Very simple question, very simple requirement, a high quality train of my thoughts is below, what do you think?
What do others think about this idea that to make GPT more human like: GPT needs to be able to change its cursor location in notepad. ? I mean GPT only writes the next token at the end, shouldn't GPT have the ability to predict GoBack and then that triggers associated notepad motors to change its cursor location so to change where it reads and writes? Or is it possible just sending the whole story through an ANN would let it choose where to focus on? [M]aybe the ANN could learn that when it reads "tell me the 1st and last letter of this story" to grab them from such locations[?] I have marked above such letters [M] and [?], so my answer here is "M?". Even if it could, wouldn't this ANN spend more compute and RAM trying to create a "neural program" instead of some learnt token<>motor assocaitions? Another thing to consider is how does one train such AI on these? So for "JumpManyLettersBack", it'd need to either randomly generate an action that leaps its notepad cursor location back far, or be guided by human by mimicking? Thinking... Whereas the ANN approach would be doing it itself... But wait 🙂 How would such AGI be able to look back on a Real Piece Of Paper in real life on a table using its eyeballs? Ok wait, the AGI could first read in the whole story on a Real Piece Of Paper in real life, and THEN send it through an ANN to be able to choose where to read and write. But, what if it is a long story and it says right at the start "Skip this story, go to end"? Though one could argue the real world is much more limited or slower maybe I mean. The AIs could just read and write in their brains to work on code upgrades to themselves, no need for real life experiments here really. This brings us back to where I made this comment above "Whereas the ANN approach would be doing it itself..." Furthermore, to be able to think of raising your hand in real life but resist doing it, requires the AGI to have a switch that allows associated motors to fire. So if the AGI has the goal/strong prediction "do it, don't withhold" + "the envision actions/ scene to do", then the senses with link to associated motor actions and the switch will open the link between them so that it will be done in Real Life and not just daydreamt. Do you agree? I am looking for collaborators and agreement. If these 2 things are added to GPT, it would give it much more agency and control. Currently GPT cannot edit its completion or prompt, or change where to read ex. the start again, nor does GPT have control over to only_think_of_doing_such in notepad or to do it for real. Don't forget that DALL-E and Facebook's Blender is a step up from GPT-3, closer to AGI. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta334a9577dc3538a-M32c770b676aa5898076cb1fc Delivery options: https://agi.topicbox.com/groups/agi/subscription
