Whoever wins AI wins.
That’s the US president, calling in live to Jensen Huang on stage at the All-In Summit, and it sums up the split that defined AI this week. I caught the whole story in the latest Friday news roundup from the creator behind Future Tools, who laid out the timeline better than anyone else I’ve seen. Here’s why it matters.
Last week, people leaving Anthropic and OpenAI went public with worries about where AI is heading. This week was the response. Dario Amodei, Anthropic’s CEO, published an essay called “We Must Pace the Frontier,” and within days almost every big lab head lined up behind it. Elon Musk said “Dario is right.” Sam Altman, who rarely agrees with Dario, said OpenAI had been discussing the same thing for weeks. Demis Hassabis called it the right direction. Satya Nadella said Microsoft welcomes the idea. Then Mark Zuckerberg pushed back: stop asking for regulation, just do the right thing. He pointed out Meta delayed shipping Muse for months over safety without telling anyone else to. And the president called the whole thing a hoax and a conspiracy that only helps China.
The key idea in Dario’s essay is a three-part plan, and the creator breaks it down simply:
- Embedded evaluators. Third-party people with desks, badges, and laptops inside the labs, working like employees but answering to someone else. Anthropic says it’s doing this no matter what.
- US regulation covering every frontier lab, including the ones that won’t cooperate voluntarily. Think chip export limits, cracking down on distillation (training a new model on a bigger model’s outputs), and protecting model weights from theft.
- Global pacing with China. Agreements on things like no bioweapons, testing each other’s models before release, and a speed limit on recursive self-improvement.
The creator’s own take is measured. He thinks labs should spend more time and money on alignment over profit, isn’t as scared of doom scenarios as some people, but says nobody should dismiss them either. I found that a refreshingly calm position in a very loud week.
The rest of the video is product news, with a running theme: companies announcing features that most people can’t actually use yet. If you want to try something this week, here’s where I’d start:
- Check your Claude app. Cowork and chat are merging, and Projects got a redesign where each thread runs as its own Claude Code cloud session on its own branch. Both are rolling out slowly to Pro and Max.
- Try the stealth model “Union Alpha” in OpenRouter or OpenCode. It’s free for a week and turned out to be a router system from Unbiased that hands tasks to different models. In the creator’s test it finished in under 5 minutes using less than a third of the tokens GPT-6 Astra needed for a near-identical result.
- Open Gemini Notebook and talk to your notes out loud. An audio recorder for lectures is coming to mobile next.
- If you’re a developer, hook the ElevenLabs MCP into your agent. It now covers voice, music, image, and video generation.
Quick hits worth knowing: Qwen3.8 Omni Flash landed with a 1M token context. Google shipped Gemini 3.8 Live for dialogue and a family agent called CC. Meta launched a Meta One subscription starting at $2.99 a month, and Muse can now make outbound calls to US businesses. Grok Build got memory. ChatGPT ads now open a chat with the advertiser. And UBTECH opened a factory that builds a humanoid robot every 10 minutes, with humanoids helping on the line. 🤖
That last one still makes my head spin. Watch the full video for the creator’s complete breakdown, including the Trump call highlights and his very relatable rant about “available now” features that aren’t.