@lcamtuf As an average: yes.
-
@lcamtuf As an average: yes. In a specific domain, if you are a true expert you might be slightly better than it. This leads to a sort of "reverse imposter syndrome" where most experts poke at the LLM until they find an "AHA! it doesn't know something!" and then in a sort of reverse Gell-Mann Amnesia trope apply it to everything the LLM does. Forgetting that most people are NOT experts in anything, let alone multiple things, so statistically the LLM is going to beat the pants off the average human.
E.g. they can crank out limericks:
If you’re an expert in your own domain,
You might find the LLM's brain a bit plain.
Reverse imposter you'll feel,
When you catch a small deal,
But forget most folks can't do the same! -
Rich Felkerreplied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf I hope you're joking.
-
kurtseifried (he/him)replied to Rich Felker last edited by
@dalias @lcamtuf not really, I’m working on an article but TLDR a combo of:
-Gartner hype cycle (trough of disillusionment)
-reverse imposter syndrome/ gell-Mann amnesia as mentioned
-uncanny valleySeem to go a long way to explaining why so many people are really hating on AI. The classic “they’re taking our jobs” is also part of it, and training on everyone hard written work, but this doesn’t explain the richness and variation in how people despise AI and LLM specifically.
-
Rich Felkerreplied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf OK but LLM's are just laughably stupid all the time. They absolutely do not "beat the pants off the average human being" at anything.
-
Ariadne Conill 🐰:therian:replied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf LLMs are not capable of “knowing” anything. they may predict sequences of text that sometimes match up to actual facts, but they’re just MegaHAL on steroids. it’s just luck (and fairly good statistical modeling).
as for “knowing” things in humans, that too is a complicated subject, as skills and facts are handled by completely different subnetworks in the DMN, and facts are remembered based on encoded opinions about them more than anything else.
-
Ariadne Conill 🐰:therian:replied to Ariadne Conill 🐰:therian: last edited by
(an interesting experiment to do, at least in the US to illustrate the contextual encoding influence on fact recall is to ask your friends about January 6.)
-
Risottoreplied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf LLMs are an amusing nick-nack.
they're worryingly confident sounding,
which can mislead people who know nothing about the subject.
(it's not reverse Gell-Mann Amnesia, it's just plain Gell-Mann Amnesia, every expert in every field knows they're a slightly more carbon intensive version of the same guessing that Bayes classifiers use. They don't "understand"; they guess with heuristics. It's the Chinese room experiment all over again, and they're usually easily spotted for making awful mistakes.)
-
kurtseifried (he/him)replied to Risotto last edited by
@risottobias @lcamtuf We don't need things that "understand" in many use cases. How many people carry out procedures that they don't understand? Hint: whenever someone complains about safety and wants to take a short cut, they probably don't understand why that safety regulation exists.
-
kurtseifried (he/him)replied to Rich Felker last edited by
@dalias @lcamtuf Ok, one real world example of what I use LLM for: can you please analyze these pet insurance policies and tell me what's good/bad/different about them, in less than 5 minutes? https://www.petsplusus.com/sites/default/files/PPU_B_11.2017%20PC%20Policy%20WordingswStat%20FA1217NC.PDF and https://www.trupanion.com/docs/trupanionwebsitelibraries/trupanion/files/pdfs/2019_policy_english_sample.pdf?sfvrsn=e8db2911_2
I bet ChatGPT knows more about contract law and pet insurance than I do. Also you can validate their output and conclusions to a reasonable degree:
"can you provide relevant sections of text and the page number to backup each item above?"
-
Rich Felkerreplied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf "I bet ChatGPT knows more about contract law and pet insurance than I do."
No, it does not know anything whatsoever about either. It "knows" how to put together words in plausible sounding ways. The output it gave you looks like content farm drivel I'd hit the back button as soon as I landed on. It's worse than 50-50 whether any particular detail on anything like that is correct. On a domain like this you HAVE TO RTFC anyway.
-
Ariadne Conill 🐰:therian:replied to Rich Felker last edited by
@dalias @kurtseifried @lcamtuf yeah there is no way i would ever enter into a contract of any significance without reviewing it
-
Corvid Cronereplied to kurtseifried (he/him) last edited by
@kurtseifried @risottobias @lcamtuf
"whenever someone complains about safety and wants to take a short cut, they probably don't understand why that safety regulation exists."
Are you familiar with what a weasel word is?
This is telling me that you are just guessing the reasons why things happen instead of actually investigating them.
-
Risottoreplied to kurtseifried (he/him) last edited by
@kurtseifried @lcamtuf @CorvidCrone ah, someone who's never been racially profiled by a machine learning algorithm.
if you expand your viewpoint outside your privilege, you might see that sometimes the things we do in the industry are harmful.
I think it's easy to dupe leadership positions and c-suite folks with potential and sales pitches, but engineers are wary because they see the actual results.
I was once in a meeting where they mentioned outsourcing some analyst work, "yay", but also mentioned outsourcing board positions, "boo" (yes, AI board members exist)
it is difficult for a man to understand that which his job depends on him not understanding.
e.g., the harm caused by LLMs, or their deficiency, or the bubble the industry's coming on. They haven't delivered any promise (besides Nvidia returns - when in a gold rush, sell shovels. Or... AIaaS)