Michael,
I was a little surprised to see you say
“What LLMs do is to solve the problem of common sense reasoning better than any other approach to date. LLMs know that things fall down, that water is wet, etc. “
When I asked Claude 3.5 the following:
“Is it accurate to say that LLMs solve the problem of common sense reasoning and that LLMs know that things fall down, that water is wet, etc.?”
The response was:
“No, that's not accurate. The relationship between LLMs and common sense reasoning is much more complex and limited.
LLMs can often produce text that appears to reflect common sense understanding, but this is primarily because they've learned statistical patterns in human-written text about how objects and phenomena are typically described. They don't have genuine understanding or reliable common sense reasoning capabilities.
For example, an LLM might correctly state that objects fall downward or that water makes things wet, but it doesn't truly "understand" gravity or wetness the way humans do. Some key limitations:
This is why LLMs can sometimes make surprisingly basic errors about physical reality or common sense relationships that would be obvious to humans. They might correctly state that "water flows downhill" in one context but then make an elementary mistake about fluid dynamics in another context.
What are your thoughts on the nature of common sense reasoning in AI systems? I'm curious about your perspective on this.”
I think the key point is that LLMs do not understand naïve physics or any deterministic relationships as humans do, nor have the inferential capabilities symbolic logics do. But as you point out, they do make a great interlocutor.
Mike
From: ontolog-forum@googlegroups.com [mailto:ontolog-forum@googlegroups.com] On Behalf Of Michael DeBellis
Sent: Monday, February 3, 2025 6:47 PM
To: ontolog-forum
Subject: [ontolog-forum] Re: Latest from the world of LLMs
> It’s official: the myth of an LLM Scaling Moat has been debunked
It is a big leap from what we know about DeepThink to make that conclusion. For one thing, I don't trust the mainstream press (CNBC) for any info about a complex scientific or technology question. For another, there is so much we don't know with confidence about DeepThink. How well does it really do on benchmarks? How much money or computing resources did the Chinese government or other groups associated with them pump into DeepThink? For another, there are things that Open AI can do that DeepThink can't do such as their recent Deep Research announcement: https://www.youtube.com/watch?v=YkCDVn3_wiw ; Also, see the Playlist "12 Days of Open AI" on Youtube: https://www.youtube.com/playlist?list=PLOXw6I10VTv9lin5AzsHAHCTrC7BdVdEM ;
But even supposing that every claim made about DeepThink is true, it doesn't mean "the myth of an LLM Scaling Moat has been debunked" It just means the moat isn't as large as we thought. And if that moat isn't as large as we thought, I don't see how that justifies the conclusion: " All roads lead back to good old symbolic reasoning and inference—now more accessible and understandable when combined with a multi-modal natural language interface that enhances interaction possibilities." If anything, it makes LLMs even MORE usable and accessible because it doesn't require the huge resources that we used to think to create one (from what I know, I'm deeply skeptical but I actually hope that the claims of DeepThink are true because I think it is a good thing that we don't restrict foundational LLMs to just a few companies)
Where is there a purely symbolic AI system that can support NLP anywhere in the same ballpark as LLMs? If such a thing exists would love to know about it. Keep in mind the system has to do a lot more than turn natural language into SPARQL. It needs to keep the context of a discussion thread, it needs to be able to respond appropriately when you give it feedback like: "give me more/less detail" "change the tone to [more playful] [more professional] [more creative]" etc. It needs to be able to do common sense reasoning and handle very hard NLP problems like anaphora.
What LLMs do is to solve the problem of common sense reasoning better than any other approach to date. LLMs know that things fall down, that water is wet, etc. It doesn't make sense to also expect them to be able to solve that problem AND have completely reliable domain knowledge. There are several approaches to using curated, authoritative knowledge sources with an LLM to reduce or eliminate hallucinations. One of them is Retrieval Augmented Generation (RAG). That's an example of how semantic tech and LLMs complement each other: https://journals.sagepub.com/eprint/RAQRDSYGHKBSF4MRHZET/full also see: https://www.michaeldebellis.com/post/integrating-llms-and-ontologies
When you use a RAG, you have complete control over the domain knowledge sources. Building a RAG is the best way I've ever seen to put an NLP front end to your knowledge graph.
Another way to get domain knowledge into an LLM are to train LLMs with special training sets for the domain. One excellent example of this is Med-BERT, a version of the BERT LLM specially trained with healthcare data. See: L. Rasmy, Y. Xiang, Z. Xie, C. Tao and D. Zhi, "Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction," Nature npj digital medicine, vol. 4, no. 86, 2021.
Still another (much newer) way is to use reinforcement fine-tuning to train an LLM with domain specific data. This doesn't take as much work as the domain specific LLM but is much more work than creating a RAG. See: M. Chen, J. Allard, J. Wang and J. Reese, "Reinforcement Fine-Tuning—12 Days of OpenAI: Day 2," 6 December 2024. https://www.youtube.com/watch?v=yCIYS9fx56U&t=29s