The “There’s no way we can stop our agents from going rogue” people, suddenly have a burst of ability to reign in their creation when it suits their interests.
Does anyone think that perhaps what they are afraid is that source code of such clone would be compared to the original and show evidence that maybe the model was already trained on the code of the game?
I think it’s very likely that these models are effectively being tricked into retrieving source code of projects from memory of seeing the actual source code in the biases. This is why some project are much easier to uncover than others.
I’m no expert, but if my understanding of LLMs as ‘probability calculators’ is correct, then if you prompted one properly to produce code which does x (a Photoshop function for example) and it had only been trained on one example of this (the Photoshop source code) it would reproduce the PS source code because that has the highest probability of that being the correct ‘answer’.
This is baasically correct, though what the models actually produce is a little more randomized than that. Though, if there aren’t any words that are anywhere cllose to being likely to follow next, then there won’t be a any other valid options to choose.
Looks like everywhere companies mandate employees to install copilot or equivalent.
I think originally they were saying they won’t be using the proprietary code for training, but I don’t trust those companies when they already broke laws.
From the stolen math research for tha NS, from the company response that they can’t guarantee that the model didn’t train on the mathematician sessions, we know for sure they are using the inputs for training lol
This is full conspiracy from me, but could the terms of service for github grant the right to “enhance tools” based on private github projects. These then get moved around as ‘anonymized’ data to train off?
It’s long been proven that LLMs can produce exact copies of their training data. Not every piece of training data, of course, but enough where spitting out exact replicas of proprietary code is a legitimate concern. Not to mention that copies don’t need to be exact in order to be a problem.
The models are essentially a giant lossy compression applied to the training data.
They don’t store the loseless version of the training data sure, but that doesn’t mean it can’t reproduce something close enough to it, it’s literally how the whole thing works.
My understanding of it is that it’s far more complex than “lossy compression”. Of course it can generate text that is close to something in its training data, but I don’t think people have the right mental model if they think the model itself somehow has whole chunks of code in it. It does not.
My understanding of it is that it’s far more complex than “lossy compression”. Of course it can generate text that is close to something in its training data, but I don’t think people have the right mental model if they think the model itself somehow has whole chunks of code in it. It does not.
I’ve never seen any proof of this? They can fetch data and present it as their own maybe? But models don’t store training data verbatim. It’s like 6+6 and 4+8 both = 12, but they just store 12 and how it got there is unknown.
I’ve never seen any proof of this? They can fetch data and present it as their own maybe? But models don’t store training data verbatim. It’s like 6+6 and 4+8 both = 12, but they just store 12 and how it got there is unknown.
The “There’s no way we can stop our agents from going rogue” people, suddenly have a burst of ability to reign in their creation when it suits their interests.
Funny that…
Does anyone think that perhaps what they are afraid is that source code of such clone would be compared to the original and show evidence that maybe the model was already trained on the code of the game?
I think this is definitely a concern, yes.
I think it’s very likely that these models are effectively being tricked into retrieving source code of projects from memory of seeing the actual source code in the biases. This is why some project are much easier to uncover than others.
They aren’t being tricked. They are doing exactly what they were designed to do
Fair fair. Ih suppose what I mean is that it’s not really reverse engineering so much as it is recollecting from training samples.
I’m no expert, but if my understanding of LLMs as ‘probability calculators’ is correct, then if you prompted one properly to produce code which does x (a Photoshop function for example) and it had only been trained on one example of this (the Photoshop source code) it would reproduce the PS source code because that has the highest probability of that being the correct ‘answer’.
This is baasically correct, though what the models actually produce is a little more randomized than that. Though, if there aren’t any words that are anywhere cllose to being likely to follow next, then there won’t be a any other valid options to choose.
Where would they get the source code from so they can include it in the training data in this scenario?
Looks like everywhere companies mandate employees to install copilot or equivalent.
I think originally they were saying they won’t be using the proprietary code for training, but I don’t trust those companies when they already broke laws.
From the stolen math research for tha NS, from the company response that they can’t guarantee that the model didn’t train on the mathematician sessions, we know for sure they are using the inputs for training lol
This is full conspiracy from me, but could the terms of service for github grant the right to “enhance tools” based on private github projects. These then get moved around as ‘anonymized’ data to train off?
The models don’t store access to their training data. So no, there’s no source code in the model to retrieve.
It’s long been proven that LLMs can produce exact copies of their training data. Not every piece of training data, of course, but enough where spitting out exact replicas of proprietary code is a legitimate concern. Not to mention that copies don’t need to be exact in order to be a problem.
The models are essentially a giant lossy compression applied to the training data.
They don’t store the loseless version of the training data sure, but that doesn’t mean it can’t reproduce something close enough to it, it’s literally how the whole thing works.
My understanding of it is that it’s far more complex than “lossy compression”. Of course it can generate text that is close to something in its training data, but I don’t think people have the right mental model if they think the model itself somehow has whole chunks of code in it. It does not.
My understanding of it is that it’s far more complex than “lossy compression”. Of course it can generate text that is close to something in its training data, but I don’t think people have the right mental model if they think the model itself somehow has whole chunks of code in it. It does not.
So why can they recreate entire articles from the New Yorker and complete functions from the training set? Don’t fall for the slop laundering.
I’ve never seen any proof of this? They can fetch data and present it as their own maybe? But models don’t store training data verbatim. It’s like 6+6 and 4+8 both = 12, but they just store 12 and how it got there is unknown.
I’ve never seen any proof of this? They can fetch data and present it as their own maybe? But models don’t store training data verbatim. It’s like 6+6 and 4+8 both = 12, but they just store 12 and how it got there is unknown.