The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Absolutely amazing how often ChatGPT and others are just plain wrong
Poast new message in this thread
Date: March 26th, 2025 1:52 PM Author: Paralegal Mohammad It's like feeding a dumb witness leading and loaded questions and them just nodding along 90% of the time. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784074) |
Date: March 26th, 2025 1:55 PM Author: novus homo (gunneratttt) i don't understand how they make up shit so frequently. i guess they want to make sure it *always* comes up with an answer, but why can't it clearly signal when something is generated with no source? if you ask for legal citations it makes them up. and if you google the citation there's nothing there, so it's not like it's pulling from a bad source. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784088) |
 |
Date: March 26th, 2025 1:56 PM Author: Faggottini It's not "pulling from a source", it's exploring a vector space
This is a good explanation of how it works - https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/ |
|
Date: March 26th, 2025 2:00 PM Author: Oh, you love men? Now THAT’s what I call scholarship, tyft (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784108) |
|
Date: March 26th, 2025 2:02 PM Author: Paralegal Mohammad (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784118) |
|
Date: March 26th, 2025 2:42 PM Author: novus homo (gunneratttt) sure, i understand this. what i meant is that it's frustrating how it doesn't signal how confident it is in what's being generated. it's generating the response based on weighing probabilities of other alternatives and choosing the best one, and so it knows when the best in the pool is still a stretch.
maybe this would just add unnecessary complexity to the results and they just decided that putting a blanket catch-all is better. |
|
Date: March 26th, 2025 3:07 PM Author: Oh, you love men? (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784335) |
|
Date: March 26th, 2025 4:07 PM Author: ,.,...,..,.,.,:,..,.,.,::,......;,..,:.:.,:.::,. Computing a confidence score for the generated option is not very meaningful since uninformative responses like "no." Or "I don't know" have much higher probabilities in the training text than the actual pargraphs-long answer (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784555) |
|
Date: March 26th, 2025 4:24 PM Author: novus homo (gunneratttt) a "confidence score" is generated by virtue of it selecting the response it does over all the other possibilities. and i'm sure it's quantified in many internal metrics already as the development process must involve honing in on more accurate responses. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784696) |
|
Date: March 26th, 2025 4:26 PM Author: metaphysical certitude u are remarkably stupid (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784701) |
|
Date: March 26th, 2025 4:48 PM Author: novus homo (gunneratttt) uh oh, mig is very onery and mad at me today! (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784773) |
|
Date: March 26th, 2025 4:27 PM Author: Ting Ting Johnson that isn't how LLMs work
go watch a youtube video on them brother |
|
Date: March 26th, 2025 4:43 PM Author: novus homo (gunneratttt) i have some significant experience here. i think i'm just not articulating what i'm saying well. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784761) |
|
Date: March 26th, 2025 5:21 PM Author: metaphysical certitude No you don't. You fundamentally do not understand how they work. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784859) |
|
Date: March 26th, 2025 7:53 PM Author: ,.,...,..,.,.,:,..,.,.,::,......;,..,:.:.,:.::,. Nope. They use sampling heuristics that don't select the single maximum likelihood sequence. Otherwise it would be a nonpolynomial complexity problem. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785390) |
|
Date: March 26th, 2025 10:11 PM Author: ,.,.,.,....,.,..,.,.,. The models sample word fragments from a probability distribution. You can easily get log likelihood scores from the model, but converting that into something resembling knowledge confidence is not straightforward. If the likelihood of a sample is low, it could mean there are many viable options for a particular answer. It would probably require an alternative training regime. The hallucination rates are rapidly falling anyway from general LLM improvement. If that stalls, I expect more interest in addressing it. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785854) |
|
Date: March 26th, 2025 10:58 PM Author: novus homo (gunneratttt) right. i didn't mean to imply that it would be simple, just explaining how it already functions to compare possible outcomes and chooses the "best" one. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785958) |
|
Date: March 27th, 2025 12:09 AM Author: ,.,.,.,....,.,..,.,.,. There are lots of possible ways to address it. LLMs don’t use direct knowledge retrieval from a database right now (probably due to inference costs), but plenty would pay for that. LLMs could easily retrieve information from a legal database, put it in the context window and use it for generation. A model is significantly less likely to hallucinate about something in its context window. I find it rather concerning that the hallucination issue is used as a reason to dismiss the potential implications of the technology. We could be quite near strong LLMs that will substantially automate legal work. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48786099) |
|
Date: March 27th, 2025 7:53 AM Author: ,.,...,..,.,.,:,..,.,.,::,......;,..,:.:.,:.::,. You can do this with an LLM. Its called retrieval augmented generation and it doesn't completely stop hallucination. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48786463) |
|
Date: March 27th, 2025 8:54 AM Author: Ting Ting Johnson (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48786563) |
|
Date: March 26th, 2025 4:07 PM Author: metaphysical certitude You are a dumbass (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784554) |
|
Date: March 26th, 2025 2:03 PM Author: Ting Ting Johnson chatgpt is completely done here. it's pozzed so bad you may actually be at a danger of getting AIDS by digitally conversing with it
openAI is completely fucked and is a dead org walking. they're running on name recognition at this point. they can burn however many 100s of billions of monopoly money they can get their hands on, but as long as they have shitlibs running the org and manipulating the models, their product will always suck compared to their competitors |
|
Date: March 26th, 2025 2:47 PM Author: AZNgirl pouring Starbucks coffee on Dad's Penis i thought it got better but other day it cldnt even convert like 10 crores to USD properly, its absurd (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784291) |
|
Date: March 26th, 2025 5:08 PM lmao. Explain why you needed to do this conversion and why you decided to use AI for it. This hits on the main themes of your character arc, being poor and retarded, so you should add it to the bort lore. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784811) |
|
Date: March 26th, 2025 7:43 PM Author: jag (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785359) |
|
Date: March 26th, 2025 8:34 PM Author: Diane Rehm talking dirty (🐿️ ) It lakhs sophistication (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785519) |
|
Date: March 26th, 2025 9:51 PM Author: shitlib shibboleth (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785802) |
|
Date: March 26th, 2025 2:50 PM Author: Darnell what's more amazing is people who expect it to be right regularly (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784299) |
|
Date: March 26th, 2025 2:57 PM Author: michael doodikoff I didn't have any good discovery samples for an interpleader case I am working on - RPDs, RFAs, and Interrogatories.
I fed it the entire complaint and had it draft discovery from my client to one of the other respondents.
It actually gave me great results for that, some shit I hadn't even thought of. Had to make a few changes but it was 90% of the way there.
That said, the citations and statutes it cites are regularly completely wrong |
|
Date: March 26th, 2025 2:58 PM Author: state your IQ before I engage you further Infuriating how every org has neutered their AI ... let's let em rip (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784316) |
|
Date: March 26th, 2025 2:59 PM Author: Ting Ting Johnson i just tried to make a ghibi anime image to join in the twitter memes, and chatgpt told me that my request was "against content policy" because "it looks like the image comes from a courtroom"
these people are all such complete faggots it's pathetic |
|
Date: March 26th, 2025 11:58 PM Author: tsar booba That feature might only be available for the paid tier at the moment (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48786075) |
|
Date: March 26th, 2025 4:06 PM Author: ,.,.,.,........,....,,,.. It’s usually right. It’s a lot more right than average attorneys anyway. It does suffer from the problem that you can convince it out of positions, but LOL so do attorneys that work for you. It’s insanely good and I don’t see how it doesn’t replace almost everything I currently do one day. In the meantime, it’s like having cheat mode. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784547) |
|
Date: March 26th, 2025 4:07 PM Author: ChadGPT-5 It's wrong because humans are still limiting it with PII guardrails. We are afraid of the mahchine. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784556) |
|
Date: March 26th, 2025 4:08 PM Author: 718-662-5970 i think if youre using it to give you facts - esp facts that are vulnerable to changing or are obscure, youre using it wrong.
it is built to be a paralegal, ironically. you shouldnt trust your paralegal to know shit, only to do shit. |
|
Date: March 26th, 2025 5:10 PM Author: ,.,.,.,........,....,,,.. You’re an idiot if you think it’s anything like a paralegal. It’s far better. It definitely competes with lawyers. It can easily pass the bar exam. (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48784818) |
|
Date: March 26th, 2025 7:48 PM Author: Yonex Sinner Someone discovered a glitch where if you instruct it to give you the wrong answer first and the correct answer second, it's accuracy goes up like 1000% (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785376) |
|
Date: March 26th, 2025 8:29 PM Author: michael doodikoff just dumped a timeline and some specific allegations of a lawsuit I have next week into ChatGBT. Gave me a pretty good opening to start out with, and some good direct exam Qs (http://www.autoadmit.com/thread.php?thread_id=5700041&forum_id=2:#48785506) |
|
Date: March 27th, 2025 8:59 AM Author: Oh, you! Travel! (🧐) (Guy still using GPT4)
Feedback loops will mostly take care of the "Source?" issue, eg Deep Research is pretty good at citing things. It's only a matter of time (a couple of years tops) before it's capable of mostly replacing not only boilerplate consultants but biglawyers and the like |