I just saw the slideshow. The gazelle image is pretty. Do you know how many possible English sentences there are that can be written? You have to use partial match technology. You mention in the slides the word 'and' can be used to clarify a meaning behind bob etc, oh boy, and we can also say bob isA person. But you still have to teach it by hand every semantic relation using isA etc and what the meaning of the sentence is ex. "ate pizza with bob" = "me and bob ate pizza" not "bob tasted so good".
Let's step back. All we can do in text is syntax and semantics. Aka cat>ate or cat=dog. With that you can do cat....dog ate.....therefore cat ate (you didn't know that prior). All these questions above like who ate the pizza and was bob in the pizza as a food item and was a fork in my hand and who is 'they' refer to and are these words rearrangeable? and where did the dragon land or who is the dragon and why does rain fall, are syntax and semantics. GPT-2 does this. BERT does this. Transformers (the architecture behind gpt2 and bert), do this. You are looking for a match in memory to answer the question as accurate as can. And I imagine if GPT-2 does deep nesting commonsense reasoning it can predict even better! Currently the best data compressors are amazing but can't talk like GPT-2 still, hence should be improvable! ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T21bdc2c440c86db7-Mc17655bdd71507dda41fd768 Delivery options: https://agi.topicbox.com/groups/agi/subscription