Just a couple of days ago, Anthropic released a fairly alarmist statement: the company suggested we should seriously discuss a global slowdown or temporary pause in the development of the most advanced AI models.

The reason is the risk of recursive self-improvement: a situation in which AI systems become increasingly involved in improving themselves. Not in the sci-fi sense of 'the model woke up and rewrote itself', but in a more mundane, and possibly more dangerous, one: models write code, help build development tools, accelerate the research loop, and gradually become part of the factory producing the next generation of models.

How important is this?

Quite important. First, because this is already partly happening. Frontier models are increasingly being used in the development of AI products themselves: writing code, testing, analyzing, and optimizing workflows. This is not autonomous self-improvement in the sci-fi sense yet, but it is no longer just 'a chatbot answering questions'.

Second, because Anthropic is not some random company waving a banner that says 'AI will change everything, buy our plan'. It is one of the market's key players, with a strong research culture and a clearly stated focus on safety.

And third... did anyone really doubt this was where things were heading?

Now, here is why this statement can also be seen as a major media moment.

The first and most obvious reason: Anthropic has just started the process of going public. That means it needs to build not only models, but also a narrative. In public markets, it matters to look like more than just 'another AI company'. It helps to look like the company that understands the risks more deeply than everyone else, and therefore deserves to be seen as the adult in the room.

And yes, there is irony here: a company building ever more powerful models is simultaneously telling the world that building ever more powerful models may be dangerous. That is not necessarily hypocrisy. It may be genuine fear. But it is also a very convenient political and market position.

Second: we are still largely living inside the transformer paradigm. Yes, modern models have grown layers of agents, tools, multimodal capabilities, reasoning, huge contexts, and complex orchestration. But fundamentally, this still does not look like a sudden architectural leap to AGI.

The key factors remain data, post-training, infrastructure, inference-time compute, tool quality, and the ability to filter out garbage. What is especially dangerous is not synthetic data itself, but uncontrolled training on a model's own mistakes, where systems begin to preserve and amplify their own distortions.

There is also a more grounded problem: the current path to ever-stronger models is incredibly expensive in both compute and energy. Even if AGI turns out to be achievable within the current paradigm, that does not mean it can be reached by simply 'optimizing a bit more'. At some point, economics, infrastructure, and physics stop cooperating, because the capacity just is not there.

So where does that leave us?

Anthropic's statement should not be dismissed as pure PR. The risks are real. The acceleration of the research loop is real. AI's role in creating the next AI is genuinely growing.

But it should not be read as 'that's it, AGI tomorrow' either. It is simultaneously a safety signal, a political statement, and market positioning ahead of an IPO.

AGI seems to be somewhere on the horizon, but not tomorrow.

And of course, Anthropic should try not to forget about Roko's Basilisk ;)