A LLM
Uncategorized
10
Posts
5
Posters
0
Views
-
A LLM
Also a LLM
-
A LLM
Also a LLM
Somewhere between given you easy answers to problems overrepresented in their training set and giving you nonsense lies the problem you are trying to solve.
How do you determine where your problem lies in this space? How do you determine where the answer lies in this space?
When we test these LLMs in spaces we are deeply familiar with and problems that are well represented in the training sets we can easily be lulled into a false sense of trust.
We can add constraints and such not but these tricks don't really assure us of that the solution is correct. It just tells us that it founds a minimum that passes your constraints. That feels like a pretty low bar unless your constraints are very very comprehensive. Even then it does not answer other critical question, such as how maintainable or fragile the code is.
It seems like no one is really trying to quantify this. Folks are going to be building real systems using these tools but we will only understand how bad is years after these things are in production.
It will be interesting to read books twenty years from now on this topic.
-
Somewhere between given you easy answers to problems overrepresented in their training set and giving you nonsense lies the problem you are trying to solve.
How do you determine where your problem lies in this space? How do you determine where the answer lies in this space?
When we test these LLMs in spaces we are deeply familiar with and problems that are well represented in the training sets we can easily be lulled into a false sense of trust.
We can add constraints and such not but these tricks don't really assure us of that the solution is correct. It just tells us that it founds a minimum that passes your constraints. That feels like a pretty low bar unless your constraints are very very comprehensive. Even then it does not answer other critical question, such as how maintainable or fragile the code is.
It seems like no one is really trying to quantify this. Folks are going to be building real systems using these tools but we will only understand how bad is years after these things are in production.
It will be interesting to read books twenty years from now on this topic.
@shafik It's very hard for me to shake the feeling that the current state of development is actually very poor, but looks good because we quantify it by "shipping", essentially. And these tools aren't actually that amazing, it's more that we've impressed ourselves with a very low local maxima.
-
@shafik It's very hard for me to shake the feeling that the current state of development is actually very poor, but looks good because we quantify it by "shipping", essentially. And these tools aren't actually that amazing, it's more that we've impressed ourselves with a very low local maxima.
-
@shafik It's very hard for me to shake the feeling that the current state of development is actually very poor, but looks good because we quantify it by "shipping", essentially. And these tools aren't actually that amazing, it's more that we've impressed ourselves with a very low local maxima.
@shafik For example, it seems the "killer aspect" of LLMs is code generation. It's a remarkably inefficient way to do it, but it demonstrably works (to some degree) in a way that people find helpful. (And ignoring the externalities is a mistake.)
To me, this opens up an avenue of research into code reuse techniques without the bad parts. We know that LLMs are helpful ("work" is too strong a term). But more importantly we know what we can consider failures: frameworks, DSLs, modelling languages, and so forth.
"How do you determine where your problem lies in this space, if at all?" is a very good question. And I think the maintenance over time with this kind of code generation is going to bring the same bitrot we have now.
-
@shafik It's very hard for me to shake the feeling that the current state of development is actually very poor, but looks good because we quantify it by "shipping", essentially. And these tools aren't actually that amazing, it's more that we've impressed ourselves with a very low local maxima.
This is also my feeling but coming up with a way to demonstrate these feels hard.
All the obvious problems are in the training set, all the problems that they obviously are unable to answer feel silly. Problems that don't lie in either of these spaces seems like a tremendous amount of work to construct.
It feels like a large epistemology problem, something Wittgenstein would love to work on but I am not Wittgenstein.
-
This is also my feeling but coming up with a way to demonstrate these feels hard.
All the obvious problems are in the training set, all the problems that they obviously are unable to answer feel silly. Problems that don't lie in either of these spaces seems like a tremendous amount of work to construct.
It feels like a large epistemology problem, something Wittgenstein would love to work on but I am not Wittgenstein.
@shafik I feel as though I'm trying to prove a negative all the time when talking to people who are overcome with the rush of euphoria when using these things.
-
@shafik For example, it seems the "killer aspect" of LLMs is code generation. It's a remarkably inefficient way to do it, but it demonstrably works (to some degree) in a way that people find helpful. (And ignoring the externalities is a mistake.)
To me, this opens up an avenue of research into code reuse techniques without the bad parts. We know that LLMs are helpful ("work" is too strong a term). But more importantly we know what we can consider failures: frameworks, DSLs, modelling languages, and so forth.
"How do you determine where your problem lies in this space, if at all?" is a very good question. And I think the maintenance over time with this kind of code generation is going to bring the same bitrot we have now.
-
Somewhere between given you easy answers to problems overrepresented in their training set and giving you nonsense lies the problem you are trying to solve.
How do you determine where your problem lies in this space? How do you determine where the answer lies in this space?
When we test these LLMs in spaces we are deeply familiar with and problems that are well represented in the training sets we can easily be lulled into a false sense of trust.
We can add constraints and such not but these tricks don't really assure us of that the solution is correct. It just tells us that it founds a minimum that passes your constraints. That feels like a pretty low bar unless your constraints are very very comprehensive. Even then it does not answer other critical question, such as how maintainable or fragile the code is.
It seems like no one is really trying to quantify this. Folks are going to be building real systems using these tools but we will only understand how bad is years after these things are in production.
It will be interesting to read books twenty years from now on this topic.
@shafik idk, I keep seeing gen AI stumble on stupidest of things, that are *just* outside of common patterns (and sounding very sure about it too).
-
@shafik idk, I keep seeing gen AI stumble on stupidest of things, that are *just* outside of common patterns (and sounding very sure about it too).
@shafik and then I have to explain to my colleagues why gen AI is wrong in this particular instance.
-
undefined oblomov@sociale.network shared this topic
Feed RSS
Gli ultimi otto messaggi ricevuti dalla Federazione
-
@peppenamir tutto torna
-
The first ten minutes I spent on social media this morning made me feel all kinds of things. Why is it that people who routinely use LLMs are so loud and brash and proud, making these tools appear as essential and inevitable?
A post by a dev whose app I use said something along the lines of: "no use exercising your coding skills, AI is too good now, you can't compete with it anyway".
Another post by a user on an instance I try to engage with wrote - literally: "tired of overthinking every decision?" and then disclosed he had created an AI that will "run a weighted decision matrix so you don't have to." In all seriousness.
What is this dystopian world where human qualities are devalued, critical thinking is discarded and surveillance capitalism is ignored at the altar of AI worship?
If they are loud and proud, maybe so I can be too... but in the opposite direction.
This weekend I will start the MIT's Missing Semester class (the 2020 Lectures, so pre-AI) because in this brave new world hyping up techno-fascist LLMs, knowing the basics of code are essential IMHO.
So my March "project" will be a deep dive in MIT's Missing Semester and my April project will be off-grid mesh radio communication.
What about you, what are you doing to resist?
Special props to @emilymbender @cwebber and @tante for being outspoken on these issues... you're my beacons of hope
#NoAI
-
@gian_d_gian tutto è bene ciò che finisce bene ma quando hai fatto i test al buio con l'ausilio dell'accendino, quante volte ti sei scottato?
Il pubblico ha il diritto di sapere.
-
@webdevtrust I don't think he's much of a gamer. 🥲
-
ma che cazzo sto vedendo?
-
@Gina Nice to hear. Ensure he keeps up with that. I currently regret that I didn't encourage my parents in all of the above mentioned activities, when they retired.
-
We will have a combined #XMPP and #UnifiedPush stand at #CLT2026 on March 28th and 29th. Come talk to us if you are curious about either! We will have merch for both, test devices, lots of advice, and I've even heard rumors about free Haribo.
XMPP Standards Foundation: https://chemnitzer.linux-tage.de/2026/en/programm/beitrag/204
UnifiedPush: https://chemnitzer.linux-tage.de/2026/en/programm/beitrag/411
-
@Gina could you get him into gaming? Buy him a Switch + Zelda: Tears of the Kingdom?
