Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Large language models reduce public knowledge sharing on online Q&A platforms

academic.oup.com

Post reply on HN
Print view View original

141–150 of 366 posts

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#141

Post by finolex1 »

insane_dreamer wrote: post #120

The problem is eventually what are LLMs going’s to draw from? They’re not creating new information, just regurgitating and combining existing info. That’s why they perform so poorly on code for which there aren’t many many publicly available samples, SO/reddit answers etc.

There is still publicly available code and documentation to draw from. As models get smarter and bootstrapped on top of older models, they should need less and less training data. In theory, just providing the grammar for a new programming language should be enough for a sufficiently smart LLM to answer problems in that language.

Unlike freeform writing tasks, coding also has a strong feedback loop (i.e. does the code compile, run successfully, and output a result?), which means it is probably easier to generate synthetic training data for models.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#142

Post by lrpanb »

appendix-rock wrote: post #90
LeadB wrote: post #57

Earlier quoted context omitted.

For the major programming languages, it must be a pretty esoteric question if it does not have an answer yet. Increasingly, the free products of experts are stolen from them with the pretext that "users need to be protected". Entire open source projects are stolen by corporations and the experts are removed using the CoC wedge. Now SO answers are stolen because the experts are not trained like hotel receptionists (wh…

I’m sorry that you ran afoul of a CoC or whatever, but this sounds like a real ‘airing dirty laundry’ tangent.

One man's tangent is another man's big picture. It may be the case of course that some people guilty of CoC overreach are shaking in their boots right now because they went further than their corporations wanted them to go.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#143

Post by roughly »

zmgsabst wrote: post #100
delichon wrote: post #85

Earlier quoted context omitted.

I've gotten answers from OpenAI that were technically correct but quite horrible in the longer term. I've gotten the same kinds of answers on Stack Overflow, but there other people are eager to add the necessary feedback. I got the same feedback from an LLM but only because in that case I knew enough to ask for it. Maybe we can get this multi-headed advantage back from LLMs by applying a team of divergent AIs to the…

You can usually even ask the same LLM: - do a task - criticize your job on that task - redo that task based on criticism I find giving the LLM a process greatly improves the results.

What’s fun is that you can skip step 1. The LLM will happily critique its own nonexistent output.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#144

Post by n_ary »

internet101010 wrote: post #110
tlogan wrote: post #67

Earlier quoted context omitted.

The main issue with Stack Overflow (and similar public Q&A platforms) is that many contributors do not know what they do not know, leading to inaccurate answers. Additionally, these platforms tend to attract a fair amount of spam (self promotion etc) which can make it very hard to find high-quality responses.

Medium is even worse about this. It's more self-promotion than it is common help.

Begin rant.

I don’t want to be that guy saying this, but 99% of the top results on google from Medium related to anything technical is literally the reworded/reframed version of the official quick start guide.

There are some very rare gems, but it is hard to find those among the above mentioned ocean of reworded quick starts disguised as “how to X”, “fixing Y”. Almost reminds me of the SEO junks when you search “how to restart iPhone” and find answers that dance around letting it die from battery drain and then charge, install this software, take it to the apple repair shop, go to settings and traverse many steps while not saying that if you are between these models use the power+volume up button trick.

End of rant.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#145

Post by jneagu »

mycall wrote: post #125
insane_dreamer wrote: post #120

The problem is eventually what are LLMs going’s to draw from? They’re not creating new information, just regurgitating and combining existing info. That’s why they perform so poorly on code for which there aren’t many many publicly available samples, SO/reddit answers etc.

I thought synthetic data is what is partially training the new multimodal large models, i.e. AlphaGeometry, o1, etc.

Yeah, There was a reference in a paywalled article a year ago (https://www.theinformation.com/articles/openai-made-an-ai-br...): "Sutskever's breakthrough allowed OpenAI to overcome limitations on obtaining high-quality data to train new models, according to the person with knowledge, a major obstacle for developing next-generation models. The research involved using computer-generated, rather than real-world, data like text or images pulled from the internet to train new models."

I suspect most foundational models are now knowingly trained on at least some synthetic data.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#146

Post by SunlitCat »

rq1 wrote: post #61

People should just share their conversations with the LLMs online no? This would blogging 5.0. Or web 7.0.

Well, I just asked ChatGPT to answer my "How to print hello world in c++" with a typical stack overflow answer.

Lo and behold, the answer is very polite, explanative and even lists common mistakes. It even added two very helpful user comments!

I asked it again how this answer would look in 2024 and it just updated the answer to the latest c++ standard!

Then! I asked it what a moderator would say when they chime in. Of course the moderator reminded everyone to stay on focus regarding the question, avoid opinions and back their answer by documentation or standards. In the end the mod thanked for everyone's contribution and keeping the discussion constructive!

Ah! What a wonderful world ChatGPT is living at! I want to be there too!

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#147

Post by elicksaur »

stickfigure wrote: post #133
insane_dreamer wrote: post #120

The problem is eventually what are LLMs going’s to draw from? They’re not creating new information, just regurgitating and combining existing info. That’s why they perform so poorly on code for which there aren’t many many publicly available samples, SO/reddit answers etc.

> The problem is eventually what are LLMs going’s to draw from? Published documentation. I'm going to make up a number but I'll defend it: 90% of the information content of stackoverflow is regurgitated from some manual somewhere. The problem is that the specific information you're looking for in the relevant documentation is often hard to find, and even when found is often hard to read. LLMs are fantastic at reading…

Following the article’s conclusion farther, humans would stop producing new documentation with new concepts.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#148

Post by jneagu »

generic92034 wrote: post #136
jneagu wrote: post #126

Earlier quoted context omitted.

Edit: OP had actually qualified their statement to refer to only underrepresented coding languages. That's 100% true - LLM coding performance is super biased in favor of well-represented languages, esp. in public repos. Interesting - I actually think they perform quite well on code, considering that code has a set of correct answers (unlike most other tasks we use LLMs for on a daily basis). GitHub Copilot had a 30%+…

The OP has qualified "code" with bad availability of samples online. My experience with LLMs on a proprietary language with little online presence confirms their statement. It is not even worth trying, in many cases.

Fair point - I actually had parsed OP's sentence differently. I'll edit my comment.

I agree, LLMs performance for coding tasks is super biased in favor of well-represented languages. I think this is what GitHub is trying to solve with custom private models for Copilot, but I expect that to be enterprise only.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#149

Post by n_ary »

insane_dreamer wrote: post #120

The problem is eventually what are LLMs going’s to draw from? They’re not creating new information, just regurgitating and combining existing info. That’s why they perform so poorly on code for which there aren’t many many publicly available samples, SO/reddit answers etc.

LLMs show their limits as you try to ask something new(introduced in last 6-12 months) being not used. I was asking Claude and GPT4o about a new feature of go, it just gave me some old stuff from go docs. Then I went to go docs(official) and found what I was looking for anyways, the feature was released 2 major versions back, but somehow neither GPT4o nor claude know about this.

Re: Large language models reduce public knowledge sharing on online Q&A platforms

#150

Post by y7 »

mycall wrote: post #125
insane_dreamer wrote: post #120

The problem is eventually what are LLMs going’s to draw from? They’re not creating new information, just regurgitating and combining existing info. That’s why they perform so poorly on code for which there aren’t many many publicly available samples, SO/reddit answers etc.

I thought synthetic data is what is partially training the new multimodal large models, i.e. AlphaGeometry, o1, etc.

Synthetic data can never contain more information than the statistical model from which it is derived: it is simply the evaluation of a non-deterministic function on the model parameters. And the model parameters are simply a function of the training data.

I don't see how you can "bootstrap a smarter model" based on synthetic data from a previous-gen model this way. You may as well well just train your new model on the original training data.

Post reply on HN