Artyom BologovUpdating my projects page <https://aartaka.me/projects.html> made me recall that I did a mostly compliant #Eliza re-implementation <https://codeberg.org/aartaka/laliza> in the past. Now I’m tempted by it again, but want to add some more features to it, like nested structures and symbol explosions. This will make it a reasonable #programming system, actually!
DougMerritt (log😅 = 💧log😄)@aartaka
That's sufficient for you to immediately go public with a $4 trillion valuation.
That's sufficient for you to immediately go public with a $4 trillion valuation.
@dougmerritt I’m right on the AI hype wave here!
Ramin Honary@aartaka
Recently I have been genuine curious as to what would happen if you told an LLM to write a bunch of grammar rules for an implementation of ELIZA , I wonder if we had an LLM recreate enough of itself as ELIZA rules would we be able to ditch the LLM and just run ELIZA? It would be really funny if an LLM-generated ELIZA rule set could actually handle 90% of what most people are using LLMs to do right now. Plus the fact of commanding an LLM to do the work that would replace itself, rather than a tech-bro boss telling their employees to train an LLM to replace themselves, is pretty ironic too.
@dougmerritt
Recently I have been genuine curious as to what would happen if you told an LLM to write a bunch of grammar rules for an implementation of ELIZA , I wonder if we had an LLM recreate enough of itself as ELIZA rules would we be able to ditch the LLM and just run ELIZA? It would be really funny if an LLM-generated ELIZA rule set could actually handle 90% of what most people are using LLMs to do right now. Plus the fact of commanding an LLM to do the work that would replace itself, rather than a tech-bro boss telling their employees to train an LLM to replace themselves, is pretty ironic too.
@dougmerritt
@ramin_hal9001 actually, @cwebber made a good write-up on one of such experiments, where an LLM was “persisting” its state to a Datalog database, and generating Datalog queries as rules for its “behaviors”:
https://dustycloud.org/blog/an-ai-called-winter-neurosymbolic-computation-or-illusion/
But yes, creating a lot of efficient Eliza rules might work as well! Weizenbaum initially planned Eliza to be a math assistant, and therapist script was a byproduct of experiments with that. Math assistant has never happened, but it well might someday!
@aartaka
Such things are only getting a fraction of the headlines, but hybrid systems of various sorts are all the rage with some.
Currently it seems like the logical way to go -- certainly as opposed to trusting LLMs to do everything.