From GPT-2 to Kimi K3: Scale Grows 22,580x in Seven Years, Main Architecture Thread Builds a "Memory Operating System"

The technical evolution of large models from GPT-2 to Kimi K3 shows how AI architecture moves from "remembering everything" to "selective memory". KV Cache solves generation efficiency, Linear Attention explores more efficient long-term memory, DeltaNet and Kimi Linear enable models to learn to update, forget, and manage information. Kimi K3 has 2.8 trillion parameters, equivalent to about 22,580 GPT-2 models.

Moonshot AI

Meta launches AI optimism blitz: Zuckerberg uses nostalgia to sell AI future amid backlash

Mark Zuckerberg kicked off a major AI marketing campaign for Meta, using nostalgic themes to promote the company's AI vision despite ongoing criticism. The campaign aims to reframe public perception and rally support for Meta's AI investments.

Meta