AI Briefing — 2026-07-10

🔬 Research

Vidu S1: A Real-Time Interactive Video Generation Model

95 upvotes · Jintao Zhang et al. · Paper

Vidu S1 generates infinite-length 540p video in real time at up to 42 FPS on consumer GPUs, with voice-driven control of digital characters at any point during generation. Built on TurboDiffusion and TurboServe, it avoids common artifacts like blur, drift, or distortion even for long sessions. Users can upload images of real people, anime characters, or pets and pick custom voice tones. The combination of real-time throughput with voice-driven interactivity on consumer hardware is a meaningful step toward practical, personalized live video synthesis; a playable demo is live at vidu.com/vidu-stream.

SciReasoner: Native Structural Reasoning Across Biology, Chemistry & Materials Science

76 upvotes · Chen Tang, Yizhou Wang et al. · Paper

SciReasoner is a multimodal foundation model that discretizes protein coordinates, molecular topologies, and crystal connectivities into a unified vocabulary, letting it reason over structural evidence rather than treating structure as a black-box input. Concretely, it raises Gene Ontology Cellular Component F-max from 0.42 to 0.55 on low-homology proteins, and single-step retrosynthesis accuracy from 0.63 to 0.72. It achieves state-of-the-art on 67 of 86 benchmarks, and expert evaluators preferred or matched its reasoning traces to a frontier LLM in 98% of cases. The interpretability angle — producing inspectable, step-by-step traces grounded in physical constraints — is what distinguishes it from prior domain-specific models.

LaMem-VLA: Dual Latent Memory for Robotic Manipulation

48 upvotes · Hongyu Qu et al. · Paper

Most Vision-Language-Action models assume Markovian observations and struggle on long-horizon tasks that require historical context. LaMem-VLA injects both short-term and long-term memory directly as latent tokens into the VLA's continuous embedding sequence, so past experience participates natively in reasoning and action generation rather than sitting in a separate retrieval sidecar. The framework uses four components — a curator, a seeker, a condenser, and a weaver — to retrieve and compress relevant history on the fly. Experiments on SimplerEnv and LIBERO show consistent gains over prior memory-augmented VLAs, making it a relevant read for anyone building robot policies on tasks with temporal dependencies.

🛠 Tools

addyosmani/agent-skills — Slash-command engineering workflows for AI coding agents

2 554 stars/day · Repo

This repo (76 k total stars, last pushed today) packages senior-engineer best practices as 8 slash commands — /spec, /plan, /build, /test, /review, /webperf, /code-simplify, /ship — each activating the right quality gates and checklists automatically. A /build auto mode lets the agent run the full plan-to-implementation loop after a single approval. Practically it's a structured prompt/workflow library for Cursor, Claude Code, and similar agents, useful for teams wanting more repeatable output from AI pair programmers.

iOfficeAI/OfficeCLI — Agent-first Office suite, no Microsoft Office required

1 929 stars/day · Repo

OfficeCLI is a free, open-source, single-binary CLI tool (C#) that lets AI agents read, edit, and automate Word, Excel, and PowerPoint files without requiring any Office installation. The "single binary, no dependencies" design makes it straightforward to drop into agent pipelines or containers. For agentic workflows that need to generate or process Office documents — reports, spreadsheets, slide decks — this removes a painful integration gap.

VoltAgent/awesome-design-md — Ready-to-use DESIGN.md files from real design systems

1 391 stars/day · Repo

DESIGN.md is a plain-text design-system format introduced by Google Stitch: drop the file in your project root and AI coding agents instantly know what your UI should look and feel like — no Figma exports, no JSON schemas. This repo (100 k total stars!) curates 73 ready-made DESIGN.md files reverse-engineered from popular brand design systems, covering tokens, patterns, and visual rules. If you're using any AI coding agent and want consistent UI generation without setting up a full design toolchain, this is the fastest on-ramp.

📡 Analysis

The New GPT-5.6 Family: Luna, Terra, Sol

9 July 2026 · Post · tags: openai, llms, llm-pricing, llm-release, gpt-5

OpenAI shipped three new flagship models: Luna ($1/$6 per 1M in/out tokens), Terra ($2.50/$15), and Sol ($5/$30), all with a 1M-token context window, 128k max output tokens, and a February 2026 knowledge cutoff. On the Agents' Last Exam agentic benchmark, Sol scores 53.6 — 13 points above Claude Fable 5 — though Fable 5 reportedly leads on SWE-Bench Pro (80% vs. 64.6%), a benchmark OpenAI simultaneously called "~30% broken." New API features worth watching include Programmatic Tool Calling (agents compose and run JavaScript to orchestrate tools) and built-in multi-agent spin-up, which could reshape how complex pipelines are structured.

Introducing Muse Spark 1.1

9 July 2026 · Post · tags: meta, llms, llm-release

Meta released Muse Spark 1.1, the first model in the Spark family to expose a public API, with claimed significant improvements in agentic tool calling and computer use. Simon Willison published llm-meta-ai, a plugin for his LLM CLI tool that wraps the new API, making it immediately accessible from the command line. The model is notable for its "Attractor States in Self-Conversation" behavior documented in the evaluation report — two instances of the model conversing converge on existential statements — a curious signal about the model's self-representation under extended dialogue.

Quoting OpenAI on ChatGPT Work's Confusing Architecture

10 July 2026 · Post · tags: openai, chatgpt, ai

Simon flags an OpenAI help-center quote attempting to clarify how ChatGPT Work handles cloud vs. desktop context — cloud conversations don't sync to desktop, and desktop threads stay local. The implicit commentary is that the clarification itself is confusing enough to be worth highlighting. For engineers evaluating ChatGPT Work for enterprise agent deployments, the data-residency and context-isolation behaviour is worth understanding carefully before assuming a unified experience.

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