I hate it when people use unnecessary terms to describe something.
It’s a script that runs a search and then the LLM takes the output of that and reformats it into an answer. It’s the same as feeding it a document and having it rephrase something.
It’s a script that runs a search and then the LLM takes the output of that and reformats it into an answer.
“I hate when people use concise, reasonably common, and understandable terminology. Why can’t we just expand everything into full sentences that are also oversimplified?”
RAG is literally just polling for information and rewriting it. It’s the same garbage that gave us Gemini telling us to put glue on pizza to prevent the cheese from slipping off.
You can, and should be more critical of where you source the information but it’s not going to magically make language models actually intelligent. It’s not going to make them reason, or be able to properly select what is relevant or not. Just because you give it a bunch of scientific papers doesn’t mean that the stuff they output will be accurate or not misleading.
Literally here. And sorry, before you posted this, I did quickly edit my comment to “oversimplified”. Because technically yes, it’s searching and using what it’s retrieved mixed with a (modified) user prompt to generate an output. But it’s searching based on a prompt (rewriting it to aid retrieval), often reranking results, stripping the query-specific context from the results into chunks, attempting to resolve contradictions between sources (which is objectively more than just rephrasing), and then synthesizing between whatever its pretraining is and what its retrieval results are (thus “retrieval-augmented generation”). That’s why I amended it to “oversimplified”: you’re, for no explicable reason, taking well-established terminology that you think people shouldn’t use (for being “unnecessary”), expanding it out to sentence-length, and even then oversimplifying the process.
LLMs do not possess the ability to reason over the information that it is fed. It converts it to numbers and performs arithmetics on it. Augmenting it with scripts won’t change the fundamental nature of how it works.
It takes information and regurgitates it. There is no analytical capability present that makes it able to distinguish the importance between a small segue and the main points. They can just as easily combine several separate facts into a single point, and phrase things in a way that a footnote has as much weight as the main subjects.
Hiding the actual workings behind silly marketing buzzwords serves to sensationalise what these things actually do. It feeds the AI hysteria and further muddles the discussion around them. It’s why laymen think these models are basically magic and buy into the idea that they’re somehow going to solve all our problems.
I love machine learning. It is, and has historically been a fantastic tool for plenty of tasks, but it isn’t magic.
If I implement a script to automate database migrations during application deployment I could definitely market that as Deployment Ready Database Optimisations or some other BS term, but that doesn’t make it more than a simple automation.
LLMs do not possess the ability to reason over the information that it is fed.
Ah, yes, I forgot that if an LLM has no conscious ability to reason, then we shouldn’t have any terminology to describe the general process it’s using to create an output. Case closed. I’m glad you’ve enlightened us about how useful jargon isn’t actually useful. Data goes in, data goes out; you can’t explain that.
That isn’t what I said. You’re doing a pretty good LLM impression yourself.
I hate it when people use unnecessary terms to describe something.
Hiding the actual workings behind silly marketing buzzwords serves to sensationalise what these things actually do.
That is why I hate marketing buzzwords.
Putting an LLM to process the output of a search in a repository of scientific papers isn’t going to automatically make the output useful or accurate. Papers aren’t necessarily high quality just because they’ve been published, just look at the garbage that Lisa Littman, Kenneth Zucker, and their ilk have shat out over the decades.
An LLM, no matter how many scripts or cleverly written prompts you augment it with, will never be able to differentiate good science from bad, and will just as easily give equal credence to garbage papers as it will to actual quality ones. That’s a problem, without “hallucinations” even entering the picture.
Edit: I think the overall idea of the site is awesome, knowledge should be freely available. I just don’t see the value add that an LLM provides. I only see problems with it.
Putting an LLM to process the output of a search in a repository of scientific papers isn’t going to automatically make the output useful or accurate. Papers aren’t necessarily high quality just because they’ve been published.
For someone who likes to get riled up about people not responding to “what you said”, this whole tangent about the accuracy of RAG and the fact scientific papers aren’t automatically 100% reliable is pretty hilarious. Literally nobody was arguing that it makes it “automatically useful or accurate” or that published papers are “necessarily high-quality”.
You’re genuinely acting like you’re taking issue with terminology describing a process because that process isn’t perfect. “RAG” adequately describes a general technique to improve the accuracy of an output of a query to an LLM, and all you’re doing now is pissing and moaning that “um, just because it’s published doesn’t mean it’s *high-quality” – which has categorical fuck-all to do with the usefulness of the term.
We’ll continue to use it, and you’re welcome to continue being annoyeed by it.
PS: I write material that LLMs are trained on as a hobby; sorry if it annoys you that my writing style is coincidentally similar.
Sure, but RAG has a Wikipedia article about the specifics of the process, history of its use, links to papers and articles about it and its advantages and drawbacks. It’s also useful as a feature on a matrix for comparing one tool or model’s capabilities to another. None of that is true of the sentence.
Virtually all of computing could be reduced to voltages across terminals changing over time, but it can still be useful to give specific terms to specific applications of this process, so we have something to talk about.
I hate it when people use unnecessary terms to describe something.
It’s a script that runs a search and then the LLM takes the output of that and reformats it into an answer. It’s the same as feeding it a document and having it rephrase something.
“I hate when people use concise, reasonably common, and understandable terminology. Why can’t we just expand everything into full sentences that are also oversimplified?”
Point it out then.
RAG is literally just polling for information and rewriting it. It’s the same garbage that gave us Gemini telling us to put glue on pizza to prevent the cheese from slipping off.
You can, and should be more critical of where you source the information but it’s not going to magically make language models actually intelligent. It’s not going to make them reason, or be able to properly select what is relevant or not. Just because you give it a bunch of scientific papers doesn’t mean that the stuff they output will be accurate or not misleading.
They’re still just token prediction engines.
Literally here. And sorry, before you posted this, I did quickly edit my comment to “oversimplified”. Because technically yes, it’s searching and using what it’s retrieved mixed with a (modified) user prompt to generate an output. But it’s searching based on a prompt (rewriting it to aid retrieval), often reranking results, stripping the query-specific context from the results into chunks, attempting to resolve contradictions between sources (which is objectively more than just rephrasing), and then synthesizing between whatever its pretraining is and what its retrieval results are (thus “retrieval-augmented generation”). That’s why I amended it to “oversimplified”: you’re, for no explicable reason, taking well-established terminology that you think people shouldn’t use (for being “unnecessary”), expanding it out to sentence-length, and even then oversimplifying the process.
LLMs do not possess the ability to reason over the information that it is fed. It converts it to numbers and performs arithmetics on it. Augmenting it with scripts won’t change the fundamental nature of how it works.
It takes information and regurgitates it. There is no analytical capability present that makes it able to distinguish the importance between a small segue and the main points. They can just as easily combine several separate facts into a single point, and phrase things in a way that a footnote has as much weight as the main subjects.
Hiding the actual workings behind silly marketing buzzwords serves to sensationalise what these things actually do. It feeds the AI hysteria and further muddles the discussion around them. It’s why laymen think these models are basically magic and buy into the idea that they’re somehow going to solve all our problems.
I love machine learning. It is, and has historically been a fantastic tool for plenty of tasks, but it isn’t magic.
If I implement a script to automate database migrations during application deployment I could definitely market that as Deployment Ready Database Optimisations or some other BS term, but that doesn’t make it more than a simple automation.
Ah, yes, I forgot that if an LLM has no conscious ability to reason, then we shouldn’t have any terminology to describe the general process it’s using to create an output. Case closed. I’m glad you’ve enlightened us about how useful jargon isn’t actually useful. Data goes in, data goes out; you can’t explain that.
That isn’t what I said. You’re doing a pretty good LLM impression yourself.
That is why I hate marketing buzzwords.
Putting an LLM to process the output of a search in a repository of scientific papers isn’t going to automatically make the output useful or accurate. Papers aren’t necessarily high quality just because they’ve been published, just look at the garbage that Lisa Littman, Kenneth Zucker, and their ilk have shat out over the decades.
An LLM, no matter how many scripts or cleverly written prompts you augment it with, will never be able to differentiate good science from bad, and will just as easily give equal credence to garbage papers as it will to actual quality ones. That’s a problem, without “hallucinations” even entering the picture.
Edit: I think the overall idea of the site is awesome, knowledge should be freely available. I just don’t see the value add that an LLM provides. I only see problems with it.
For someone who likes to get riled up about people not responding to “what you said”, this whole tangent about the accuracy of RAG and the fact scientific papers aren’t automatically 100% reliable is pretty hilarious. Literally nobody was arguing that it makes it “automatically useful or accurate” or that published papers are “necessarily high-quality”.
You’re genuinely acting like you’re taking issue with terminology describing a process because that process isn’t perfect. “RAG” adequately describes a general technique to improve the accuracy of an output of a query to an LLM, and all you’re doing now is pissing and moaning that “um, just because it’s published doesn’t mean it’s *high-quality” – which has categorical fuck-all to do with the usefulness of the term.
We’ll continue to use it, and you’re welcome to continue being annoyeed by it.
PS: I write material that LLMs are trained on as a hobby; sorry if it annoys you that my writing style is coincidentally similar.
Need the deets asap with all that hot tea low key context? Get on the RAG!
Pre-order access for $5.99/USD month for your first 12 months. You know the next one comin’ soon!
Sure, but RAG has a Wikipedia article about the specifics of the process, history of its use, links to papers and articles about it and its advantages and drawbacks. It’s also useful as a feature on a matrix for comparing one tool or model’s capabilities to another. None of that is true of the sentence.
Virtually all of computing could be reduced to voltages across terminals changing over time, but it can still be useful to give specific terms to specific applications of this process, so we have something to talk about.
is way easier to search then:
So if people want to look into it further and research what it is, instead of taking some persons 1 sentence explanation, they can.
Ironically trying to search for that phrase would work better in a RAG then a standard key word search.
So… Search… Assisted… Generation?
RAG is a name from a research paper that very accurately describes what happens, but your argument seems to say you just don’t like acronyms.
Looking at the whole thing as a workflow, you’d be correct.
But RAG can be a bit more than just running a search, which implies keyword based regex style search.
“RAG”’s tough in acronym form though concept is quite popular right now - decent summary btw I’d say (fully non-expert)