Using ChatGPT for estimating

sw_ross

Senior Member
Location
NoDak
Had a younger co-worker that was playing around with ChatGPT, playing around with it to put together an estimate for wiring up a shop.
I have no experience with it but it was an interesting experience. It seemed pretty thorough. As you would gradually give it more information it would modify the estimate price and actually give you feedback about whether your hourly rate was too high or low, etc.
I would never submit a bid/estimate based on it but I would definitely use it to check my estimate before submitting.
Where does it get its information?
 
Tools such as ChatGPT are really helpful for catching errors in stuff you write, and really helpful for generating ideas. But you absolutely should _not_ trust them. Only use them in situations where you have the ability to verify their output or they will surprise you with just how wrong they can be.

Did you ever meet someone who can just make things that sound factual up with a perfectly straight face? Someone who can weave a convincing story that step by step sounds perfectly believable but is actually meaningless? That is _exactly_ what large language models do. They have no _understanding_ of what they are saying, they are just connecting words in ways that statistically match the corpus of text that they were trained on.

The thing is that this is very useful. So much of the text we have to produce is boilerplate stuff, wrapped around a few key ideas that require human input. So having a language model adding in all the surrounding fluff saves a huge amount of human time.

But you _must_ verify the output.

Last week I needed to add some text to a government contract. I told ChatGPT to craft me a statement protesting a new requirement. The system crafted a beautifully worded statement, referencing some federal code that I'd never heard of, almost exactly what I needed. I could have just pasted it into my contract. Fortunately I decided to look up the code reference. The code book existed, but with a different date than ChatGPT used. And the topic was no where to be found. So I asked ChatGPT to show me the exact text in the reference. The response "Good catch, it appears that that code doesn't contain that topic...." So happy, so helpful, so _wrong_.

When I ask ChatGPT to do calculations (not electrical estimating), sometimes it grabs the wrong units and gives just out there answers; thins off by orders of magnitude. But very often it will help me come up with a good framework for doing the calculations; I just have to be careful to go through and double check the steps.

I'll say it again: It can be a useful tool, but Always Verify The Output.
 
Better figure out how to use it properly cause if you don’t, you’re gonna be swept away in the current.

As long as you know the information you’re feeding it and you should know approximate outcome. It’s a good way to pinpoint your trajectory. You can’t just blindly give it information and you know nothing about it. It will lead you astray that way.
 
My city is going to widen the road in front of my house and needs part of my land to do so. They sent an appraiser, and then a negotiator, and made what i felt was a very insulting offer. I replied with an equally insulating offer and their response was “see you in court”.

So two weeks ago i used ChatGPT to write a response that i thought could better articulate my points about the value of my property, and damages I’m going to incur. I spent a couple of hours over the course of a day feeding it information, asking questions, but also asking very specific questions like “find cases similar to mine in this state and what was the outcome?” And “Did you consider xyz in your answer and if not how would that change the value”. Not those words exactly but something along those lines. Each time it’d give me an answer, I’d think on it for awhile to come up with another line of questioning. Mostly what i wanted to do was lead it to the answer i wanted but give relevant facts to get me there.

I tweaked the final product from ChatGPT, added some of my own language, and sent back to the DOT. They called me back and doubled their original offer, which is still comically low, but it was interesting how quickly they changed their offer after reading that response. But i have some worry that if it gets to the point an actual attorney reads what I’ve written that they’ll immediately know its ChatGPT b.s.
 
Had a younger co-worker that was playing around with ChatGPT, playing around with it to put together an estimate for wiring up a shop.
I have no experience with it but it was an interesting experience. It seemed pretty thorough. As you would gradually give it more information it would modify the estimate price and actually give you feedback about whether your hourly rate was too high or low, etc.
I would never submit a bid/estimate based on it but I would definitely use it to check my estimate before submitting.
Where does it get its information?
An LLM like ChatGPT does not get information in the human sense of the word. It's simply been fed a lot of language which may contain examples from the internet and can pick out examples similar to your prompt, and it parrots that language with a probability algorithm that guesses what you likely want the next word in the sentence to be. There's no underlying reasoning and (ironically for a computer) it may not do math as well as most humans are capable of. It is, in essence, a booksmart bullsh----r. It has no comprehension of the real world, and lacks the ability to evaluate its own output. It doesn't actually know what it's talking about, just how to sound like it does.

Considering that for various reasons successful small businesses tend to keep their financials and strategies private, public info on the internet is probably not a great source of estimating advice to begin with. (True of many things.) And so an LLM will be all the more handicapped producing reliable advice in this situation.
 
An LLM like ChatGPT does not get information in the human sense of the word. It's simply been fed a lot of language which may contain examples from the internet and can pick out examples similar to your prompt, and it parrots that language with a probability algorithm that guesses what you likely want the next word in the sentence to be. There's no underlying reasoning and (ironically for a computer) it may not do math as well as most humans are capable of. It is, in essence, a booksmart bullsh----r. It has no comprehension of the real world, and lacks the ability to evaluate its own output. It doesn't actually know what it's talking about, just how to sound like it does.

Considering that for various reasons successful small businesses tend to keep their financials and strategies private, public info on the internet is probably not a great source of estimating advice to begin with. (True of many things.) And so an LLM will be all the more handicapped producing reliable advice in this situation.
I don’t understand how you guys don’t think it’s viable for estimating. You upload your Neca labor unit PDF it processes it you tell it in detail what you’re doing for your job and very descriptive terms and it pulls it up. You can double check if you don’t think it’s right, but I’ve done my job enough. I know approximately how long jobs take. It gives me a price I ask it to break it down into each labor unit material cost it pulls it off from platt.com and I know it’s bullshit and if it says $3000 for a 200 amp meter main residential and then I asked for a link and it links back to some large Commerical multi gang
But still in 30 seconds I can get an estimate . Any of you guys use? Talk to chat the same premise it dictates for you most of the time it’s good sometimes it screws up words then you proofread it. Edit it and submit it the same thing like I’m being lazy. I did talk to speak if the stuff is messed up cause I didn’t proofread it. I just hit post same thing.
 
My city is going to widen the road in front of my house and needs part of my land to do so. They sent an appraiser, and then a negotiator, and made what i felt was a very insulting offer. I replied with an equally insulating offer and their response was “see you in court”.

So two weeks ago i used ChatGPT to write a response that i thought could better articulate my points about the value of my property, and damages I’m going to incur. I spent a couple of hours over the course of a day feeding it information, asking questions, but also asking very specific questions like “find cases similar to mine in this state and what was the outcome?” And “Did you consider xyz in your answer and if not how would that change the value”. Not those words exactly but something along those lines. Each time it’d give me an answer, I’d think on it for awhile to come up with another line of questioning. Mostly what i wanted to do was lead it to the answer i wanted but give relevant facts to get me there.

I tweaked the final product from ChatGPT, added some of my own language, and sent back to the DOT. They called me back and doubled their original offer, which is still comically low, but it was interesting how quickly they changed their offer after reading that response. But i have some worry that if it gets to the point an actual attorney reads what I’ve written that they’ll immediately know its ChatGPT b.s.
You may think it’s lawyer bs but those are gonna be the jobs that go away cause all they all lawyers do this site case law.

The lawyers that learned to adopt chat and can spit out for their client information quick and cheap are gonna be the ones that are still around and take over. It’s gonna be the Matthew principal. And those lawyers need to know what they’re talking about and then it works.
 
Plus, you gotta create your own custom chat bot, and you feed the information and just keep learning and learning off your processes as you go so the more you train it the better it grows accustomed to you it’s just a fancy cheaper estimating program. Estimating program suck if you don’t know how to use them right same with this.

Nice thing about it is all the information I give it and it gives me a quote. I like it and I say submit a quote for Customer. It fits it all in nice language. I can say make it simple make a detailed and it does it in less than 10 seconds I don’t have to spend 40 minutes writing it up. It’s time saver
 
A friend of mine just got her PhD and basically used this technique.

She fed the LLM a ton of relevant journal articles, and used the LLM to create a custom model of the information in the papers, find relevant parts that she needed to work from, used it to tune her literature searches, etc.

But every step of the way she was validating the output, and it sounds like you are doing the same. I'm not saying LLMs are bad, if fact I agree with you that the people who will do well are the ones who learn how to use them well. I'm just saying that you have to use them carefully or they will trip you up.

-Jonathan
 
One of my clients aCisco engineer and that’s what he basically just use self check his work and he double checked its work. I mean all it is just a tool. I don’t rely on my whole hog not to hit a wire. I gotta know we’re not the drill. Just gotta know how to use the tool.
 
My city is going to widen the road in front of my house and needs part of my land to do so. They sent an appraiser, and then a negotiator, and made what i felt was a very insulting offer. I replied with an equally insulating offer and their response was “see you in court”.

So two weeks ago i used ChatGPT to write a response that i thought could better articulate my points about the value of my property, and damages I’m going to incur. I spent a couple of hours over the course of a day feeding it information, asking questions, but also asking very specific questions like “find cases similar to mine in this state and what was the outcome?” And “Did you consider xyz in your answer and if not how would that change the value”. Not those words exactly but something along those lines. Each time it’d give me an answer, I’d think on it for awhile to come up with another line of questioning. Mostly what i wanted to do was lead it to the answer i wanted but give relevant facts to get me there.

I tweaked the final product from ChatGPT, added some of my own language, and sent back to the DOT. They called me back and doubled their original offer, which is still comically low, but it was interesting how quickly they changed their offer after reading that response. But i have some worry that if it gets to the point an actual attorney reads what I’ve written that they’ll immediately know its ChatGPT b.s.

The real question here is if you cited any court cases or other facts, did you verify them outside of the chat, to make sure they weren't made up and actually support your position? Because if you don't that's what will trip you up down the line. Actual lawyers and judges have been caught with fake stuff in their filings. It's bad.

Honestly I think if Google and other regular search wasn't so ensh--itified, it would do the work better than an LLM.
 
I don’t understand how you guys don’t think it’s viable for estimating. You upload your Neca labor unit PDF it processes it you tell it in detail what you’re doing for your job and very descriptive terms and it pulls it up. You can double check if you don’t think it’s right, but I’ve done my job enough. I know approximately how long jobs take. It gives me a price I ask it to break it down into each labor unit material cost it pulls it off from platt.com and I know it’s bullshit and if it says $3000 for a 200 amp meter main residential and then I asked for a link and it links back to some large Commerical multi gang
But still in 30 seconds I can get an estimate . Any of you guys use? Talk to chat the same premise it dictates for you most of the time it’s good sometimes it screws up words then you proofread it. Edit it and submit it the same thing like I’m being lazy. I did talk to speak if the stuff is messed up cause I didn’t proofread it. I just hit post same thing.

If you are using the NECA method then that's widely documented and described enough that the LLM will approximate. Our business cannot do our sales that way so it's not going to be as helpful.

And note that you still have to bring your knowledge and experience to bear on things like equipment pricing.

What I've found the few times I've used ChatGPT is that after the verification and thinking required to ensure the quality of the output to my satisfaction, in the end it did not save me time.
 
If you are using the NECA method then that's widely documented and described enough that the LLM will approximate. Our business cannot do our sales that way so it's not going to be as helpful.

And note that you still have to bring your knowledge and experience to bear on things like equipment pricing.

What I've found the few times I've used ChatGPT is that after the verification and thinking required to ensure the quality of the output to my satisfaction, in the end it did not save me time.
Yeah, your job might not be best suited for it at the moment, but I’ve been using it for two years and and it has changed drastically over those two years and another two years, it’s gonna be another beast that to get on the bandwagon now we’ll figuring it out. It’s not like it’s going back in the bottle.
 
The real question here is if you cited any court cases or other facts, did you verify them outside of the chat, to make sure they weren't made up and actually support your position? Because if you don't that's what will trip you up down the line. Actual lawyers and judges have been caught with fake stuff in their filings. It's bad.

Honestly I think if Google and other regular search wasn't so ensh--itified, it would do the work better than an LLM.

I did not cite any specific cases. It said something along the lines of ".... these claims are well established in Georgia case law....." etc etc.
 
Top