Comment by tempest_ | Hacker News Reader

Comment by tempest_ | Hacker News Reader

Parent Full thread

It is very weird. 6 months ago I generated some documentation for a service that was largely undocumented. When it was done my only thought was "who is this for?" Is any one going to actually read this or are they just going to ask the LLM? If someone didnt write it, why would I read it? It just becomes something to pipe into the context for the next person. It seems though a lot of people don't have this experience, its like they think they are the only ones with access to this technology.

Additionally I am so, so ,so sick of reading claudes overly wordy technical writing ish bullshit. Its exhausting trying to find the fuck ups or hidden assumptions in the pages of text it produces for everything.

Every time I try and search for ways to make it stop I just find people trying to save tokens which I don't give a shit about. I just want it to communicate things to me without having to constantly read 4 paragraphs, dump the context or fork it or start a new session so it "remembers" or re-weights that is was told to be concise and terse in like 7 different places. Maybe I am using it wrong.

In the rare occasion I use LLMs, I always add at least one of the following:

    - Do not include opinions, just official documentation.
    - Add all references you have used as a bibliography section at the end.

Then I skim the result, and jump to the references it has given. I don't have time for other way.

You have reinvented the search engine.

That’s not meant to be glib; Google and friends are pretty trash now.

That’s the point. I only use LLMs if and only if Kagi can’t unearth the thing I’m looking for for a week or so.

This means it finds me a single video which is not in a language which I know, or similar.

An LLM-vector-based search engine has the potential to be fantastic. No autocomplete - just make it search for similar vectors.

For my medical notes assistant I had to set a global "write in a telegraphic, minimal style" because it was fluffing up every aspect of the notes. Nobody wants to read that BS and it's impossible to find information in a series of walls of text.

I wonder if reading cooking recipe sites in training contributed to this.

9 pages about how it felt to be a kid eating the dish, for 7 ingredients and a one line recipe.

It can also go the other way. It outputs a paragraph so full of telegraphic technobabble that it’s just unreadable, inventing new jargon every other sentence, using terminology invented in its thinking stream and never explained. And when it combines the two issues and it outputs 4 pages of technobabble, it gets just exhausting to read. Paradoxically, given that 2 years ago everyone was using LLMs to summarize web pages, their ability to condense text (including their own thought stream) is still lackluster. I dream of some post-processing diffusion optimizer just condensing all of their slop away. Until then, I just tell it to _avoid_ telegraphic and newspaper-title speech and to output one or two paragraphs of prose only.

In medical notes you definitely want the telegraphic. It's both easier to find the keywords, and easier to manually prune extraneous notes.

To be fair, LLMs didn't really change this that much: Throughout my career, I've been generating docs that nobody is going to read, only because some process somewhere says "In order to get engineering approval to commence work, an Engineering Risks Document (ERD) must be produced." Everyone knows that nobody reads these, but someone with a rubber stamp checks to see that the doc exists and it looks sufficiently long, and then he stamps the stamp. LLMs just make producing this shit faster.

Reply on news.ycombinator.com