Comment by jsheard
10 months ago
I'm far from the first to make this observation but LLMs are like anti-compression algorithms when used like that, a simple idea gets expanded into a bloated mess by an LLM, then sent to someone else who runs it through another LLM to summarize it back to something approximating the original prompt. Nobody benefits aside from Sam Altman and co, who get to pocket a cool $0.000000001 for enabling this pointless exercise.
> LLMs are like anti-compression algorithms when used like that, a simple idea gets expanded into a bloated mess by an LLM,
I think that's the answer:
LLMs are primarily useful for data and text translation and reduction, not for expansion.
An exception is repetitive or boilerplate text or code where a verbose format is required to express a small amount of information.
There is one other very useful form of "expansion" that LLMs do.
If you aren't aware: (high-parameter-count) LLMs can be used pretty reliably to teach yourself things.
LLM base models "know things" to about the same degree that the Internet itself "knows" those things. For well-understood topics — i.e. subjects where the Internet contains all sorts of open-source textbooks and treatments of the subject — LLMs really do "know their shit": they won't hallucinate, they will correct you when you're misunderstanding the subject, they will calibrate to your own degree of expertise on the subject, they will make valid analogies between domains, etc.
Because of this, you can use an LLM as an infinitely-patient tutor, to learn-through-conversation any (again, well-understood) topic you want — and especially, to shore up any holes in your understanding.
(I wouldn't recommend relying solely on the LLM — but I've found "ChatGPT in one tab, Wikipedia open in another, switching back and forth" to be a very useful learning mode.)
See this much-longer rambling https://news.ycombinator.com/item?id=43797121 for details on why exactly this can be better (sometimes) than just reading one of those open-source textbooks.
I use LLMs a huge amount in my work as a senior software engineer to flesh out the background information required to make my actual contributions understandable to those without the same background as me. eg, if I want to write a proposal on using SLO's and error budgets to make data driven decisions about which errors need addressing and which don't, inside a hybrid kubernetes and serverless environment, I could do a few things:
* Not provide background information and let people figure it out for themselves. This will not help me achieve my goals.
* Link them to Google's SRE book and hope they read it. Still not achieving my goals, because they won't.
* Spend 3 hours writing the relevant background information out for them to read as part of my proposal. This will achieve my goals, but take an extra 3 hours.
* Tell the LLM what I'm looking for and why, then let it write it for me in 2 minutes, instead of 3 hours. I can check it over, make sure it's got everything, refine it a little, and I've still saved 2.5 hours.
So for me, I think the author has missed a primary reason people use LLMs. It saves a bunch of time.
4 replies →
Yeah you're totally right with this use case.
It feels like the information is there strewn across the internet, in forums, Reddit posts, stack overflow, specs, books. But to trawl though it all was so time consuming. With an LLM you can quickly distill it down to just the information you need.
Saying that, I do feel like reading the full spec for something is a valuable exercise. There may be unknown unknowns that you can't even ask the LLM about. I was able to become a subject expert in different fields just but sitting down and reading through the specs / RFCs, while other colleagues continued to struggle and guess.
2 replies →
This is how I currently am relearning upper high school math. It’s tremendously helpful as I am a why guy.
Why is the angle called m? Why is a combination nPr * (1/r)? What is 1/r doing there?
I use mathacademy.com as my source of practice. Usually that’s enough but I tend to fall over if small details aren’t explained and I can’t figure out why those details are there.
In high school this was punished. With state of the art LLMs, I have a good tutor.
Also it’s satisfying to just upload a page in my own handwriting and it understands what I did, and is able to correct me there.
1 reply →
They are also useful for association. Imagine an LLM trained on documentation. Then you can retrieve info associated with your question.
This can go beyond just specific documentation but also include things like "common knowledge" which is what the other poster meant when they talked about "teaching you things".
In my experience, if the information I need is in the documentation, then I don't need the LLM. If it is not in the documentation, then the LLM will invent stuff that could be there but that isn't, and it's actually a loss of time.
4 replies →
Yeah, this is the usual interaction with LLMs when coding: Ask GPT/Claude to write me a simple function. It writes 100 lines of code, trying to infer the rest of your codebase. Tell it to only solve my issue and do it in 5 lines of code. Get something close to working.
IMO the main uses of LLM are guided problem solving and a better Google search that you can ask questions in a natural way.
Producing text as output is not the way
Actually, pretty sure Sam Altman gets to spend that money on power and compute, not sure they’ve figured out how to turn it into an income stream yet.
I think they actually lose the $0.000001
Yep. They're very closely linked.
http://prize.hutter1.net/
Note the preamble, FAQs, and that all of the winning entries are now neural networks.
I blame humans. I never understand why unnecessarily long writing is required in a lot of places.
Rituals are significant because they are long. A ritual that consisted of the words “Rain please” wouldn’t convince the gods, much less their human followers.
"I have made this longer than usual because I have not had time to make it shorter." — attributed to many[1]
The funny thing is that people use LLMs to do the opposite instead of what is implied to be a smarter thing to do with the above quote.
[1] https://quoteinvestigator.com/2012/04/28/shorter-letter/
Depends on what you are looking for. I’ve turned half baked ideas into white papers for plenty of praise. I’ve used them to make my Jira tickets seem complicated and complete. I’ve used them to get praised for writing comprehensive documentation.
Part of my performance review is indirectly using bloat to seem sophisticated and thorough.
> comprehensive documentation
Documentation is an interesting use case. There are various kinds of documentation (reference, tutorial, architecture, etc.) and LLMs might be useful for things like
- repetitive formatting and summarization of APIs for reference
- tutorials which repeat the same information verbosely in an additive, logical sequence (though probably a human would be better)
- sample code (though human-written would probably be better)
The tasks that I expect might work well involve repetitive reformatting, repetitive expansion, and reduction.
I think they also might be useful for systems analysis, boiling down a large code base into various kinds of summaries and diagrams to describe data flow, computational structure, signaling, etc.
Still, there is probably no substitute for a Caroline Rose[1] type tech writer who carefully thinks about each API call and uses that understanding to identify design flaws.
[1] https://folklore.org/Inside_Macintosh.html?sort=date
Yes, but none of the current LLMs are even remotely useful doing that kind of work for even something moderately complex. I have a 2k LOC project that no LLM even "understands" *. They can't grasp what it is: It's a mostly react-compatible implementation of "hooks" to be used for a non-DOM application. Every code assistant thinks it's a project using React.
Any documentation they write at best re-states what is immediately obvious from the surrounding code (Useless: I need to explain why), or is some hallucination trying to pretend it's a React app.
To their credit they've slowly gotten better now that a lot of documentation already exists, but that was me doing the work for them. What I needed them to do was understand the project from existing code, then write documentation for me.
Though I guess once we're at the point AI is that good, we don't need to write any documentation anymore, since every dev can just generate it for themselves with their favorite AI and in the way they prefer to consume it.
* They'll pretend they understand by re-stating what is written in the README, then proceed to produce nonsense.
1 reply →
Delivering a library with an llm to explain the api and idiomatic usage seems like an interesting use case.
I’d rather be homeless in Philadelphia than work where you work
This kind of "perf review" hacking works ~everywhere; how well it works correlates with how entrenched the organization is (i.e., how hard it is for new players to disrupt).
You don't have to play the game the same way to work there. But it helps to accept that others will play it, and manage your own expectations accordingly.
11 replies →
Would be nice to fix the performance reviews so we don't end up in a arms race of creating bloat until it becomes so unproductive it kills the host.
Over-fitting proxy measures is one of the scourges of modernity.
The only silver lining is if it becomes so wide spread and easy it loses the value of seeming sophisticated and thorough.
> creating bloat until it becomes so unproductive it kills the host
Maybe we should let/encourage this to happen. Maybe letting bloated zombie-like organisations bloat themselves to death would thin the herd somewhat, to make space for organisations that are less “broken”.
3 replies →
I fully believe you and I am saddened by the reality of your situation.
At the same time, I strive really hard to influence the environment I am in so it does not value content bloat as a unit of productivity, so hopefully there are at least some places where people can have their sanity back!
If your organisation is such that you have to do this even though you are competent for your job, then they deserve it. They lose money because they do it wrong.
If your organisation is functional and you are abusing it by doing that, then you deserve to get fired.
...thinking about it, there are probably situations where making something more verbose makes it take less effort to read. I can see how an LLM might be useful in that situation.
[flagged]