Topic
Explore the latest artificial intelligence news and research.
We examine reports of a 30% cut in ABF supply to China against Ajinomoto's official IR disclosures and mass-production qualification processes, and outline the allocation factors and substitute-material conditions that determine the impact on advanced packaging.
A paper titled "Lost in Compaction" from Pennsylvania State University researchers demonstrates via the COMPINT benchmark that AI context compaction causes an average of 83% of user instructions to be lost. The study also points out that Anthropic's official documentation does not guarantee automatic instruction retention.
In Vercel's July usage data, Anthropic accounted for 65.1% of list-price-equivalent spend while processing just 29.8% of total tokens. A mass shift toward low-cost models and a concentration of high-value processing on Claude are happening at the same time.
Examining materials claiming Qwen3.8-27B inference at 30+ tokens/s on a hypothetical 64-core XuanTie C950 configuration. We unpack the SHL optimizations behind the performance, the undisclosed test conditions, and the distance to mass production.
OpenAI has automatically applied a dedicated ChatGPT experience for 13- to 17-year-olds, integrating learning support with interpersonal protections. The accuracy of age estimation and safety metrics in actual use remain undisclosed, and constraints on parental linking persist.
Google won a $10 million bid for the internal data of bankrupt, liquidating Spirit Airlines — including 100 million emails. This piece examines the bidding war with runner-up Mercor, the 100x gap versus typical Asset Hub pricing for liquidation data brokerage, and how effective the line drawn around excluded personal information really is.
In a Qualcomm-commissioned comparison of retail laptops, the Snapdragon X2 Elite Extreme machine posted major gains in CPU and AI performance. Intel's machine retains an edge in GPU performance and battery life.
OpenAI cannot rule out that Astra has reached "Critical" cyber capability, and has held off its largest frontier RL run as a result. Monitoring now covers roughly 20% of targeted inference compute, as safety standards are being updated mid-development.
Etched has raised $700 million, pushing its valuation to $21 billion. It has delivered its first inference rack to Jane Street, but performance comparisons and disclosure of production scale remain to come.
Claude Code's /design has moved into an early preview that lets users compare and edit multiple UI options before handing off to implementation. We break down the differences from the feature that has existed since June, along with the undisclosed availability and Artifact handling conditions.
