KnowComp HKUST รีทวีตแล้ว
What kind of research should I do in the era of LLMs? Nowadays, some people get so excited to work on LLMs and the other feel so suffering as they don't have GPU cards to work on LLMs. IMHO, there are two ways to think about this: (1) What LLMs cannot do and (2) What LLMs have enabled. In this post, I will focus on the second problem, what they have enabled.
If you browse recent arxiv papers, many are working on Agent, Tool use, RAG. I agree they are very interesting and important problems. But at university, we'd better think of something one step further. For example, for RAG, it would be better for us to think about what if the external resource can also make predictions. Particularly, with LLMs, we can improve many semantic matching results in traditional database queries with free text attributes. Nowadays, we are in the era of "neuralizing everything", compared to what we were doing to "digitalize everything". With those neuralized models, agents, agent-tools, tool-tool's interactions can be all memory interactions, which naturally fit Yann LeCun's world model. In this regard, instead of worrying about having no machines to train LLMs, we can focus on improving the predictability of tools. Now it's time for us to leverage LLMs to improve the prediction power of DBs/KBs.
In particular, I love the idea of neural graph databases (ngdb.org), which not only do DB/KB embeddings, but are also equipped with logical query power. I believe this is one of the most important developments in the field of data mining! In this year's KDD, we will present our work on Privacy-Preserved Neural Graph Databases(arxiv.org/abs/2312.15591). This could be particularly useful for the next generation of RAG systems. You are welcome to attend our talk and have more discussions with us then.
English
