AI Infrastructure: Software IP Devalued, LLMs on 4GB GPUs
Chamath declares software IP effectively worthless after Adobe's Big 5 products were decompiled, while a new tool called AirLLM enables 70B LLMs to run on a single 4GB GPU — compressing the AI stack from both ends.
What we know
Two AI-structural signals emerged. The All-In Pod quoted Chamath calling software IP 'effectively worthless' after all five major Adobe products were decompiled — a sweeping claim about proprietary software moats. Separately, an account on X highlighted AirLLM, a layer-wise inference method that loads, computes, and flushes one model layer at a time — enabling a 70B model on 4GB VRAM and Llama 405B on 8GB VRAM. Aaron Levie noted the governance implication: as agents do 'vastly more task execution than humans,' the need to secure, audit, and control that work becomes paramount.
By the Global Eye 24 desk, with AI assistance. Updated as the story develops. Spot an error? Write to corrections@globaleye24.com.