OpenCog by Ben Goertzel working on AGI since 2008

opencog.org

2 points·seansh··3 comments

I don't know if the project is still active or not but its architecture wasn't based on neural networks. From the website:

"The AtomSpace is a generalized graph store, with a sophisticated graph query engine. This query engine far exceeds the features and capabilities of popular commercial and open source graph stores. Atomese is a graph language, optimized for manipulation by algorithms. The intent is that automated agents, reasoning subsystems and processing pipelines can work with a rich and sophisticated language system. Think of Atomese is "assembly code", or, better yet, an "intermediate language"."

I recall there was a video somewhere of a virtual agent that was walking around in a virtual world, figuring out creative ways to find food, seeking shelter etc.

Does anyone know how far they got? and if they gave up after ML revolution?

Development seems to be alive, according to:

https://github.com/opencog/

Edit: Even more interesting is this repo of theirs:

https://github.com/opencog/opencog-cycl

They’ve rearchitected the approach to incorporate neural networks alongside their existing symbolic approach, in a new design called OpenCog Hyperon. This work is currently at: http://hyperon.opencog.org. Their main wiki hasn’t been updated yet for some reason. Development is definitely active! The founder shared his thoughts on LLMs here: https://youtu.be/MVWzwIg4Adw. The TL;DR is LLMs currently do not generalize well, so they aren’t sufficient for AGI unless they end up generalizing better somehow (perhaps at scale).

Reply on news.ycombinator.com