A question to all AGI professionals reading this: What is your Evaluation method for testing/improving your AGI? Typically most use Perplexity or chatting with their generator to see if the results/predictions look good.
I use Lossless Compression instead of Perplexity though because 1) you can understand the same training data as test data and hence compress/understand the training data the most because of better feedback, 2) forces Online Learning, 3) can't accidentally get the test data in the training data, and 4) is funner compressing data. AGI is all about Patterns, and compression is all about finding patterns. If there is only noise (random) left over, no further compression can be done, no hints. So yes, prediction/statistics is AGI; using past experience that shares contexts. The "recognition" part of AGI is used for this, to do induction during tasks like Entailment and Translation. Needs: Often we want related or desired or common (popular) content/answers. Tasks: And often we want entailments, translation, or summarization. I don't know of any other need or task besides those above. Hence, for Evaluation, if you can add something more, I'd be very interested! Improving frame rate of videos falls under the above, and walking on a tight rope falls under Prediction because you want a specific outcome and know how to get it. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T2a0cd9d392f9ff94-M831272b0908808532824a4b4 Delivery options: https://agi.topicbox.com/groups/agi/subscription
