Hacker News Reader

Comment by Zambyte | Hacker News Reader

I think the "killer use case" is slightly more general: operations in linguistic space, of which summarization is one. Some other useful uses include:

- the inverse of summarization; given a bulleted list of facts, write a sensible set of paragraphs presenting those facts

- semantically searching in large documents

- make some text sound more professional / casual

- translate some text between human languages (they're not perfect at this yet, but they're surprisingly good, and I think they will be the way to do translations in the future).

- expand abbreviations given some context

The key here is that none of these depend on the language model injecting factual information into its output. All of the facts are encoded in your input - the language model just does some operation in linguistic space. If you don't give it the opportunity to give you incorrect assertions, it won't give you incorrect assertions.

View on HN

> The key here is that none of these depend on the language model injecting factual information into its output.

This feels like an under-appreciated insight. There are lots of valuable (and in some cases previously difficult, expensive, or impossible) linguistic manipulations that fall under this lower-risk end of spectrum.

everything there sounds great (though i suspect for actual human-passing linguistic register-transfer we will need something different than the bog standard GPT models because everyone has this six sense for things being written by it and, particularly in English, register is one of oldest forms of 'discrimination' in the language/culture, and certainly persists to the current day. And there's a huge competitive advantage to being able to have an LLM write things for you but make it sound like you yourself wrote it.)

For translations considering how 'mature' the field is relative to everything else I'm a little surprised we haven't seen any rigorous benchmarks on round-trip translations, especially given that you aren't even limited to natural languages. You can take text, flip it into Klingon, and back, if you so chose.

>- the inverse of summarization; given a bulleted list of facts, write a sensible set of paragraphs presenting those facts

Except this one, this sounds unintentionally evil for some reasons i can't entirely put my finger on. Perhaps the inevitable conclusion of countless GPU farms spamming terabytes of text at each other predicated upon some short bullet point lists and the fact that "those are our marching orders"

I guess Fundamentally I would like to live in a future where LLMs can only decrease the length of text written by humans, never increase it or fabricate it.

Another application is to use the LLM to build an embeddings database, and then do semantic search only within that database. Don't just let the LLM talk about whatever it wants to hallucinate up, force it to use a set of known good answers while accepting natural language input and summarizing documents in the known good dataset.

Reply on news.ycombinator.com