cross-posted from: https://lemmy.world/post/34416839
The fundamental idea of this paper is for ChatGPT-like apps to lose natural language for less energy consumption and more determinism in their answers based on controlled natural languages like ACE; for the user to be able to modify this trade-off-ratio at will based on LLMs (which is not possible when starting from a ChatGPT-like app); and to capture this new paradigm in a new type of browser that has natural language as its primary interface, here called a semantic web-first browser.
this paper isn’t really about llms at all, they’re just using the natural language interface of chatgpt as an example for how to interact with a web browser. they suggest restricting to ACE in order for the language parsing to be made simpler.
It does use LLMs with the precision controller - which determines the degree to which ontologies are applied to not-semantic-web-data. This is done with LLMs as translation tools if the precision controller is adjusted with a low precision degree (ergo more fuzziness, ergo more LLMs).
But yeah, it’s basically meant as a contra-approach to ChatGPT-like apps, giving users back control. And because rule-based systems have always failed in the past, LLMs are used if the user wants to make the app more fuzzy and access more data this way.
yeah but that’s sort of external to the core of the paper. since they would be invoked by the browser to figure out inconsistencies in parsing, they’re not really part of the described interface.
Right, but the interface is only one part of the program (which has already been implemented in practice). Another part is the precision controller.
right, i didn’t clock that you were the author.
"
Basically with an app like ChatGPT, you have before you a black box that you can send commands to and that gives you unpredictable answers and consumes huge amounts of energy.
Instead, the semantc web browser with the precision controller starts with a complete white box, where you can control the ontology, have full control of the language, the outcome and consumes much less energy; but you can move on the ratio towards a ChatGPT-like black box with the precision controller.
With the default level, the ACE-syntax is enforced very strictly and semantic web data is expected to have the same syntax as defined. With a lower precision level, an LLM is stuck inbetween, syntax is not enforced as strictly and the ontology is enforced even if the data does not match it, also propagating to other web pradigms like MCP/AI-web and traditional web-services with REST-APIs.
The precision controller basically let’s the user move between a very strict semantic web browser and the lose cannon of a ChatGPT+MCP-app. And I think this moving of ratio is only possible if you start developing a strict semantic web browser, which has a precision controller integrated.
Another merit is that the energy consumption can be adjusted at will. If money/energy is low, for example in a state’s administration, the semantic web browser can still be used, while ChatGPT-like apps become unfeasible. "
the main issue here is still one of use case. this is a text-interface for the semantic web, but the semantic web is built to be easily parsed so you wouldn’t need a specific interface. llms, meanwhile, are data transformers, which you don’t want loose on strict content because the integrity can no longer be guaranteed. so what are you left with?
also, if you think this can be used on government code you might need to adjust your expectations.
I find this interesting, and ACE is a neat concept, but it’s so hard to write. I feel like having an entire encyclopedia written in ACE would be kinda cool as you could search through it using natural language without needing to use an AI.
The semantic web data would already be there: https://www.wikidata.org/wiki/Wikidata:Main_Page
I was mostly talking about ACE. It’s a neat concept, but it’s really hard to rewrite the world’s knowledge into.
