OpenAI and Anthropic Set Up New Companies in Japan — Is This a "Black Ships" Moment for Domestic SIers?

Anthropic and OpenAI have both announced the establishment of new companies in Japan to deliver their AI services. Each is building out a structure to serve Japanese enterprises directly, staffed with engineers such as Field Deployment Engineers (FDEs) and Applied AI Engineers.

According to ITmedia, AI adoption in Japan has so far been dominated by indirect delivery through SIers. Now that AI vendors are providing engineering support from their own Japanese bases, the existing SIer business model could take a major hit. In particular, if the trend of "AI vendors providing direct support from requirements definition through operations" accelerates, SIers will be forced to move beyond their role as mere middlemen and redefine their own value-add.

Japan's SIers now find themselves at a crossroads: they must deepen their technical understanding of AI while figuring out how to leverage their strong customer relationships.

Meta Begins Testing an AI Subscription Service

Meta has begun testing a subscription service offering its AI features. According to CNBC, the cheapest plan starts at $7.99 per month.

Meta has historically offered its AI features for free, predicated on an ad-supported revenue model. The move into subscriptions appears to be an effort to find a new revenue stream to recoup the enormous compute costs of AI development. With competitors like OpenAI (ChatGPT Plus at $20/month) and Anthropic (Claude Pro at $20/month) already running subscription services, Meta's entry at a lower price point could further intensify price competition in the consumer AI market.

Critical Vulnerability Found Affecting VLLM and Numerous MCP Servers

According to reports on Reddit's LocalLLaMA community, a critical vulnerability has been discovered in an open-source framework widely used by VLLM, numerous MCP servers, and other LLM tools.

Ars Technica reports that the vulnerability could put millions of AI agents at risk. VLLM is one of the de facto standards for inference servers, and MCP (Model Context Protocol) servers are spreading rapidly, which makes the potential blast radius a serious concern. If you use the affected framework, applying patches and checking for mitigations as soon as possible is recommended.

Dario Amodei's View on Employment Shifts — A Jevons Paradox Perspective

Fortune reports that Anthropic CEO Dario Amodei is revising his views on AI's impact on white-collar employment.

Amodei had previously issued stark warnings that "AI will cause mass unemployment among white-collar workers," but lately he has been invoking the concept of the "Jevons paradox," suggesting that efficiency gains could actually increase demand. The Jevons paradox is an economic phenomenon in which improvements in the efficiency of resource use lead to an increase in total consumption; applied to AI, it reads as "if AI makes work more efficient, the volume of tasks people want to use AI for will actually grow."

That said, the short- to medium-term impact on employment remains unclear, so his position should be taken with the caveat that this is "as of now."

NEO-ov: A Native Vision-Language Model That Needs No Encoder

NEO-ov, which has been attracting attention on Hugging Face Daily Papers, is a native vision-language foundation model with no external image encoder, adapter, or post-hoc fusion whatsoever.

Conventional VLMs (Vision-Language Models) have typically followed a modular architecture in which an image encoder and a language decoder are trained separately and then connected through multi-stage alignment. NEO-ov does away with this structure, learning pixel-level correspondence and cross-frame association end-to-end. As a result, it reportedly delivers performance competitive with modular VLMs across diverse tasks — single-image, multi-image, and video understanding, as well as spatial intelligence — while exhibiting finer-grained visual perception.

This "one vision" design philosophy could prove to be an important research direction, hinting at what multimodal AI will look like going forward.

---