3 ms·

Has anyone investigated to see if OpenCyc can be converted to natural language (presumably English) and then injested into this? Cyc made an attempt years ago t

Has anyone investigated to see if OpenCyc can be converted to natural language (presumably English) and then injested into this? Cyc made an attempt years ago to "encode common sense" and a subset called OpenCyc was released. That might be a great way to kickstart information representation of the real world. The latest version of Cyc is proprietary but I think there OpenCyc is an open subset (though I'm having trouble confirming that, so the licensing may not be good). Some links: https://github.com/bovlb/opencyc https://github.com/bovlb/opencyc https://github.com/asanchez75/opencyc https://github.com/asanchez75/opencyc

sp332 3y ago
I've been wondering this for a while now. Cyc has tons of knowledge in a white-box, formal system. If it just had a front-end that could convert from natural language to Cyc knowledge queries and back, we wouldn't have to worry so much about hallucinations, catastrophic forgetting, or trying to fit the entire database in VRAM.
speed_spread 3y ago
My understanding is that LLM and Cyc are fundamentally different forms of AI. Even if you could turn OpenCyc into text rules, once ingested it would just dissolve into the ocean of training text data and would not significantly gain more apparent "common sense" than it already had. Maybe a more interesting combination could be to have both Cyc and LLM working side by side and comparing notes before agreeing on a result.
- btown 3y ago
  GPT-4 seems capable of creating CycL output from a text prompt. It might be an interesting guard against hallucination - much like a student being asked to show their work, you're forcing the LLM to go through the steps of framing the problem logically, in a way that's interpretable by the teacher independent from the student's thought processes. That said, it certainly seems like there hasn't been recent work on hosting the OpenCyc knowledge graph in a reasonably modern way, much less the more recent closed-source work by Cycorp (https://cyc.com/ https://cyc.com/). And it's likely GPT-4 doesn't know the full capabilities beyond whatever tutorials were on the web at the time of its training. If I were Cycorp I'd be seriously looking at developing this kind of hybrid model, with an agent model having access to recall their closed-source examples, as a paid cloud offering; there would likely be many who would desire this best-of-both-worlds.

sp332 3y ago

I've been wondering this for a while now. Cyc has tons of knowledge in a white-box, formal system. If it just had a front-end that could convert from natural language to Cyc knowledge queries and back, we wouldn't have to worry so much about hallucinations, catastrophic forgetting, or trying to fit the entire database in VRAM.

speed_spread 3y ago

My understanding is that LLM and Cyc are fundamentally different forms of AI. Even if you could turn OpenCyc into text rules, once ingested it would just dissolve into the ocean of training text data and would not significantly gain more apparent "common sense" than it already had. Maybe a more interesting combination could be to have both Cyc and LLM working side by side and comparing notes before agreeing on a result.

btown 3y ago
GPT-4 seems capable of creating CycL output from a text prompt. It might be an interesting guard against hallucination - much like a student being asked to show their work, you're forcing the LLM to go through the steps of framing the problem logically, in a way that's interpretable by the teacher independent from the student's thought processes. That said, it certainly seems like there hasn't been recent work on hosting the OpenCyc knowledge graph in a reasonably modern way, much less the more recent closed-source work by Cycorp (https://cyc.com/ https://cyc.com/). And it's likely GPT-4 doesn't know the full capabilities beyond whatever tutorials were on the web at the time of its training. If I were Cycorp I'd be seriously looking at developing this kind of hybrid model, with an agent model having access to recall their closed-source examples, as a paid cloud offering; there would likely be many who would desire this best-of-both-worlds.