Hacker News Reader

Comment by jcims | Hacker News Reader

This is my take as well.

There was a story a couple days ago about a neural network built on a single photonic chip. I fed the paper to ChatGPT and was able to use it to develop a much more meaningful and comprehensive understanding of what the chip actually delivered, how it operated, the fundamental operating principles of core components and how it could be integrated into a system.

The fact that I now have a tireless elucidator on tap to help explore a topic (hallucination caveats notwithstanding) actually increases my motivation to explore dense technical information and understanding of new concepts.

The one area where I do think it is detrimental is my willingness to start writing content on a provebial blank sheet of paper. I explore the topic with ChatGPT to get a rough outline, maybe some basic content and then take it from there.

View on HN

On the other hand you might be getting worse at reading those papers yourself.

The more youngsters skip the hassle of banging their heads on some topic the less able they will be to learn at later age.

There's more to learning than getting information, it's also about processing it (which we are offloading to LLMs). In fact I'd say that the whole point of going through school is to learn how to process and absorb information.

That might be the cognitive laziness.

What if the LLMs are teaching us that long form prose/technical writing is just a really bad, unnatural format for communication but natural dialogues are a good format?

If that was the case every scientific paper would be written as socratic dialogue. But it’s not. Because that’s a good format for beginners, but not for science.

the reason the current format exists and is used is because it's very information dense. i think scientific papers would be better if they were socratic dialogues.

but the limitation in publishing a dialogue is that you'd just get to publish one of them and each reader is going to come in with different questions and goals for what they want out of the paper.

The way I see it it is sort of like debugging code you're not well accustomed with.

While you're still going to learn whether you go through the hassle of understanding the system, develop a method for debugging it and learning about it along the way...

Of course a senior could point you to the issue right away, probably an llm too, and even provide a learning opportunity, but does it hold the same lasting impact of being able to overcome the burden yourself?

Which one makes a more lasting effect on your abilities and skills?

Again, LLMs are a tool, but if people in school/college start using it to offload the reasoning part they are not developing it themselves.

I do read the paper, but when you run into dense explanations like this:

>To realize a programmable coherent optical activation function, we developed a resonant electro-optical nonlinearity (Fig. 1(iii)). This device directs a fraction of the incident optical power ∣b∣2 into a photodiode by programming the phase shift θ in an MZI. The photodiode is electrically connected to a p–n-doped resonant microring modulator, and the resultant photocurrent (or photovoltage) detunes the resonance by either injecting (or deplet-ing) carriers from the waveguide.

It becomes very difficult to pick apart each thing, find a suitable explanation of what the thing (eg. MZI splitter, microring modulator, how a charge detunes the resonance of the modulator) is or how it contributes to the whole.

Picking these apart and recombining them with the help of something like ChatGPT has given me a very rapid drill-down capability into documents like this. Then re-reading it allows me to intake the information in the way its presented.

If this type of content was material to my day job it would be another matter, but this is just hobby interest. I'm just not going to invest hours trying to figure it out.

Sure, same as I'm probably pretty bad at going to the library and looking up information there, with the advent of the internet.

In practice, this lets you reasonably process the knowledge from a lot more papers than you otherwise would, which I think is a win. The way we learn is evolving, as it has in the past, and that's a good thing.

Though I agree that this will be another way for lazy children to avoid learning (by just letting AI do the exercises), and we'll need to find a good solution for that, whatever it may be.

Not being able to glean information from a paper is wildly different than being unable to use a card catalog. The former is basic reading comprehension; the latter is a technology.

You AREN'T learning what that paper is saying; you're learning parts of what the LLM says is useful.

If you read just theorems, you aren't learning math. You need to read the proof too, and not just a summary of the proof.

> (hallucination caveats notwithstanding)

This is a pretty big caveat to the goal of

> develop a much more meaningful and comprehensive understanding

Which is still my biggest issue with LLMs. The little I use of them, the answers are still confidently wrong a lot of the time. Has this changed?

I use ChatGPT a lot each day for writing and organizing tasks, and summaries/explanations of articles etc.

When dealing with topics I'm familiar with, I've found the hallucinations have dropped substantially in the last few years from GPT2 to GPT3 to GPT4 to 4o, especially when web search is incorporated.

LLMs perform best in this regard when working with existing text that you've fed them (whether via web search or uploaded text/documents). So if you paste the text of a study to start the conversation, it's a pretty safe bet you'll be fine.

If you don't have web search turned on, I'd still avoid treating the chat as a search engine though, because 4o will still get little details wrong here and there, especially for newer or more niche topics that wouldn't be as well-represented in the training data.

I've found them to be quite accurate when given enough context data. For ex, feeding it an article into it's context window and asking questions about it. Relying on the LLM's internal trained knowledge state seems to be less reliable.

I've found that whatever powers Kagi.com's answer seems to be pretty accurate. It cites articles and other sources.

Trying a share link, hope it works:

https://kagi.com/search?q=what+factors+affect+the+freezing+p...

What powers it is Kagi Search :) All chatbots have access to similar models, what distinguishes the answer quality is/will be the quality of search results fed to them.

Excellent! :)

I'm really happy about being able to share Kagi results. It's allowed me to slip Kagi into a few discussions to raise awareness. Also, being on bluesky helps because so many folks complain about google but they're not aware of better options.

One thing I wanted to raise: please keep the Kagi homepage as light as possible. It's tempting to keep adding little things and over time you get a big mess.

I agree in general but the way this has worked for me in practice is that I approach things hierarchically up and down. Any specific hallucinations tend to come out in the wash as the same question is asked from different layers of abstraction.

Reply on news.ycombinator.com