Which is exactly what the paper recommends! As long as you have something that isn’t an LLM in the pipeline to vet the output and you’re aware is the tech’s limitations, they can be useful tools. But some of those limitations might be a more solid barrier than some sales departments would like us to believe.

For those of you who didn’t read the paper, the argument they’re making is similar to Godel’s Incompleteness Theorem: no matter how you build your LLM, there will be a significant number of prompts that make that LLM hallucinate. If the proof holds up then hallucinations aren’t a limitation of the training data or the structure of your particular model, they’re a limitation of the very concept of an LLM. That doesn’t make LLMs useless, but it does mean you shouldn’t ever use one as a source of truth.

Agreed! Especially since the behavior on clicking a thumbnail isn’t very consistent on mobile. Sometimes I click the thumbnail and it expands the image. Sometimes it follows a link. Sometimes going back from a link takes me back to the same Lemmy page I was on, but other times it reloads the page and I can’t find the post I was looking at anymore. An option to open all links in a new tab would really help me not lose the post I was looking at when I clicked.

Prior to the API fiasco, Reddit Inc had demonstrated a pattern of promising changes to the mods which they failed to deliver timely if at all. They’ve acknowledged this pattern, promised to do better, then failed to deliver time and again. That part isn’t new.

Then the API changes were announced and the Reddit community gave Reddit Inc the loudest and most decisive rebuke they ever have. That was the feedback conversation. And Reddit Inc went forward with their plan unchanged. No concessions were made. No concerns were addressed or alleviated. Reddit Inc was informed of what this decision would break and they went ahead and broke it anyway.

As a former mod, there is nothing left to discuss. There is no reason to believe Reddit Inc will act on anything that doesn’t agree with what they’ve already decided to do. I’m not going back to that kind of abusive relationship. They had their chance to listen to feedback and made it clear that they won’t.

Agreed. I find Bing chat is really good when I know almost nothing about what I’m searching, or when I know a whole lot about what I’m searching. Like in your example, if I know exactly what I need but can’t remember its name Bing will read all the spammy beginners’ guides for me and get the answer. And on the opposite end, if I’m looking to buy a gift in a hobby I don’t remotely understand Bing does a pretty good job of holding my hand through the search process.

Weirdly, medium knowledge questions seem to still do better as a basic Google search. If I need to fix an appliance I’ve fixed before, but it’s been a long time so I really need a full walkthrough, the first few results on Google are faster than waiting for Bing to talk through it.

I use the term “autocomplete on steroids” because it gets across a vaguely accurate idea of what an LLM is and how it works to people who are thinking of it like sci-fi movie AI. Sorry if it came across that was my whole reason for considering them not intelligent.

LLMs do seem to pass a lot of intelligence tests we’ve come up with. Talking with one for the first time is a really uncanny experience, it’s a totally different thing than the old voice assistants. But they also consistently fail at tasks that would indicate an understanding of a topic. They produce good looking equations, but the math underneath doesn’t make sense. They hallucinate facts that don’t fit with the rest of what they themselves are saying, but look similar to the way right answers are written and defended. They produce really convincing responses, but when they fail they betray some really basic failures to understand what they’re saying.

I feel that LLMs are brute-forcing the tests people designed to measure intelligence. They can pass the bar exam, but they also contain thousands of successful bar exams to consult and millions of bits of text to glue those answers together with. But if you ask the LLM to actually do the job of a lawyer, they start producing all kinds of garbage that sounds good but doesn’t stand up to scrutiny when someone looks up the hallucinated case references.

Part of the problem is that AI research likes to use terminology that sounds like what people do, when that’s not what the AI actually does.

Large language models are not intelligent in any sense. They are autocomplete on steroids. This is a computer program that was fed a book someone wrote, then mathematically tweaked to be able to guess the next word in a sentence in a way that resembles that book. That’s all it does. It does not think or learn in any sense we’d apply to a human.

To me, LLMs sound like a massive plagiarism engine, and I think they should need to get a license from the authors whose works they used to make the LLM under whatever terms that author wants to give, just like a publisher needs to get permission to print a copy of the work. But copyright law has no easy “bright line” for what counts and what doesn’t. So the courts will have to decide whether what the AI “creates” is similar enough to the original works to count as a violation, or if the AI and its results are transformative enough to count as something new.

I’m just using the mobile website. It works and it doesn’t pester me to use something else, which is really refreshing!