Hi EatingNoone!

I see your desire for figuring out how to build AGI sooner instead of later.

You mention above what Hinton calls "equivalence" for recognizing rotated/ 
scaled/ offset image objects/ features (or text objects like w a l k, w a  l   
k, wwaallkk, klaw, and "walk too fast" <--> "walk just too very fast").

But why would you think this is a problem or THE problem when 1) DALL-E looks 
like it solves this as seen on openAI.com, and 2) I know of much bigger 
requirements to make DALL-E into AGI, and it isn't about exactly improving 
Prediction/Recognition abilities, it's something even more ground breaking and 
unifying.

We are close to AGI if it thinks like us, so far if you look at openAI.com you 
can see it nearly does. And Jukebox generated me human level music completions. 
Perplexity and the Hutter Prize's Compression evaluations also help BUT they 
the Prediction Scores do not tell you how close to AGI you are, without looking 
at the actual generated "prompt completions" as openAI.com calls them. The best 
evaluation is how long the machine survives in the universe, but that takes way 
to long to test. If you look at DALL-E, you can see where it needs to improve 
on to "talk and think" more like a human brain. Notice it seems a bit stagnant. 
I can say but not here in public.

BTW notice DALL-E can be asked to give the same image on the right but with a 
word on it, or broken, or super-resolved, or deblurred, so it is much more 
general than text only etc. This is becoming the G in AGI more now.

Currently OpenAI.com released a word-based-math problems dataset in a blog 
post, this would be the perfect testbed for my innovating discovery I 
pinpointed we need add. I guess it would work on many of the long standing hard 
prompts though, but maybe this would be more helpful for testing on.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tabd1763b276303cf-Mb6acbd49d642a69c44bde4ed
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to