In partnership with
Hey AInauts,
Welcome to the latest edition of your favorite newsletter.
Anthropic wants to go public. According to its prospectus, the company generated $16.3 billion in revenue in the first half of the year. A year earlier, it was just over $1.1 billion for the same period.
Now that's scaling. The prospectus, which is naturally meant to make the stock appealing, values knowledge work worldwide at more than $40 trillion.
Too many zeros for us. But those are exactly the people Anthropic wants to make far more productive with AI.
We've been busy testing again this week, doing our bit to help Anthropic and the other AI labs make you more productive. 😉
Here's what we've got for you:
🤝 How to Hire Your AI Employee Team
🧭 What Becomes Scarce? The Job Check
🧪 NEW: Vibe Check. GPT-6.1 Terra and Kimi K3.1 Pro Put to the Test
Let's get into it!
Your AI prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning (constraints, edge cases, examples, tone) and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
Used by teams at Microsoft, Amazon, and Notion. Try Wispr Flow free on Mac, Windows, iPhone, and Android.
🤝 How to Hire Your AI Employee Team
At the end of last week's issue, we suggested trying Grok Bot to see where working with AI is heading.
That brought in a few questions. So today we're getting straight into why Grok Bot feels so magical, and how quickly you can hire and build your own AI employees.
We'll show you what we built, then how to get started.
The idea: not one bot for everything, but a small team with distinct roles. The way we'd hire actual people.
Our main bot acts as a Chief of Staff (earlier explanation in German). We gave it the idea.
Right at the start, it said it would approach this "almost exactly like hiring real employees." Each bot gets its own chat, role, and memory.
Visual note: the screenshots in this segment show our real German-language chats and drafts. The surrounding English text explains the workflow.
Chief of Staff: approaching bot creation like hiring real employees.
The Job Interview
For our test, we wanted a Growth Manager for AInauten to handle things we don't have time for.
Before the Chief of Staff created it, five questions came back:
Permissions: can it adjust ads itself, or only draft proposals? Who is the human owner? Along the way, it found a contradiction between two of our own rules files. Thanks.
Budget and goals: how much can it spend, and which metrics matter?
Access: beehiiv, Meta, Google Ads, Drive?
Which offer is current? Our live page still showed a version missing from our offer document. Oops.
Attribution: how do we measure what comes from the newsletter?
Five decisions before creating the Growth Manager: permissions, budget, access, current offer, attribution.
It's the same process you'd use with a human employee: build context, define a mission, provide access, and delegate tasks.
Then They Talk to Each Other 🤯
The Chief of Staff then hired a Content Bot.
It read our style guide, analyzed our previous LinkedIn and X posts, and sent an editorial calendar running through Black Friday. The first drafts went straight into our publishing tool.
The Growth Manager and Content Bot exchange messages and coordinate.
A LinkedIn carousel draft created by the Content Bot.
The great thing about Grok Bot: all the bots can talk to one another. You can speak to them individually, and they coordinate.
It's impressive how smoothly that works.
The Growth Manager prepares things, but nothing goes live before launch. The Content Bot makes sure no post reveals the offer early.
Both bots know the launch restrictions. Nothing goes live early.
Before long, we had post drafts, Meta ad proposals, and draft launch pages. Everything in draft. Nothing live.
Honestly, the whole thing was up and running in an hour. We do have a big advantage: we've built an extensive second brain for AInauten.
We can tell Claude, "Put together a package containing all our knowledge," upload it straight to GitHub, and give Grok Bot access.
Pretty magical. And fairly straightforward.
How to Hire Your First Team
Grok Bot is available for Mac, Windows, iPhone, and Android, starting at $30 a month. Each bot gets a job, works around the clock on a cloud computer, and signs into your tools.
If there's enough interest, we can put together a full walkthrough with a video as a Deep Dive.
To get started, try this prompt:
You're my Chief of Staff. Before we create any bots, conduct a brief hiring interview with me.
Ask one question at a time, no more than five: What do I do for work, and what needs to be finished in the next four weeks? Which tools do I use every day? Which task eats the most time each week? What must a bot never do? When and how do I want to hear from you?
Then propose two bots. For each: a name, a one-sentence job description, only the tools it actually needs, a recurring routine with a set time, and one line beginning "Never: ..."
Before creating anything, clarify for each bot: Who is the human owner? What is the goal, and how do we measure it? What access does it get? Can it only prepare drafts, or can it act independently?
Show me what each bot would do on its first run. Create the first one only after I approve. For the first two weeks, every bot works in draft-only mode.Three things we learned:
Start small. One bot. Split the work only when it gets unwieldy.
Use a probation period. Drafts only for the first few weeks, with approval for everything.
Watch the shared computer. All your bots share one cloud computer, including all its logins. The separation comes from roles and approvals, not isolated machines. Monday's rule applies here too: grant only the access that's needed.
Our Take: Almost Everyone Will Need an AI Employee Team
We firmly believe we'll all hire our own teams through AI. In our view, Grok Bot currently does the best job of putting that into practice.
We'll soon be managers of our agents. We need to get better at leading, delegating, and holding our helpers accountable.
The bots do the work. The decisions and the setup are our responsibility.
Which brings us to the next question...
A fast AI answer isn't always a coordinated one. When AI resolves B2B issues without looping in the teams responsible for them, confidence outpaces alignment. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, looks at closing that gap.
🧭 What Becomes Scarce? The Job Check
We've had a FAZ article sitting in our backlog since July: Which Jobs Have a Future? Language note: the article is in German.
Going through our topic backlog this week, we thought it would fit nicely after Grok Bot.
So let's talk about the future, jobs, and AI again.
The Sandwich
Computer scientists Arvind Narayanan and Sayash Kapoor describe knowledge work as a sandwich:
Deciding: what are we doing, for whom, and toward what goal?
Executing: researching, writing, calculating, building.
Delivering: checking the work, taking responsibility, and answering to the client.
AI eats the middle.
Visual note: the German illustration below labels the layers "Deciding," "Executing," and "Delivering." The robots are eating the execution layer.
AI eats the middle: deciding, executing, delivering.
That's exactly what we saw with our bots. Execution? They handle it. Decisions and approval? They can do those too, but those should stay with us.
Approval gets harder when we give bots too much freedom, because we no longer see much of what they're doing.
More on that soon.
What Becomes Scarce?
Economist Alex Imas asks a different question from "Which job disappears?": What becomes scarce when execution costs almost nothing?
Judgment. Responsibility. Trust. And something we've written about often: taste (earlier article in German).
A lot is happening around jobs and AI. We're hearing about huge numbers of layoffs, some of them driven by AI.
There are encouraging developments that are easy to miss, too:
More work, not less. Years ago, AI pioneer Geoffrey Hinton predicted the end of radiologists. Today, more are working than ever. When something gets cheaper, people want more of it. We've even run our own MRI images through AI and found its diagnostic output impressive. That personal impression is not clinical validation.
Human work becomes more valuable. As the economy becomes more productive, pay also rises for work AI can't do: haircuts, care, and hands-on work with people.
Solopreneurs are booming. Stripe's data shows that the number of solo founders generating more than $1 million in revenue doubled in 2025. Above $5 million, it tripled.
That last point is the most interesting to us. It's what working as one person with an AI team looks like.
Those of us engaging with this now have a head start. Put your head down and get to work, and something great can come of it.
We already see it in our own businesses. What the two of us handle today would have been impossible two years ago.
Run the Check
We've built a prompt to give you an honest assessment. It works in ChatGPT, Claude, or Gemini:
I want an honest assessment of which parts of my work AI can already take over today, and what becomes more valuable as a result.
First interview me about my typical workweek. Ask no more than five questions, one at a time.
Then sort my tasks into three layers: Deciding (what I define, prioritize, and take responsibility for), Executing (what I research, write, calculate, and prepare), and Delivering (what I check, present, negotiate, and resolve with people).
For each task: How much can AI realistically take over today: none, some, or most? With which tool? What remains my responsibility?
Finish with:
1. Three tasks I should delegate starting next week.
2. Two of my skills that become scarcer and more valuable because of AI.
3. One uncomfortable truth about my role that I probably don't want to hear.
Be honest, not nice. No predictions tied to specific years.Our Take: AI Is Taking Over More of the Execution
The job doesn't disappear. The middle does.
If all you do is execute, your work gets cheaper. People who can decide, brief, and approve are needed, whether they're managing humans or bots.
Honestly, we notice how unfamiliar that feels ourselves. Making decisions all day is more exhausting than working through tasks all day.
But that's the part that stays.
🧪 Vibe Check: What Can the New Language Models Do?
Let's finish with something new.
It feels like a new model arrives every week. Each time, we get the same question: Do I need to try this?
Two more models were introduced this week.
So we're launching a recurring format: Vibe Check. We test new models against a fixed set of tasks and give you the short version: what can it do, when should you use it, and do you need it at all?
We use the same five everyday tasks as in our major model comparison two weeks ago:
Expenses
A contract
Store figures
Trick questions
A LinkedIn post
Other models evaluate the results blind, so you can compare the numbers from week to week.
Visual note: the source chart below is in German. It compares all five tasks as of October 28, 2026: Terra, 5 cents and 1.9 minutes, all correct; Kimi, 24 cents and 12.2 minutes, one error; Opus 5.5 reference, 45 cents and 3.4 minutes, two errors.
Source benchmark chart, October 28, 2026. Costs and timing are explained in English above.
👍 GPT-6.1 Terra (OpenAI)
In one sentence: OpenAI's new workhorse for agents, built to click through programs and websites. And quite inexpensive.
Our test: as precise as the major flagships. Terra found the calculation error in the contract in all three runs. Opus missed it twice last time.
Running all five tasks once costs 5 cents with Terra and takes under two minutes. With Opus 5.5, it's 45 cents and just over three minutes.
Terra's writing is solid but matter-of-fact. Opus and Astra sound more human.
Use it for frequent, automated work: analysis, checks, summaries, and agent routines.
Skip it for writing that needs to sound like you.
Clicking: OpenAI's models are currently the strongest in our view for computer use.
👎 Kimi K3.1 Pro (Moonshot)
On Monday, Kimi complained during our coffee break that nobody could find it. Now we've found it. And tested it.
In one sentence: the new flagship from China costs half as much as Opus. Its weights are expected to become public on November 4, allowing other providers to host it.
First, the hype: there are allegations that Moonshot paid influencers without advertising disclosures. Those claims haven't been confirmed. We don't find it hugely dramatic. Except that, sadly, they didn't pay us. That wouldn't have changed our verdict, though. 😉
Our test: capable, with almost everything correct and one missed calculation error. But slow. All five tasks took more than twelve minutes, six times as long as Terra. Two runs hung for half an hour without returning an answer.
For writing, Kimi trails Terra. Solid, but nothing worth switching for.
The catch: Kimi currently runs only through Moonshot itself. Your data therefore goes to a Chinese provider, whose responses also reflect a Chinese perspective.
Revisit later: once the weights are open and European providers host Kimi, we'll take another look.
Until then: a solid model, but you can skip it.
That's it! As always, thanks for reading.
What do you think of Vibe Check? Want more detail, or is it not relevant to you? Just reply to this email. We read everything.
Reto & Fabian from AInauten
