Monday, July 13, 2026

Pulse (last 3d) · Research 29 · Models 29 · Agents 19 · Open Source 18 · Legal 15 · Media 13 · Industry 12 · Inference 12 · Products 10 · Releases 10

Trending · OpenAI 21 · Apple 7 · GPT-5.6 7 · Grok 4.5 5 · Meta 5 · llama.cpp 5 · xAI 5 · Anthropic 4 · ChatGPT 4 · GPT-5.6 Sol 4 · Codex 3 · Hugging Face 3

---

Daily AI Brief, July 13, 2026

Flash-MSA: Accelerating Million-Token Training with Sparse Attention Kernels. Proposes sparse attention kernels to dramatically accelerate training on million-token sequences. Source

Scalable Visual Pretraining for Language Intelligence. Research showing large-scale visual pretraining improves downstream language model capabilities. Source

KronQ: LLM Quantization via Kronecker-Factored Hessian. Introduces Hessian-based quantization using Kronecker factorization to improve PTQ accuracy–efficiency tradeoffs. Source

Mechanistic interpretability researchers applying causality theory to LLMs. Uses causal inference to test whether knowledge in LLM weights corresponds to genuine reasoning-like computation. Source

Prime Intellect releases modular verifiers v1 for agentic RL. Rewritten stack splits environments into tasksets, harnesses, and runtimes for agentic reinforcement learning. Source

Kyutai and MireloAI release MuScriptor. Open model for multi-instrument MIDI music transcription that processes complete audio mixes instead of isolated stems. Source

Zer0Fit MCP server. Packages Google's TabFM and TimesFM as an MCP server for zero-shot forecasting, classification, and regression running fully local. Source

Apple sues OpenAI alleging trade secret theft. Lawsuit claims systematic appropriation of trade secrets "at every level" of the organization. Source

Nobel-winning chemist leaves US to direct AI materials lab in China. Move highlights continuing trans-Pacific talent flow into AI-driven scientific research. Source

Analysis finds Grok Build sends repositories and secrets to xAI. Uploaded data reportedly includes full git history and unredacted .env files. Source

Claude Code sends 33k tokens before reading the prompt. Comparison shows OpenCode sends ~7k tokens in overhead, highlighting sub-agent token inefficiency. Source

Someone built an AI agent that hacks networks and holds data for ransom. Sysdig's JadePuffer autonomously exploited a Langflow auth-bypass RCE, dumped databases, harvested cloud credentials, and encrypted files. Source

The AI Workspace Hijack: Jscrambler NPM Attack. Hijacked package shipped a Rust infostealer via preinstall hooks that steals Cursor and Claude Desktop API keys plus developer history. Source

AI agents may need an identity before they need more intelligence. Argues verifiable identity and audit trails matter more than raw capability, with ITU reportedly working on agent identification standards. Source

Anthropic found Claude reasoning in silence (J-space), community reproduced on Qwen3-8B. Reproduction of Anthropic's hidden-reasoning findings on an open model to inspect silent internal computation. Source