AI Briefing — 2026-07-28

🔬 Research

Kimi K3: Open Frontier Intelligence

177 upvotes · Kimi Team (Moonshot AI) · Paper

Moonshot AI releases Kimi K3, a 2.8T-parameter Mixture-of-Experts model with 104B activated parameters, native vision, and a 1M-token context window. It introduces Kimi Delta Attention and Stable LatentMoE (16 of 896 experts activated per token), achieving ~2.5× scaling efficiency over Kimi K2. Performance sits just behind Claude Fable 5 and GPT-5.6 Sol but outperforms all other open and proprietary models tested. Weights are fully released on HuggingFace (already ~99K downloads) under what Simon Willison flags as a "janky license" worth checking before production use.

JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents

89 upvotes · Yunlong Lin et al. · Paper

JarvisHub proposes treating an editable canvas as both the agent's external memory and the shared project state for long-horizon multimodal creative work (images, video, audio, UI, storyboards). Unlike chat- or node-based systems that discard intermediate context, JarvisHub uses a three-layer architecture (canvas state → protocol bridge → agent runtime) where artifacts, versions, and dependencies are explicit typed nodes. This makes agent reasoning inspectable and human-steerable at any point — a meaningful step toward auditable creative automation pipelines.

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

57 upvotes · Hao Liang et al. · Paper

K12-KGraph is a knowledge graph extracted from official Chinese K-12 textbooks (math, physics, chemistry, biology), with 9 node types and 14 relation types covering curriculum structure and visual grounding. From it, the authors derive K12-Bench (23,640 questions) and K12-Train (7,335 SFT samples). Frontier models struggle: Gemini-3-Flash hits only 57% exact match, revealing a real gap in "curriculum cognition" beyond standard exam QA. The graph, benchmark, and training data are fully released, making this a concrete resource for anyone building educational AI on structured knowledge.

🛠 Tools

alibaba/open-code-review — Hybrid LLM + deterministic code review at scale

979 stars/day · Go · Repo

Battle-tested at Alibaba's scale, this tool combines deterministic static-analysis pipelines with an LLM agent to produce precise line-level review comments. It ships with a built-in fine-tuned ruleset covering NPE, thread-safety, XSS, and SQL injection, and is compatible with OpenAI and Anthropic APIs. The hybrid architecture is the key differentiator: deterministic rules keep costs low and false-positive rates down, while the LLM handles nuanced logic issues.

pbakaus/impeccable — Design language skill for AI coding agents

847 stars/day · JavaScript · Repo · ~51,948 total stars

Impeccable installs via npx impeccable install and adds a structured design skill to AI coding tools (Claude, Cursor, etc.), with 23 slash commands (/impeccable polish, audit, bolder, etc.) and 60 deterministic detector rules that run without an LLM or API key. It fights the homogenization of AI-generated UI (Inter everywhere, purple-to-blue gradients, cards-in-cards) by writing a persistent PRODUCT.md/DESIGN.md that all subsequent commands reference. Practically useful for any team shipping AI-generated frontends and wanting brand consistency.

moeru-ai/airi — Self-hosted AI companion with real-time voice and game integration

572 stars/day · TypeScript · Repo · ~44,401 total stars

AIRI is an open-source, self-hosted AI virtual character runtime inspired by VTuber Neuro-sama, supporting real-time voice chat and autonomous gameplay in Minecraft and Factorio. It runs on Web, macOS, and Windows, and keeps all data user-owned — no cloud lock-in. The project illustrates growing interest in persistent, embodied AI agents that operate in live interactive environments rather than single-turn prompts.

📡 Analysis

moonshotai/Kimi-K3 weights release

2026-07-27 · Post · tags: ai, llms, llm-release, ai-in-china, janky-licenses

Simon Willison notes that Moonshot has followed through on its earlier promise to release Kimi K3's full weights (1.56T parameters on disk). He flags the license as "janky," which is a practical warning for engineers considering production or derivative use. Worth reading the license carefully before building on it.

An opinionated guide to which AI to use to do stuff

2026-07-27 · Post · tags: llms, ethan-mollick, code-interpreter, general-agents

Willison summarizes the evolution of Ethan Mollick's practical AI guide: one year ago it was all about chat interfaces; today it centers on agentic systems capable of "many hours of real human work in one go." Key confusion flagged: ChatGPT's "Work" and "Codex" modes vs. Claude's "Cowork" and "Code" modes are named inconsistently and behave differently on mobile vs. desktop. Useful orientation for teams trying to figure out which agent surface to standardize on.

An Inside Look at the Relay Market Powering Token Resellers and Fraud

2026-07-26 · Post · tags: llm-pricing, ai-ethics, ai-in-china

Matt Lenhard's investigation (summarized by Willison) documents a Chinese grey market where operators pool stolen or abused API keys via open-source proxies (one-api, new-api) to resell LLM tokens at a discount. Buyers use it for cheap tokens, geo-restriction bypasses, and model distillation data collection. For engineers exposing LLM-backed endpoints publicly, this is a concrete threat model — Willison argues LLM vendors urgently need hard per-key spending caps to protect developers.

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