Exciting AI Updates Weekly - May 22, 2026

Exciting AI Updates Weekly - May 22, 2026 Presented by Lev Selector Slides on GDrive: https://drive.google.com/drive/folders/1dmj00iilSP-JYmriMUUaHCKz5X54e3UX?ths=true Slides on GitHub - https://github.com/lselector/seminar (click on pptx file, then on "raw" or download button on the right) ------------------------------------------------------- Question to you: Do you use a personal Digital "Employee"? If not, then why not? It is so easy to start! ------------------------------------------------------- Need AI Consulting ? - Enterprise AI Solutions - https://EAIS.ai - Linkedin - https://www.linkedin.com/in/levselector - GitHub - https://github.com/lselector ------------------------------------------------------- Our books: "Artificial Intelligence In Business: Strategies That Work" https://www.amazon.com/dp/B0FKZ3364G/ "AI Product Manager: Building Tomorrow" https://www.amazon.com/dp/B0FMS9NFK9/ "How to Learn with AI?: A Practical Guide for Students, Professionals, and Self-Learners" https://www.amazon.com/dp/B0FWLNVQMJ/ "Professional + AI: Future-Proofing Your Career" https://www.amazon.com/dp/B0GBXLCTZQ "FP&A + AI: How to Work Faster, Smarter, and Deliver More Value" https://www.amazon.com/dp/B0FY6Z5BBW --------- Contents of today's video: 00:00 Harness vs Model, MDASH vs Mythos 00:45 Crowd-sourced "LM Arena" Leaderboard 01:56 Microsoft MDASH Harness 02:18 Elon Musk lost his lawsuit against OpenAI 02:58 Grok 4.3 was released in April 2026 03:32 Anthropic got $30 Bln at $900 Bln Valuation 04:37 Codex vs Claude Code 05:16 AI Companionship - Long-Term Harm 06:23 Google I/O 2026 10:07 Thinking Machines - Real-Time Interaction 11:18 Karpathy's CLAUDE md - open source 12:36 Perplexity Opens Up Its Skills Design Core 13:40 Productise Yourself 14:10 Anthropic's Claude for Legal - May 12 14:52 Connecting Claude AI to Obsidian and OMI 16:30 Krea 2 Update - Generate Visuals 17:02 Hermes vs OpenClaw 19:15 RAG and RAE 22:46 Gemini API file search as multimodal RAG 23:12 Hybrid RAG using Python 23:30 Google AI pointer with Visual Understanding 24:00 Zero Programming Language 25:03 Agents' core iterative process 26:03 OpenAI ChatGPT Personal Finance using Plaid.com 26:26 Tencent "Marvis" Personal AI assistant 26:42 MiniMax Mavis ("MiniMax as a Jarvis") 27:48 Agent sandboxing 33:44 Recursive Intelligence - Use AI to Design Chips 35:00 Modal AI - Infrastructure Startup 35:30 Recursive Superintelligence - Richard Socher 35:57 Julia McCoy - How to Start a Business 36:26 A Thin Prompt on a Frontier Model Is Not a Business 37:05 Why Complexity Kills Faster Than Competition 37:58 AI Avatar in YouTube videos 38:06 Rebuild AlphaGo from scratch on a modest budget 38:33 Jobs & Layoffs ---- Takeaways: 1. Harness matters more than the model. Microsoft MDASH on a regular model beat Anthropic's secret Mythos on Cyber Gym. 2. RAG no longer needs a vector database. Keyword search, graphs, BM25, grep, or full context work well. 3. Productize yourself and keep things simple. Thin prompts are not businesses; complexity kills projects. ---- Quotes: "Hardness is more important than model." "A thin prompt on a frontier model is not a business." "The most effective skill system is lightweight bespoke or one built from your own practical engineering experience." ---- Summary: In this Friday May 22nd update I cover the week in AI. Headline: harness beats model. Microsoft MDASH topped a cybersecurity benchmark using a regular model, beating Anthropic's Mythos. LM Arena shows Muse Spark and Qwen 3.7 Max strong, Qwen ran 35 hours autonomously. Musk lost his OpenAI lawsuit. Anthropic closed 30B at 900B valuation, Karpathy joined Anthropic. Google IO 2026 brought Gemini Omni, Gemini 3.5 Flash default, Anti-Gravity IDE, Gemini Spark, agentic search. Thinking Machines released a full-duplex real-time model. Karpathy open-sourced his CLAUDE md, Perplexity opened its skills design core, Anthropic added 12 legal plugins. I cover RAG vs RAE, 13 ways to do RAG without a vector DB, Gemini file search, hybrid RAG, Google AI pointer. Also Zero language with JSON diagnostics, agent sandboxing with microVMs and Docker, Tencent Marvis, MiniMax Mavis, Recursive Intelligence designing chips 100000x faster, Julia McCoy on productizing yourself, complexity kills projects, rebuilding AlphaGo cheaply, Meta layoffs of 8000. ---- Keywords: harness, MDASH, Mythos, LM Arena, Qwen, OpenAI, Anthropic, Karpathy, Google IO, Gemini, Anti-Gravity, Thinking Machines, CLAUDE md, Perplexity, Obsidian, OMI, Krea, Hermes, OpenClaw, RAG, RAE, AI pointer, Zero language, Marvis, MiniMax, sandboxing, Docker, Recursive Intelligence, Julia McCoy, AlphaGo, Meta layoffs ---- ---- ----