3 ms·

I have a family member that is an attorney in housing law. She claimed that LLMs are not particularly useful for her work. If she asked a simple question like,

by hirvi74 2mo ago

I have a family member that is an attorney in housing law. She claimed that LLMs are not particularly useful for her work. If she asked a simple question like, "Find all the <insert specific housing laws> for all 50 states," then she still has to go and check every single one of the laws. Since the legislature is modified so often, she cannot look at, say, Maryland's law and know if it the LLM output was the 1990, 2014, 2018, or 2026 version of the codified law. In order to fact check the law, she has to look it up, and by that point in time, she has the answer she did the work of the LLM.

ipaddr 2mo ago
This is my experience if you need the data to be accurate often it isn't and you only know that because you had to check because it was important. Things not important you never check.
nottorp 2mo ago
You can reproduce these results even if you're not a lawyer, if you've dabbled in any MMO that has been around for a few years and had multiple patches. Ask any LLM about a random game mechanic and then be prepared to verify if the answer is for the patch in 2025, 2023 or 2020...
nrmitchi 2mo ago
This doesn’t sound as much that “LLMs can’t help me” as much as “trying to one-shot a complicated answer in a chatbot can’t help me”. Output straight from a generative model definitely should not be trusted. But that doesn’t mean that an appropriate harness (even Claude code and a dynamic workflow) can’t complete, and validate, an answer.
- suddenlybananas 2mo ago
  Seems like it would be a massive market to make such a harness, and yet you're making it sound very easy to do. It would be surprising no one has done it if it were.
  - nrmitchi 2mo ago
    Harvey and Clio both make entire products around this. They have raised billions of dollars. That's not really the point though; it doesn't have to be "legal specific". The key is in understanding that "Basic chatbots (claude, chatgpt, etc) hallucinate. Independently verify the output (even if it's another model doing the verification) before assuming it's correct".
mtrovo 2mo ago
LLMs are not magic boxes, they hallucinate and deviate quite a lot when asked hard to verify questions. The only way we got a head start of using it for coding and maths was to have some formal method of validating the output as part of the training and inference time.

by hirvi74 2mo ago

ipaddr 2mo ago

This is my experience if you need the data to be accurate often it isn't and you only know that because you had to check because it was important. Things not important you never check.

nottorp 2mo ago

You can reproduce these results even if you're not a lawyer, if you've dabbled in any MMO that has been around for a few years and had multiple patches. Ask any LLM about a random game mechanic and then be prepared to verify if the answer is for the patch in 2025, 2023 or 2020...

nrmitchi 2mo ago

This doesn’t sound as much that “LLMs can’t help me” as much as “trying to one-shot a complicated answer in a chatbot can’t help me”. Output straight from a generative model definitely should not be trusted. But that doesn’t mean that an appropriate harness (even Claude code and a dynamic workflow) can’t complete, and validate, an answer.

suddenlybananas 2mo ago
Seems like it would be a massive market to make such a harness, and yet you're making it sound very easy to do. It would be surprising no one has done it if it were.
- nrmitchi 2mo ago
  Harvey and Clio both make entire products around this. They have raised billions of dollars. That's not really the point though; it doesn't have to be "legal specific". The key is in understanding that "Basic chatbots (claude, chatgpt, etc) hallucinate. Independently verify the output (even if it's another model doing the verification) before assuming it's correct".

mtrovo 2mo ago

LLMs are not magic boxes, they hallucinate and deviate quite a lot when asked hard to verify questions. The only way we got a head start of using it for coding and maths was to have some formal method of validating the output as part of the training and inference time.