Hacker News Reader

Comment by reticulates | Hacker News Reader

Your curiosity is enabled but is it being best served? You are being fed a single perspective on every question, a singular world view. You do not have the serendipitous exposure to a broad range of ideas and opinions we experience as we meander through information.

You Google about why tungsten is in light bulbs, you find yourself on a website about the history of lightbulbs written by a hobbyist that has spent decades thinking about lightbulbs. You click around, link after link, your world is expanded to include this lightbulb fanatic’s.

Personally, I find LLM answers to questions completely devoid of the substance that actually matters to my curious mind. Sure, when I’m working and need a concrete answer to a specific problem, an LLM can be a great help, but for personal curiosity, to expand my world, they suck.

To me, what you’re describing sounds like learning lots of trivia, not expanding your understanding of the world.

View on HN

To each their own, I guess. If I find out that yarn works was a type of textile mill and spinning factory, I then read on about other parts of the textile industry in New England. Maybe how it relates to the progress of steam engines vs. hydro power, or its place in the Industrial Revolution overall. It gives me an on-ramp to understand one thing and branch out to others. I don't know what AI models or agents you're using, but they must be weak sauce. Mine give me sharp and specific answers, at least 2-5 references per question, and I can branch out to that lovingly curated blog, to Wikipedia, to local historical societies and their curated collections. Et cetera.

I end up reading more sources and exploring more things. If you want to dismiss that as "trivia," or dismiss the topics I want to know about, okay. But it's important to me. You could equally deride search engines as wanton seas of advertising, and tell me I need to head to a library and its paper-based card catalog, the only true and proper way to do research. :shrug:

It's true there are a lot of weak models out there. I tend to use the frontier models, and they are great research assistants IME.

I didn’t mean your topics of interest are trivial, I meant that LLMs reproduce information devoid of human perspective which turns it into just a bunch of trivia, facts without substance. A blog post about tungsten’s use in lightbulbs by a lightbulb fanatic is a fascinating insight into the subject built on their years of fanaticism, every word is meaningful. An LLM’s words mean nothing, it is bland, it is soulless, an LLM’s summarization of a blog post is not a more efficient way to consume the blog post, it is something altogether different.

That the answers of an LLM are “sharp” and “specific” is not a good thing. You are only experiencing the subject through your own lens, through your own questions, through your own sycophantic robot. The imperfect way that humans communicate is so important to broadening our horizons. Some of the most influential learnings I’ve experienced came from tangential lines in books and blogs. And when I’m reading, sitting, thinking, letting the words bounce around my head, I am forced to reconcile ideas myself, to consider, to think. Just saying “tell me more about x” or “why is x?” every time I have a thought leaves no room to really learn and grow.

Anyway, we are all different, if your curiosity can be satisfied by LLMs summarizing, great :)

Could you define what you mean by human perspective? If LLM models are trained on data that's generated by humans and the resultant weights represents a gestalt of that data, isn't the resulting output going to be token patterns that represent the human perspective? For code, I could get that argument so I suppose it depends on the data you're trying to get.

> Could you define what you mean by human perspective?

They can't. This is an "anti-consciousness" qualia argument adapted to a general purpose anti-ai argument.

They are arguing that meaning doesn't come from the accurate reproduction of information, but from the subjective, lived experience behind the words. That subjectivity can't be defined, it's the whole point. It will never be defined, no matter how far along we come.

For me, I don’t stop at the LLMs. I use them to do a deep semantic search of the internet. If all I want is a summary, then that may be as far as I go.

But frequently, I want more than just a summary, so the LLM is just a launching point for me to go do a deep dive on the sites it turns up.

It increases signal to noise ratio in comparison to raw google results.

You might or might not miss that hobbyist web page. I can't say for certain, but I do know that most of my self learning explicitly includes going over the history of the topic.

Or you know, same scenario, but you are looking for information on how quanta work. You really can't make sense of the first 10 pages of google results, but you happen on "Motion Mountain" on the 11th page and you end up spending the next two years telling people the universe is made of little knots.

If that sounds like a pretty specific example, it's because it's a real one of someone I knew. I don't think the tool here is going to massively impact the quality. The person's judgement will play a way heavier role in that.

Reply on news.ycombinator.com