An AGI needs to be able to recognize abc when it sees azbzc, or cba, and when 
it sees arc. Less if delay and less if missing letters. Thsankivngi turkeey was 
so _?_. So that you can get a text-in text-out natural GUI and can just 
communicate to the AI and allow it to understand data to other data and not be 
so weak in seeing things, for predicting better.

Among the biggest things, currently my AI uses multiple length exact matches 
ex. [a[b[c]]]>predictions, and recent stats are boosted for recency (Matt calls 
it ICM). But not fuzzy matches yet.

GPT USES DROPOUT TO SOLVE THIS. It makes it not depend on certain inputs so 
much. It's a regulation technique.

GPT uses its word2vec/ seq2seq (related words tree) to recognize translated 
matches too cat=dog>p to get predictions for cat ex. cat>barked.

Others appear to not know WHY gpt uses those methods. In my AI and Matt's as 
you can see ours is directly tackling the biggest nodes in the tree storage and 
handling all the probabilities of the distribution, it doesn't need complex 
backprop etc. I still have to prove this but really it should not require much 
much more code than my current code yous saw to do the 2 mechanisms above.

Once I add those 2 things above my AI should predict much more like gpt. I also 
think I'll need to add a long dependency mechanism so it can complete the rest 
of a memory at far away ex. the cat which was sold during the night when i was 
home jumped over the couch. Another example: And he said "this was cool, ya, 
and if we sort it out we can buy more", where the " are at the start and end of 
the passage, see? "passage".

The reason I'm not using GPT architecture is because it is seemingly more 
complex code and more inefficient to run and seems to do things like above 
where nobody says it IS doing that hence how do you know it is fully doing 
Translation and holed/delay matches if you don't know how ANNs work 
inside?????, I know they know hoe backprop works and error but they son't talk 
about related matches and fuzzy matches, just "Dropout" and hill climbing, and 
I need to understand the code fully anyway so that I can improve the prediction 
ability.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T7bd650af76f955ee-Mea25894e7d905efcc44f167c
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to