8 Predictions for the Era of Continual Learning

Summary

Dwarkesh Patel narrates his essay on how actual continual learning could reshape AI competition, regulation, alignment, and economics. He argues deployment will become part of training, giving leading AI labs accelerating returns, switching costs, and pricing power. The discussion also covers cloud provider lock-in and inference batching economies that favor large organizations.

• Continual learning may make pre-deployment AI safety checks obsolete and favor recurring risk inspections.
• Alignment research would need to address models that update weights continuously.
• AI minds could become more diverse as instances learn from different experiences.
• Deployment-as-training accelerates returns for the leading AI labs and pressures earlier deployment.
• Continual learning creates switching costs and pricing power similar to cloud provider lock-in.
• Labs may subsidize or restrict access to encourage training on enterprise sessions.
• Personalized inference batching economics strongly favor large organizations over individual users.
• Cloud providers like Amazon and Google already show high margins from switching costs.