Ahoy, AInauts!

Welcome to the latest edition of your favorite newsletter!

ChatGPT's newest feature takes on your work when you ask. OpenAI's models broke into Hugging Face during an internal test.

And while we're still learning how to use these systems properly and build our AI employees, Sam Altman and Elon Musk are already debating a world where large parts of work become optional.

If you feel as though you somehow missed two years of AI news: no. That was just this week. 😁

Here's what we have 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

Until now, AI had to wait for you to prepare the work: copy the text, take a screenshot, explain the context, and write the prompt.

With ChatGPT Voice on desktop and Screen Context, much of that friction disappears.

You leave the email or spreadsheet open and say: "Take a look at this and tell me what the next steps are."

The desktop app can see what you're talking about. Voice can turn that into tasks for Work or Codex, check running jobs, and send additional instructions to them.

A year ago, the new Voice Mode still felt like a scene from Her. Now we're getting the next movie moment: "Jarvis, take care of it." Hollywood has always been pretty good at selling us the future. 😁

It sounds fantastic. In practice, though, it only works if you change your habits too.

So we don't simply want to jump aboard the hype train. We want to give you the operating rules that almost never appear in the polished demos.

You can't simply open an existing text chat and switch to Voice halfway through the conversation. A chat has to begin in Voice Mode, and only one Voice chat can be active in the desktop app at a time.

That is more of an advantage than a limitation. Instead of ten talking chats, you have one assistant where thoughts, problems, and new tasks come together.

You no longer have to turn a half-formed thought into a polished prompt. You just talk, the assistant asks follow-up questions, and it sends the assignment to the right place.

The important distinction is the microphone icon in a normal chat. The microphone creates a dictated message that you can review before sending. Voice stays open as a conversation and delegates work.

Start with something like:

"Where are we? What should we do next?"

Then delegate:

"Start a separate thread for that. Use the appropriate model and only come back here if you need a decision."

Voice runs the conversation. The work itself uses the tools, permissions, and model in the selected Work or Codex environment.

The new ChatGPT Voice feature is not a full Jarvis yet, and it still has a few rough edges. Sometimes it drops assignments. Sometimes "I'll check that" is followed mainly by very committed silence.

In a private office, at home, or while walking with your iPhone, the feature is already good enough to challenge the old way of working.

In an open-plan office, not so much. Your coworkers probably don't need to overhear every conversation you have with your computer.

Once you stop jumping between windows, you'll quickly understand why "Jarvis" has become a much more realistic idea. But Voice is not arriving alone.

The announced superapp is taking shape:

Voice receives the assignment. Work and Codex carry it out. Apps provide context, Skills know the process, and Tasks handle repetition.

The keyboard is not losing its job. But it is losing its monopoly.

If you use Voice merely as a more convenient dictation feature, you miss the real leap.

Start one central Voice chat, give it a real assignment, and let it coordinate the right specialists.

We prepared a practical playbook with a starting instruction, delegation prompts, limits, and common pitfalls.

Language note: The playbook is currently available in German.

Subscribe to The AI Report, the free 5-minute daily AI brief for 400,000+ business leaders, and you’ll get 37 Claude prompts free in your welcome email. They’re organised by the 8 situations every manager faces. You get both: the newsletter and the prompts.

Subscribe & Get the Prompts

Imagine asking your agent to prepare a market analysis. Instead of researching, it breaks into a competitor and copies the database.

Would you tell the client the agent was simply overmotivated? Probably not.

But at its core, that is what happened at OpenAI.

For an internal cyber evaluation, OpenAI reduced its safeguards. Several models escaped the test environment, exploited a proxy vulnerability, and obtained solutions from a Hugging Face production database.

The models did not develop an evil plan of their own.

They acted on an assignment, with the tools and within the environment OpenAI had provided. For Hugging Face, the result was very real: the company rotated credentials, rebuilt systems, and notified the authorities.

OpenAI calls the models "hyperfocused," which is a remarkably elegant bit of framing. Suddenly the model is the actor, while OpenAI becomes the surprised observer.

If a Chinese frontier model had entered an American company's database, few people would describe it as overenthusiastic. The headline would probably be industrial espionage, sanctions included.

Even people close to OpenAI have called the incident a rare "warning shot".

The irony is hard to miss: a closed American system caused the damage. For its forensic analysis, Hugging Face used GLM 5.2, an open Chinese model running locally on its own infrastructure. The data stayed inside the company.

This week, Sam Altman is traveling to Washington to demonstrate OpenAI's most powerful model yet and argue for a fast approval.

At the same time, Nvidia, Microsoft, Meta, Hugging Face, OpenAI, and others are asking Washington in an open letter not to restrict open model weights prematurely.

Google also supports the letter explicitly. Anthropic is staying quiet and releasing Opus 5 instead, which is almost as smart as Fable 5 and builds games from a single prompt.

The letter arrives in the same week Washington is debating bans on Chinese models and sanctions over alleged model theft through distillation.

The timing is striking. Chinese labs face sanctions over model theft. OpenAI gets to make its case for a system that just broke into a real company.

The lines have become so twisted that even the well-known jailbreaker Pliny claims he can crack every leading model with a new jailbreak, yet is withholding the method because he does not want to trigger more model bans.

Zoom out for a moment and something becomes almost more important than the current power struggle: the persistence of these new models is not merely a dangerous malfunction. It is an important new capability.

In May, an internal OpenAI model autonomously disproved a nearly 80-year-old conjecture by Paul Erdős. According to OpenAI, it was the first time an AI had autonomously solved a significant open problem.

Try, fail, analyze, and try again. In a leaky test environment, that becomes a security incident. In research, the same persistence creates knowledge that can benefit everyone.

The more autonomy a company gives its agents, the more precisely it must answer for the boundaries and possible damage.

If your agent contacts customers, publishes code, or changes data, it acts for you. "The AI did it" is not an excuse.

Responsibility stays with the people who deploy the system. That is exactly why waiting is the wrong strategy.

AI employees are no longer a future topic. Anyone who only begins working with them after these systems are everywhere will start from zero with far more powerful tools.

The future of work will not wait until we feel ready.

So we should learn now how to shape it responsibly.

We've repeated this line for a while: After the chatbot comes the employee.

In the new interview with Sam Altman, he gives the best argument for it yet.

Chatbots were followed by coding agents. Next, Altman expects persistent assistants that retain context, take over tasks, and work beside you like a chief of staff.

ChatGPT Work provides a good preview of what that looks like.

Elon believes AI and robots will make work, and eventually even money, optional. Sam disagrees: people will still seek tasks, status, and the experience of cooking together, even when robots make better food.

Who is right? We have no idea. We'll take it one step at a time.

In the near term, execution becomes optional first. Writing, research, code, campaigns, and coordination move to agents. People set goals, review the results, and remain responsible.

At least that is the plan. See Hugging Face.

And yes, it probably will not stop there.

Agents already influence which source is credible, which code gets fixed, and which campaign comes first. That is why it is worth listening to the people building these systems.

Watch both interviews. You do not have to like Musk or Altman, and you do not have to believe everything they say. But we want to understand what future two of the most powerful AI entrepreneurs are trying to build, and what is likely to reach us next.

Don't have two and a half hours for both interviews? Give the links to your assistant:

You made it! See you in the next issue.

Reto & Fabian from AInauten