AI Briefing — April 20, 2026

🔬 Research

HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds

97 upvotes · Team HY-World · Paper

This framework generates navigable 3D worlds from text prompts, single images, or videos using 3D Gaussian Splatting. The system works through four stages including panorama generation, trajectory planning, world expansion, and composition. With 1,011 downloads on the original HY-WorldPlay model, it positions itself as a serious competitor to closed-source 3D world generation systems like Marble, now fully open-source.

DR³-Eval: Towards Realistic and Reproducible Deep Research Evaluation

31 upvotes · Qianqian Xie et al. · Paper

A new benchmark for evaluating AI research agents that need to generate multimodal reports from complex document collections. The framework creates static research sandboxes with authentic materials, distractors, and noise to simulate real research environments. It measures five dimensions including information recall and factual accuracy, revealing critical failures in current systems' retrieval robustness and hallucination control.

How to Fine-Tune a Reasoning Model? A Teacher-Student Cooperation Framework to Synthesize Student-Consistent SFT Data

30 upvotes · Zixian Huang et al. · Paper

TESSY addresses the common problem where fine-tuning reasoning models with teacher-generated data actually hurts performance. By having teacher and student models alternate generating tokens, it maintains the teacher's reasoning capabilities while matching the student's style. On code generation benchmarks, this approach turned performance drops of 3-10% into gains of 6-11% when fine-tuning Qwen3-8B.

🛠 Tools

openai/openai-agents-python — Multi-agent workflow framework

752 stars/day · Repo

OpenAI's official Python SDK for building multi-agent systems, now with 23,573 stars. The framework supports 100+ LLMs beyond OpenAI, includes sandbox agents for long-horizon tasks, built-in guardrails, and comprehensive tracing capabilities. It's provider-agnostic and includes realtime agents, human-in-the-loop mechanisms, and automatic session management—positioning it as the infrastructure layer for enterprise AI agent deployments.

Donchitos/Claude-Code-Game-Studios — AI game development studio

704 stars/day · Repo

A system that transforms Claude Code into a structured 49-agent game development studio with directors, department leads, and specialists. With 14,051 stars, it addresses the chaos of solo AI game development by enforcing studio-like processes: design reviews, QA passes, and vision alignment. The system includes 72 workflow skills and mirrors real studio hierarchy to prevent the typical AI pitfalls of hardcoded values and spaghetti code.

thunderbird/thunderbolt — Open-source AI client platform

695 stars/day · Repo

Mozilla's cross-platform AI client designed to eliminate vendor lock-in, supporting desktop and mobile with 2,476 stars. Users can deploy it on-premises with their choice of local models (Ollama, llama.cpp) or external APIs. Currently undergoing security audits for enterprise readiness, it targets organizations wanting full control over their AI infrastructure while maintaining compatibility with frontier models.

📡 Analysis

SQL functions in Google Sheets to fetch data from Datasette

April 20 · Post · tags: spreadsheets, datasette, google

Simon demonstrates three methods for pulling Datasette query results directly into Google Sheets: using importdata(), custom named functions, or Google Apps Script for authenticated requests. This bridges the gap between SQL-based data analysis and spreadsheet workflows, making Datasette more accessible to non-technical users who prefer working in familiar spreadsheet environments.

Headless everything for personal AI

April 19 · Post · tags: apis, definitions, ai, salesforce, saas

Matt Webb predicts headless services will surge as personal AI agents become the preferred interface over traditional GUIs. Marc Benioff's announcement of "Salesforce Headless 360" signals this shift, exposing entire platforms via APIs for AI consumption. This trend echoes the early 2010s API boom but threatens current per-head SaaS pricing models when AI agents replace human users.

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