ChatGPT 4 QuAnTiTaTiVe FiNaNcE

[citizenofdoom]

As we - Vietnamese people are suffering through a real estate related bond crisis and potentially crashes the entire economy; this is not funny tho, many people have lost their work and some of the big textiles companies here bankrupted already as our economy consists of 30% real estate related things (it looks totally unrelated but trust me, more than 60% of real estate filled bankruptcy and the biggest real estate companies have share price dropped by more than 50%, it’s somewhat in a domino effect now).

Today I read Vietnamese economic news, I saw HCM Real Estate Association proposed to the government that laws should be constructed for allowing ChatGPT usage in updating the real estate price as well as determine the price of real estate. I kinda laughed a lot at these clowns-like people because (seriously why r3@l politik seems so damn stupid and humorous; they totally brightened my miserable days of final year experimentin) and probably that bill should be rejected immediately but who knows.

I think it’s so funny so I guess I just share the news here. Try to imagine if that bill is accepted by the general assembly; share to me some of the funniest outcome of what would happen?

source(use google translate for translation); this is really not some kind of fake news, it’s real: https://plo.vn/de-xuat-ung-dung-chat-gpt-de-xay-dung-bang-gia-dat-post724067.html

 

[TheMeInTeam]

for me, it depends what one means by "allow chatgpt usage". to what extent is it not allowed presently?

using it to "determine the price"...not sure how that would even work. chatgpt is very good at predicting what comes next in language sequences, probably superhuman. it's varying degrees of much less capable for other tasks.

if they just mean using it to make write-ups quickly then sure? if they mean to have it "try to analyze the market and set prices" or something, that sounds odd.

 

[citizenofdoom]

for me, it depends what one means by "allow chatgpt usage". to what extent is it not allowed presently?

using it to "determine the price"...not sure how that would even work. chatgpt is very good at predicting what comes next in language sequences, probably superhuman. it's varying degrees of much less capable for other tasks.

if they just mean using it to make write-ups quickly then sure? if they mean to have it "try to analyze the market and set prices" or something, that sounds odd.

truthfully, they want a dummy to put the blames on; and chatgpt is not a surrogate model but a memorizer. i don’t trust the authorities.

however, there is fishy thing you might not know; transformers in general are turing machines; and chatgpt is known to be able to solve heuristic problems better than human designed algorithm; evidenced by https://arxiv.org/pdf/2302.12170.pdf.

 

[TheMeInTeam]

imo chatgpt is a model (of language/future language), but that's not something that can transform into a (reasonably correct) model of pricing/markets. i'm sure you can get it to say something on the topic, though making that something useful is a higher bar.

is it really fishy that chatgpt can do better than writing an algorithm for the same thing? we see that with training ais vs trying to script them in games too. chatgpt has an enormous amount of compute and training data right?

 

[citizenofdoom]

I think it depends on whether the task can be done by non-specialized humans since i think chatGPT memorizes more language based data than any of us. For example, comparing its performance in weather prediction with a SDE-based model trained for weather prediction is meaningless; however, asking it to play something verbal like Faster Than Light game, might cost less effort than training a full fledged RL model; on transfer learning settings, where models trained in vanilla game and tested in game + mods, maybe ChatGPT or other LLMs (large language models) can do better solely because of their abilities to interpret human languages? I don’t have money nor time to test the hypothesis. GPT-4 has some cool new techs in logic reasoning; I haven’t tried yet; don’t know how it would be.

What objectives you think GPT or other LLMs can do better than a dedicated algorithm?

 

[Lexicus]

Why would you use a LLM to do anything that involves doing math

 

[citizenofdoom]

Why would you use a LLM to do anything that involves doing math

That’s exactly something I wouldn’t dare to do despite people have some success with that. I don’t think Neural SDE has capacity to solve mathematic since neural things are just approximations. And I pretty sure that LLMs are the best math solvers currently (under human performance ofc).

 

[Lexicus]

That’s exactly something I wouldn’t dare to do despite people have some success with that. I don’t think Neural SDE has capacity to solve mathematic since neural things are just approximations. And I pretty sure that LLMs are the best math solvers currently (under human performance ofc).

My understanding was that trying to use ChatGPT to design rockets would have grim results and I can't see why this is any different.

 

[Verbose]

My understanding was that trying to use ChatGPT to design rockets would have grim results and I can't see why this is any different.

Absolutely. If you already know exactly what to ask ChatGPT to get a quality answer – i.e. you can design the rocket anyway – you can probably get the right kind of answers.

The problem is that ChatGPT's answers tend to have roughly the quality of the questions put to it.

If played around a bit with it, and found that if you know and ask rather specific stuff, you can get OK outcomes. But the darn thing will simply mirror your question and happily make things up that sorta kinda might sound OK if you don't.

 

[GoodEnoughForMe]

Yeah I have gotten ChatGPT to literally provide opposite answers to the same question back-to-back based on how I asked it each time. Feels particularly ripe for abuse for those who would have a financial motive for certain answers, which is exactly what this proposed usage is.

 

[mdl5000]

so when human actors screw something up they can just blame the robot?

 

[citizenofdoom]

Absolutely. If you already know exactly what to ask ChatGPT to get a quality answer – i.e. you can design the rocket anyway – you can probably get the right kind of answers.

The problem is that ChatGPT's answers tend to have roughly the quality of the questions put to it.

If played around a bit with it, and found that if you know and ask rather specific stuff, you can get OK outcomes. But the darn thing will simply mirror your question and happily make things up that sorta kinda might sound OK if you don't.

Try ask ChatGPT about diffusion models for image synthesis.
ChatGPT keeps denying its existence due to too big time complexity so many times.

Feels particularly ripe for abuse for those who would have a financial motive for certain answers, which is exactly what this proposed usage is.

Something like ChatGPT API? It works like this: the people having power feed into it fraudulent data via some unreal transfers and try to fool others.

so when human actors screw something up they can just blame the robot?

responsibility -> robot -> collective wanting it => responsibility divided -> no responsibility !!!