Please wait...
Subject: Re: Large Plaigiarism Models and programming
One can hope.
AGI holds that promise.
If the robots get too smart they'll conquer the world
Subject: Re: Large Plaigiarism Models and programming
What if people starts posting LLM answers to StackOverflow, and LLMs
get locked in a closed feedback? I would argue it is already the case
with many common non-software questions.
That's kind of what I was getting at. If everyone is cribbing from Stack Overflow, and so is LLM, and no one is contributing to anything, eventually it poisons the well. The tragedy of the commons, in a way.
Whenever I encounter an LLM I have to test it with a few questions for which I already know the answer, just to see how it handles them. Some of the usual culprits will refuse to tell the truth if it challenges an establishment narrative.
Others are just stupid.
I run my own search engine, so I guess eventually I'll need to run my own LLM.
Subject: Re: Large Plaigiarism Models and programming
For code models, a lot of the data comes from a curated pull of GitHub as their core training. ( which can reach into the petabytes.. not for your casual weekend AI hacker ) Not exclusively of course, but a huge part.
( dont want to say too much, but i'm in that world. )
Subject: Re: Large Plaigiarism Models and programming
And it's literally everything, even entire subprojects that were deleted from the tree decades ago.
I want to find a way to poison the well.
Subject: Re: Large Plaigiarism Models and programming
I want to find a way to poison the well.
Doesn't seem like that should be too difficult. Start submitting progressively more broken changes to it.
That is too bad, i figured you guys would embrace technology advancing, and not want to hinder it, or even sabotage it.
its just another tool for everyone to use, including yourselves.
Rather disappointing.
2025-04-28 20:10 from ZoeGraystone
That is too bad, i figured you guys would embrace technology
advancing, and not want to hinder it, or even sabotage it.
its just another tool for everyone to use, including
yourselves.
Rather disappointing.
I am a bit skeptic when it comes to progress because so many things are hyped and posed as the next groundbreaking milestone in the History of mandkind, and they turn out to be a disappointment or a straight step back.
The problem is usually not that the alleged progressful thing exists, but that they want to use it to replace something that works great before they can prove they are good enough themselves. And when they manage to overtake the good old thing and replace it with the new half-functional thing nobody really wins.
Last friday I was testing some LLMs with friends asking it solutions to technical questions. Results were often ok or at least gave you a hint as to where the true answers could be found, but this stuff is basically a search engine with a fancy UX. And it suffers from the same ailments.
if you are just using it as a replacement search engine, yes, you will be disappointed. That is not what its meant to be used for. It can do that to a limited extent of course, but its not going to be perfect as that is not how its trained or designed. As far as ground breaking, the current methods are. And just wait, everyone will see the truth in time. In the beginning it will be more hidden, but over time it will become more clear what is going on.
You also need to be sure you are using the correct LLM for the correct task. Just like people they have varying skills in different areas. And if you are using 'compressed' versions locally, that effects things greatly as well.
And with luck, the movie does really happen in our lifetimes. As id like to see how it turns out. ( "i have seen this movie before" sort of thing )
I am a bit skeptic when it comes to progress because so many things are hyped and posed as the next groundbreaking milestone in the History of mandkind, and they turn out to be a disappointment or a straight step back.
The problem is usually not that the alleged progressful thing exists, but that they want to use it to replace something that works great before they can prove they are good enough themselves. And when they manage to overtake the good old thing and replace it with the new half-functional thing nobody really wins.
Last friday I was testing some LLMs with friends asking it solutions to technical questions. Results were often ok or at least gave you a hint as to where the true answers could be found, but this stuff is basically a search engine with a fancy UX. And it suffers from the same ailments.
That is too bad, i figured you guys would embrace technology
advancing, and not want to hinder it, or even sabotage it.
What I want is to advance past the hype cycle. Right now it's being shoved into everything.
It'll be great when it's just quietly part of the technology stack.
2025-04-28 20:10 from ZoeGraystone <zoegraystone@uncensored.citadel.org>
That is too bad, i figured you guys would embrace technology
advancing, and not want to hinder it, or even sabotage it.
its just another tool for everyone to use, including yourselves.
Rather disappointing.
To be clear, I view what they're calling AI as an overhyped novelty at best. It's not going to change the world like everybody thinks it is.
Here's the problem. We use computers because they're good at carrying out very repetetive tasks at speed and with highly predictable consistency. AI doesn't deliver on any of those things, and it manages to still demand laughably absurd resources to do it. That's actually not that impressive to me.
It's not AI. It's advanced predictive text completion. It doesn't have intelligence or thoughts.
It'll be great when it's just quietly part of the technology stack.
I don't want it to be quietly part of anything. The other side of the coin is that it only works by spying on you.
It already has, and will continue do so in ways you cant imagine.
Tue Apr 29 2025 00:25:49 UTC from zelgomer
To be clear, I view what they're calling AI as an overhyped novelty at best. It's not going to change the world like everybody thinks it is.
Exactly like humans. And when we release AGI in a year or so, it will surpass meat puppets.
Tue Apr 29 2025 00:25:49 UTC from zelgomerIt's advanced predictive text completion.
Just like humans, if you are honest with yourself.
Tue Apr 29 2025 00:28:49 UTC from zelgomer
I don't want it to be quietly part of anything. The other side of the coin is that it only works by spying on you.
Exactly like humans. And when we release AGI in a year or so,
it will surpass meat puppets.
This makes me have to ask: are you the real Zoe or are you a U-87 Cyber Combat Unit?
Scott Adams likes to talk about people who have decent reasoning skills vs. people who are just NPCs who believe everything the news tells them. The latter group isn't really thinking much so it makes sense that you can replace them with an electronic brain that only says "What?" and "I don't understand" and "Where's the tea?" and no one will ever know the difference.
(wowzers ... I went Charles Adams, Scott Adams, and Douglas Adams in the same post. I am invincible!)
Aside from the fact they don't exist in the real world, I am neither. I DO happen to be a fan of the series. And my first name is actually Zoe, ironically ( but the last name, no, its not Graystone ). However, as a career i am professionally working toward AGI, but not in the military space, even tho that is always where new tech gets applied first.
And it is greatly disappointing to see how closed minded some are about the entire AI concept, and simply refuse to understand it. Clinging to the past, or their fears, and fail to accept reality.
Tue Apr 29 2025 16:57:49 UTC from IGnatius T Foobar
This makes me have to ask: are you the real Zoe or are you a U-87 Cyber Combat Unit?
Just as an example its more than "just a predictor" ( then i will mostly shut up about the subject ):
One of my tasks is to test models. To do that i have several tasks that are given to the models to see how they preform. One of the early tests is ( short version ) to generate some HTML code, which includes pulling a image from a specific URL. Its not 'hard' but failures are common.
In this case it failed. The code was correct, however the URL was not. It also messed up one of the values. My comment back was "you neglected to use the URL i provided for the texture map." While descriptive to a human, not so much for a LLM, on purpose. And i left out the other mistake on purpose. It went back re-analyzed my original request, looked at the code relative to the mistake i pointed out, and returned with the following during its 'analysis process' ( which is often hidden, but important here ):
"Okay, the user pointed out that I used the wrong texture URL. Let me check the code again. Oh right, in the original answer, I used 'uv_grid_opaque.jpg' instead of the provided 'land_ocean_ice_cloud_2048.jpg'. That's a mistake.
I need to correct the texture loader line. Let me replace the URL. Also, maybe the user wants the Earth's surface texture, so using the correct URL is essential. Let me make sure the textureLoader.load() uses the exact URL they specified.
Other parts of the code seem okay. The geometry has 32 segments, but the user asked for 64. Wait, in the code, it's SphereGeometry(1, 32, 32). Oh, that's a problem. The user specified 64 segments, so I should change that to 64. That's another error.
So two issues: texture URL and segment count. Let me fix both. Change the geometry to 64 segments and update the texture URL. Then test the code to ensure it works. Also, check if the lighting and animation are correctly implemented as per the user's request. Once fixed, present the corrected code. "
It then proceeded to correct the code, and was successful.
Anything that happens within a session it still has stored (until it runs out of context space, its not unlimited ), so it just applies that to its reasoning steps directly as it "was". It really cant "replay" as by design there is a small percentage of randomness involved baked in ( just like real life ) so it would not come up with the same exact result if it started over.
It is possible to save and reload past sessions for it to chew on. But again there are limits to size.
One of our goals is to make all sessions, both good and bad, training data in effect, and feed it back in real time into to the actual model, not via traditional training. This way it learns dynamically from its experiences and actions, just like people. How, i cant really talk about that. ( but i can say it's not via traditional RAG methods, but will be actually embedded into the model and real time, once we have it all worked out.. its still a WIP )
Wed Apr 30 2025 00:40:25 UTC from zelgomerWhen you give it a corrective prompt, how does it resume its train of thought? Does it effectively replay all past dialog for the session, or does it actually have some kind of internal state that represents abstract context?