3 ms·

This is a step towards what I think AGI would be: a model trained on visual and language data from the Internet used as a prior for an action model (finetuned p

by GaggiX 3y ago

This is a step towards what I think AGI would be: a model trained on visual and language data from the Internet used as a prior for an action model (finetuned perhaps with reinforcement learning) that would be able to learn how to use the frozen prior to make useful actions; I would separate the prior from the action model so there is no possibility of catastrophic forgetting (this is dealt by co-fine-tuning the model in the RT-2 paper), and because a more advanced robot would need to control more actions quickly, so it would be expensive to run the entire prior just to do basic movements in real time. Edit: also, the model should be smart about how to use its "context window", when the robot is taught how to do a task and it needs to do that task, it must retain the knowledge.

valine 3y ago
That seems like a real step in the direction of AGI but not anywhere near the full solution. The current context windows are far too small to replicate human intelligence. I can’t quite put my finger on it but it feels like we are missing a sort of bridge between learning done in-context and offline training. A true AGI would be able to learn in-context and then quickly apply those lessons to the base model. If in context learning is analogous to a person’s short term memory, we need a mechanism to move short term memory to long term memory. In the near term I expect we will see much more general robotics that know how to do lots of tasks and can follow basic instructions, but lack the ability to develop complex new skills over time. Robots doing dishes and laundry will soon be feasible, just don’t expect unbounded self improvement.
- GaggiX 3y ago
  I agree that it would be much better if a model could move the knowledge to the weights, but I don't think it's necessary. Models like RNN have practically infinite lossy "context window", and I think it would be much more reasonable to find an RNN architecture that would scale like a transformer than what you propose, while still achieving our goal. An alternative analogy you can think of is that the frozen weights in the model are our DNA, while the state or "context window" of the model is our state.

by GaggiX 3y ago

valine 3y ago

That seems like a real step in the direction of AGI but not anywhere near the full solution. The current context windows are far too small to replicate human intelligence. I can’t quite put my finger on it but it feels like we are missing a sort of bridge between learning done in-context and offline training. A true AGI would be able to learn in-context and then quickly apply those lessons to the base model. If in context learning is analogous to a person’s short term memory, we need a mechanism to move short term memory to long term memory. In the near term I expect we will see much more general robotics that know how to do lots of tasks and can follow basic instructions, but lack the ability to develop complex new skills over time. Robots doing dishes and laundry will soon be feasible, just don’t expect unbounded self improvement.

GaggiX 3y ago
I agree that it would be much better if a model could move the knowledge to the weights, but I don't think it's necessary. Models like RNN have practically infinite lossy "context window", and I think it would be much more reasonable to find an RNN architecture that would scale like a transformer than what you propose, while still achieving our goal. An alternative analogy you can think of is that the frozen weights in the model are our DNA, while the state or "context window" of the model is our state.