A mega-round out of Europe, Meta's cross-app subscription, top executives sounding off on AI safety — and a research survey on how "AI improves AI." Here are five stories from today's news.

Mistral AI Raises Over ¥530B — Samsung-Led, at a €21B+ Valuation

On September 14, French company Mistral AI announced it had raised €3 billion (approx. ¥530 billion) in a Series D round, at a post-money valuation of over €21 billion. Reaching that mark barely three years after founding, the company is positioning it as "the largest equity funding round ever by a European tech company."

Samsung Electronics led the round, with co-leads from the "Scaleup Europe Fund" managed by EQT and existing investor PSG Equity.

What's noteworthy is not just the amount but the context in which the money came together. In an industry where U.S. players such as OpenAI and Anthropic compete backed by enormous capital, a European open-weight AI company has pulled off a round of this scale. As the term "sovereign AI" symbolizes, the value of holding homegrown AI capabilities that don't depend on the models of any single country or company is increasingly recognized on both national-security and industrial-policy grounds. For Samsung, it also deepens a relationship that stretches from semiconductors up to the model layer.

Meta's "Meta One" — Spanning Instagram/Facebook/WhatsApp, From ¥949/Month in Japan

Meta has launched "Meta One," a new subscription that spans Instagram, Facebook, and WhatsApp. Centered on expanded AI usage limits, it bundles more than 50 features — including creative tools and business tools — into a single plan. It's rolling out in Japan too, with bundle plans starting at ¥949 per month, and it reportedly includes business-oriented features such as verification badges.

The key point is that Meta is signaling it will sell AI features not as standalone products but as a "cross-app experience bundle." The sub-¥1,000 price tag reads as a statement of intent: to normalize expanded AI usage as a default experience. Many AI features that companies have been charging for individually could end up absorbed into bundles like this.

Zuckerberg: "Each Company Should Act on Its Own"; Altman Says "Trust" Is Warranted — But Acknowledges "the Right to Be Afraid"

Meta CEO Mark Zuckerberg has explicitly distanced himself from the "coordinated slowdown" argument — the idea that the industry should move in lockstep on AI safety. According to ITmedia, Zuckerberg stated that "each company should address these issues individually, at its own responsibility and pace," arguing that ensuring proper alignment and bringing in external evaluators will happen naturally through market competition and legal accountability, and that "the decentralization of power is the foundation of safety."

Around the same time, OpenAI CEO Sam Altman reportedly told the BBC that while "the world has a right to be afraid," people should "trust" AI companies. Meanwhile, The Guardian reported on the split between Anthropic's CEO, who is once again calling for an AI slowdown, and NVIDIA, which insists on acceleration — the slowdown-versus-acceleration divide continues to smolder.

What's interesting is that the leaders of the two dominant U.S. powers — Meta and OpenAI — chose this moment to simultaneously emphasize "each company's autonomy and trust over regulation and coordination." Just as institutionalization efforts advance, such as standing third-party evaluators and shared safety standards, the argument that safety should be left to competition and legal liability is moving to the foreground — a meaningful signal for where the debate heads next.

Frontier Models' Edge Is a "4-Month Lead" — Chinese Open Models Are Closing In

An analysis reported exclusively by Ars Technica is gaining traction on Hacker News. Its claim: the advantage of sticking with paid frontier models amounts to a "4-month lead" for "5x the cost."

The argument is that open models out of China are closing the gap with Silicon Valley's frontier models. While the detailed measurement methodology couldn't be verified at the time of collection, this "4 months, 5x" framing makes it easy to oversimplify model-selection debates. Considered alongside Mistral's mega-round, the balance of power between closed top-tier models and open-weight models looks set to be re-evaluated on cost grounds as well.

"The Last AI Built by Humans" — An L1–L5 Roadmap for Recursive Self-Improvement (RSI)

A survey titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" is drawing attention on Hugging Face Daily Papers.

Recursive self-improvement (RSI) is defined as a system's ability to convert the experience and feedback it acquires into permanent changes to itself, with those changes influencing the subsequent generation, evaluation, selection, and integration of improvements. The targets of those updates range from model weights, prompts, code, memory, skills, and task distributions to the improvement mechanism itself.

The backbone of the research is a roadmap that organizes the scope of RSI's autonomy into five levels.

L1** Autonomy in "executing" improvements — the system executes improvement procedures defined by humans, and the results persist for subsequent tasks

L2** Autonomy over improvement "strategy" — the system chooses how to improve (objectives and evaluation criteria remain external)

L3** Autonomy over future learning experiences — the learner's state influences which experiences, tasks, and curricula it takes on next

L4** Autonomy over deployment and environmental adaptation — reusable memory and skills are retained, changing later behavior

L5** Meta-improvement — the system improves the very mechanism that generates improvements

The authors have also released an auditable collection that organizes 491 baseline papers and 28 extensions under this taxonomy, and introduced a metric called the Headroom-Closed Index (HCI). They reportedly examine the current state of RSI in scientific discovery, embodied intelligence, and software engineering as well.

As the provocative title suggests, this is a foundational piece of groundwork for thinking about how the human role shifts once "AI improving AI" gets serious — and it seems likely to become a reference point in future debates on both safety and capability.