Preventing AI proliferation seems hard

I’m wary of solutions to AI risk which involve a small number of organizations maintaining exclusive control over these systems:

• This seems like an unstable equilibrium prone to defection and slips.
  • LLaMa’s weights leaked in roughly a week. And Yann LeCun thinks that was good, wants to open-source Llama’s weights (original).
• Steady algorithmic advances will continuously lower the barriers to training “indie models”.
• Massive security breaches of high-value computer systems occur somewhat routinely.
• What about all the ex-employees of the big AI labs?