AI Briefing — 2026-06-30

🔬 Research

LiveEdit: Towards Real-Time Diffusion-Based Streaming Video Editing

55 upvotes · Xinyu Wang, Chongbo Zhao, Fangneng Zhan, Yue Ma · Paper

LiveEdit is a streaming video editing framework that performs causal, frame-by-frame editing at 12.66 FPS — fast enough for interactive and AR use cases. The core novelty is a three-stage distillation pipeline that transfers editing capability from a bidirectional foundation model to a lighter unidirectional streaming model, paired with an AR-oriented mask cache that cuts redundant computation across frames. The key challenge it solves — preserving unedited regions stably over long horizons — has been a persistent blocker for real-world deployment. A new dedicated benchmark for streaming video editing is also introduced alongside the method.

Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent

54 upvotes · Lei Bai, Zongsheng Cao, Yang Chen et al. · Paper

Agents-A1 is a 35B Mixture-of-Experts model that claims to match or beat trillion-parameter models (Kimi-K2.6, DeepSeek-V4-pro) on long-horizon agentic benchmarks by "scaling the agent horizon" rather than raw parameter count. Training uses a three-stage recipe: broad supervised fine-tuning, domain-level teacher models, then multi-teacher domain-routed distillation unifying six heterogeneous domains into one student model. Average trajectory length is 45K tokens. Results include SEAL-0 56.4, BrowseComp 75.5, HLE 47.6 — competitive with models estimated to be ~30× larger, which has real deployment cost implications.

TUA-Bench: A Benchmark for General-Purpose Terminal-Use Agents

35 upvotes · Shoufa Chen, Luyuan Wang, Xuan Yang et al. · Paper

TUA-Bench fills a gap between GUI-centric computer-use benchmarks and narrow shell/coding benchmarks by evaluating agents on 120 real-world terminal tasks across five families: document editing, email management, live-web search, and PhD-level scientific/engineering workflows. Tasks run in a real terminal with deterministic setup scripts and execution-based scoring. The current best agent — Claude Code with Claude Opus 4.8 at max reasoning effort — scores only 65.8%, highlighting that general-purpose terminal competence remains a hard open problem.

🛠 Tools

msitarzewski/agency-agents — Curated library of specialized AI agent prompts

1,425 stars/day · ⭐ 119,789 total · Repo

A collection of production-ready AI agent "personalities" (frontend developers, community managers, reality checkers, etc.), each with a defined voice, workflow, and deliverable format — not generic prompt templates. Originally born from a Reddit thread, it now has a companion native app (agencyagents.app) for macOS/Linux/Windows that installs agents into Claude Code, Cursor, Codex, Gemini, and others with one click. With nearly 120K stars it is one of the fastest-growing prompt engineering repositories to date, reflecting strong demand for structured, role-specific agent definitions.

xbtlin/ai-berkshire — Multi-agent value investing research framework

1,386 stars/day · ⭐ 7,199 total · Repo

Built for Claude Code / Codex, this Python framework structures equity research around the methodologies of four investors (Buffett, Munger, Duan Yongping, Li Lu) running in parallel as adversarial agents — each forced to output a concrete buy/pass verdict with price ranges, not hedged prose. Built-in anti-bias mechanisms include information confidence ratings (A/B/C) and a mandatory Munger-style inversion check. The author claims real-money returns of +69% in 2024 and +66% YTD 2025, outperforming the S&P 500 by ~46–50 percentage points per year — extraordinary claims worth scrutinising, but the framework's structured decision discipline is the genuinely replicable part.

browser-use/video-use — AI coding agent for video editing

967 stars/day · ⭐ 12,263 total · Repo

Drop raw footage in a folder, describe what you want to Claude Code (or compatible agents), and get final.mp4 back. Handles filler-word cuts, color grading, 30ms audio fades, subtitle burning, and animation overlay generation via parallel sub-agents. It self-evaluates rendered output at every cut boundary before returning results, and persists session memory so work can resume across sessions. Requires an ElevenLabs API key for audio; otherwise fully open source.

📡 Analysis

Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

2026-06-29 · Post · tags: ai, local-llms, llm-release, qwen, gemma, lm-studio

Ornith-1.0 is the debut model from DeepReinforce — MIT-licensed, built on Gemma 4 and Qwen 3.5 (both Apache 2.0), with variants at 9B Dense, 31B Dense, 35B MoE, and 397B MoE. Simon ran the 35B Q4_K_M GGUF (20 GB) in LM Studio at 103 tokens/second and found it handles multi-step tool-calling agentic tasks well. The clean licensing stack (no restrictive Gemma Terms of Use this time) makes it practically deployable, and it reportedly achieves SOTA among open-source models of comparable size on coding benchmarks.

HTML table extractor

2026-06-29 · Post · tags: html, tools, wikipedia, cors

Simon released a browser-based tool that accepts pasted rich text (including HTML tables copied from any web page) and exports detected tables as HTML, Markdown, CSV, TSV, or JSON — no upload needed. A subsequent update, built with Codex, adds a Wikipedia search integration that uses Wikipedia's open CORS API to auto-import tables from any Wikipedia article directly. Useful for data engineers who regularly scrape structured data from the web without wanting to spin up a full pipeline.

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