Daily Brief

Only 3.46B of 78B parameters activate, and the 1M-token context model ships under Apache 2.0.X：Testing Catalog (@testingcatalog)

78.1B total parameters with just 3.46B activated per token (4.4%), released as a German-English model under Apache 2.0.Hacker News 热门（buzzing.cc 中文翻译）

781B total params but only 3.46B active per token, cutting German tokens 15% versus GPT-5.Hacker News：AI 热帖

Bernstein projects Ascend at around 50% this year, versus Nvidia's claimed 95% pre-export curbs.X：X.PIN (@thexpin)

AI itself screens roughly 50PB of data after six Australian government sites were notified.IT之家（RSS）

The three-and-a-half-year safety report lead's public exit marks the most direct insider condemnation of OpenAI's culture to date.TechCrunch：AI（RSS）

In The Atlantic, he criticized the company's safety culture, a rare public exit from OpenAI's safety team.The Decoder：AI News（RSS）

Direct water use is just 0.5% of US industrial water, and Amazon has poured over $9B into affected US communities in three years.TechCrunch：AI（RSS）

GPT-5.5 mentioned the new method losing to the baseline in only 2 of 200 summaries.X：Rohan Paul (@rohanpaul_ai)

A murder sentence was reversed because an AI-generated video of the victim was played at sentencing—a likely U.S. first.Hacker News：AI 热帖