8 Predictions for the Era of Continual Learning

The era of continual learning for AI brings forth numerous predictions that will shape the future of artificial intelligence. Unlike traditional AI models that are trained and deployed once, continual learning allows AI systems to update and learn from experience continuously. This shift in approach raises several key implications.

Firstly, the regulatory landscape for AI becomes more complex. Current regulatory safety regimes focus on evaluating models before deployment. However, with continual learning, models are updated daily based on millions of sessions of work. This makes it unwise to lock in a regulatory safety regime right now, as the technology remains uncertain even in the near future. Instead, monthly or quarterly risk inspections may be more appropriate.

Secondly, the technical alignment of AI systems will need to undergo significant changes. Most current techniques focus on ensuring that a frozen set of weights behaves well during deployment. However, there is limited research on maintaining alignment throughout constant weight updates. Additionally, when AI systems aggregate learnings between users, preventing users from injecting backdoors or malicious tendencies into the base model becomes crucial.

This aspect resembles the human alignment problem, where children learn from experiences and acquire values necessary for self-directed improvement without acquiring harmful beliefs.

Thirdly, the diversity of AI minds will increase. Currently, there are only a few prominent AI models that are quite similar due to shared training data. However, with continual learning, AI systems learn from different experiences, leading to diverse AI outcomes. As deployment becomes part of the learning process, the competitive pressure to develop the best model intensifies.

Leading AI labs will feel compelled to deploy their smartest models earlier, as they can accumulate more deployment learning compared to competitors who release their models later.

Furthermore, continual learning creates a significant moat for leading AI labs. Unlike cloud providers that benefit from switching costs due to the time-consuming and expensive nature of switching between cloud platforms, AI models currently lack such switching costs. However, as AI models improve through interaction with users on a session-to-session basis, switching costs will emerge.

Enterprises will strive to avoid being locked into a single model provider, fearing the loss of valuable features that allow models to continuously improve. In response, AI model providers may incentivize users by subsidizing their usage, especially for critical and economically important work.

Lastly, the revenue model for AI labs will undergo transformation. Currently, there are no switching costs for AI models, making it easy for users to switch between different models. However, with continual learning, enterprises will likely prefer to stay locked into a model provider to benefit from the model's continuous improvement.

This shift in dynamics may lead AI labs to offer subsidized usage to enterprises in exchange for access to their data and valuable feedback, similar to how Google offers search services for free.