3 ms·

Somewhat? On the one hand, replacing a deterministic system with a probabilistic one is generally bad. Why replace something that's 100% accurate all the time w

Somewhat? On the one hand, replacing a deterministic system with a probabilistic one is generally bad. Why replace something that's 100% accurate all the time with something that is... somewhat less than that? On the other hand, the semantic web has been a complete failure. Even present attempts at semantically understanding the web (aka what search engines do all day) require highly complex heuristic and machine-learned solutions... so the "fully deterministic" thing is only technically true, but not meaningfully useful in any sense. Even the actual semantic web standards that exist (and are poorly adopted) don't seem to really solve the core problem with the semantic web: there's no such thing as a universal ontology. Every attempt to produce a broad-based shared ontology has failed - there are fundamental disagreements about how to organize information, and information organization needs turn out to nearly always be application-specific, and be highly unstable over time, preventing there from being a useful, durable universal standard for various topics. So I guess long story short: we need AI (or just highly advanced heuristic methodologies) to scrape useful information out of natural content. We always have, because the deterministic alternative has never actually worked.

alm1 3y ago
I think the deterministic requirement was what made the semantic web nearly impossible at scale. Aligning on the schema for all the entities can be challenging within a small group, yet alone the world. Doing it continuously is impossible today. Wikidata has tons of abandoned and branches which are no longer relevant or even wrong. It seems like LLMs and vector representations for words are more practical ways to explore semantic data. Then the specific, reccuring queries may be optimized through graph representations.
- potatolicious 3y ago
  A good way to think about it is that the "semantic web" was all about finding some kind of universal data representation for the world's data, and that has failed in large part because data representation is inherently application-specific. The only practicable version of the semantic web needs to embrace that reality. We need tools for converting, extracting, and interpreting data from one application-specific data representation to another. Sometimes that's LLMs, sometimes it can be done heuristically, but either way, the notion of the universal data structure simply isn't realistic.

alm1 3y ago

I think the deterministic requirement was what made the semantic web nearly impossible at scale. Aligning on the schema for all the entities can be challenging within a small group, yet alone the world. Doing it continuously is impossible today. Wikidata has tons of abandoned and branches which are no longer relevant or even wrong. It seems like LLMs and vector representations for words are more practical ways to explore semantic data. Then the specific, reccuring queries may be optimized through graph representations.

potatolicious 3y ago
A good way to think about it is that the "semantic web" was all about finding some kind of universal data representation for the world's data, and that has failed in large part because data representation is inherently application-specific. The only practicable version of the semantic web needs to embrace that reality. We need tools for converting, extracting, and interpreting data from one application-specific data representation to another. Sometimes that's LLMs, sometimes it can be done heuristically, but either way, the notion of the universal data structure simply isn't realistic.