If the goal is to read what actual humans think, it's hard to see how an LLM filter can do anything but obscure and degrade the content.

LLMs, as we know them, express things using the patterns they've been developed to prefer. There's a flattening, genericizing effect built in.

If there are people who find an LLM filter to be an enhancement, they can run everything through their favorite LLM themselves.

▲

GMoromisato 7 hours ago | parent [-]

I think it's a spectrum:

1. I enter "Describe the C++ language" at an LLM and post the response in HN. This is obviously useless--I might as well just talk to an LLM directly.

2. I enter "Why did Stroustrup allow diamond inheritance? What scenario was he trying to solve" then I distill the response into my own words so that it's relevant to the specific post. This may or may not be insightful, but it's hardly worse than consulting Google before posting.

3. I spend a week with creating a test language with a different trade-off for multiple-inheritance. Then I ask an LLM to summarize the unique features of the language into a couple of paragraphs, and then I post that into HN. This could be a genuinely novel idea and the fact that it is summarized by an LLM does not diminish the novelty.

My point is that human+LLM can sometimes be better than human alone, just as human+hammer, human+calculator, human+Wikipedia can be better than human alone. Using a tool doesn't guarantee better results, but claiming that LLMs never help seems silly at this point.

▲

Avicebron 7 hours ago | parent | next [-]

> 3. I spend a week with creating a test language with a different trade-off for multiple-inheritance. Then I ask an LLM to summarize the unique features of the language into a couple of paragraphs, and then I post that into HN

I think where you are getting hung up is the idea of "better results". We as a community don't need to strive for "better results" we can easily say, hey we just want HN to be between people, if you have the LLM generate this hypothetical test, just tell people in your own words. Maybe forcing yourself to go through that exercise is better in the long run for your own understanding.

▲

GMoromisato 6 hours ago | parent | next [-]

My example was not great.

But my point is that I read HN partly because people here are insightful in a way I can't get in other places. If LLMs turn out to ultimately be just as insightful, then my incentive to read HN is reduced to just, "read what other people like me are thinking." That's not nothing, but I can get that by just talking with my friends.

Unless, of course, we could get human+LLM insightfulness in HN and then I'd get the best of both worlds.

▲

xenophonf 6 hours ago | parent | prev [-]

If someone can't explain something in their own words, then they don't _really_ understand it. The process of taking time to think through a topic and check one's understanding, even if only for oneself and the rubber duck, will reveal mistakes or points of confusion.

	▲	Avicebron 6 hours ago \| parent [-]
		Which gets to the core of the issue nicely, I want to go on to HN and talk to people who know things or have thought about things to the degree that they don't need a cheat sheet off to the side to discuss them.

▲

jmull 6 hours ago | parent | prev [-]

How is it not better, in your third scenario, if you described what you think are the important and interesting aspects of your idea/demo?

And what motivated you to make it -- probably the most interesting thing to readers, and not something an LLM would know.

Believe me, I don't care what an LLM has to say about your thing. I care about what you have to say about your thing.