Jim, I look forward to hearing your thoughts. immortal.discoveries "equivalence" seems to indeed be talking about same problem. Where something is different by one measure like position but the same in another measure. The simplest solution is to approximate the difference away by accumulating more data, but that makes large assumptions on the availability of data, time, and resources.
To be clear I don't think invariance/equivariance is the important problem that needs to be solved to achieve AGI. Only that it is the problem that holds back the model I referenced from being considered as a candidate for AGI. Some of these more important problems to solve include: planning, motivation, abstraction, memory, real-time learning, temporal causation, and the ability to operate with real world input data. I only invested so much time into the aforementioned model due to it seeming to possess many of these capabilities. My understanding is that large transformers (like DALL-E) are able to achieve a kind of contextual invariance/equivariance if fed enough data. It is an emergent property that comes with the amount of data being provided. That is ultimately the same answer that was given by the author of the aforementioned work. Large quantities of well labeled data, containing enough variation required to achieve this emergent property are not common in the real world and are not efficient. DALL-E cannot update what it knows in real-time, you cannot reuse the same trained model for different tasks, training it is extremely resource intensive, its generality is precluded upon massive datasets which have to be carefully curated by humans, and OpenAI describes various situations where it is brittle. This is not to mention that due to its designated purpose it has almost no agency. That is not to say that DALL-E and GPT-3 are not amazing achievements. The models that OpenAI produces generate results that I never thought would be possible on the tasks they are designed for. Its just that these things alone do not make me think of AGI. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tabd1763b276303cf-Mec331d1331fb989f95c237b5 Delivery options: https://agi.topicbox.com/groups/agi/subscription
