AI News Today | Julian Goldie Podcast

Sakana Fugu Ultra vs OpenRouter Fusion: Side-by-Side Tests (Beating Fable 5?)The video compares two new “panel” APIs—Sakana’s Fugu (Ultra and Mini) and OpenRouter Fusion—both claiming benchmark performance above Fable 5, by testing them side by side on visual/code generations like a landing page, raycaster maze, living spiral galaxy, inner solar system, and other simulations. In these examples, Fugu Ultra consistently produces the best-looking, smoothest outputs, Fusion is generally second, and Fugu Mini is faster but more variable and sometimes buggy. The script notes practical tradeoffs: Sakana/Fugu is cheaper with a flat plan but is heavily limited by daily generation/time caps and token limits, while Fusion is easier to use via API but costs per token. Both are one-prompt systems with long waits, and both are integrated into the creator’s Agent operating system and AI Profit Boardroom community.00:00 Sakana vs Fusion Intro00:32 How Panel APIs Work01:35 Landing Page Showdown02:17 Maze and Galaxy Tests03:42 Solar System Quality Gap04:10 Limits Pricing and Workflow06:09 Fusion Open World Example07:07 Fugu Mini Results08:00 Liquid Simulation Comparison08:55 Final Rankings and Tradeoffs09:35 Agent OS and Community Plug11:12 Wrap Up

Stop Prompting, Start Loop Engineering: Build Self-Iterating AI Agent SystemsThe script explains why “loop engineering” is replacing manual prompting: instead of iterating with an AI yourself, you define what “done” looks like, let an agent act, and use an independent judge model to verify results and repeat until the goal is met. It cites industry figures and examples (including Anthropic and NVIDIA’s GTC keynote) to argue that better loops, not better prompts, will power future self-iterating agent systems. The speaker describes a doer/judge loop structure and shows four implementations inside an “agent operating system”: a Fusion loop with configurable rounds and separate builder/judge models, an agent Kanban board with planner/builder/reviewer roles, a Fusion “boardroom” combining multiple model outputs via a judge, and a Sakana/Fugu council approach. It also describes an Obsidian-based memory system and promotes the AI Profit Boardroom community and resources.00:00 Why Loops Beat Prompts00:48 What Loop Engineering Means03:39 Prompting vs Looping04:44 Doer and Judge Framework05:37 Fusion Loop Walkthrough06:26 Agent Kanban Loops07:21 Fusion Boardroom Method08:31 Sakana Fugu Council09:05 Agent OS Bundle Overview10:01 Memory Feedback Loop11:26 Getting Started and Proof12:01 Join the Community12:35 Final Wrap Up

Agent OS Q&A: Fixing Kanban Blockers, Orchestrating Agents, Paperclip Paths, and Model SetupsThe episode answers recent questions about building and using an agent operating system (Agent OS) to organize multiple AI agents (e.g., Hermes, Jarvis) in one place, automate pipelines from idea to implementation, and deploy outputs like SEO content. It demonstrates using a Kanban board to triage tasks and complete an SEO keyword research report, then explains ways to resolve blocked tasks (commenting, moving cards, asking Hermes to sync and self-unblock, or using Claude to improve the system). It outlines options for agent orchestration via group chat, pipelines, and Paperclip teams, and recommends Agent OS over Hermes Workspace due to sync issues. The script also covers OpenRouter Fusion for high-stakes one-shot answers, local vs VPS hosting for Hermes, switching from local models to CLI/API, enforcing Paperclip output paths via working directory, simplifying Agent OS for clients, and integrating Obsidian memory and NotebookLM into the workflow.00:00 Agent OS Overview01:24 Kanban Swarm Setup02:06 Fixing Blocked Tasks04:15 Orchestrating Agents Together05:13 Why Workspace Sync Fails06:05 OpenRouter Fusion Explained08:07 Where to Run Hermes08:24 Game Studio Model Fixes09:18 Paperclip Output Paths10:18 Client Ready Agent OS11:46 Custom Upgrades Showcase13:31 Best Model Stack Choices14:15 Obsidian Shared Memory16:09 NotebookLM MCP Workflow

Hermes Agent OS: Build Your Own Autonomous AI System (No Code)

Discover how to transform standard AI models into a powerful Agentic Operating System featuring voice control, autonomous loops, and specialized agents. Learn why prompting is dead and how loop engineering allows non-technical users to build complex video, music, and SEO workflows automatically.

00:00 - Intro: The Hermes Agent OS

00:13 - Real-Time Voice Control & Jarvis

01:24 - Studio & Personal AI Butler

01:49 - Goal Mode: Why Prompting is Dead

04:10 - Specialized Music & Video Agents

06:24 - Orchestrating Agent Organizations

08:31 - Fully Automated SEO Loops

10:16 - How to Build it Without Code

Fugu Ultra vs Fable 5 & Fusion: New Multi-Agent Panel Model Benchmarks (GoldyBench)The video reviews Sakana.ai’s newly released Fugu Ultra, a multi-agent panel API that runs prompts across multiple models in parallel and fuses results via a judge, aiming for Fable 5-level outputs comparable to Fable and Mythos. The presenter shows one-shot examples (websites, animated galaxy, inner solar system), explains integration into their Agent OS alongside Fusion, and compares speed and reliability: Fugu is faster for latency while Fugu Ultra is much slower but optimized for benchmark performance. Benchmarks shown indicate Sakana outperforming Fable on several tests (e.g., Terminal Bench 2.1, GPQA Diamond, Live Code Bench), and it beat Opus 4.8 on most creations except a voxel game that failed due to 16K token truncation after a long wait. They also note strict API limits that can block usage for hours, regional access issues, and recommend using multiple models in an agentic OS for parallel work and fallbacks, with Fusion remaining top for their outputs.00:00 Fugu Ultra Arrives00:32 One Shot Demos01:19 How Multi Agent Fusion Works02:13 Agent OS Integration03:16 Benchmarks And Strengths03:57 Tradeoffs Limits And Truncation05:18 Leaderboard And Fusion Examples05:59 Speed Vs Smooth Workflow07:22 Pricing Access And Availability09:16 Best Setup And Wrap Up10:21 Community And Final Outro

Sakana Fugu (Fugu Ultra) Just Dropped: Multi‑Agent API vs Fusion, Fable 5 & GLM 5.2 BenchmarksThe video covers a new release from Japanese AI lab Sakana called Sakana Fugu, a multi-agent orchestration system accessible via a single model API, including a higher-end option called Fugu Ultra that reportedly matches frontier models like Fable and Mythos. The host explains the panel-style approach (similar to Fusion) where multiple models produce responses that are synthesized into one answer, then shares early hands-on tests generating a polished website, a maze game, a galaxy simulation, and an orbit simulation. They review benchmark comparisons (noting SW Bench Pro favors Fable 5) and side-by-side output quality against GLM 5.2, Opus 4.8, and Fusion, while warning about self-reported benchmark claims and recommending independent testing via their Goldy Bench. They show integrating Sakana into their Agent OS, discuss latency and use cases (Fugu vs Fugu Ultra), pricing differences versus Fusion, and how to access the API via sakana.ai, then promote the AI Profit Boardroom community and daily Agent OS updates.00:00 Sakana Fugu Launch01:11 Early Demos Showcase02:06 Fusion Style Overview02:23 Benchmark Performance03:06 Head To Head GLM04:45 Versus Opus Comparison05:12 Benchmark Hype Warning05:52 Agent OS Integration06:42 Using The Sakana API07:37 Fugu Versus Ultra08:32 Orbit Demo Results09:04 Fugu Versus Fusion09:37 Saving Outputs Workflow10:05 Cost And Pricing11:07 Wrap Up And Community

Claude Sonnet 5 LEAKED? New Anthropic Model Rumors!

Rumors are swirling about a new Anthropic release following the sudden shutdown of Fable 5. Learn about the leaked Claude Sonnet 5 and Mythos 6 developments, and discover why you must build model-agnostic systems to survive the volatile AI landscape.

00:00 - Intro: The Anthropic Mystery

00:50 - Claude Sonnet 5 Leaked?

01:39 - New Mythos Model Training

02:54 - Why Fable 5 Was Banned

04:00 - Confirmed vs Rumors

04:53 - The Trap of Model Dependency

05:34 - Future-Proofing with Agent OS

08:20 - AI Profit Boardroom Update

Fusion by OpenRouter: One-Prompt Game Builds, Panel-of-Models AI, and Benchmark Results vs Opus 4.8The video introduces OpenRouter’s new Fusion API, designed to reach Fable 5-level intelligence by running five models as a panel and using a judge to critique and fuse their responses into one output from a single API call. The creator demos multiple projects generated from one prompt, including a 3D maze, Dragon Realm open-world game, block breaker, neon racer, pool emulator, voxel dash, raycast game, and Outrun, then compares Fusion outputs side-by-side against Claude Opus 4.8, showing Fusion’s versions as more playable and polished. The episode reviews OpenRouter benchmark claims, notes Fusion can reach comparable performance at about half the cost, and explains premium vs budget panels across 42 tasks. It also shows integrating Fusion into an agent operating system workspace, while highlighting the main drawback: slow response times and workflow friction.00:00 Fusion Game Demos01:03 How Fusion Works01:44 Benchmarks and Cost03:58 More One Prompt Creations04:21 Fusion vs Opus Comparisons06:26 Agent OS Integration07:18 Limitations and Verdict07:55 Community and Resources09:16 Final Wrap Up

The script demos “Hermes Jarvis,” a fast, voice-activated Jarvis-style assistant that can open apps like Obsidian and Google, respond in real time, be stopped mid-response, teach basics like Japanese greetings, and discuss building a snake game. It gives a daily briefing (dated 20 June 2026) pulling from the creator’s context, including YouTube performance, tasks like restructuring an LLM wiki, transcript fixes, and a note about US consumers disliking AI in brand messaging. The creator shows an Agent OS interface with conversation history, previews of created outputs, wall mode with briefing history, and deep integration with an Obsidian-based “memory galaxy” that logs conversations automatically and blends personal and business context. They describe plans to automate SEO using stored case studies and promote getting the setup via the AI Profit Boardroom with daily updates, tutorials, community Q&A, coaching calls, and local meetups.00:00 Hermes Jarvis Demo00:14 Voice Commands Showcase01:25 Daily Briefing Preview02:21 Why It Feels So Fast02:42 History And Real Time03:44 Obsidian Memory Context05:10 Projects Themes Universe06:08 Wall Mode Workflow06:41 Real World Use Cases07:51 Automating SEO Next08:39 Inside Agent OS Suite09:20 Get The Setup Boardroom09:46 Community Coaching Outro10:27 Final Thanks

Agent OS Q&A: Hardware, Windows Setup, Local Models, SEO Workflows, NotebookLM MCP + MoreThe video tours an Agent OS “mission control” that connects AI agents, tools, CLIs, and custom workflows (including an agent Kanban, SEO setup, video automation, and an Obsidian-based memory system), then answers community questions. It covers local hardware and model suggestions (Mac Mini/Studio not ideal for local models unless high spec; RTX setups like RTX 590 and DGX Spark; Qwen 3.6 for text and Ernie for images) while noting Agent OS is lightweight if using APIs/CLIs. For Windows, it recommends native setup over WSL. It outlines agency use cases (video pipeline, SEO content deployment, and organizing clients as Obsidian project folders), shows SEO ranking examples, explains connecting NotebookLM via MCP, discusses OS/VPS options, email safety (use a sandboxed email), model choices and a personal benchmark site (goldybench.com), and describes AI Profit Boardroom access, support, updates, and coaching.00:00 Agent OS Overview00:56 Local Hardware Models02:14 Windows Setup Choices02:45 Agency Use Cases04:45 SEO Wins Examples05:43 NotebookLM MCP Integration07:02 OS VPS Hosting Options08:44 Model Picks Benchmarks10:44 Custom Builds Showcase11:23 Automation Email Safety13:30 More Community Q&A15:12 Wrap Up Join Boardroom