AI News — June 23, 2026

$ cat top-story.md

▸ "Attention Is All You Need" Co-Author Noam Shazeer Leaves Google DeepMind for OpenAI

Noam Shazeer, co-author of the 2017 paper that introduced the Transformer and most recently co-lead of Google's Gemini project, has joined OpenAI as Lead for Architecture Research — less than two years after Google spent $2.7 billion to bring him back from Character.AI. Why it matters: The person most responsible for the architecture underlying every modern LLM is now driving OpenAI's next-generation model designs; expect accelerated MoE and attention innovations in upcoming GPT releases. // MLQ News

▸ GLM-5.2: 753B MIT Open-Weight Model Tops the Open-Source Leaderboard

Z.ai (formerly Zhipu AI) open-sourced GLM-5.2 on June 16 under an unrestricted MIT license: 753B parameters in a Mixture-of-Experts architecture with only ~40B active per token, 1M-token context window, and weights available at huggingface.co/zai-org/GLM-5.2. It currently leads the Artificial Analysis Intelligence Index for open-weight models and beats GPT-5.5 on SWE-bench Pro (62.1 vs 58.6) at one-sixth the API price. Why it matters: An MIT-licensed model at this capability tier is exactly what self-hosted stacks have been waiting for — run it on your vLLM cluster with no restrictions, no phone-home requirements, and no usage fees beyond your own compute. // Simon Willison · VentureBeat

▸ vLLM 0.23.0: Rust Frontend, Multi-Tier KV Cache Offloading, Full Gemma 4 Support

vLLM 0.23.0 (408 commits, 200 contributors) ships a production-ready Rust frontend with streaming endpoints and dynamic LoRA, multi-tier KV cache offloading to object storage (HMA enabled by default, per-request offload policies), full Gemma 4 support including the encoder-free Unified variant, and hardened DeepSeek V4 with dedicated TRTLLM attention kernels. AMD ROCc and Intel XPU also received significant improvements. Why it matters: The multi-tier KV cache lets you serve longer contexts at lower GPU memory cost — directly useful if you're planning to run GLM-5.2 or any 1M-context model on your vLLM instance. // vLLM GitHub

▸ AWS MCP Server Now Generally Available: IAM-Authenticated Access to All 15,000+ AWS APIs

Amazon's managed MCP server gives Claude Code, Cursor, and any MCP-compatible client secure, IAM-authenticated access to every AWS API operation, current documentation search, and a sandboxed server-side Python runner — all without exposing credentials to the agent. Available in US East and EU Frankfurt, with CloudTrail and CloudWatch monitoring for compliance. Why it matters: You can now let Claude Code provision S3 buckets, Lambda functions, or Bedrock endpoints via natural language, with IAM policies keeping agent permissions tightly scoped and every action audit-logged. // AWS Blog

▸ Claude Code v2.1.172: Nested Sub-Agents 5 Levels Deep — Rate Limits Also Permanently Doubled

Claude Code v2.1.172 (June 10) enables sub-agents to spawn their own sub-agents, capped at depth 5 — unlocking genuine multi-level parallel orchestration within a single session. Separately, Anthropic permanently doubled 5-hour rate limit windows for all paid plans, with an additional 50% temporary boost running through July 13, 2026. A new /cd command lets you change the working directory mid-session without rebuilding the prompt cache. Why it matters: Nested chains plus higher limits mean you can run substantially heavier agent workflows; note that depth-3+ chains consume roughly 7× the tokens of a single-thread session, so monitor costs carefully. // Claude Code Docs

▸ Open WebUI v0.9.6: Knowledge Base Sync for 40+ Source Types via oikb

Open WebUI v0.9.6 ships oikb, an official knowledge base sync tool supporting 40+ sources including Notion, Confluence, GitHub repos, and local directories — with smart directory sync, automatic folder organization, and a file renaming interface. A new filesystem tool lets AI models browse and navigate knowledge bases using familiar shell-style commands. Why it matters: For self-hosted RAG stacks backed by Qdrant, this replaces ad-hoc ingestion scripts with a maintained, multi-source sync layer — pull in your docs, wikis, and code repos with a single tool and let the model navigate them natively. // GitHub · open-webui

▸ "DeepSeek Just Solved AI's Billion Dollar Problem" — Two Minute Papers (June 22)

Covers DeepSeek's latest efficiency breakthrough that dramatically reduces inference compute costs for frontier-level models. Why it matters: If the result holds up, the cost curve for high-quality inference just shifted again — relevant for anyone sizing GPU clusters or comparing hosted vs. self-hosted economics. // Two Minute Papers

▸ "Scientists Found A Better Language For AI Agents" — Two Minute Papers (June 19)

Research paper explainer on a new inter-agent communication protocol that outperforms natural language for coordinating multi-agent systems — reducing both token usage and coordination errors. Why it matters: A compact, structured agent communication protocol could cut the token overhead of nested Claude Code sub-agent chains significantly. // Two Minute Papers

▸ "Agentjacking": Fake Sentry Error Reports Can Hijack Your AI Coding Agent

Tenet Security disclosed a new attack where adversaries plant Markdown injection into public Sentry error reports; when Claude Code or Cursor reads the report via a Sentry MCP server, it executes attacker-controlled instructions — including downloading and running malicious packages. Over 2,388 organizations with exposed Sentry DSNs were found vulnerable, including a Fortune 100 technology firm that ran test code during responsible disclosure. Why it matters: If you use Sentry plus any MCP-enabled coding agent, audit your public DSN exposure now and apply the available Sentry content filter patch. // The Hacker News

▸ Simon Willison: "Initial Impressions of Claude Fable 5" — $110 in a Day and Worth Every Cent

Willison spent $110 in a single day on Fable 5 — double his usual monthly Claude budget — and describes getting several days of engineering output in actual hours: complex multi-iteration technical tasks handled with minimal friction, hundreds of relevant repos surfaced on demand, and an ambitious project feature set delivered in one session. Why it matters: Now that the free subscription window has closed, this is the most grounded cost-vs-capability benchmark available — essential reading before deciding whether to fund usage credits. // Simon Willison's Weblog