Hacker News Reader

Comment by OfSanguineFire | Hacker News Reader

I would if the arrival of ChatGPT has sort of taken the wind out of Marginalia's sails as a human-directed search engine. It seems likely that the future of answering one's questions using the internet, is an LLM giving a straightforward, concise answer that is free of the quirks of a human author. Therefore, there is less motivation to search for personal websites and read website makers' own writing.

For example, imagine an old-school hobbyist website that contains information about some obscure band or author that can't be found elsewhere on the web, and doesn't readily show up in a Google search. Yet at the same time the author writes terrible prose, uses annoying HTML/CSS, or goes into tiresome political rants, etc. Instead of using a Marginalia-like search engine to discover those sites and read them directly, wouldn't it be a superior experience to have an LLM gorge on all those sites and then tell you just the facts that you care about?

View on HN

Based on my experience with Large Language Models (LLM) so far, I see more value in niche sites that marginalia will help the user find.

An LLM is regurgitating what it reads, true or false. There's a bias for things it sees more often. There's randomness in the response. Sure, humans are error-prone but a problem with known computer responses is that humans tend to think "a computer did this so it must be right." And LLMs are not pumping out true thoughtful answers but simply putting a string of probably words together.

The internet is vast and is increasingly filled with garbage, stolen and duplicated data, and monetization. It's nice to be able to look in the nooks and crannies of the internet for data.

I'm actually more interested in sites that do personal curation of interesting links that are personally vetted and of interest to someone than what's found through google or chatgpt.

I do agree that the question-answering search engine is probably headed the way of the dodo, but on the other hand, no search engine ever did question-answering particularly well in the first place (even beyond that, such projects are deeply problematic from an epistemological point of view).

LLMs aren't search engines though. They offer a new way of interacting with information, but they need to be provided with the appropriate context to answer questions well (e.g. as in RAG), which is where a traditional search engine still has a role even in an LLM maximalist future.

Though I'd note there are other aspects of the web that are criminally under served these days. The itch I'm trying to scratch with Marginalia Search is just letting humans with websites find each other.

> ChatGPT has sort of taken the wind out of Marginalia's sails as a human-directed search engine.

Not at all. ChatGPT (I've never used it) seems to me to be about answering questions or generating some text.

But for me the real pleasure and enjoyment of the web in general is reading essays and posts with a persons individual voice, and organically exploring topics, skipping from one web-site to the next. In that regard, Marginalia is fast becoming for me an essential part of the web-experience, as much as Wikipedia or the Internet Archive is.

> Not at all. ChatGPT (I've never used it) seems to me to be about answering questions or generating some text.

You have never used chatGPT but you are convinced it hasn't taken the wind out of Marginalia's sails... I would suggest you try it so you can see their perspective as well? I get that your use case might be different, but I don't think we know if that is the main use case of most users.

Is it functionally any different to Bard? Because I've used that.

If so, my reasoning remains entirely unchanged.

> that is free of the quirks of a human author

Uh, is that supposed to be a bad thing? Humans don't want to read uncanny-valley output devoid of personality and subjectivity. Even a Wikipedia article tends to have a small amount of flavor in the prose.

LLM output is currently designed to be entirely without personality, because its developers are well aware of how disturbing it would be if the AI tried to sound like a person. But human people generally want to read and watch and talk to other people, genuine people--which is the problem new-old-web projects like Marginalia are trying to solve, what with most of the Internet nowadays being so full of SEO junk and for-profit websites and corporate-curated walled gardens... and now machine-generated "AI" content.

The Internet can be (used to be) much more than just machine-generated summaries of raw data, and content that's just there to sell you things you don't need.

(Also, if everyone just asks an AI instead of searching the Web, why would anyone bother posting new content for those AIs to digest? No human would ever visit their site...)

> free of the quirks of a human author

Sometimes those quirks are what matters.

The pendulum always swings.

10, 15 years later and there will be a widespread “rediscovery” of the human made web by those who grew up with LLMs.

If that were true, they would be well advised to get out; nobody's going to ride out a fifteen year cycle.

LLMs have no concept of facts. They can hide this to an extent with alignment when it comes to sacred platitudes, but they certainly can't tell you if a singleton website with archaic HTML/CSS is accurate or not.

Reply on news.ycombinator.com