A licensing-based alternative to banning open-weight AI, argued via cracked software analogy
DivideHorror3217 · reddit · 2026-09-13
The poster argues open-weight AI is physically impossible to ban, likening it to pirated software: regulators can only meaningfully target enterprises, many of which already run Chinese open models in sandboxes — revenue OpenAI and Anthropic are losing.
He warns that cloud giants process B2B data on unencrypted memory, meaning your data could be harvested for training, and proposes an alternative:
- Allow only verified open-source models with registered hashes/signatures
- Allow only verified harnesses (llama.cpp, vLLM) validating offline-verifiable JWT license. files
- Restrict usage to credentialed technical professionals
Claimed benefits: blocking harmful models, safe offline enterprise use, price pressure on frontier labs, democratized research.
More from AGI Musings
- Claim: Dario Wants to Slow Research to Curb Runaway Compute Costs Ahead of IPO — pdamodaran · 2026-09-13
- Sherry Turkle's 'Artificial Intimacy' makes New Scientist's best science books of September — AnnaCiaunica · 2026-09-13
- MIRI's Nate Soares says for the first time in a long time, humanity has a real chance — UltraRareAF · 2026-09-13
- Ethan Mollick: x-risk shouldn't be the only AI policy focus — job impacts need prep now — emollick · 2026-09-13
- Elvis Saravia hits back at AI control proposals: open-source AI matters more than ever — omarsar0 · 2026-09-13
- Economist proposes diff-based open letters, with AI mapping where signers agree and disagree — soumitrashukla9 · 2026-09-13