AI News Today | Julian Goldie Podcast

Headroom: Free Open-Source Tool to Cut AI Agent Token Use by 60–95%This episode introduces Headroom, a free open-source project trending on GitHub that compresses what AI agents read to reduce token usage by 60–95% without losing meaning. It’s described as a “zip file” for agent context—files, tools, search results, logs, and conversation history—helping agents run faster, cost less (especially via APIs), and forget less due to context-window limits. The script claims Headroom can plug into many agents and tools (including Claude Code, Codex, Cursor, OpenClaude, and Hermes) and offers proof examples such as 92% token savings and compressing 10,144 words to 1,260 while finding the same log error. It outlines a three-step approach: crush, keep (shared reversible memory), and compound (learn from failures).00:00 Token Saving Breakthrough01:07 Why Agents Burn Tokens03:00 Headroom Zip Compression03:03 Quick Install Demo03:27 Three Key Benefits03:39 Goldie Framework Steps04:31 Proof and Benchmarks05:31 Common Objections Answered06:03 Recap and Next Steps06:45 Boardroom Offer and Outro

Odysseus (PewDiePie’s Open-Source AI Agent): Local ChatGPT-Style Workspace TestedThe video tests Odysseus, PewDiePie’s free, open-source, self-hosted AI workspace that runs locally (offline) like ChatGPT or Claude but on your own computer with your data. It shows the app’s retro, somewhat messy UI and key features including chat, an agent for tasks, deep research, document writing, memories/skills, notes, tasks, calendar, email, a library for documents, scheduled task triggers, and model switching. Setup is shown via GitHub instructions (Docker or native macOS), creating an admin account, adding an API “brain,” and running on a local address. The script demonstrates using local models (including Gemma) or cloud models via OpenRouter, the “cookbook” for selecting downloadable local models, and a side-by-side model comparison, concluding that it works and is lightweight aside from local model requirements.00:00 Meet Odysseus00:21 What It Is00:43 Setup Options01:12 Tour The Interface02:38 Tasks And Scheduling03:39 Chat And Agent Test04:38 Cookbook And Compare05:16 Six Step Setup05:56 Who Its For06:32 Recap And Next Steps06:51 Community And Training07:51 Final Thanks

Run Gemma 4 Locally in Hermes Agent (Free AI Automation + New Web UI Setup)The script explains how to combine Google’s newly released Gemma 4 12B model with Hermes Agent to run a free, local AI agent designed for agentic reasoning and automation. It shows examples built with Hermes Agent and an agentic operating system, including games, a Pomodoro timer, a color palette, and animations. Setup is demonstrated via Ollama (download Ollama, run the command to launch Hermes with Gemma 4) or through Hermes Agent’s new web UI to select Gemma 4 as the model. It also suggests configuring a stronger main model with Gemma 4 as a sub-agent to save tokens, and notes an alternative free API option via OpenRouter for Gemma 4 26B/31B if local hardware is limited.00:00 Free Agents With Gemma00:36 What You Can Build00:54 Local Setup With Ollama01:15 Web UI Model Switching01:32 Main Model And Subagent02:07 Free API Option03:01 Body And Brain Explained04:09 Automation Use Cases05:13 Offline And Token Savings06:15 Chat Versus Agent06:34 Hermes Dashboard Tour07:20 SOP Pairing Steps07:53 Benchmarks And Ecosystem08:24 Recap And Offline Travel08:51 Training And Community Pitch09:46 Final Thanks

Headroom: Free Open-Source Tool to Cut AI Agent Token Use by 60–95%This episode introduces Headroom, a free open-source project trending on GitHub that compresses what AI agents read to reduce token usage by 60–95% without losing meaning. It’s described as a “zip file” for agent context—files, tools, search results, logs, and conversation history—helping agents run faster, cost less (especially via APIs), and forget less due to context-window limits. The script claims Headroom can plug into many agents and tools (including Claude Code, Codex, Cursor, OpenClaude, and Hermes) and offers proof examples such as 92% token savings and compressing 10,144 words to 1,260 while finding the same log error. It outlines a three-step approach: crush, keep (shared reversible memory), and compound (learn from failures).00:00 Token Saving Breakthrough01:07 Why Agents Burn Tokens03:00 Headroom Zip Compression03:03 Quick Install Demo03:27 Three Key Benefits03:39 Goldie Framework Steps04:31 Proof and Benchmarks05:31 Common Objections Answered06:03 Recap and Next Steps06:45 Boardroom Offer and Outro

Gemma 4 12B Just Dropped: Free Agentic Open-Source Model You Can Run LocallyGoogle has released Gemma 4 12B, a new free, open-source, agentic model you can access now and run locally, and the creator demonstrates building real apps inside an agent operating system by plugging Gemma 4 into Hermes Agent. Examples shown include a mouse-follow animation app, a color palette designer, a Pomodoro timer, generative art, a simple game, a website, and a wallpaper, emphasizing improved usefulness versus earlier Gemma versions. The script explains how to get Gemma 4 via Ollama (recently updated), or use free Gemma 4 API options like a 26B model via OpenRouter if you lack powerful hardware, and highlights core specs such as 16GB memory and a 256K context window along with benchmark claims around 77% reasoning and ~72% coding.00:00 Gemma 4 Drops00:32 Demos Built in Hermes01:35 Get It via Ollama02:20 What Gemma 4 Is02:56 Using It in AOS03:50 Goldie Pocket Genius06:50 Setup and Integrations07:33 Use Cases and Limits09:08 Benchmarks and Multimodal10:04 Recap and Community11:23 Final Goodbye

Gemma 4 12B: Google's New Private & Free Local AI Model

Discover Google's latest Gemma 4 12B model, a lightweight yet powerful multimodal AI you can run privately on your own laptop. We break down its Mixture of Experts architecture, impressive benchmarks, and how to set it up for your AI agents for free.

00:00 - Intro to Google Gemma 4 12B

01:03 - Multimodal Intelligence Explained

02:24 - Hardware Requirements & Size

03:44 - Mixture of Experts Architecture

04:39 - 256K Context & Privacy Benefits

05:42 - Performance Benchmarks vs Larger Models

07:04 - How to Run Gemma 4 Locally

09:12 - Free API & Agent Integration

10:31 - Recap: Is Gemma 4 the New Standard?

China’s MiniMax M3: The AI “Night Shift” That Works 24/7 for a Fraction of the CostThe script introduces MiniMax’s new Chinese AI model M3 (released June 1) and claims it beats major U.S. models like GPT 5.5 and Gemini 3.1 Pro on certain coding benchmarks at a much lower price. It highlights an internal MiniMax test where M3 worked autonomously for 24 hours, making nearly 2,000 moves across 147 attempts without human intervention, noting the results are from MiniMax’s lab and not yet independently verified. The core idea is shifting from using AI as a single chatbot to running it as a coordinated “team” that can operate a “night shift,” continuing work while you’re offline. It describes integrating M3 into an agent operating system with shared memory and pairing it with Hermes as the “hands” to execute tasks, enabling image, video, and voiceover creation from one dashboard.00:00 MiniMax M3 Bombshell00:21 Benchmarks and 24 Hour Test01:02 Chatbot to AI Team01:48 The Night Shift Framework02:15 Agent OS Demo Setup02:50 Million Word Memory03:20 Multimodal Studio Powers03:50 Hermes Hands Integration04:15 Live Workspace Walkthrough05:32 Not Techy Setup Objection05:51 AI Profit Boardroom Offer06:39 Why This Model Matters06:58 Stubbornness Beats Quitting07:30 Putting the System Together08:07 Don't Wait Start Now08:43 Final Call to Build Yours

My Claude Agent Operating System (Mission Control) + New Anth CLI UpdateThe video walks through the creator’s “Claude agent operating system,” a mission-control dashboard that centralizes tools and AI agents to reduce context switching and keep work organized. They explain why they prefer one workspace over jumping between Claude, ChatGPT, terminals, and other apps, and emphasize keeping the system minimal by removing rarely used features like UltraCode. A new update released today—Anth CLI—lets users access Claude’s console from a terminal to talk to cloud-hosted, managed agents that can run 24/7 and be scheduled for recurring tasks, such as generating SEO strategy and link-building ideas. The setup links agents like Hermes with an Obsidian memory vault, and the creator highlights Claude’s workspace for easily revisiting past projects. They also mention using Claude to troubleshoot and update components, and offer the full system via AI Profit Boardroom.00:00 Agent OS Overview00:27 Why Mission Control Matters01:08 Anth CLI Update01:50 Managed Agents In Cloud02:52 Keep It Minimal03:28 Linking Tools And Memory04:09 Chrome Workflow And Focus05:39 UltraCode For Big Workflows06:17 Claude Workspace Benefits07:28 Future Trend And Community08:12 Maintenance And Plug And Play09:11 Get The Full System09:47 Final Thanks

Why Hermes Needs an Agent Operating System (Mission Control Dashboard Demo)The script argues that using Hermes Agent in a terminal or even Hermes Desktop limits its value because it becomes a single-chat experience without strong organization, artifact previews, or team/agent management. The speaker showcases a custom-built “agent operating system” as a mission-control dashboard that centralizes multiple agents and tools, previews outputs like websites and media, saves everything in one workspace, and enables autonomous “goals mode” for long-running tasks. They highlight faster access to features (like voice mode), the ability to quickly plug in new releases (e.g., Minimax M3, Claude’s Ant CLI, MCPs), and a memory system via an Obsidian vault to avoid repeatedly re-explaining context. The episode concludes by promoting the AI Profit Boardroom as a place to get the system, tutorials, updates, and community support.00:00 Why Agents Need Systems00:42 Terminal Use Is Limiting01:03 Mission Control Dashboard Demo02:03 Hermes Desktop Problems03:22 Voice Mode Instant Setup03:58 Plug In New Tools Fast05:31 Memory With Obsidian Vault06:47 Kanban Made Simple07:35 Build Systems Not Models08:52 Get the Operating System09:19 Community Proof Anyone Can10:10 Boardroom Features and Wrap

Anthropic Files for IPO: Claude’s Maker Nears a Trillion-Dollar Valuation as the AI IPO Race Heats UpAnthropic, the company behind Claude, says on its website (June 1, 2026) that it has filed paperwork to sell shares to the public, beginning the IPO process with no price or timing set yet. Coverage cited in the script claims Anthropic’s annual revenue run-rate rose from about $9B a year ago to $47B today, it recently raised $65B, and its valuation reached roughly $965B. The script frames the filing as a strategic move ahead of OpenAI’s expected IPO and alongside SpaceX’s filing, potentially opening the “floodgates” for major AI listings. It highlights Anthropic’s business-heavy revenue (about 80%), including over 1,000 businesses paying $1M/year, strong growth from Claude Code, and investor interest tied to advances like the unreleased Claude Mythos model.00:00 Anthropic Files IPO00:28 Explosive Revenue Growth01:04 AI IPO Race Heats Up01:37 Business Model and Claude Code02:12 Rivalry With OpenAI03:01 What IPO Reveals03:49 Claude Mythos Breakthrough04:35 Why This IPO Is Historic05:38 IPO Timeline and SEC Review05:57 Market Impact and Concentration07:35 Bubble Fears and Profitability08:39 The Great Insider Handoff09:08 Final Takeaways09:50 Sponsor and Community Pitch