AIHOY, AInauts,

Welcome to a new edition of your favorite newsletter!

This week, several agents wanted more responsibility from us at once.

Claude builds a finished video from a single prompt. Meta is putting Muse in glasses, your ears, and your pocket. And at his book launch, Seth Godin asked us an uncomfortable question: Who is actually leading, you or the AI?

That is what today's issue is about. Here is what we have for you:

Let's go!

Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

Since Opus 5.5 launched last week, our feeds have filled with videos that Claude supposedly built entirely on its own. People are genuinely excited.

One prompt in particular is everywhere: paste it in, walk away, and two hours later a finished MP4 is waiting in the folder.

That sounded too good to be true, so we tested it ourselves with an ad for this newsletter.

A few examples that impressed us:

For a deeper look, this thread collects the strongest examples of the week and sums up the moment well.

A reality check is necessary. Not every “one-shot video” is as simple as it sounds. For one of the best-known music videos, Claude ran for twelve hours, with a very long prompt, custom source material, and paid image and video services running in the background.

Claude cannot generate video by itself. It writes code: a browser animation with voice, music, and effects, then renders the result as an MP4.

You do not need an editing app, only Claude Code. Claude can also edit existing footage, for example with the new MCP for DaVinci Resolve.

The simplest cases are the ones in which everything is generated from code. That is what we tested first.

Twenty-four minutes after starting, a 50-second video was sitting in the folder. Claude asked zero questions.

It told a story in nine scenes: 50 new tools every week, no time, 30 open tabs, and then us. It even animated the newsletter signup form at the end.

Claude generated the music and sound effects in code and timed them cleanly. The weakest part was the voice. We deliberately started without an API key, so it used the built-in Mac voice. More on that in a moment.

Then we pushed it further. We gave Claude photos of us and Nauti and asked it to build three versions in parallel: a collage storybook, a cinematic trailer, and a vertical pixel game.

Our favorite is the pixel game. Together with Nauti, we fight the final boss, “The Hype.” You can find the prompt here.

For the audio, Claude compared three German voices through the ElevenLabs API and chose a game-show announcer. The smartest trick came from Claude itself.

AI voices often mispronounce “AInauten” and “Reto.” Claude transcribed its own audio with speech recognition and tested spellings until the names sounded right. “Eh-Ai-Nauten” works, and so does “Räto.” You can use the same trick for any brand name.

Visual note: the interface and voice labels in this screenshot are in German.

Here is a prompt from X that you can use to get started. Launch it in Claude Code, grab a coffee, and see what happens. ElevenLabs is useful for the voice, or simply tell Claude to make the video without narration.

You can also try this slightly longer version of the prompt.

Here is the result:

We have one more, and we think the result turned out really well. What do you think?

Here is the one-shot result:

For a launch video, explainer, or social teaser, you no longer need an agency or editing software for the first draft.

Take the prompts, add your URL or product, and see what comes out. Bonus tip: turn the reasoning level up. Have fun!

DNSMOS gives you a score, not whether that data fits your model. Treat it as pass/fail and you'll train on audio that looks clean but hurts performance, while tossing good data for no reason. Voices' CTO DJ Jalali just published a white paper with the four-step framework the team uses to set internal thresholds instead.

Download the White Paper

“Help us save money,” we told Muse a few days ago. It worked beautifully and saved us about $490.

At Connect, Meta showed us what to spend those savings on: glasses starting at $349, a pocket-sized pendant, and $1,300 VR glasses. 😄

Joking aside, our test showed that Muse can do more than chat. That matters now, because Connect on September 23 was largely about where an agent like this will live.

Meta's answer: not only inside an app on your home screen, but everywhere.

Meta is building four access points:

The agent also gets its own email address, avatars with real-time facial expressions, and new connectors ranging from Notion and GitHub to PayPal and Instacart. The $1,299.99 VR glasses coming in spring 2027 are the extra.

The agent is the main event.

For those of us in Europe, there is one decisive catch: there is no launch date. Muse is currently available only in the US, Canada, and Mexico.

The larger problem affects everyone. An agent that can see and hear, operate your Mac, and pay through PayPal is only as safe as its permissions. Meta has built safeguards for that.

Each user gets a separate virtual cloud machine, and the agent never sees the real login credentials. Meta says a Confidential VM arriving later this year will also prevent Meta itself from technically accessing the data.

Security researcher Patrick Wardle showed how thin the ice can be two weeks after launch. Through an undocumented Mac-app setting, malware could redirect dictation and steal Muse's authentication token. Meta patched the issue.

Meta's strength is distribution. It can put the agent into glasses, a pendant, a Mac, and a huge user base across Instagram, Facebook, and WhatsApp. If Muse becomes reliable enough for everyday work, Zuckerberg has already claimed a broad playing field.

OpenAI will certainly try as well. Dev Day is this week, and rumors say the new personal assistant “o” may launch there. We are watching.

You are not supposed to pay for Muse, at least not directly. It will remain free for most users. Zuckerberg plans to make money through a small fee on every transaction the agent completes for you.

The first major providers are already connected, and the bet is simple: our agents will buy a great deal on our behalf.

That closes the loop with our test. The agent designed to help us save money earns money when we spend it. Meta has not said how large the fee will be or whether merchants or customers will pay it.

We will test voice, glasses, and computer use as soon as we get access. Deep Dive incoming. For now, here is the official highlight reel and the definitely unofficial commentary.

Last week marked the launch of Seth Godin's new book, The Knot: Problems Can Be Solved.

We were there up close, thanks to our reader Eike, who gave us his ticket. Thank you, Eike! ❤️

One lesson stayed with us: technology is often not the biggest obstacle. New tools do not automatically change our habits.

Seth describes status, belonging, and convenience as forces that shape behavior. An agent can work technically and still go unused in everyday life.

The afternoon with Seth opened a second question.

One image has stayed in our minds ever since: the reverse centaur.

A centaur normally has a human head and the strength of a horse.

But when AI does the thinking and the person merely supplies the legs, things become problematic.

As a warning, Seth points to Amazon drivers whose every move is timed by a system.

That concerns us more than the usual question of which jobs AI will replace. AI is still often sold as a cost-cutting program: fewer jobs, lower expenses.

Saving time and money is useful, but it does not tell us who benefits. If that is the whole goal, we have missed the most interesting part.

Language note: the following clip is dubbed in German using HeyGen; it is not Seth's voice.

In “The Real AI”, Seth writes about what remains ours: good questions, judgment, taste, and responsibility for the result.

We should invest the time AI frees up in exactly those things.

Bad work remains bad work, whether a human or an AI created it.

Let AI work. But do not hand over the decision about who the work serves and why.

If you imagined the race toward AGI as a sober benchmark table, how about a Mad Max-style desert race instead?

Seedance 2.5 shines, and almost four million people have already watched. We do not know who will win the race either, but this trailer makes us want to see what comes next.

That is it. See you in the next issue.

Reto & Fabian
AInauten Team

X · LinkedIn · Facebook · Instagram · YouTube · TikTok