Well, my post was a bit of humor, but yes I know Brian and have a lot of respect

Well, my post was a bit of humor, but yes I know Brian and have a lot of respect for him. As for our work, it’s extremely rigorous and there are challenges incorporating it into the LLM architecture because by default the reason it’s weak in reasoning is the same reason it’s weak at my work. We need to work hard to (a) proceduralized a chain of thought, and (b) train it to use a set of terms as invariants which is counter to the current LLM architecture. Lastly (c) LLMs are very shallow at reasoning, (terribly) so we need more attention nodes so to speak and to consume more working memory. Unfortunately human intuition is ‘off’ in that math is simple, programming a little less so, physcal operations much less so, behavioral operations much much less so, and verbal reducibility effectively infinite. So there is a great emphasis in the research community on what is effectively simple, and zero research on what is effectively hard. We do what is hard. So we are always working against the rather shallow limits every AI. And only ChatGPT is even worth trying to use.

Reply addressees: @MetaPrime001 @BrianRoemmele @ceobillionaire