Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I am not an expert in modern NLP SOTA but from what I see these models do not work that well.

You didn't summarize the article but took some of the most important points and shared what really mattered. It may not be perfect but required a real life experience. I am afraid there will always be a chance of getting "techno-babble" from our models, at least until they we get to some sort of AGI.

Look at your first line:

> Most of the article appears to be about what does not work.

No model will decide to include that info and it is vital for me to understand why you chose just a few lines out of it. We are still long way out from trusting in reliable LLMs.



There are two things to that answer; first, a summary:

> Most of the article appears to be about what does not work.

and then a segmentation, where they list the content.

This is very much in line with LLMs can do. Segmentation is part of a standard QA benchmark, and summaries are obviously something we have.

You can also ask a model to not only identify a summary, but segment only the relevant parts given the summary.

The "difficulty" is in pairing the two LLMs together.


I asked chatgpt to pick the most important 2 or 3 lines of text from the article. It returned:

    1. "Fortunately, there is advice out there on how to break the ice with strangers. Unfortunately, it’s abysmal."
    2. "Making contacts on a site like LinkedIn is a lot less stressful."
    3. "The real secret is to save your energy for the people who are most likely to be interesting to you."
It got 2 out of 3 wrong basically. I'm not sure how to objectively measure/improve that in an automated way. I suppose some prompt might do better though?

I guess a second pass of the results through another LLM could potentially help like you suggest?




Consider applying for YC's Winter 2026 batch! Applications are open till Nov 10

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: