3 ms·

I’d be interested in modern perspectives on Cyc and Eurisko. Hopefully I’ll be forgiven for this impression, but I’ve only heard of Cyc being legendary for basi

by thom 5y ago

I’d be interested in modern perspectives on Cyc and Eurisko. Hopefully I’ll be forgiven for this impression, but I’ve only heard of Cyc being legendary for basically wasting Lenat’s career after the promising results from Eurisko. It’d be an interesting debate, whether recent advancements like GPT-3 confirmed or refuted the intuition that architecture isn’t enough, we need to encode knowledge. Feels like we have settled on the idea that a good enough architecture (and enough compute time) can take care of the latter. But it’d still be valid to claim GPT-3 lacks common sense. I still think Eurisko is fascinating. Stochastic, heuristic systems seemed to reach their highest point with random forests and subsequent gradient boosting approaches. But we’re still obsessed with quite static models over living, learning agents.

gumby 5y ago
> I still think Eurisko is fascinating Note that Doug later wrote a paper “Why AM and Eurisko appear to work” (emphasis mine). People are still using semantic KBs like OpenCyc for research. > …confirmed or refuted the intuition that architecture isn’t enough, [that] we need to encode knowledge For my position on this: I left the project for a number of reasons, but one fundamental one was that I perceived the corpus to be largely pointless (note the countervailing position above, in that people still get value from them today ). Architecture alone is clearly inadequate as can be seen from the fragility / inadequacy of current systems, yes, including GPT-n. They remind me of things like the X-15: providing valuable data but ultimately not on the path to space flight. Knowledge acquisition is ad hoc, and more importantly the graph connections are highly as hoc, but dependent on architecture and prior state. I no longer believe (perhaps I never did) that you can meaningfully have one and not the other. After leaving Cyc I went into a different direction (e.g. Cygnus) and only returned to symbolic AI a few years ago working from a constructivist model. Unfortunately I had to step away from that work last year due to a more pressing problem. But I think it’s ultimately a more powerful approach.

by thom 5y ago

gumby 5y ago

> I still think Eurisko is fascinating Note that Doug later wrote a paper “Why AM and Eurisko appear to work” (emphasis mine). People are still using semantic KBs like OpenCyc for research. > …confirmed or refuted the intuition that architecture isn’t enough, [that] we need to encode knowledge For my position on this: I left the project for a number of reasons, but one fundamental one was that I perceived the corpus to be largely pointless (note the countervailing position above, in that people still get value from them today ). Architecture alone is clearly inadequate as can be seen from the fragility / inadequacy of current systems, yes, including GPT-n. They remind me of things like the X-15: providing valuable data but ultimately not on the path to space flight. Knowledge acquisition is ad hoc, and more importantly the graph connections are highly as hoc, but dependent on architecture and prior state. I no longer believe (perhaps I never did) that you can meaningfully have one and not the other. After leaving Cyc I went into a different direction (e.g. Cygnus) and only returned to symbolic AI a few years ago working from a constructivist model. Unfortunately I had to step away from that work last year due to a more pressing problem. But I think it’s ultimately a more powerful approach.