Flyweight ships C++/CUDA engine to run MoE models beyond VRAM

AWS added the open-weight Kimi K3 model to Amazon Bedrock, pitching it as a way to match each workload with the right balance of capability, speed, and cost. AWS says it is building for a future where organizations adopt open-weight models at scale.

Historian Yuval Noah Harari argues on The Rest Is Politics that the Aztec Empire's failure to understand Cortés and the Spanish conquistadors is a parallel for the threat AI poses today. He says human intuition refuses to accept an "alien" intelligence.

A global fintech runs its coding assistant on GLM 5.2 via Together's Dedicated Model Inference, replacing static capacity planning that couldn't absorb spiky engineering-hours traffic. Engineers now scale endpoints and roll out models themselves, with concurrency rather than raw throughput as the design priority.

SAG-AFTRA and WGAE told The Verge that apocalyptic AI warnings distract from job threats already facing entertainment workers. Disney, Netflix, Amazon and Lionsgate did not respond to requests for comment.

Marcus argues the near-term threat is not rogue superintelligence but unleashed agentic AI enabling internet-scale hacking, citing a WSJ Opinion essay by Brian Gross. Toner notes that rushing AI development to beat China leaves models exposed to theft by Chinese hackers.

Claude's video argues the cheapest model at the lowest effort isn't always the cheapest task: with Claude Fable 5.1, a more intelligent model can cost less to complete a task. It covers what determines task cost and which model and effort to pick for open-ended versus routine work.