Hacker News Reader

Nice, makes sense to chain models so you don't waste the attention of the big smart model on the grunt work of JSON structure.

I bet there are some good post-processing heuristics you could also apply for hallucination with a flag for "this should be in the text verbatim" & then string matching whether the answer it outputted was a string from the text or not.

View on HN

interesting idea! yeah next big challenge for us is hallucination in null fields (IE: if you ask for a "name" from text that doesn't have a name you usually get "John") so need to add more sampling heuristics to doublecheck against the input blob for a sense of if it's an accurate extraction

Awesome, agree on the nulls! Any plans to add support for image inputs?

we have no plans for image or other modalities. I'd like to keep it just text-to-text so it can be as sharp of a tool as possible

Reply on news.ycombinator.com