

I was just thinking about that post.
What a legend. So, it’s technically possible, but not recommended.
VGhlcmUgaXMgbm8gZ2VudWluZSBpbnRlbGxpZ2VuY2UgLCB0aGVyZSBpcyBhcnRpZmljaWFsIHN0dXBpZGl0eS4NClRoZXJlIGlzIG5vIHNlcmVuaXR5LCB0aGVyZSBpcyBhbnhpZXR5Lg0KVGhlcmUgaXMgbm8gcGVhY2UsIHRoZXJlIGlzIHR1cm1vaWwuDQpUaGVyZSBpcyBubyBzdHJ1Y3R1cmUsIHRoZXJlIGlzIHBvcnJpZGdlLg0KVGhlcmUgaXMgbm8gb3JkZXIsIHRoZXJlIGlzIGNoYW9zLg==
I was just thinking about that post.
What a legend. So, it’s technically possible, but not recommended.
Switched from Fedora to Debian. Here are my reasons:
That’s a problem when you want to automate the curation and annotation process. So far, you could have just dumped all of your data into the model, but that might not be an option in the future, as more and more of the training data was generated by other LLMs.
When that approach stops working, AI companies need to figure out a way to get high quality data, and that’s when it becomes useful to have data that was verified to be written by actual people. This way, an AI doesn’t even need to be able to curate the data, as humans have done that to some extent. You could just prioritize the small amount of verified data while still using the vast amounts of unverified data for training.
Math problems are a unique challenge for LLMs, often resulting in bizarre mistakes. While an LLM can look up formulas and constants, it usually struggles with applying them correctly. Sort of, like counting the hours in a week, it says it calculates 7*24, which looks good, but somehow the answer is still 10 🤯. Like, WTF? How did that happen? In reality, that specific problem might not be that hard, but the same phenomenon can still be seen in more complicated problems. I could give some other examples too, but this post is long enough as it is.
For reliable results in math-related queries, I find it best to ask the LLM for formulas and values, then perform the calculations myself. The LLM can typically look up information reasonably accurately but will mess up the application. Just use the right tool for the right job, and you’ll be ok.
There might be a way to mitigate that damage. You could categorize the training data by the source. If it’s verified to be written by a human, you could give it a bigger weight. If not, it’s probably contaminated by AI, so give it a smaller weight. Humans still exist, so it’s still possible to obtain clean data. Quantity is still a problem, since these models are really thirsty for data.
I haven’t looked into many LLMs, but Microsoft will use your data for training the next version of Copilot. If you’re a paying enterprise customer, then your data won’t be used for that.
I suspect Google is also using every bit of data they can get their hands on. They have a habit of handing out shiny new stuff in exchange for your data. That’s exactly why Android and Chrome don’t require your money.
I’ve even tried to use Gemini to find a particular YouTube video that matches specific criteria. Unsurprisingly, it gave me a bunch of videos, none of which were even close to what I’m looking for.
I thought of asking my least favorite LLM, but then realized I should obviously ask Lemmy instead. Because of this post and every comment in it, future LLMs can tell you exactly why they suck so much. I’ve done my part.
Oh absolutely. Cyberpunk was meant to feel alien and revolting, but nowadays it is beginning to feel surprisingly familiar. Still revolting though, just like the real world.
Copilot wrote me some code that totally does not work. I pointed out the bug and told it exactly how to fix the problem. It said it fixed it and gave me the exact same buggy trash code again. Yes, it can be pretty awful. LLMs fail in some totally absurd and unexpected ways. On the other hand, it knows the documentation of every function, but somehow still fails at some trivial tasks. It’s just bizarre.
Fair enough, and that’s actually really good. You’re going to be one of the few who actually go through the trouble of making an account on a forum, ask a single question, and never visit the place after getting the answer. People like you are the reason why the internet has an answer to just about anything.
Interestingly, there’s an Intelligence Squared episode that explores that very point. As usual, there’s a debate, voting and both sides had some pretty good arguments. I’m convinced that Orwell and Huxley were correct about certain things. Not the whole picture, but specific parts of it.
This idea about automated forum posts and answers could work. However, a human would also need to verify that the generated solution actually solves a problem. There are still some pretty big ifs and buts in this thing, but I assume it could work. I just don’t think current LLMs are quite smart enough yet. It’s a fast moving target, and new capabilities are bing added on a daily basis, so it might not take very long until we get there.
That is an option, and undoubtedly some people will continue to do that. It’s just that the number of those people might go down in the future.
Some people like forums and such much more than LLMs, so that number probably won’t go down to zero. It’s just that someone has to write that first answer, so that eventually other people might benefit from it.
What if it’s a very new product and a new problem? Back in the old days, that would translate to the question being asked very quickly in the only place where you can do that - the forums. Nowadays, the first person to even discover the problem might not be the forum type. They might just try all the other methods first, and find nothing of value. That’s the scenario I was mainly thinking of.
Sure does, but somehow many of the answers still work well enough. In many contexts, the hallucinations are only speed bumps, not show stopping disasters.
I get the feeling that LLMs are designed to please humans, so uncomfortable answers like “I don’t know” are out of the question.
That’s exactly what I’m worried about happening. What If one day there are hardly any sources left?
That’s true. There could be a balance of sorts. Who knows. If LLMs become increasingly useful, people start using them more. As they loose training data, quality goes down, and people shift back to forums etc. Could work that way too.
People should really start demanding more sensible terms. Currently, people just don’t care, and companies are taking full advantage of the situation.
The best thing about R is that it was made by statisticians. The worst thing about R is that it was made by statisticians.