I’m not using LLMs often, but I haven’t had a single clean example of hallucination for 6 months already. This recursive calls work I incline to believe
I got hallucination from trying to find a book I read but didn’t know the title of. And hallucinated NBA play off results of the wrong team winning. And gotten basic math calculations wrong.
Its a language model so its purpose is to string together words that sound like sentences, but it can’t be fully trusted to be accurate. Best it can do is give you source so you can got straight to the resource to read that instead.
It’s decent at generating basic code, and testing yourself to see if it outputs what you want. But I don’t trust it as a resource when it comes to information when even wrong sports facts have been provided.
The three things I’ve found search engine LLMs to be useful for. Searching for laptop since it’s absurdly good at finding weird fucking regional models or odd configurations that arnt on the main pages of most shops.
Like my current laptop wasnt on newegg Amazon or even msi’s own shop. It was on a fucking random ass page on their website that nothing linked to and was some weird ass model that wasn’t searchable even.
The second most useful one was generating a metric crapload of boiler plate json files for a mod.
The third thing is bad dnd roleplaying while I’m bored at work. The hallucinations are a upside lol
Either you’re using them rarely or just not noticing the issues. I mainly use them for looking up documentation and recently had Google’s AI screw up how sets work in JavaScript. If it makes mistakes on something that well documented, how is it doing on other items?
Just a few days ago I tried to feed my home automation logs to copilot in hopes that it might find a reason why my controller jams randomly multiple times per hour. It confidently claimed that as my noise level reported by controller is -100dB (so basically there’s absolutely nothing else on that frequency around, pretty much as good as it can get) it’s the problem and I should physically move the controller to less noisy area. A decent advice in itself, it might actually help on a lot of cases, but in my scenario it’s a completely wrong rabbit hole to dig in. I might still move the thing around to get better reception on some devices but it doesn’t explain why the whole controller freezes for several minutes on random intervals.
I use them at work to get instructions on running processes and no matter how detailed I am “It is version X, the OS is Y” it still gives me commands that don’t work on my version, bad error code analysis, etc.
Hallucination is not just a mistake, if I understand it correctly. LLMs make mistakes and this is the primary reason why I don’t use them for my coding job.
Like a year ago, ChatGPT made out a python library with a made out api to solve my particular problem that I asked for. Maybe the last hallucination I can recall was about claiming that manual is a keyword in PostgreSQL, which is not.
It’s more the hallucinations are due to the fact we have trained them to be unable to admit to failure or incompetence.
Humans have the exact same “hallucinations” if you give them a job then tell them they aren’t allowed to admit to not knowing something ever for any reason.
You end up only with people willing to lie, bullshit and sound incredibly confident.
We literally reinvented the politician with LLMs.
None of the big models are trained to be actually accurate, only to give results no matter what.
I’m not using LLMs often, but I haven’t had a single clean example of hallucination for 6 months already. This recursive calls work I incline to believe
I got hallucination from trying to find a book I read but didn’t know the title of. And hallucinated NBA play off results of the wrong team winning. And gotten basic math calculations wrong.
Its a language model so its purpose is to string together words that sound like sentences, but it can’t be fully trusted to be accurate. Best it can do is give you source so you can got straight to the resource to read that instead.
It’s decent at generating basic code, and testing yourself to see if it outputs what you want. But I don’t trust it as a resource when it comes to information when even wrong sports facts have been provided.
The three things I’ve found search engine LLMs to be useful for. Searching for laptop since it’s absurdly good at finding weird fucking regional models or odd configurations that arnt on the main pages of most shops.
Like my current laptop wasnt on newegg Amazon or even msi’s own shop. It was on a fucking random ass page on their website that nothing linked to and was some weird ass model that wasn’t searchable even.
The second most useful one was generating a metric crapload of boiler plate json files for a mod.
The third thing is bad dnd roleplaying while I’m bored at work. The hallucinations are a upside lol
Either you’re using them rarely or just not noticing the issues. I mainly use them for looking up documentation and recently had Google’s AI screw up how sets work in JavaScript. If it makes mistakes on something that well documented, how is it doing on other items?
Just a few days ago I tried to feed my home automation logs to copilot in hopes that it might find a reason why my controller jams randomly multiple times per hour. It confidently claimed that as my noise level reported by controller is -100dB (so basically there’s absolutely nothing else on that frequency around, pretty much as good as it can get) it’s the problem and I should physically move the controller to less noisy area. A decent advice in itself, it might actually help on a lot of cases, but in my scenario it’s a completely wrong rabbit hole to dig in. I might still move the thing around to get better reception on some devices but it doesn’t explain why the whole controller freezes for several minutes on random intervals.
I use them at work to get instructions on running processes and no matter how detailed I am “It is version X, the OS is Y” it still gives me commands that don’t work on my version, bad error code analysis, etc.
Hallucination is not just a mistake, if I understand it correctly. LLMs make mistakes and this is the primary reason why I don’t use them for my coding job.
Like a year ago, ChatGPT made out a python library with a made out api to solve my particular problem that I asked for. Maybe the last hallucination I can recall was about claiming that
manualis a keyword in PostgreSQL, which is not.It’s more the hallucinations are due to the fact we have trained them to be unable to admit to failure or incompetence.
Humans have the exact same “hallucinations” if you give them a job then tell them they aren’t allowed to admit to not knowing something ever for any reason.
You end up only with people willing to lie, bullshit and sound incredibly confident.
We literally reinvented the politician with LLMs.
None of the big models are trained to be actually accurate, only to give results no matter what.
What is a hallucination if not AI being confidently mistaken by making up something that is not true?