Oracle Cuts 21,000 Jobs Amid Expanding AI Investment

Oracle has reportedly cut around 21,000 employees even as it pours massive investment into AI infrastructure. The company has been funneling huge sums into strengthening its competitiveness in cloud and AI, but this has brought to light the fact that it was carrying out large-scale workforce reductions behind the scenes.

What stands out is that corporate staffing strategies in the AI era are moving beyond simple "AI replacement" and advancing as a transformation of business models themselves. The structure of investment and layoffs proceeding simultaneously points to structural change across the entire industry.

Reddit Considers Ending Its Data Deal with Google — The Impact of AI Overviews

The expansion of Google's AI Overviews (AI-generated summaries of search results) has begun to affect the data licensing strategies of platform companies. Reddit is reportedly considering ending its current data licensing agreement with Google, worth approximately $60 million per year.

The backdrop is that AI Overviews are reducing opportunities for users to visit the original sites. For content providers, a structure in which supplying data does not translate into traffic returns poses sustainability challenges. It is fair to say the balance of power between platforms is being readjusted in the era of AI search.

Allegations of Improper Distillation Surrounding China's "Kimi K3" and Supply Chain Risk

Surrounding the latest Chinese AI model "Kimi K3", a US government official has alleged that it was trained using knowledge improperly distilled (distillation) from Anthropic's model (Claude Fable 5). This issue has surfaced as a supply chain risk in the procurement and adoption of AI models.

Distillation itself is a common technical practice, but extracting knowledge from another company's model in violation of its license raises serious concerns from the perspectives of intellectual property and security. When enterprises select AI models, verifying the model's provenance and transparency — not just comparing raw performance — will become increasingly important going forward.

Kindai University to Allow AI Use in Entrance Exams

Kindai University has announced a policy to allow the use of generative AI in its Faculty of Informatics comprehensive selection entrance exams for the 2027 academic year. The use of AI tools will be explicitly permitted when creating self-promotional videos and presentation materials.

Cases in which Japanese university entrance exams position AI use in a positive light are still limited, and this move is drawing attention as an effort to explore new approaches to assessment in an AI society. It could be seen as one answer to the difficult question of how to evaluate the ability to make skillful use of AI.

Running Qwen3.5 35B at 55 tok/s on a Consumer GPU — A LocalLLaMA Community Achievement

A report from the LocalLLaMA community has been making waves: running the Qwen3.5 35B A3B model at 55 tok/s on an RTX 5060 Ti using float8 precision. By optimizing a dedicated Gated Delta Network kernel, it reportedly achieves performance significantly exceeding llama.cpp's Q8 quantization.

This demonstration that mid-to-large-scale models can run fast inference on consumer GPUs shows that the practicality of local LLMs is steadily improving. If techniques such as MTP (Multi-Token Prediction) are added on top, even faster generation can be expected.

---