AI

Task completion is replacing per-step accuracy as embodied AI's core KPI. Why WAM hits a wall and how WTM changes the game.

GPT-5.6 flagship Sol nearly triples detection accuracy and tops coding benchmarks; Terra and Luna match rivals at a fraction of the cost.

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.