This artificial pseudointelligence exists because there’s the “gee whiz, that’s cool” of a computer talking like a person, and a bunch of hype chasers looking to cash in. Much like cryptocurrency before it, and the dot-com boom before that, there is little substance to it, and most of it will be commercially irrelevant a decade from now.
"to keep the quality of answers high, we may arbitrarily close questions, regardless of how many upvotes it gets and how helpful it is" - stackoverflow
I’m not liking the announced changes to search. That sounds like we will be losing the lexical search and in exchange we will be getting the same technology that allows google to answer questions different to the one we asked.
How many minutes between starting to use OverflowAI until we get something like “As a large language model trained by the Stack Exchange Network i can not answer duplicated questions”.
Can someone tell me what their angle is? Are user’s supposed to curate and help train the model for free? Is it just a model trained on stackoverflow data?
All their data is open so what edge do they over the already established competition.
This type of Q&A interface is very popular and stealing traffic away from sites like Google and Stack Overflow. Stack Overflow can train it on their data and has a feature where it links to every answer it pulled from. I think that’s a nice feature and like that I can troubleshoot further on my own, as AI can often hallucinate an answer or lose a piece of context I need.
Hah, good to know that even on [email protected] there are people who agree that stack overflow moderation is too draconian to ask questions in anymore. It’s a good resource, though, so an LLM will probably be the answer to make the knowledge base more usable without angering its elder gods.
I’ve found it great for asking documentation questions. It saves me a ton of time having to search through documentation myself. The problem is when it encounters something it doesn’t have information on, it’ll just confidently make shit up, and if you’re not enough of an expert to recognize when that happens, you can be mislead. It still saves me time, but I use it as a recall tool to get me started when I’m learning to do something new, I’d never use the code it puts out without reading through it line by line. I’m also experienced enough to know when it’s wrong and how to refactor its examples. People new to programming could get set down the wrong path by over relying on gpt to teach them.
Yeah works well, as long as the code is rather simple and it occurred rather often in the training set. But I seldom use it currently (got a little bit more complex stuff going on). It’s good though to find new stuff (as it often introduces a new library I haven’t known yet). But actual code… I’m writing myself (tried it often, and the quality just isn’t there… and I think it even got worse over the last couple of months as also studies suggest)
I’ve gotten really good results asking chat gpt for programming help. Problem is that it’s wrong like 10% of the time, and when it’s wrong it’s very confidently incorrect. That wasn’t a problem for me because I knew when it was wrong and could course correct it and get the correct solution and it still saved me time and helped me eventually get to the right solution. But if someone who’s still getting started is trying to use chat gpt to learn, they could easily be mislead because they won’t know when its output is wrong.
Definitely depends on the type of question. I find for documentation type questions I get the 90% good answers, like how do I do something with this library, it’s good, which makes sense because that libraries documentation is probably in the training data. But for more open ended questions, like how do I solve this problem, I see similar performance to what you’re saying. I think it’s a good retrieval and synthesises tool which can really save a ton of time if you already have a high level plan of action and just use it to fill in some specific details.
Well, using ai-generated answers to train their own ai would bring down the quality of answers and worse quality means lesser money. Don’t you want them to make any money??!!
Stack Overflow is unique as a page, in the sense that its contributions are under a license that allows for reuse (Creative Commons Share-Alike) as long as the individual users are properly credited. Does this mean that OverflowAI keeps the credit metadata and knows who wrote each individual part of an answer?
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.
Then I’m guilty of breaking the license. I have always been stealing code from Stack Overflow. Well, since I’m a senior dev right now I steal only from answers.
I understand Google and Microsoft getting into it as it makes sense as a “better” Google search but for StackOverflow that sounds like they have just given up on their current platform.
Add comment