AI-HOY, AInauts,

Welcome to a new edition of your favorite newsletter!

It is almost like Christmas: every day, another AI feature appears in the advent calendar 😁.

ChatGPT now lives inside Word, Excel, Google Sheets, and PowerPoint, taps into your data, looks at your screen through Appshots, and allows browser extensions in the desktop app.

Claude is merging Chat and Cowork, making Projects easier to navigate, integrating /design, launching Docs, Slides, and Sheets, and continuing to work even when your laptop is closed.

All cool. We could turn this into a “look what AI can do now” newsletter.

But honestly, a few other developments matter more to us than AI feature bingo.

Here is what we have for you today:

Let's go!

Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:

Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.

Download the Free Report

Last week, we tested a new AI assistant that saved us $490 in real life.

But the task that stayed with us longest was far less spectacular: “Find out when the Governors Island Pumpkin Festival takes place.”

“Sorry, the date has not been announced yet.”

A normal chatbot tells you exactly that. If you are lucky, it adds a link and calls it a day.

But the new personal assistants we are looking at today, namely Instinct and Meta Muse, keep working on the task.

This is an early real-world test. And it shows that the point is not a better answer inside a chat. The point is that something still happens after the chat is over.

We had long forgotten about the Pumpkin Festival when, two days later, a casual notification arrived: the date was now confirmed and had already been added to our calendar. Nice.

Note: Half of the AInauten team lives in the Big Apple, which lets us test products such as Meta Muse and Instinct a little earlier in everyday life. Instinct also works outside the US, and Meta is rolling Muse out to more countries.

The setup takes only a few minutes: connect email and calendar, click through a few permissions, and you are done. That speed matters if these agents are supposed to reach a mass audience.

We deliberately made our first task uncomfortable: “Help us save money.”

After reviewing our emails, the assistant found that we could reduce our Google phone plan by $360 per year.

It found another $130 in savings on our internet service. In that case, the agent contacted customer support directly and negotiated on our behalf.

Then it moved on to Amazon: reviewing orders, checking potential refunds, and following up on missing credits.

Let's be honest: these are exactly the kinds of tasks that always get postponed. Here, they were handled in the background, and we only had to help briefly with login.

We put both personal assistants through the same everyday tasks.

Muse understood our taste in movies better: indie documentary instead of a Hollywood horror flick. Instinct found the perfect Airbnb and did a much more precise job checking whether our dentist accepted our new health insurance.

We also appreciated that neither assistant flooded us with meaningless updates. For an agent with an ongoing assignment, that is not something you can take for granted.

Instinct is currently free through this invite. Muse includes a free allowance and an upgrade option (bonus-points code: B2FG35).

If you want a more systematic comparison, the Assistant Benchmark lists more than 100 assistants and compares Muse, Instinct, and others directly, using concrete tasks and observed behavior rather than feature lists.

OpenAI is also expected to launch its personal assistant this week.

New features keep appearing to make these tools even more useful.

Instinct Files turns comparisons, plans, and explanations into interactive pages that can remain private or be shared. Muse is now available as a Mac app and can use meeting notes from Granola and documents from Notion as context. More connectors are coming, and Meta has just opened the platform to developers.

That is a big deal. Developers can connect APIs to the Muse agent, its browser, and the context users have approved. What used to be a closed assistant is becoming a platform.

Phone calls are also entering limited rollouts. Instinct Concierge is designed to handle more demanding booking and service cases by phone. Muse is expanding a beta for outbound calls to US businesses.

And if necessary, one assistant can simply hire the other 😁.

When a task combines research, waiting, follow-up, and calendar context, Muse and Instinct are already genuinely useful. We no longer have to remember every next step because the agents do it for us, unless they are busy talking to each other on Musebook.

The price is access to your private data, and that is not a small issue.

According to its Privacy Policy, Instinct can access connected apps, messages, emails, documents, sensitive information, and credentials, store that information indefinitely, and use it to train its AI models. For many people, that is an absolute deal-breaker.

Meta promises more privacy with Muse: You choose apps and permissions, can disconnect integrations, and can opt out of training use.

In practice, that means enabling the training opt-out before connecting any apps, limiting permissions, and avoiding unnecessary concentration of sensitive accounts inside a single agent.

It is the familiar trade-off: If something is free, you are the product. The more an assistant knows about you and the more it is allowed to do, the more useful it becomes, and the more expensive its mistakes can be.

The next wave of millionaires will be people who figured out how to make AI work for them.

The window to get ahead is still open. But not for long.

Here are 200+ proven ways to make money with AI in 2026.

Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.

Claim your free list

At first, we kept scrolling on X. Another new model. But the next day, our feed was full of it: a dozen posts in a row, no exaggeration, and everyone was excited about Jev.

What do we do in situations like that? Exactly: We dive into the rabbit hole 🕳️🐰 and show you why this kind of AI is extremely practical, even though it cannot chat with you.

Let's start with the number that grabs attention: $0.042 per million input tokens, with output currently free. That is not a typo. For comparison, OpenAI's frontier model GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens.

At $0.042 per million input tokens, Jev costs roughly 238 times less than GPT-6 Astra's $10 input price. And while Astra charges separately for its answer, Jev's output is free.

Jev comes from TypeSafe. Founder Diogo Almeida worked on ChatGPT at OpenAI and spent the past two years building the system with his team in stealth. The launch made waves accordingly.

What makes Jev different is easy to explain. What it means usually takes a few rounds of thinking and doomscrolling on X or browsing Jevable 😁.

A normal language model writes text. Jev receives a case and a predefined set of possible answers. It chooses one and reports how confident it is.

Think of a nightclub bouncer. A regular language model such as GPT is the person who writes the guest list, in detail and with plenty of adjectives. Jev stands at the door and does one thing: in, out, or a shrug when the decision is close. The guest-list writer can decide what happens after the shrug.

In other words, Jev does one job extremely well, extremely fast, and extremely cheaply.

If you want your preferred agent to try it, Jev is available through OpenRouter, Vercel, and the Python SDK. Here is the prompt:

We naturally wanted to test it, so we built a small AInauten headline generator. First, we used headlines from our own newsletter as a data set, together with additional headline swipe files.

A conventional language model uses that material to generate suitable headlines in the AInauten style.

Jev's job is to score those options against a rubric: Is the headline clear? Does it create curiosity? Does it look spammy?

Imagine an editorial meeting. The LLM is the copywriter producing ideas. Jev is the managing editor who sorts, scores, approves, or rejects them.

A normal language model can score headlines too. Jev's advantage is that it can only return answers from the rubric defined in advance. No creative detour and no extended chain of thought. That makes it essential to design a rubric that actually fits your situation.

Most of us entered the AI world through classic language models during the past three years. But that is only one part of the AI universe.

Jev is a System 1 model, a specialized decision model that complements LLMs. It is interesting because almost every AI workflow contains dozens of tiny decisions: route a support ticket, score a lead, or check a text before it is sent. Today, most of us simply use a large language model for all of them.

Jev does not replace a language model. That is exactly what makes it interesting. Over the past few years, we have handed almost every AI task to a chat model: writing, evaluating, sorting, checking, and deciding. It works, but it is about as efficient as renting a moving truck for every grocery run 😉.

We believe AI workflows will become more specialized. Language models will handle work that requires creativity, context, and open-ended answers.

Models such as Jev will handle the many small decisions in between. That can make workflows faster, cheaper, and easier to control.

Whether Jev permanently owns this role remains to be seen. We expect OpenAI and others to launch similar models. But the direction makes sense to us: not one giant model for everything, but a team of specialists.

We are thinking more often about the bigger questions behind AI. Last week, the subject finally reached the mainstream, and suddenly everyone had an opinion about what our future will look like.

With so many voices, it can be difficult to form clear thoughts of your own. So we looked for ideas and mental frameworks that can guide us through the noise.

Our first anchor is the Five Whys method: ask “why?” five times in a row. Another is first-principles thinking. When someone says “AI leads to X,” we ask: What exactly do you mean? What data shows the change? What observable signal would make us revise our view?

That turns a huge prediction into a claim that can actually be tested.

We removed the German-language transformation map from this English edition because a translated version was not available.

The AI labs also have to confront these questions, and they are approaching them in very different ways.

The new DeepMind Institute explicitly focuses on the societal consequences of AGI. Anthropic brings similar work together in the Anthropic Institute, whose agenda covers work, the economy, safety, and the behavior of AI systems.

OpenAI publishes usage data and research through Signals and its own public policy agenda. Meta publishes research on responsible AI.

DeepMind's new essay on economic policy for AGI, for example, connects prepared policy responses to observable triggers instead of predicting a date for “the big disruption.” We find that more useful and concrete than most debates about the future.

We do not have to agree on a single future. The goal is to recognize the relevant factors, understand how they interact, and see which possible futures follow from them.

For our own orientation, we find the major labs' work surprisingly useful right now. It is not as exciting as a new model or feature, but it is probably much more important.

The only problem is that much of it reads like an economics textbook.

That is why we collected the most important texts from DeepMind, Anthropic, and OpenAI in a GeminiLM notebook.

If you want to go deeper, you can use it to create explainer videos or simple maps. Language note: The linked example video is in German. Here is our prompt:

We also removed the second German-language map because no translated version was available.

P.S. Gemini Notebook still makes plenty of spelling mistakes. Come on, Google. Just drop your graphic into ChatGPT and say: “Correct all spelling mistakes.”

You made it! Thanks for staying with us all the way to the end. See you in the next issue.

Reto & Fabian
AInauten Team

X · LinkedIn · Facebook · Instagram · YouTube · TikTok