AI doesn’t work that way. No one wrote “part of the answer.” It’s more like each contributor casted a vote on what the next token should be and it randomly picks one of the top ten voted tokens. (Very very roughly.)
Edit: definitely read the other responses because apparently there are some techniques I wasn’t aware of and don’t understand nearly as well as I understand the underlying AI technology - and I’m only an enthusiast layman.
I don’t think there is any way of doing that. AI is like a huge matrix that says ‘if (’ is followed by
’ x’: 60%
’ foo’: 19%
’ person’: 9%
Etc.
And then it does it all over again for the next token based on randomly selecting one of the tokens and then saying ‘if ( person’ is followed by
‘.id’: 30%
‘.name’: 27%
Etc.
So just to write a simple ‘if person.name.startsWith(“foo”) {’ is the aggregate result of thousands of contributors - really pretty much every author of every code snippet ingested from the training material.
There is no single author even if the code matches existing code token for token. The only exception would be code that is so esoteric that there is only a single author writing code that does a particular thing. But even in that case, there is nothing in the probability matrix to indicate that a particular sequence of tokens is unique to a certain author. Best you could do is full text search a line of code to see if it matches anything in the training data and if there is a very small set of authors to whom credit might be assigned. That might be possible, but it would be an add-on (and significant performance hit) to the actual AI itself. Sort of like how browser integrated AI just runs a search and feeds the result into the context to make the output more likely to contain information in the top results.
It depends. The base model, sure you can’t really figure out what percentage of it came from which data source since there’s just too many data sources and that information is lost along the way. They’re likely not using the entirety of SO to generate answers though. Retraining LLMs is ungodly expensive, so they can’t retrain it every time a new Q or A is created, and even retraining on a regular basis would be impractical.
Instead, without knowing exactly how they’re doing it of course, my guess is they’re pulling relevant Q&As from their database, then using those results to improve the response (for example by providing them as context). If you’re interested, look into retrieval-augmented generation.
Check out the article and feature video. It does appear to link to answers it pulled from. Bing and Bard do the same. Posters saying it’s impossible are mistaken.
If it’s doing a search for the code, pulling it in to the context, and then spitting it back out in slightly modified form, then it can attribute the source it pulled in. That’s a very different thing from the AI because code that is pulled into context by a search had a strong influence on the output. The output is still generated the same way but it would be reasonable to credit the author of the code that is pulled in. However, the code in the training data cannot be credited. How you would pull in just the right piece of code in the first place though is a bit of a mystery to me.
There are a few ways of finding which code is relevant, but one way is to use some sort of vector database to perform the search using embeddings generated from the Qs, As, and query.
Embeddings are essentially semantic representations of the text which can be compared to each other for similarity.
Posters aren’t saying that its impossible to put search results through an LLM and ask it to cite the source it reads. They’re saying that the neural networks, as used today in LLMs, do not store token attribution in the vocabulary or per node. You can implement a system for the neural network to work in that provides it the proper input (search results) and prodding (a prompt that encourages the network to biasing toward citation), not that the single LLM can conceptualize of that on its own.
You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !programming@programming.dev
Welcome to the main community in programming.dev! Feel free to post anything relating to programming here!
Cross posting is strongly encouraged in the instance. If you feel your post or another person’s post makes sense in another community cross post into it.
Hope you enjoy the instance!
Rules
Rules
Follow the programming.dev instance rules
Keep content related to programming in some way
If you’re posting long videos try to add in some form of tldr for those who don’t want to watch videos
AI doesn’t work that way. No one wrote “part of the answer.” It’s more like each contributor casted a vote on what the next token should be and it randomly picks one of the top ten voted tokens. (Very very roughly.)
Fair enough, but at least there should be a way for OverflowAI to list which contributors had the strongest link to the given answer, right?
Edit: definitely read the other responses because apparently there are some techniques I wasn’t aware of and don’t understand nearly as well as I understand the underlying AI technology - and I’m only an enthusiast layman.
I don’t think there is any way of doing that. AI is like a huge matrix that says ‘if (’ is followed by
’ x’: 60%
’ foo’: 19%
’ person’: 9%
Etc.
And then it does it all over again for the next token based on randomly selecting one of the tokens and then saying ‘if ( person’ is followed by
‘.id’: 30%
‘.name’: 27%
Etc.
So just to write a simple ‘if person.name.startsWith(“foo”) {’ is the aggregate result of thousands of contributors - really pretty much every author of every code snippet ingested from the training material.
There is no single author even if the code matches existing code token for token. The only exception would be code that is so esoteric that there is only a single author writing code that does a particular thing. But even in that case, there is nothing in the probability matrix to indicate that a particular sequence of tokens is unique to a certain author. Best you could do is full text search a line of code to see if it matches anything in the training data and if there is a very small set of authors to whom credit might be assigned. That might be possible, but it would be an add-on (and significant performance hit) to the actual AI itself. Sort of like how browser integrated AI just runs a search and feeds the result into the context to make the output more likely to contain information in the top results.
It depends. The base model, sure you can’t really figure out what percentage of it came from which data source since there’s just too many data sources and that information is lost along the way. They’re likely not using the entirety of SO to generate answers though. Retraining LLMs is ungodly expensive, so they can’t retrain it every time a new Q or A is created, and even retraining on a regular basis would be impractical.
Instead, without knowing exactly how they’re doing it of course, my guess is they’re pulling relevant Q&As from their database, then using those results to improve the response (for example by providing them as context). If you’re interested, look into retrieval-augmented generation.
I am interested, thank you!
Check out the article and feature video. It does appear to link to answers it pulled from. Bing and Bard do the same. Posters saying it’s impossible are mistaken.
If it’s doing a search for the code, pulling it in to the context, and then spitting it back out in slightly modified form, then it can attribute the source it pulled in. That’s a very different thing from the AI because code that is pulled into context by a search had a strong influence on the output. The output is still generated the same way but it would be reasonable to credit the author of the code that is pulled in. However, the code in the training data cannot be credited. How you would pull in just the right piece of code in the first place though is a bit of a mystery to me.
There are a few ways of finding which code is relevant, but one way is to use some sort of vector database to perform the search using embeddings generated from the Qs, As, and query.
Embeddings are essentially semantic representations of the text which can be compared to each other for similarity.
Posters aren’t saying that its impossible to put search results through an LLM and ask it to cite the source it reads. They’re saying that the neural networks, as used today in LLMs, do not store token attribution in the vocabulary or per node. You can implement a system for the neural network to work in that provides it the proper input (search results) and prodding (a prompt that encourages the network to biasing toward citation), not that the single LLM can conceptualize of that on its own.
Thanks for the TLDW - I could ogle a bit of the article but since I was at work, I couldn’t just play the video out loud.