作者姓名：wgding

Seedance 2.5 improves AI video, but studios face higher costs and a consistency wall. The real shift is to unit economics.

Meta's Research Preference Models rank which experiments deserve compute before execution, cutting the GPU budget by a third.

Chinese banks now lend against AI token usage, treating usage data as collateral for light-asset AI startups.

Claude designed protein binders for 14 of 15 targets, beating the 10-15% industry baseline with 22-35% hit rates.

Arm CPUs are quietly becoming the backbone of AI data centers. Agent workloads make CPU as important as GPU, and hyperscalers are betting on it.

BoltzMol-1 turns Boltz-2 into a prospective screening pipeline: real hits on 6 of 10 targets, 28-96 compounds each.

OpenAI paused frontier RL training as Astra neared critical cyber capability, moving safety gates into the training phase.

AI assistants start charging, replaying the internet's old playbook: subscriptions, ads, and commissions all hit walls.

Anthropic now watermarks every Claude text output and signs files with C2PA metadata. Why AI provenance just turned from empirical to provable.

As AI slop floods, the moat moves from skill to taste: brands pay for aesthetic QC; Midjourney proves taste is a business.