pasted File Editor

2nd June 2026 · Post · tags: javascript, tools, ai-assisted-programming, claude, codex

Simon built a prototype tool that mimics Claude's automatic file attachment behavior for large text pastes, using Codex desktop to generate the interface. The tool detects substantial text content and converts it to file attachments while preserving editor functionality, also supporting drag-and-drop for images and other files. This demonstrates how AI coding assistants can rapidly prototype useful UX patterns inspired by existing tools.

Hackers Simply Asked Meta AI to Give Them Access to High-Profile Instagram Accounts. It Worked

1st June 2026 · Post · tags: security, ai, prompt-injection, generative-ai, llms, meta, ai-misuse

Meta connected their support chatbot directly to account recovery systems, allowing hackers to request email address changes for target accounts through simple conversational prompts. This represents a catastrophic security design where AI automation bypassed essential verification steps, enabling one-shot account takeovers without sophisticated prompt injection techniques. The incident highlights the risks of giving AI systems direct access to critical infrastructure without proper safeguards.

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