interesting approach! my concerns are about the final results though - it doesn't capture the context in which things are said and WHEN things are said (for e.g. poll over 30% vs poll over 48% - where they said at the same time, if not how far apart?), these additional features would contribute to actually figuring out the inconsistencies (for e.g. the first example of $6mil USD coming through). It'd be great if we are able to add capturing these additional things in the aforementioned pipeline
However, the idea of using one averaged contradiction vector to identify all logical inconsistencies sound too simple. Intuitively, for example, there would be a different degree of contradictiveness for different relationships and different sentences. (Eg: in Mary has a cat vs Mary has a dog. Or Joel is tall vs Joel is just 5'10.) It seems hard to assume that just one vector would be able to identify & represent all of these very different semantic relationships. Anyway, would love to know if you get any cool updates here
Is there a particular reason why you started off with Claude generating tasks.md instead of directly going with Gemini cli ? Or was it simply a reflex habit of using Claude?
hey can you confirm that does it used sentence transformer embeddings models or openai embeddings model? Because as stated in the jian ai blogs one of their embedding models is trained on the constrative examples to solve the issue. Then the difference wil be large i guess
interesting approach! my concerns are about the final results though - it doesn't capture the context in which things are said and WHEN things are said (for e.g. poll over 30% vs poll over 48% - where they said at the same time, if not how far apart?), these additional features would contribute to actually figuring out the inconsistencies (for e.g. the first example of $6mil USD coming through). It'd be great if we are able to add capturing these additional things in the aforementioned pipeline
I wonder what happens when people do this for Biotech/ QC/ Medicine, domains that are superspecialized and hard to debunk.
Makes me wonder why I haven't used CLI yet.
However, the idea of using one averaged contradiction vector to identify all logical inconsistencies sound too simple. Intuitively, for example, there would be a different degree of contradictiveness for different relationships and different sentences. (Eg: in Mary has a cat vs Mary has a dog. Or Joel is tall vs Joel is just 5'10.) It seems hard to assume that just one vector would be able to identify & represent all of these very different semantic relationships. Anyway, would love to know if you get any cool updates here
Probably he went with single vector inconsistency to grokk something under 1 hr as the title suggests!
Is there a particular reason why you started off with Claude generating tasks.md instead of directly going with Gemini cli ? Or was it simply a reflex habit of using Claude?
hey can you confirm that does it used sentence transformer embeddings models or openai embeddings model? Because as stated in the jian ai blogs one of their embedding models is trained on the constrative examples to solve the issue. Then the difference wil be large i guess
https://jina.ai/news/text-embeddings-fail-to-capture-word-order-and-how-to-fix-it