AI News Today | Julian Goldie Podcast

Le Chaton Fat Just Broke the Internet...

Why AI Benchmarks are Fake (And How to Actually Test Models) A fake French AI model recently went viral for beating the industry's top benchmarks, proving how easy it is to manipulate performance data. This video explains why you should stop chasing hype-filled charts and start evaluating AI based on your own real-world business workflows. 00:00 - Intro: The Le Chatton Fat Joke 01:08 - Why AI Benchmarks Can Lie 02:42 - The Problem with Self-Reported Tests 04:18 - Real Work is the Only Benchmark 05:20 - How to Avoid AI Overwhelm 06:34 - The New Way to Evaluate AI 07:31 - 3 Key Takeaways for AI Testing 08:45 - Testing AI Systems Yourself

NEW Grok Build AI is INSANE!

Grok Build CLI Just Surprised Me: Side-by-Side Tests vs Claude, Kimi & GLM + Agent OS IntegrationThe script reviews the new Grok Build coding agent and CLI, saying it’s a major upgrade over the old version and the best results the creator has seen from the same prompt, especially for games and 3D builds. It showcases multiple outputs (a dark first-person game, clean websites, 3D/MMO/RPG-style demos, animations like glowing fish trails and lava-lamp liquid that follows the mouse) while noting quality is inconsistent and sometimes disconnected from prompts. The creator compares Grok Build against Kimi Code, GLM 5.2, and Claude Opus, claiming Grok Build’s game beat Claude’s for the same prompt and outperformed Kimi and GLM on some tests, though GLM looked better for certain animations. They highlight OAuth login and that X Premium Plus/Super Grok subscribers get access, plus parallel sub-agents, self-check loops, and integration into an agent operating system alongside other models, then promote an Agent OS package via AI Profit Bot.00:00 [https://www.youtube.com/watch?v=rNzH-meiOD8] Grok Build Surprise00:37 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=37s] Touring Early Demos01:12 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=72s] OAuth and CLI Access02:08 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=128s] Studio Media Features02:36 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=156s] Beta Announcement Context03:11 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=191s] Side by Side Comparisons04:37 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=277s] Animations Mixed Results05:50 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=350s] Parallel Sub Agents06:22 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=382s] Dragon Realm Benchmark07:02 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=422s] Self Check Improvement Loop07:58 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=478s] Using Grok in Agent OS08:48 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=528s] Agent OS Offer Breakdown09:43 [https://www.youtube.com/watch?v=rNzH-meiOD8&t=583s] Community Coaching and Close

Hermes + Jarvis + Obsidian Memory is INSANE!

I Turned My Mac Into a Voice-Activated AI Butler (Hermes + Jarvis + Obsidian Memory)The script demonstrates a voice-activated AI assistant running on a Mac inside an agent operating system, using Hermes Agent with Jarvis and connected to an Obsidian vault for persistent memory. The AI provides daily and weekly briefings that summarize work themes, open action items, priorities, and recent news headlines with clickable links, and it can be triggered by tapping a panel or a wake word. It can open websites and apps, show dashboards and live numbers, run “wall mode” as a command-center display, build working apps by voice with outputs saved to a gallery, and save user commands like “Remember that” directly into Obsidian so it can recall past context across days. The episode ends by promoting the Agent Operating System setup available through AI Profit Boarding.00:00 [https://www.youtube.com/watch?v=aUNzI0lN2wU] Talking Butler Demo01:33 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=93s] Voice Commands and Builds02:19 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=139s] What Hermes Jarvis Is03:52 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=232s] Six Powers Overview04:40 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=280s] Briefings and History05:42 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=342s] Memory and Agent Mode06:21 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=381s] How Commands Route08:22 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=502s] Recap and Setup Offer09:33 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=573s] Agent OS and AI Profit Boarding11:04 [https://www.youtube.com/watch?v=aUNzI0lN2wU&t=664s] Final Thanks and Goodbye

NEW GLM 5.2 BEATS Claude?

GLM 52 vs Qwen 37 Max vs Claude Opus 48: Real-World Tests vs Benchmarks (No Second Chances)The episode compares GLM 52 (ZAI), Qwen 37 Max (Alibaba), and Claude Opus 48 (Anthropic) head-to-head on five one-shot tasks, arguing that benchmark rankings didn’t match real usability. In coding-focused tests like a voxel runner game, a liquid-in-a-bowl animation, a business landing page, and an arcade game, GLM 52 produced the most fun, polished, and feature-rich results, while Claude’s outputs were often basic and Qwen’s were sometimes buggy or incomplete; Claude clearly won the solar-system orbit map task. The script also notes Qwen’s strong reported benchmarks and faster replies, GLM’s slower responses in agents but strong CLI coding, and highlights limitations integrating Claude into agent workflows compared to Qwen/GLM in Hermes and the creator’s agent operating system.00:00 [https://www.youtube.com/watch?v=MnFwz8O3F-U] Head To Head Setup01:27 [https://www.youtube.com/watch?v=MnFwz8O3F-U&t=87s] Coding Tests Results04:09 [https://www.youtube.com/watch?v=MnFwz8O3F-U&t=249s] Arcade Game Showdown04:50 [https://www.youtube.com/watch?v=MnFwz8O3F-U&t=290s] Benchmarks Versus Reality06:01 [https://www.youtube.com/watch?v=MnFwz8O3F-U&t=361s] Agents Workflow Tradeoffs07:59 [https://www.youtube.com/watch?v=MnFwz8O3F-U&t=479s] Final Recommendations

China’s Kimi K2.7 is ABSURD!

White House Forces Anthropic to Pull Mythos & Fable 5 Worldwide in 90 MinutesThe script recounts how Anthropic’s private security-focused model Mythos and its public, guardrailed version Fable 5 were taken offline worldwide after the Trump administration ordered Anthropic to block all foreign nationals—including non-citizen staff—from using them, with one report saying the shutdown happened within 90 minutes. It describes a weekend escalation beginning with Fable 5’s public launch, an Amazon-discovered jailbreak reported by CEO Andy Jassy, and a White House “fix it or pull it” demand using export-control-style rules typically reserved for weapons. Reporting from Semaphore suggests unconfirmed White House concerns that a China-linked group may have accessed Mythos, raising fears of “distillation” to copy capabilities. The narrator frames this as evidence AI is now treated like critical power infrastructure and urges business owners to adopt practical, publicly available AI tools, promoting the AI Profit Boardroom.00:00 [https://www.youtube.com/watch?v=UwPAwCl31PY] Mythos Shutdown Shock01:11 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=71s] Timeline to Blackout02:07 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=127s] Why Mythos Matters02:31 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=151s] Distillation Explained03:39 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=219s] Jailbreak and Standoff04:35 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=275s] AI as National Power05:17 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=317s] Business Gap Warning06:07 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=367s] Boardroom Pitch and Fusion07:00 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=420s] Three Issues Clarified07:58 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=478s] Earlier Mythos Leak08:29 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=509s] Weekend DC Negotiations08:44 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=524s] Everyday AI Keeps Advancing09:23 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=563s] Stop Waiting Start Using09:52 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=592s] Final Takeaways and CTA11:04 [https://www.youtube.com/watch?v=UwPAwCl31PY&t=664s] Closing Thanks