AI News Today | Julian Goldie Podcast

Claude Mythos & GPT-5.5 Leaks: Unreleased Frontier Models Exposed in the Same WeekThe script covers two major AI leaks happening in the same week: Anthropic’s unreleased Claude Mythos (also referenced as MFO), described as so powerful at finding software bugs and zero-day vulnerabilities that Anthropic said it was too dangerous to release publicly, and OpenAI’s apparent GPT-5.5 leak via a Codex UI model picker. It recounts how a CMS misconfiguration previously exposed documents about Mythos, then describes a Bloomberg-reported incident where a private Discord group allegedly gained access by guessing model endpoint names and using third-party contractor credentials, prompting Anthropic to investigate. Separately, Codex users reportedly saw at least eight internal OpenAI model names (including GPT 5.5, OAI 2.1, Arcanine, Glassier Alpha variants, Heisenberg, and GPT Rosalyn), with some developers briefly testing GPT-5.5 and describing it as faster and more token efficient.00:00 Two Major Model Leaks00:27 Claude Mythos Exposed00:51 Why Mythos Is Dangerous02:00 Discord Access Breach03:23 What It Means For You04:20 OpenAI Codex Model List05:39 GPT 5.5 First Impressions07:24 Mystery Models Explained08:04 Industry Moving Too Fast09:13 Predictions And Wrap Up09:39 Community And Final Thanks

GPT-5.5 Leaks: Internal OpenAI Model Names Spotted in Codex & ChatGPT (OAI 2.1, Arcanine, Glassier Alpha)The script covers new unconfirmed leaks suggesting GPT-5.5 may release soon after internal OpenAI model names briefly appeared in Codex and ChatGPT model dropdowns for some Pro users on April 22, indicating a UI-level exposure of an internal testing environment that was quickly pulled but screenshotted and prompted by developers. Leaked entries include GPT-5.5 (described as an agentic coding model with codename “Spud,” pre-training completed March 24 and reportedly in safety evaluation/final testing), OAI 2.1 (same description as GPT-5.5 but different naming), GPT 5.4/5.2 Codex, Arcanine, and multiple Glassier Alpha variants (CY3/CY4 checkpoints), plus mentions of Heisenberg and Rosalind. The speaker interprets the list as evidence of parallel internal tracks and a shift toward agentic, domain-specific models, while cautioning that performance claims may be exaggerated, and ends by promoting the AI Profit Boardroom community and training.00:00 GPT 5.5 Leak Buzz00:47 Codex Dropdown Evidence02:14 How the Leak Happened03:01 GPT 5.5 Details and Hype04:15 OAI 2.1 Naming Mystery05:18 Arcanine and Agent Workspaces06:44 Glassy Alpha Variants Explained08:31 What the Leak Signals09:49 Release Timing and Speculation11:42 Wrap Up and Community Pitch

ChatGPT Image 2.0: Best AI Image Generator Yet (Side-by-Side vs Gemini + Agents, Codex, API)The episode tests the new ChatGPT Image 2.0 update in ChatGPT, showing examples like a hyper-realistic movie poster (“The Last Noodle”), comics, a fantasy world map, and UI-style images, and compares results side-by-side with Gemini, concluding Image 2.0 generally looks more detailed, realistic, and has better text rendering, while logo results are similar. It highlights workflow features like quick aspect-ratio switching, editing parts of existing images, generating multiple images, and typical generation times (about 43 seconds in one test). The script also explores using Image 2.0 via the OpenAI Playground and API, connecting it to agents (OpenClaw and Hermes), and trying Codex 2.0 for mockups and page generation, noting the mockup can look better than the resulting implemented page. It cites ELO rankings claiming Image 2.0 leads by a wide margin.00:00 Image 2.0 Reveal01:00 Movie Poster Demo01:27 Gemini Comparison02:10 More Examples API04:08 Comic Strip Test06:18 Aspect Ratio Controls07:16 Codex Integration08:33 Logo Prompting Claude12:21 What Image 2.0 Is14:03 Codex Website Mockup15:21 Reasoning Better Details16:10 More Prompts Timing16:24 Image Speed Test17:10 Poster Results Review17:40 Comic Prompt Plus Rankings18:22 Codex Page Reality Check19:35 Detailed Map Stress Test20:56 Wild Examples And Editing23:14 Book Mockup UGC Test24:48 API Setup For Agents27:28 Why Image Two Matters29:18 Wrap Up And Prompt Workflow

Space Agent Review: Build AI Dashboards in Your Browser (Open Source + Time Travel)The video tests Space Agent, a new open-source AI agent from the founders of Agent Zero that runs in the browser or a desktop app and is positioned as an easier alternative to tools like OpenClaw and Hermes. The presenter shows a quick setup, connects models via local options or an API key (e.g., OpenRouter), and demonstrates creating widgets and mini apps inside “spaces,” including a weather widget, Pomodoro timer, to-do list, and a Bangkok clock, plus basic browser automation. Key features highlighted include a cleaner UI, multiple spaces for different use cases, file and user/team support, Git-tracked changes, and a “time travel” rollback system to revert broken updates. It also emphasizes token efficiency via tool calling, structured output, and YAML. Availability is shown for browser and Mac, not mobile.00:00 Meet Space Agent00:50 Quick Setup and LLM Options02:26 Time Travel and App Download04:40 Desktop App Weather Widget Demo07:53 Browser Automation in One Space09:00 Space Agent vs OpenClaw10:39 Multiple Spaces Productivity Build12:53 Why It Stands Out15:13 How Spaces and Chat Work17:08 Teams Git Tracking and Rollbacks17:57 Ways to Use and Final Thoughts18:57 AI Profit Boardroom Plug20:45 Q&A and Wrap Up

OpenClaw 4.21: GPT Image 2 Integration, 4K Outputs, Image Editing + Critical Security FixOpenClaw 4.21 adds same-day support for OpenAI’s new GPT Image 2 model (used in ChatGPT Images 2.0), letting agents generate higher-quality images with accurate, readable text and new size options up to 4K (3840×2160), plus image editing, reference-image handling, and masked partial edits. Setup requires adding your OpenAI API key and pointing OpenClaw to the new model, after which it becomes the default image provider. The release also includes six fixes: a smarter plugin “doctor,” clearer image provider failover logging, a closed owner-command security hole (especially important for shared Slack/Discord channels), correct Slack thread replies, faster browser actions when elements are missing, and cleaner installs. The script advises backing up before updating, updating immediately for the security fix, then testing real business assets like menus, flyers, and social posts.00:00 OpenClaw 4.21 Overview00:20 Setup GPT Image 200:40 Better Text Rendering01:22 4K Sizes and Formats02:00 Business Use Example02:15 Image Editing and Masks03:05 AI Profit Boardroom Pitch03:53 Plugin Doctor Fix04:29 Image Logging Backup05:10 Security Owner Patch06:09 Slack Threads Fixed06:37 Browser Speed and Cleanup07:46 What It Means for Business08:17 Update Steps Checklist09:19 Bigger Picture and Wrap

Build & Automate Anything with Hermes AI Agent (Full Tutorial)Discover how to build and automate anything using Hermes, the powerful open-source AI agent from Nous Research. This step-by-step guide covers everything from local setup and memory integration with Obsidian to managing multi-agent swarms for maximum productivity.00:00 - Intro & Hermes Overview01:23 - Step-by-Step Installation02:13 - Real-World Use Cases03:12 - Enhancing AI Memory with Obsidian07:38 - Building AI Agent Teams13:11 - Using the Hermes Dashboard16:27 - One-Click Setup with Ollama18:52 - Final Thoughts & Resources

How to Setup Kimi k2.6 with Ollama & OpenClaw (Easy AI Agents)Learn how to deploy the agentic Kimi k2.6 model using Ollama and OpenClaw for high-speed AI automation. We cover both terminal-based setups and user-friendly GUI alternatives like Atomic Chat for seamless agent integration.00:00 - Intro & One-Click Setup00:18 - Accessing Kimi Models in Ollama00:48 - Web Search & Performance Test01:22 - Kimi Cloud vs. Local Gemma Models02:10 - Running Hermes with Ollama02:40 - Atomic Chat: The Easy UI Alternative03:17 - Connecting API Keys for Local Models04:43 - AI Agent Training & Resources

Kimi K2.6 + Hermes Agent: One-Click Free AI Agent SetupLearn how to set up the powerful Kimi K2.6 model for free using Olama to power AI agents like Hermes and Claude Code. This guide covers one-click configurations for autonomous web research, coding, and background tasks without needing technical expertise. Discover how to leverage Kimi K2.6 as a free alternative to top-tier models for long-horizon tasks.00:00 - Intro: Setting up Kimi K2.6 & Olama00:58 - Running Kimi K2.6 with Hermes Agent02:10 - Using Kimi K2.6 with OpenClaw02:27 - Kimi K2.6 in Claude Code02:56 - Performance Benchmarks & Comparisons04:23 - What is Hermes Agent?04:52 - Training & Community Resources

China’s Kimi K2.6: The Most Powerful Open Source AI Agent?Explore Kimi K2.6, a groundbreaking open-source AI agent that outperforms Claude and GPT-4o in complex benchmarks. Learn how to automate your workflow using agent swarms, cloud-hosted Kimi Claw, and local coding tools.00:00 - Intro to Kimi K2.600:21 - Benchmarks vs. Claude & GPT01:47 - Kimi Agent & Swarm Modes04:25 - Building Websites & UI06:58 - Automating Sheets & Systems08:57 - Kimi Claw & Cloud Agents10:32 - Kimi Code: CLI & Coding15:01 - Creating Custom AI Skills

Run Claude Code Locally for FREE: Offline & Private AI Agent

Learn how to set up and use Claude Code Local to run powerful AI agents entirely on your machine for free. This guide covers offline usage, model switching between Llama and Gemma, and even controlling your agent via voice or iMessage.

00:00 - Intro to Claude Code Local

00:34 - Running Claude Code Offline

01:47 - Comparing Local Models (Llama, Gemma, Qwen)

02:33 - Local Performance & Private Data

03:23 - Fixing Tool Call Reliability

04:19 - iMessage & Hands-Free Voice Control

05:14 - Hardware Requirements (M1/M2/M3)