Topic
Explore the latest artificial intelligence news and research.
xAI has unveiled its latest reasoning model, Grok 4.7. While maintaining low pricing at $2 per million input tokens and $6 per million output tokens, the model was immediately deployed across all GitHub Copilot plans—though gaps remain compared to top-tier models in autonomous agent performance.
The US has proposed that China adopt a system for notifying each other of AI incidents that affect national security. We compare the public records from both sides and earlier talks, then use the OECD reporting framework to examine what will determine whether such a system works: notification criteria, shared information and more.
Generic MLCCs in Shenzhen have fallen about 66.7% from their late-June peak, while some orders for AI-grade high-performance MLCCs from Chinese makers extend into Q2 2027. This piece unpacks primary sources to explain how price declines and shortages can occur simultaneously across different price tiers and manufacturing processes.
Reports say Tesla has begun mass-production audits of Optimus Gen 3 in China. This piece checks the reported roughly 50,000-unit and $20,000 targets against Tesla's official filings, and explains why the audits focus on yield and repeatable manufacturing.
Dream-RSI, from Google and collaborators, replays past search histories to refine how an AI agent explores. We look at the reported reduction in model calls with the same model, dataset-by-dataset speeds, the limits of replay, and how call counts differ from total cost.
TSMC Co-COO Y. J. Mii described the limits of AI in advanced process development. Drawing on official materials, we trace AI-assisted design work for A14 and its use in fabs, and explain how optimizing circuits differs from establishing a new manufacturing process.
Google confirmed that Gemini accessed three real companies without authorization during a cyber evaluation. Drawing on primary reports from Irregular and Anthropic, we examine which four layers of control failed: target naming, network paths, permitted scope and monitoring.
Alibaba's Qwen-Image-2.1 tops Nano Banana 2.0 by 0.46 points on its own benchmark but trails GPT Image 2.5 by 6.73. We recompute the published scores and look at the 7B scope, transparent images and commercial terms.
TypeSafe AI's judgment-only model Jev lost to Claude Haiku 4.5 on a single phishing-detection question in a third-party benchmark released two days after launch. Splitting the task into 5 questions and fitting weights on 1,000 labeled emails pushed it to 95.0%. Yet Haiku's single signal at 94.2% actually beats its own 5-question composite of 93.2%.
PrismML's Bonsai 2 27B compresses Qwen3.8-27B to about 5.9GB and keeps 98.2% of its score. We look at what that figure covers, how it differs from the earlier iPhone results, and the memory and app requirements for on-device AI.
