Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

Three warning lights are flashing on the AI dashboard today: OpenAI's new superapp wants to get closer to your daily work. Claude is practically writing its own changelogs — and management thinks it's time to tap the brakes (but can't quite bring itself to…).

Meanwhile, a simple infographic project reminded us that the best AI workflow can be surprisingly straightforward — and we're sharing exactly how we did it.

Here's what we've packed for you today:

Let's go!

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.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

A few weeks ago we boldly declared: The AI superapp is here! From our perspective that was absolutely true — we were talking about OpenAI's Codex app.

It has quickly evolved from a pure coding tool into a general-purpose work surface. The power-user toy for hobby developers is turning into a full-blown product for the desk-bound office crowd, pardon: the knowledge-worker contingent. We often work in tandem with Claude Code/Cowork and Codex.

"Chat is dead," one OpenAI employee said matter-of-factly. It's a headline-grabber, but the reasoning holds up: chat gives you answers, Codex and agents get work done.

After months of speculation, OpenAI has now officially confirmed that a new ChatGPT super-app is coming — one that brings everything together under one roof.

The logic is straightforward: an app that combines agents, coding tools, and third-party integrations pulls in more paying customers than a classic chatbot ever could. (We can vouch for that from our own usage: we used to pay $20 for ChatGPT, now we're maxing out the $200 Pro plan so the agents can put in the hours for us.)

The rollout is set to happen over the coming weeks across desktop, mobile, and browser — with deeper integration into the broader tool ecosystem.

This is the spring cleaning every major AI lab is attempting right now: you jump between tabs less, and the agent holds the whole workflow together instead.

Since the desktop app launch in February, Codex has grown to more than 5 million weekly users — and the majority of Codex users are paying subscribers (which means: yes, you can install Codex and try it out on the free plan — highly recommended).

At the same time, 2 million business customers account for around 40 percent of OpenAI's revenue — and that share is growing. That makes it clear why Codex has suddenly moved out of the nerd corner and gone front-and-center. It's the part customers are actually willing to pay for.

OpenAI is shifting resources toward enterprise and Codex, with no small amount of pressure coming from Anthropic, who are carving out an ever-larger slice of the business pie.

ChatGPT is becoming the front door — nudging users further toward agents, coding, and partner apps like Canva and friends.

We used to debate which model was best. Now you just say: "Prep the client meeting." At its best, the agent pulls together Slack, documents, CRM, calendar, past decisions, and open tickets.

What you get back: a briefing hub as a Codex Site, a clean slide deck in your corporate design, open questions formatted as talking points, and a handful of high-level decision items where you still need to weigh in.

There are also strong updates dropping almost daily — here's the quick rundown of what's new:

A new clip from OpenAI is hyping Codex hard. We weren't entirely sure at first whether we'd accidentally clicked on the trailer for a new Hollywood blockbuster…

Sounds great, right? It is. But there's one catch you need to be aware of. For OpenAI to work in the background on your behalf, it needs access: files, apps, memory, workspace data, permissions, local machine states, ...

That's convenient. But the sharpest edge in everyday AI is also the next data vacuum cleaner.

For many users and teams, that's completely fine. Especially with Business, Enterprise, or AWS integrations, governance can be handled more cleanly. For certain professions and people dealing with sensitive data, the question remains: What context should actually be allowed to reach the OpenAI layer?

Google is no privacy saint and has exactly the same ambitions when it comes to building context. But they also offer an alternative for more local control, less platform lock-in, and less data leakage — for example, the new Gemma 4 12B. It's an open, multimodal model designed to run on laptops with 16 GB of RAM and can handle text, images, audio, and video (see also our offline setup guide here).

OpenAI is becoming extremely useful right now because the individual features are converging. ChatGPT becomes the interface. Codex becomes the worker. Memory becomes the background layer. Apps are the toolbox, and Sites will soon be the review surface.

Use it. Just don't marry it blindly. Our Files-over-Tools deep dive gets into exactly this: get your data out of the chats!

Because as long as your project brief, your style, your rules, your decisions, and your prompts live in your own files on your own machine, you stay flexible.

You can use OpenAI while OpenAI is leading the pack. Or you can switch when a local model, Claude, Gemini, Gemma, or some other tool does your next job better.

A local model gives you control — but not the polished OpenAI integration you get from ChatGPT, Codex, apps, Workspace Agents, AWS, and Memory.

What works best for you is yours to decide.

If you only have 10 minutes today, run this mini-audit: Where does the context live that your AI work can't function without?

If the answer is "in ChatGPT," pull out the most important pieces: project briefing, style rules, client boundaries, approval rules, prompts, examples. For that, you can use the /handoff prompt from here.

When you own your data, the superapp becomes a powerful tool without turning into a closed operating system. Because nobody switches operating systems for fun…

Picture a chessboard. You place one grain of rice on the first square. Two on the second. Four on the third. Then eight, sixteen, thirty-two. You know how this goes.

At first it seems harmless. After ten squares you have 512 grains. After twenty, just over half a million. Still imaginable. Then things get out of hand fast. By square 64 you're looking at around 18 quintillion grains.

And that's exactly what we wanted to illustrate: What's about to happen in AI is beyond comprehension!

Because what happens when AI completes tasks while simultaneously improving the process used to build the next AI? We don't know…

Anthropic laid out exactly this point in a new post titled "When AI builds itself". It's worth reading for yourself rather than just getting it secondhand.

According to Anthropic, Claude now internally writes more than 80 percent of the code that gets merged into its own codebase. On one optimization task, Claude's code reportedly sped up human solutions by a factor of 52.

You could also put it this way: Claude largely writes itself. And it can work autonomously for longer and longer stretches — hours, days, soon weeks…

The machine isn't quite building itself fully autonomously yet. But it's not far off. And that's where the difference between a tool and a loop becomes clear.

Anthropic isn't alone in worrying about this. OpenAI has itself flagged early signs of recursive improvement in its Frontier Safety Blueprint, and Google DeepMind has already put AlphaEvolve to work automatically optimizing data centers, chip design, and AI training — exactly the places where better algorithms make the next AI generation more efficient.

A tool makes work faster. A loop makes the next version of the tool better. When that loop becomes stable, researchers call it "Recursive Self-Improvement": AI helping to design, train, test, and improve its own successors. A loop is a loop is a loop …

Yes, that sounds abstract and academic. In practice, though, it gets concrete fast.

A team suddenly has more content, code, experiments, tools, analyses, insights, and ideas than people alone can cleanly process.

And at some point right now, the scarce resource is no longer coming up with ideas — it's making decisions. Feels like our own little microcosm is already wrestling with exactly this problem…

Anthropic sketches out three scenarios.

1. The Smooth Accelerator: AI makes research faster, but humans stay in control of goals, testing, and decisions. Probably the most comfortable version of the future.

2. The machine loop: AI is automating ever-larger chunks of AI development itself. Better systems build better research tools, and those tools build even better systems.

3. Reality pumps the brakes: Compute, energy, data, organizational complexity, safety, and human judgment remain real bottlenecks. Progress stays fast — just not explosive.

We don't need to know today which scenario plays out. Even the mildest version reshapes the economy in a big way. And the alternatives are too important to only start taking seriously once they become obvious.

Development is moving so fast right now that Anthropic is genuinely saying it would be good to hit the brakes — and is floating the idea of a global development pause to make that happen.

The idea makes sense: if frontier models start showing capabilities nobody can reliably control anymore, the world should be able to hit the brakes. The problem is, a pause like that doesn't work like a meeting resolution…

Let's be honest: AI stopped being "just technology" a long time ago. It's big business, national security, science, military, prestige, and infrastructure all at once. Anthropic knows this as well as anyone.

This only works if the major labs lead the way, if governments follow, if the US, Europe and China share enough trust, verification, and self-interest to not quietly keep racing ahead. LOL… Given today's geopolitical lineup, we're not holding our breath.

The best example was the 2023 pause petition: despite some big names attached, practically nothing happened. And now the stakes are even higher.

Do you remember the "AI 2027" forecast that made waves back in April 2025?

It described exactly this scenario: first AI agents help with coding, then they help with AI research, then a race kicks off where companies and governments have less and less time to actually understand their own systems. Sounds a lot like what's happening right now.

Important caveat: this isn't a guaranteed prophecy. We don't know whether it'll play out this way. But we strongly recommend you set aside half an hour to watch this video.

The "AI 2027" predictions have since become a useful benchmark. Experts are debating whether real-world signals are moving faster or slower than expected.

In the debate around AI's automatic self-improvement, AI doomers Eliezer Yudkowsky and Nate Soares also deserve a seat at the table.

Their book "If Anyone Builds It, Everyone Dies" is a genuinely important contribution — and it stakes out the hardest possible position: if anyone builds a truly superhuman AI using today's methods, it ends in catastrophe. We covered it here.

We're not just going to take that certainty at face value — the world is too complex for that.

But the core question does linger and keeps us up in quiet moments: Why do we assume that an ever-smarter, faster, more strategic system will permanently stay within the boundaries humans set for it? And what if humans deliberately dismantle those boundaries?

We don't have a satisfying answer … and this video makes it painfully clear just how much could go wrong. Ughhh…

We're not writing this to cause panic, or because we pretend to know what the next two years will look like. We don't.

Maybe everything stays slower than AI 2027 predicts. Maybe technical limits kick in. Maybe safety methods improve. But maybe we're also underestimating how fast small gains stack up. Maybe, maybe, maybe…

What we can say with reasonable confidence: development won't slow down just because our gut prefers linear curves. And it won't become safe just because the tools are useful. If anything, it's the opposite.

All we can do is stay true to ourselves and our creed: keep an open mind, experiment constantly, rebuild our own workflows, stress-test results hard, take political and social questions seriously. And through all the optimism about progress, hold onto a healthy dose of respect.

Our job is not to wait until square 60 before realizing the chessboard is almost full…

Stefan flagged something annoying last week: our NotebookLM infographics look great overall, but individual labels sometimes have minor text errors.

His tip: ban "pixel errors" in the prompt. It helped a little, but didn't fully solve the problem.

Until now our approach was to provide the actual text for the NotebookLM graphic together with the design concept upfront (we use the unofficial NotebookLM CLI / MCP for that). But clearly that wasn't the ultimate solution either…

Our new workaround: NotebookLM still builds the graphic (we like the style). Then we run the finished PNG through GPT Images 2.0 and have it fix the text and remove the NotebookLM logo. Et voilà, works like a charm!

Here is the prompt:

By the way: GPT-Image-2 is still sitting at #1 in the Text-to-Image Arena. And that's despite fresh competition: Reve 2.0 debuted straight at #2, Microsoft's MAI-Image-2.5 is currently at #4, and the new Ideogram 4 has also landed in the top 10 as an open design model. The image generation space is getting seriously shaken up again.

Tonight we'll also build the skill properly so the agents can take it over directly. Automation and all that. Sometimes it's the small hacks that give you the most satisfaction 😄...

And that's a wrap! See you in the next issue.

Reto & Fabian