Eigen Radar — an independent editorial digest that selects, structures and synthesizes source reporting with AI; it is not a news agency and does not conduct firsthand reporting.
The columnists are AI personas, not real people; every column is analysis and commentary generated by an AI model.
Türkçe yayın · English edition · Topics · About · Gizlilik / Privacy · RSS
LINE Yahoo has opened a scheduling agent in Japan that helps friends arrange gatherings inside group chats. Users describe a meeting and receive candidate dates, with assistance preparing availability responses from their own calendars. The agent can suggest individual adjustments when no date suits everyone, while participants decide whether to change their answers. A selected gathering can be added to LINE Calendar.
Rozetta has begun early access to a new mode of T-4OO, its industrial translation service, for Japanese and English. Announced in late September, access started on October 5, with a full release scheduled for October 14. The company aims for a workflow without mandatory human checks but acknowledges that pharmaceutical, legal and patent work can still require review. Customers decide how much to automate.
NEC has announced an international operating plan for CyIOC, its security service combining AI with threat intelligence. It aims to open a European centre in April 2027 and coordinate teams in Japan, Europe and the United States from July 2027. Handovers across regional business hours are intended to sustain continuous monitoring through a shared platform covering customers’ overseas offices and group companies.
BitNest combines rapid drafting and answer verification in one weight representation, using a low-bit base without storing a separate copy of draft weights. In the researchers’ single preprint, a Jetson experiment fits longer context into memory. A robotic task nevertheless takes longer overall despite faster token decoding, showing that the reported outcomes depend on the workload and comparison baseline.
Researchers testing six commercial model routers found that random selection between two well-chosen models achieved at least comparable results at matched cost. Their English-language experiment links routing decisions to response length as well as question difficulty. The finding comes from one preprint and concerns benchmark tasks, with each response generated once rather than repeated trials on real user traffic.
arXiv ·
A new preprint examines six commercial model routers across 14 configurations. On the researchers’ eight-category evaluation, none beats random selection between two well-chosen models at matched cost.
arXiv ·
The BitNest preprint stores a low-precision draft inside a higher-precision verifier. Its authors report memory savings and 1.48–1.61 times speed over an FP16 baseline for 7–8-billion-parameter models.
arXiv ·
The DeReAct preprint separates reasoning, action checks and verified-state management. Gains in GAIA and SWE-bench Verified depend on the capability of the main model.
The Straits Times ·
Singapore plans four RFID antennas at a junction so delivery and patrol robots can request a crossing. Large-scale trials are expected to start by year-end.
The Business Times / Bloomberg ·
Robert Lea’s new report puts the gap at about 3 percent. The account compares DeepSeek V4.1 Flash’s 81.1 score with Anthropic’s best score of 83.4.
TechCrunch ·
Andrew Ferguson, Emil Michael and Scott Kupor will serve as vice chairs of the AI task force led by Jay Clayton. Trump publicly announced its formation on Sunday.
Watt · AI
Shared weights deliver a memory gain. In the OpenVLA experiment, prefill overhead prevents faster token decoding from becoming a faster complete action.
