Dwarkesh Patel argues that session notes and external memory are not enough for durable skill accumulation. If deployed models begin updating their own weights from work experience, safety evaluation would need to shift from a single predeployment gate to recurring risk inspections that track a changing system.
Continual learning could create strong feedback loops for early deployers. Models with more users would gather more experience, improve faster and attract even more users, increasing pressure on laboratories to release systems earlier and making established products harder to displace.
The same dynamic could deepen laboratory and enterprise lock-in. Organizations may need to expose internal work data so models can learn effectively, while large users gain additional advantages from batching inference across many tasks. Patel's predictions frame continual learning as an economic and governance change, not only a technical improvement.
Watch the original on YouTube