Ask HN: Is Web 3.0 close to be solved by LLMs (+ logic frameworks)? | Hacker News Reader

Ask HN: Is Web 3.0 close to be solved by LLMs (+ logic frameworks)?

4 pointswslh3 comments

Web 3.0 (the semantic web) [1] never had traction or impressive results. DBpedia and Freebase are/were clear examples that if we structure realities (whatever it means) into data structures such as RDFs we can create more impactful queries and results than normal search engines and databases. IMHO the main obstacle for the semantic web was transforming content to a standard data structure. Once this is [quasi]solved the other processes are well known in computer science.

I was playing with LLMs to convert content to RDF or JSON structures with good results. You can interact with an LLM and ask to create those structures and insist on going deeper with some guidance.

Do you think we are there or close enough?

[1] https://en.wikipedia.org/wiki/Semantic_Web

I believe that LLMs are quite useful for sentiment analysis, and could be quite helpful as an aid to search engines. However... the best way to do it would be for LLMs to ingest the entire web and give those analysis for each and every bit of text and image out there.... which is computationally infeasible at present.

If Moore's law can hold for a few more generations of chips, it might become possible.

LLMs are not going to be reliable search engines, but most people will likely be happy with them, especially due to the consistently falling quality of Google's results in comparison. Things that would have been laughed at a decade ago will now be seen as an improvement.

We won’t need semantic web if machines are able to understand natural language.

Eh, I could see this but I think having well-structured data will still be useful for both cutting down on 'prompt clutter' and giving it a better chance to come up with something useful.

It might get better in later versions but even so, I think passing well-structured data will always result in cleaner output than passing it data with lots of noise.

If we are going to work alongside AGIs, we should consider them equal to humans and by that I mean, passing them good data rather than assuming they'll figure out what we meant. Obviously it won't be _as_ big of an issue or requirement but if you can put into such a format, why wouldn't you?

Reply on news.ycombinator.com