What do you think is the most impressive capability that the system displays atm which you think most systems struggle with?

Most of this wiki is written up was the theories that are completely different from current neuroscience/psychology... But what you ask is much more important for people to understand.. So going to take some time to write up the answer and share it.

One reason LMs are so popular is that despite they are not as capable as GOFAI techniques , they require so little "human in the loop" work,  to do so much!    So they assume they have tapped into some important substrate (some building block of [general?] intelligence).  And it  it seems to agree with Hebbian theory. This mistake could easy trick our project (or Cycorp's) down our road with a sort of "false flag" results. 

Sorry, the "they are not as capable as GOFAI techniques" are fighting words...  comparing the two is wrong though .. it be like trying to compare a LM trained to do long division with a pocket calculator.. (of course the pocket calculator will always win!)  we see using LMs at NLP to have had this same problem when trying to compete with GOFAI's NLP at problems like https://nlp.stanford.edu/~wcmac/downloads/fracas.xml

GOFAI people (like us) tended to pretend the ML leaderboards are silly and dont bother to compete in them. we had intended to spend a few months at competing to get funding.. On a short timeline we wouldn't have time to also do our own research.. then again we still might have to participate in the leaderboard competitions for the sake of money or build a team?

We believe less than 100% (except maybe test cases where multiple humans could disagree) found here https://nlp.stanford.edu/projects/snli/ is not actually using the correct mental substrates

Logicmoo system is designed to get 100% on these tests... Today it gets 10% .. 6 months from now 100%.. we am sure of this due to the questions it gets correct... also on ones it gets wrong it shows us proof of why it gave the answer it did which allows us humans to fix its misunderstandings.. todays ML wont ever do that since they believe they must allow rolling of dice.



 

 

Tags:
     
Copywrite © 2020 LOGICMOO (Unless otherwise credited in page)