Comment by bitexploder | Hacker News Reader

Comment by bitexploder | Hacker News Reader

Parent Full thread

This. I am very picky about how I use ML still, but it is unsurpassed as a virtual editor. It can clean up grammar and rephrase things in a very light way, but it gives my prose the polish I want. The thing is, I am a very decent writer. I wrote professionally for 18 years as a part of my job delivering reports of high quality as my work product. So, it really helps that I know exactly what “good” looks like by my standards. ML can clean things up so much faster than I can and I am confident my writing is organic still, but it can fix up small issues, find mistakes, etc very quickly. A word change here or there, some punctuation, that is normal editing. It is genuinely good at light rephrasing as well, if you have some idea of what intent you want.

When it becomes obvious, though, is when people let the LLM do the writing for them. The job search bit is definitely rough. Referrals, references, and actual accomplishments may become even more important.

As usual, LLMs are an excellent tool when you already have a decent understanding of the field you're interested in using them in. Which is not the case of people posting in social media or creating their first programs. That's where the dullness and noise come from.

The noise ground has been elevated 100x by LLMs. It was already bad before but it's accelerated the trend.

So, yes, we should have never been trusting anything online but before LLMs we could rely on our brains to quickly identify the bad. Nowadays, it's exhausting. Maybe we need a LLM trained on spotting LLMs.

This month, I, with decades of experience, used Claude Dev as an experiment to create a small automation tool. After countless manual fixes, it finally worked and I was happy. Until I gave thr whole thing a decent look again and realized what a piece of garbage I had created. It's exhausting to be on the lookout for these situations. I prefer to think things through myself, it's a more rewarding experience with better end results anyway.

LLM's are a great onramp to filling in knowledge that may have been lost to age or updated to their modern classification. For example, I didn't know Hokkien and Haka are distinct linguistic branches within the Sino-Tibetan language and warrants more (personal) research into the subject. And all this time, without the internet, we often just colloquially called it Taiwanese.

How is this considered "lost" knowledge there are (large) Wikipedia pages about those languages (which is of course what the LLM is cribbing from)?

"Human-curated encycolpedias are a great onramp to filling in knowledge gaps", that I can go with.

It is lost in a sense that you had no idea about such possibility and you did not know to search it in the first hand, while I believe that in this case LLM brought it up as a side note.

Such fortuitous stumblings happen all the time without LLMs (and in regular libraries, for those brave enough to use them). It's just the natural byproduct of doing any kind of research.

Most of my knowledge comes from physical encyclopedia and download the wikipedia text dump (internet was not readily available). You search for one thing and just explore by clicking.

How often do you go back to your encyclopedia hard copies only to find whatever knowledge you may have absorbed have already been deprecated? Or that information from Wikipedia may have changed at moments without notice, have never read or, dare I say, included a political bias to them?

Maybe I should have worded it better as a "beginner" or "intermediate" knowledge onramp and/or filler. For example, I have asked it on occasion to translate into traditional Mandarin in parallel for every English response. It helps tremendously in trying to rebuild that bridge that may have been burned long ago.

Not to sound too dismissive, but there is a distinct learning curve when it comes to using models like Claude for code assist. Not just the intuition when the model goes off the rails, but also what to provide it in the context, how and what to ask for etc. Trying it once and dismissing it is maybe not the best experimental setup.

I've been using Zed recently with its LLM integration so assist me in my development and its been absolutely wonderful, but one must control tightly what to present to the model and what to ask for and how.

It's not my first time using LLMs and you're assuming too much.

This is my go-to process whenever I write anything now:

1. I use dictation software to get my thoughts out as a stream of consciousness. 2. Then, I have ChatGPT or Claude refine it into something coherent based on a prompt of what I'm aiming for. 3. Finally, I review the result and make edits where needed to ensure it matches what I want.

This method has easily boosted my output by 10x, and I'd argue the quality is even better than before. As a non-native English speaker, this approach helps a lot with clarity and fluency. I'm not a great writer to begin with, so the improvement is noticeable. At the end of the day, I’m just a developer—what can I say?

Yeah, this is how I use it too. I tend to be a very dry writer, which isn't unusual in science, but lately I've taken to writing, then asking an LLM to suggest improvements.

I know not to trust it to be as precise as good research papers need to be, so I don't take its output, it usually helps me reorder points or use different transitions which make the material much more enjoyable to read. I also find it useful for helping to come up with an opening sentence from which to start writing a section.

Active voice is difficult in technical and scientific writing for sure :)

Reply on news.ycombinator.com