The real action is on-device inference and agent reliability, not scale. PalmClaw and Apple's PrismML moves signal the hardware industry is finally forcing the software layer to solve actual problems—can agents know when they're in over their head, can models fit in phones without becoming useless—rather than chasing token counts. IBM's collapse and the noise around OpenAI's speaker are symptoms of the same disease: companies betting on general-purpose AI "experiences" when the margin is in solving *specific, constrained problems* that don't need to be smart everywhere. Watch the papers on task complexity awareness and on-device frameworks; ignore the venture funding press releases until someone ships something that actually runs offline.