beat · 194 stories
Lai's government has turned the National Security Council, built under Chiang-era strongmen, into the central coordinator for Beijing pressure. The 1993 law governing the body has not kept pace.
The April 2026 cutoff is set. Mindgrove and BigEndian are still inside the multi-step chip qualification process (design, prototype, integration, validation, revenue). Four dates will tell you which clock wins.
Swancor's Qiyuan robotics brand opened reservations on two home humanoids: an 88 cm, 15 kg Q1 with body-wide touch sensing and a 100 cm, 16 kg T1 that switches between a wheel-legged and a four-legged form, with a 1,000 yuan deposit credited as
The Telegraph's anonymous sourcing, the UK National Cyber Security Centre's silence, and the withheld plant identity are doing the threshold work. Each is a question mark, not a confirmation.
Nvidia is passing a 15% AI server price hike to Microsoft, Google, and Oracle, blaming memory chips. That signals where the AI cost story is now being set.
India's Akashteer, an AI-enabled air-defense command system, paired incoming threats to weapons in 2–3 milliseconds during Operation Sindoor, the May 2025 India–Pakistan cross-border strikes; the MoD has signed Rs 2,400 crore (~$285M) to scale it.
Maharashtra is putting a 164-record sugarcane result from one district in front of 10,000 farmers across eight crops. The real test is whether AI agronomy travels.
On DeepSeek's own table, its experimental vision model V4-Flash-Vision-Exp beats Anthropic's Claude Opus 4.8 on 3 of 11 benchmarks. The yardstick DeepSeek measured against was already eclipsed six weeks earlier.
Meta's AI coding agent Muse Code launched August 5 with output priced at 20 cents per million tokens of AI-generated text for users sharing feedback — matching China's DeepSeek and undercutting OpenAI's discounted older models.
AI cloud operator Nebius just raised $5B in debt that can be swapped for stock, at stated interest rates as low as 0.50%.
Poolside shipped Laguna S 2.1, an open-weight AI coding model whose mixture-of-experts design routes each piece of text through only a slice of its 118 billion parameters (learned weights), so inference stays cheaper than a full 118B run.
A small Army Cyber Command unit, Task Force Lexington, runs each agent through the Army's Job Qualification Readiness pipeline and slots them into four named cyber work roles, with commanders holding risk.