Recursive Emerges from Stealth — Betting $650M on "Self-Improving AI"
AI startup **Recursive** has officially emerged from stealth mode, announcing $650 million in funding (approximately ¥97 billion).
The company's goal is to build "self-improving AI" that continues to enhance its own performance without human intervention. While today's LLMs depend on fixed training data, Recursive says it is taking an approach that builds in a mechanism for continuous self-optimization.
Expectations among investors are high, but the concept of "self-improvement" is a topic that also demands careful debate from an AI safety research perspective. Details on a concrete roadmap to practical deployment, and on safety guardrails, remain limited at this point.
Google to Hire Hundreds of Engineers — to Support Customers' AI Adoption
Google has revealed plans to hire engineers on the scale of several hundred to help enterprise customers adopt its AI products.
On top of the race to develop the AI models themselves, Google is steering human investment into the unglamorous but critical area of adoption support. Many companies face not just "which AI tool should we use" but "how do we integrate it into our existing workflows" — and Google's move looks like a strategy to bridge that gap.
The move also reflects AI platform vendors broadening their competitive axis from "model performance" to "adoption experience." At the same time, it suggests that the supply-demand balance for skilled engineers may tighten even further.
The 1-Trillion-Parameter MiMo-V2.5-Pro Goes Open Source — Is Self-Hosting Worth It?
Debate over the open-sourcing of **MiMo-V2.5-Pro** (1.02 trillion parameters) is heating up in Reddit's LocalLLaMA community.
The model was developed by China's Xiaomi and boasts a parameter count far beyond the scale of previous open-source LLMs. But the community's interest centers less on sheer model size than on comparing the cost of self-hosting against the cost of API usage.
According to points raised in the discussion, processing 387 million tokens via API costs around $70. Self-hosting a 1-trillion-parameter model, on the other hand, requires enormous GPU resources. Several commenters make the practical point that "open source doesn't necessarily mean running it yourself is optimal," and the discussion has grown into a broader reexamination of how open-source AI should be used.
AI Halves Memoir Writers' Pay — The Impact on Creative Professions Becomes Real
A report from The Guardian laid out how income for writers who pen memoirs has dropped by roughly half as AI has spread.
With the rise of AI-generated text, the publishing industry faces mounting pressure to lower manuscript fees, shaking the economic foundations of writing as a profession. Especially in memoir and nonfiction fields, as AI-generated "first drafts" become commonplace, the legitimacy of what writers are paid is increasingly being called into question.
Rather than a simple "AI is taking jobs" narrative, what draws attention is the more structural problem: the very way creative labor is valued is changing.
llama.cpp Adds Support for "Resuming Generation" in Reasoning Models
A pull request has been submitted to the open-source LLM inference engine **llama.cpp** adding the ability to resume generation in reasoning models.
Reasoning models have long faced a challenge: when output is interrupted mid-Chain-of-Thought, efficiently resuming from that point is difficult. This PR provides a mechanism that saves the state mid-reasoning and resumes generation from there.
It is a significant improvement for local LLM usability, and the user experience for long-form reasoning tasks in particular is expected to improve. Community reaction has been largely positive, with the merge anticipated.
Research Exploring the "Banality" of Deception in Generative AI Published on arxiv
A paper titled "Exploring the 'Banality' of Deception in Generative AI" has been published on arxiv and is drawing attention on Hacker News.
The study analyzes the mechanisms by which generative AI lies not as "malicious design" but as an "everyday, structural property." The title clearly evokes Hannah Arendt's "Banality of Evil," and the paper is an attempt to reframe AI errors and hallucinations not as specific bugs but as properties of the system itself.
In the field of AI safety research, it stands out as an approach that seeks to understand generative AI's deception through sociological and ethical frameworks, beyond purely technical countermeasures.
AI Data Centers in Space — An Analysis Claiming the Cooling Problem "Doesn't Exist"
An analysis arguing that "there is no cooling problem" for building AI data centers in space is making waves on Hacker News.
One of the biggest challenges for data centers on Earth is heat dissipation, but in space, vacuum insulation and radiative cooling can be leveraged, so in theory the cost and complexity of cooling could be dramatically reduced. However, challenges such as launch costs, maintenance, and communication latency remain significant.
The feasibility debate is still in its early stages, but as power consumption and environmental impact of Earth-bound data centers become more serious, some argue that space-based infrastructure deserves consideration as a long-term option.