Tech experts are starting to doubt that ChatGPT and A.I. ‘hallucinations’ will ever go away: ‘This isn’t fixable’
Tech experts are starting to doubt that ChatGPT and A.I. ‘hallucinations’ will ever go away: ‘This isn’t fixable’
Even OpenAI CEO Sam Altman was skeptical a few weeks ago: "I probably trust the answers that come out of ChatGPT the least of anybody on Earth."
Tech experts are starting to doubt that ChatGPT and A.I. ‘hallucinations’ will ever go away: ‘This isn’t fixable’::Experts are starting to doubt it, and even OpenAI CEO Sam Altman is a bit stumped.
You're viewing a single thread.
I was excited for the recent advancements in AI, but seems the area has hit another wall. Seems it is best to be used for automating very simple tasks, or at best used as a guiding tool for professionals (ie, medicine, SWE, …)
Hallucinations is common for humans as well. It's just people who believe they know stuff they really don't know.
We have alternative safeguards in place. It's true however that current llm generation has its limitations
Not just common. If you look at kids, hallucinations come first in their development.
Later, they learn to filter what is real and what is not real. And as adults, we have weird thoughts that we suppress so quickly that we hardly remember them.
And for those with less developed filters, they have more difficulty to distinguish fact from fiction.
Generative AI is good at generating. What needs to be improved is the filtering aspect of AI.
Hell, just look at various public personalities - especially those with extreme views. Most of what some of them say they have "hallucinated". Far more so than what GPT chat is doing.
Sure, but these things exists as fancy story tellers. They understand language patterns well enough to write convincing language, but they don't understand what they're saying at all.
The metaphorical human equivalent would be having someone write a song in a foreign language they barely understand. You can get something that sure sounds convincing, sounds good even, but to someone who actually speaks Spanish it's nonsense.
Calculators don't understand maths, but they are good at it.
LLMs speak many languages correctly, they don't know the referents, they don't understand concepts, but they know how to correctly associate them.
What they write can be wrong sometimes, but it absolutely makes sense most of the time.
but it absolutely makes sense most of the time
I'd contest that, that shouldn't be taken for granted. I've tried several questions in these things, and rarely do I find an answer entirely satisfactory (though it normally sounds convincing/is grammatically correct).
This is the reply to your message by our common friend:
I understand your perspective and appreciate the feedback. My primary goal is to provide accurate and grammatically correct information. I'm constantly evolving, and your input helps in improving the quality of responses. Thank you for sharing your experience. - GPT-4
I'd say it does make sense
Song written by an Italian intended to sound like american accented english but its intentionally gibberish.
Here is an alternative Piped link(s): https://piped.video/-VsmF9m_Nt8
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I'm open-source, check me out at GitHub.
GPT can write and edit code that works. It simply can't be true that it's solely doing language patterns with no semantic understanding.
To fix your analogy: the Spanish speaker will happily sing along. They may notice the occasional odd turn of phrase, but the song as a whole is perfectly understandable.
Edit: GPT can literally write songs that make sense. Even in Spanish. A metaphor aiming to elucidate a deficiency probably shouldn't use an example that the system is actually quite proficient at.
Sure it can, "print hello world in C++"
#include int main() { std::cout << "hello world\n"; return 0; }"print d ft just rd go t in C++"
#include int main() { std::cout << "d ft just rd go t\n"; return 0; }The latter is a "novel program" it's never seen before, but it's possible because it's seen a pattern of "print X" and the X goes over here. That doesn't mean it understands what it just did, it's just got millions (?) of patterns it's been trained on.
Because it can look up code for this specific problem in its enormous training data? It doesnt need to understand the concepts behind it as long as the problem is specific enough to have been solved already.
I can tell GPT to do a specific thing in a given context and it will do so intelligently. I can then provide additional context that implicitly changes the requirements and GPT will pick up on that and make the specific changes needed.
It can do this even if I'm trying to solve a novel problem.
But the naysayers will argue that your problem is not novel and a solution can be trivially deduced from the training data. Right?
I really dislike the simplified word predictor explanation that is given for how LLM's work. It makes it seem like the thing is a lookup table, while ignoring the nuances of what makes it work so well.
But the naysayers will argue that your problem is not novel and a solution can be trivially deduced from the training data. Right?
Yes, obviously. Unless @Serdan is publishing papers about their solutions to previously unsolved computation problems, we should assume that by "novel problem" they actually just mean "a mundane problem for which every step of the solution is trivial, even if they've never been combined in that exact order before".
If that were true, it shouldn't hallucinate about anything that was in its training data. LLMs don't work that way. There was a recent post with a nice simple description of how they work, but I'm not finding it. If you're interested, there's plenty of videos and articles describing how they work.
It doesn't have the ability to just look up anything from its training data, that stuff is encoded in its parameters. Still, the input has to be encoded in a way that causes the correct "chain reaction" of excited/not excited neurons.
Beyond that, it's not just a carbon copy from what was in the training either because you can tell it what variable names to use, which order to do things in, change some details, etc. If it was simply a lookup that wouldn't be possible. The training made it able to generalize what it learned to some extent.
Yes, but it doesnt do so because it understands what a variable is, it does so because it has statistics as to where variables belong most likely.
In a way it is like the guy that won the french scrabble championship without speaking a single word of french, by learning the words in the dictionary.
You are two - CGP Grey us a good video about it.
Here is an alternative Piped link(s): https://piped.video/wfYbgdo8e-8
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I'm open-source, check me out at GitHub.
Humans can recognize and account for their own hallucinations. LLMs can't and never will.
It's pretty ironic that you say they "never will" in this context.
They can't... Most people strongly believe they know many things while they have no idea what they are talking about. Most known cases are flat earthers, qanon, no-vax.
But all of us are absolutely convinced we know something until we found out we don't.
That's why double blind tests exists, why memories are not always trusted in trials, why Twitter is such an awful place
Well to be honest it is the best way, I mean, I'm pretty sure their purpose was a tool to aid people, and not to replace us... Right?
Yeah I fully expect to see genre specific LLMs that have a subscription fee attatched squarely aimed at hobbies and industries.
When I finally find my new project car I would absolutely pay for a subscription to an LLM that has read every service manual and can explain to me in plain english what precise steps the job involves and can also answer followup questions.
That's what I'm expecting too.
I've been using chatGPT instead of reading the documentation of the programming language I am working in (ABAP). It's way faster to get an answer from chatGPT than finding the relevant spots in the docs or through google, although it doesn't always work.
If you take an LLM and feed it documentation and relevant internet data of specific topics, it can be a quite helpful tool. I don't think LLMs will get much farther than that, but we'll see.
It will just take removing the restrictions so people can make porn, then monetizing that to fund more development.
A story as old as media.