8 Predictions for the Era of Continual Learning
Locking in AI safety regulation now is a mistake.
TL;DR
- Continual learning is essential for AI to perform jobs competently, similar to how humans accumulate experience.
- Current AI regulation is ill-suited for continually learning models, necessitating flexible, periodic risk inspections instead of pre-deployment checks.
- Technical alignment research must shift focus from static models to ensuring safety during constant weight updates and preventing malicious influence from users.
- Deployment becoming part of training will accelerate the advantages of leading AI labs, as models improve with more usage and feedback.
- Continual learning will establish strong competitive moats for leading AI labs by creating significant switching costs for users.
- Enterprises face a choice between being locked into a continually improving AI model or missing out on its benefits.
- AI labs may subsidize users or mandate training on their sessions to gather data for continual learning, akin to Google's search model.
- Continual learning can lead to economies of scale in inference, strongly favoring large organizations due to efficient serving of personalized weights.