Korean Startup XCENA Raises $135 Million — AI's Real Bottleneck Is Memory, Not Compute

South Korean chip startup XCENA has raised $135 million in a Series B round (at a $570 million valuation). The company argues that AI's biggest bottleneck lies not in compute power but in memory bandwidth, and is focusing on developing a new memory architecture to replace HBM.

While NVIDIA commands an overwhelming share of the current AI semiconductor market, memory bandwidth constraints severely limit LLM inference performance. If XCENA's approach proves practical, it could unlock more of the performance of existing GPUs. That said, the HBM market is a domain where Korean companies such as SK Hynix and Samsung hold strong positions, making the competitive landscape a complicated one.

Google Makes Nano Banana 2 and Nano Banana Pro Generally Available

On May 29, Google announced the general availability of its image generation models "Nano Banana 2" (Gemini 3.1 Flash Image) and "Nano Banana Pro" (Gemini 3 Pro Image). Nano Banana 2 also supports image generation from video.

The codenamed "Nano Banana" series embodies Google's unified approach: handling everything from text-to-image to video-to-image with a single model. The Flash variant prioritizes speed and lightweight performance, while the Pro variant specializes in high quality. Against competitors such as OpenAI's DALL-E and Midjourney, Google is making it increasingly clear that it intends to double down on API-based delivery.

AI Medical Answers Are 76% Accurate — A Penn State Study

A research team at Penn State University examined the accuracy of AI responses to medical queries and found that roughly 76% were accurate. This points to AI's potential as a tool for delivering medical information, but it is also a data point that demands caution: about a quarter of the answers contained errors.

As AI adoption in healthcare advances, guaranteeing accuracy is a life-and-death matter. A "76% accuracy rate" is by no means low, but challenges remain before AI can genuinely complement physicians' judgment. The study also analyzed patterns in the cases where AI got things wrong, which could contribute to our understanding of how hallucinations arise.

A Debate Warning of an Ungovernable "AI Bureaucracy"

A Substack article titled "The Age of Ungovernable AI Bureaucracy" is drawing attention. It argues that as AI systems grow more complex and interdependent, their decision-making processes are slipping beyond human understanding, effectively forming an "AI bureaucracy."

In a world where multiple AI agents work in tandem to carry out tasks, tracing which AI made which judgment becomes difficult. This debate symbolizes the concern that AI governance design is failing to keep pace with technological progress.

Will AI Repeat Frontend's "Lost Decade"?

An article titled "Is AI causing a repeat of Front end's Lost Decade?" has stirred debate in the developer community. The worry is that AI might trigger a repeat of the "Lost Decade" — the turmoil frontend development went through in the 2010s as the JavaScript ecosystem shifted at a breakneck pace.

As AI coding tools spread, there is a danger that developers end up spending more time figuring out how to use the tools than on essential design work. If the cycle of frameworks and tools appearing one after another only to be abandoned keeps accelerating, on-the-ground productivity could actually decline. This argument deserves attention as a prompt to rethink the balance between speed and quality in AI-driven development.

Claude Code's Late-May Turmoil — The Fallout from the Skills Integration and False AUP Flags

The story has also been making the rounds on Qiita in Japan: Anthropic's Claude Code has been hit by a string of issues in the second half of May. Reports include changed slash command behavior following the consolidation into the Skills feature, bugs in v2.1.150, and access restrictions caused by the erroneous enforcement of the Acceptable Use Policy (AUP).

The false AUP enforcement in particular has abruptly blocked legitimate development work — an issue that goes straight to developer trust. With Anthropic's responses struggling to keep up, users are actively sharing information about workarounds they can take.

---