I think this offers an interesting possibility for another way to safely allow users to get benefit from a strong AI that a company wishes to keep private. The user can submit a design specification for a desired task, and the company with a strong AI can use the strong AI to create a custom dataset and train a smaller simpler narrower model. The end user then gets full access to the code and weights of the resulting small model, after the company has done some safety verification on the custom dataset and small model. I think this is actually potentially safer than allowing customers direct API access to the strong model, if the strong model is quite strong and not well aligned. It’s a relatively bounded, supervisable task.
Existing large tech companies are using approaches like this, training or fine-tuning small models on data generated by large ones.
For example, it’s helpful for the cold start problem, where you don’t yet have user input to train/fine-tune your small model on because the product the model is intended for hasn’t been launched yet: have a large model create some simulated user input, train the small model on that, launch a beta test, and then retrain your small model with real user input as soon as you have some.
I think this offers an interesting possibility for another way to safely allow users to get benefit from a strong AI that a company wishes to keep private. The user can submit a design specification for a desired task, and the company with a strong AI can use the strong AI to create a custom dataset and train a smaller simpler narrower model. The end user then gets full access to the code and weights of the resulting small model, after the company has done some safety verification on the custom dataset and small model. I think this is actually potentially safer than allowing customers direct API access to the strong model, if the strong model is quite strong and not well aligned. It’s a relatively bounded, supervisable task.
Existing large tech companies are using approaches like this, training or fine-tuning small models on data generated by large ones.
For example, it’s helpful for the cold start problem, where you don’t yet have user input to train/fine-tune your small model on because the product the model is intended for hasn’t been launched yet: have a large model create some simulated user input, train the small model on that, launch a beta test, and then retrain your small model with real user input as soon as you have some.