Video summary
This AI Agent Merges ChatGPT And Claude Into One Smarter AI (+18 AI Updates)
Main summary
Key takeaways
Summary of the video’s main points (AI +18 updates)
1) “Free agent” approach: combine models for better results
- The video argues that relying on a single “best” model is becoming risky due to:
- locked features,
- pricing changes, or
- sudden bans.
- It highlights Hermes, a tool that can run two AI models at once and then merge their outputs into a single answer (a “mixture of agents” setup).
- Reported benchmark gains:
- ~8% above Opus
- ~11% above GPT
- ~5.5% above another baseline
2) Meta allegedly restricts engineers from using rival AI coding tools
- Citing a report from The Information, the video says Meta told its engineers not to use Claude Code and OpenAI’s code tooling.
- The reason is framed less as classic security and more as distillation risk:
- high-quality rival code could end up in training data for Meta’s models,
- potentially creating legal exposure (noting Anthropic has previously accused some Chinese companies of similar behavior).
3) New hardware and input methods
- Whispering ring / “WhisperFlow”
- Enables silent dictation by capturing very soft speech and displaying typed text on screen.
- Also works as a tiny trackpad for cursor movement and editing without a keyboard.
- OpenAI Codeex hardware
- OpenAI teased a physical macro pad for coding shortcuts to reduce repetitive keypresses/clicks.
4) Chinese AI ecosystem updates (GLM and “Claude replacement”)
- A Chinese company (Z.A.I.) is described as launching Zcode, a dedicated coding app built around GLM 5.2.
- The video frames it as meaningful because GLM-based coding previously often required plugging into other tools; now it has a standalone IDE-like environment.
- Additional note: Zcode may offer extra quota if you already have a GLM coding subscription (usable inside Zcode).
5) Anthropic’s “Fable 5” returns—after a government ban—but weaker
- Fable 5 was previously pulled globally after US government intervention over security concerns, following a launch and then a quick takedown.
- It reportedly returned after Anthropic tightened security (around June 30).
- However, the video claims it came back weaker.
- Offered explanation: DeepMind released two models the same day designed to work together:
- Nano Banana 2 Light (fast image generation)
- Gemini Omni Flash (text/image-to-video with conversational editing and chained workflows via API)
6) AI becomes more “agentic” for research and learning
- Claude Science (Anthropic)
- A research system connected to 60+ scientific databases.
- Uses specialist agents, can run code in notebooks, and supports compute-on-demand.
- Emphasizes traceability (what ran and where claims came from).
- Google Notebook LM
- Upgraded learning tool that generates short vertical “video overviews” from user-provided materials, supplementing/augmenting text explanations with narrated visuals.
7) Productivity + creation tools expand across platforms
- Gamma inside ChatGPT
- Presentation/document/website generation accessible from within ChatGPT’s app ecosystem.
- Gemini Spark (Google)
- A desktop app that performs tasks across apps (email monitoring, note cleanup, spreadsheet tracking) in a “spark mode.”
- Comfy UI + “Comfy MCP”
- Lets AI agents control image/video pipelines via chat (e.g., generating assets from Figma/Notion inputs).
- Notion “HTML block”
- AI can convert page content into interactive elements such as timelines, dashboards, prototypes, diagrams/quizzes, and even mini-games.
- Google AI Studio
- Adds design variations so creators can generate multiple UI looks quickly.
- Google AI Studio sharing improvement
- When sharing an app, users can include the full Gemini chat trail (prompts and iteration steps), not just the final output.
- Riplet moves to desktop
- Native desktop app version for Mac/Windows for multi-workspace project organization.
8) AI phone agents and app-building agents
- XAI Voice Agent Builder
- Create an AI phone agent in minutes using Gro’s voice.
- Supports document-based answering and actions (including live refund calls).
- OpenClaw
- A free AI agent now has its own iOS/Android app for managing tasks from a phone.
Zcode walkthrough / evaluation (core demonstration)
- Zcode is positioned as a coding environment for GLM 5.2 with:
- a project panel + chat + “add context” (attach files, commands, past sessions),
- model selection (GLM 5.2 and GLM 5 Turbo by default),
- support for adding other models via providers (example uses OpenRouter, requiring your own API key),
- “thinking” settings that control how much effort the model uses before responding.
What the demo tests
- Personal portfolio website generated from an exported LinkedIn PDF
- Same prompt → three models produce different styling approaches:
- dark,
- minimal,
- emerald-gradient variants.
- Same prompt → three models produce different styling approaches:
- A 2D Jungle Run game
- GLM: game runs, but core mechanics are broken (shots pass through enemies).
- Sonnet: character falls through the ground (not playable).
- Fable: game mechanics work correctly; shooting and upgrades function (declared the “clear winner”).
Demo conclusion
- Zcode isn’t necessarily the best overall replacement for tools like Cursor/Codeex/Claude Code yet.
- It stands out for easy model switching using a single OpenRouter key.
- Pricing (as described):
- a free tier with generous daily token allotments for GLM 5.2 and GLM 5 Turbo,
- paid tiers up to higher monthly costs depending on usage.
Presenters / contributors
- The narrator/presenter is the primary speaker (no specific name given in the subtitles).
- Organizations/contributors mentioned:
- Meta (internal engineering guidance; Llama ecosystem)
- Anthropic (Claude, Claude Science, Fable 5, Claude Sona 5, Claude Code)
- Google DeepMind (Nano Banana 2 Light, Gemini Omni Flash, Gemini Spark, Notebook LM, AI Studio, Gemini chat trail sharing)
- Hermes (mixture-of-agents system)
- XAI (Voice Agent Builder)
- Z.A.I. (Zcode, GLM)
- OpenAI (Codeex and teased hardware)
- Riplet, Notion, Comfy UI, Gamma, OpenClaw, XAI Gro (voice backend)
- OpenRouter (model provider)