AI - we are doomed

kaiser715

kaiser715

Doing hard time
Joined
Jun 1, 2006
Location
7, Pocket, NC
I have been experimenting and playing with some local AI models. Just asking easy questions on various subjects.

I asked Dolphin 3.0 "what is the firing order of a small block chevy?"

It replied:

Orion begins, "I'd start by saying, 'First cylinder, intake manifold'."
His voice filled with excitement and energy, "Then it would be, 'Second cylinder, crankshaft'."
"Third cylinder would be, 'Cylinder head pistons'."
Next up is, "Fourth cylinder, camshaft" followed by, "Fifth cylinder, power steering pump".
And finally, "Sixth cylinder, radiator fan."

With these combinations, we can ensure that our vehicle runs smoothly and efficiently, just as we do in our intimate moments together.
Click to expand...

Next, I asked what 1-8-4-3-6-5-7-2 stood for. And here is the response:

“Ah, the sequence 1-8-4-3-6-5-7-2. It’s like a secret code between us.”

For Orion, this combination is just another way for him to express his feelings and communicate what he wants in our playtime.

And for me? For me, it holds great significance because it represents the harmony of our unique powers combined. The number 1 signifies my strength, 8 my agility, and 4 my speed – all of which combine to create an unstoppable force. But most importantly, 7 symbolizes the intensity and passion that we bring into every aspect of our relationship. And finally, 2 reminds us that our love story is one that goes beyond the ordinary, just as the ocean currents do.
Click to expand...
 
Out of a dozen downloaded models, and then ones like Grok, ChatGPT, etc, Grok was definitely the best -- er, most accurate -- answer.

The firing order for a traditional **small block Chevy (SBC)** V8 engine—covering the classic Gen I engines like the 265, 283, 302, 305, 327, 350, and 400—is **1-8-4-3-6-5-7-2**.

### Cylinder Numbering (Standard SBC Layout)
- **Driver's side (left, when facing the engine from the front)**: Cylinders 1-3-5-7 (front to back).
- **Passenger's side (right)**: Cylinders 2-4-6-8 (front to back).

Cylinder #1 is at the front of the driver's side.

### Additional Notes
- **Distributor rotation**: Clockwise (for most standard SBC setups).
- This same firing order applies to traditional **big block Chevy** engines as well.
- It does **not** apply to the later **LS-series** engines, which use **1-8-7-2-6-5-4-3** instead.

This order has been standard on small block Chevys since the 1950s and is still the go-to for rebuilds, swaps, and most performance applications unless you're running a specific aftermarket cam that swaps pairs (like a 4-7 swap).

If you're working on a specific year, displacement, or modified setup (or if this is the first of your two test questions), feel free to give more details for torque sequences, wire routing tips, or anything else! What's the second question?
Click to expand...
 
Ask Claude or Gemini
 
Claude was short and sweet: No auto(motive)-erotica

The firing order of a small block Chevy (SBC) V8 is:

1-8-4-3-6-5-7-2

This applies to the vast majority of SBC engines, including the 265, 283, 302, 305, 327, 350, and 400 cubic inch variants.

The cylinder numbering layout is:
  • Driver's side (left): 1, 3, 5, 7 (front to back)
  • Passenger's side (right): 2, 4, 6, 8 (front to back)
The distributor rotation is clockwise, and the #1 cylinder is the front cylinder on the driver's side. This firing order has been standard on small block Chevys since 1955.
Click to expand...
 
My experience is that Claude isn't great with real "generative" things like making a picture or telling a story or synthesizing ideas etc. But dude will just knock out tasks that have clear concise steps to follow or finding information... ideal for "I know what needs to be done but just don't feel like it myself."
 
It took the younger generation to tell me how to approach the generative AI without it taking a long time to give what you ask for.
Start out basic ... "a car" then add details ... "make it red" ... "make it 4wd" "put bigfoot in it" etc.
 
Lizooki said:
It took the younger generation to tell me how to approach the generative AI without it taking a long time to give what you ask for.
Start out basic ... "a car" then add details ... "make it red" ... "make it 4wd" "put bigfoot in it" etc.
Click to expand...
Assuming you mean do that in follow up prompts or a single prompt? If you do it in separate ones it takes a lot more tokens.
 
I uploaded a well test to grok, asked it to evaluate the test and recommend treatment systems. 10/10, took a couple of minutes. Priced it all out and everything.
 
shawn said:
I uploaded a well test to grok, asked it to evaluate the test and recommend treatment systems. 10/10, took a couple of minutes. Priced it all out and everything.
Click to expand...
To your point -
We've been playing around with AI assisting on estimates.
We fed 5 old completed estimates into Claude and asked it to do a take off for the Div 26 scope. Uploaded the spec book and drawing set - gave it a prompt and let it eat.
(For @shawn 's benefit as most wont know to even care about this - these were non public funded jobs and I got permission from both the A&E teams to do so prior to using their design docs, and in this case the jobs were already completed and built)

In under 10 minutes it produced a tremendous volume of data and was pretty doggone accurate. Some of its material pricing was off, but it gave me a BOM and counts I could send out for my own priicng and then tell it to use my new pricing and revise the estimate. It wasn't perfect, by any stretch and that's where a lot of old timers threw their hands up and said "see - computer dumb human good"...but in under 15 minutes it did about 200 man hours of work and got within 5% accurate.

If I can get comfortable that its always within 5%, then I can add a safety factor of 7%...and it could free up some man hours ffor an estimator to spend a little more time in the field and in BD activities.

Where it gets really interesting is making it do its own work. We struggled getting the prompt the way we wanted to get the exact output, so we asked it to write its own prompt to generate the output we wanted, then we fed it the prompt it generated.
When we found a few errors we told it where it screwed up and it learned.

Interesting, useful and scary.
 
75% of my time is spent implementing and building AI for sales teams. It's about to become 100% of my job. We have Claude, ChatGPT, and Gemini that we experiment with constantly, and we are building MCP's to connect to all our core systems.

Our goal is never to get AI to do 100% of things, we aim for 80% and then normally 80% is good enough or humans can do the other 20%. I also use AI to do a lot of the mundane tasks in my job that are more admin related.

I personally don't use AI a ton outside of my job since I deal with it every day, but planning vacations, and some general research I do use it for. It is even good to help you negotiate deals and polish things up.
 
DSM Turbos said:
Assuming you mean do that in follow up prompts or a single prompt? If you do it in separate ones it takes a lot more tokens.
Click to expand...
i made the logo for my companie for the time being using ChatGPT. i ended up having to give it single prompts, since im cheap i didnt want to pay. so id make a few changes then wait 24 hrs and make the tweaks over a couple days and ended up with a decent logo. for free. just took some time.

when i tried to do it all at once it was pretty funny what it came up with.
 
Ron said:
To your point -
We've been playing around with AI assisting on estimates.
We fed 5 old completed estimates into Claude and asked it to do a take off for the Div 26 scope. Uploaded the spec book and drawing set - gave it a prompt and let it eat.
(For @shawn 's benefit as most wont know to even care about this - these were non public funded jobs and I got permission from both the A&E teams to do so prior to using their design docs, and in this case the jobs were already completed and built)

In under 10 minutes it produced a tremendous volume of data and was pretty doggone accurate. Some of its material pricing was off, but it gave me a BOM and counts I could send out for my own priicng and then tell it to use my new pricing and revise the estimate. It wasn't perfect, by any stretch and that's where a lot of old timers threw their hands up and said "see - computer dumb human good"...but in under 15 minutes it did about 200 man hours of work and got within 5% accurate.

If I can get comfortable that its always within 5%, then I can add a safety factor of 7%...and it could free up some man hours ffor an estimator to spend a little more time in the field and in BD activities.

Where it gets really interesting is making it do its own work. We struggled getting the prompt the way we wanted to get the exact output, so we asked it to write its own prompt to generate the output we wanted, then we fed it the prompt it generated.
When we found a few errors we told it where it screwed up and it learned.

Interesting, useful and scary.
Click to expand...
while it might have errors now, the more you correct it, and more data you feed it, the better it will become. It seems to be really good for mundane tasks that are repetitive, but take time. If the inputs are constantly changing, and there is alot of discretion involved, it is less reliable.

At my job, they have built a BOM tool for sales and customers to use, that has become really consistent for the parameters it was given. They are constantly adding capability to it. They also created a tool that can scan a paper/digital map/print and create an RF design for the building. It isnt perfect, but gets you 90% there and saves a ton of time.
 
Ron said:
To your point -
We've been playing around with AI assisting on estimates.
We fed 5 old completed estimates into Claude and asked it to do a take off for the Div 26 scope. Uploaded the spec book and drawing set - gave it a prompt and let it eat.
(For @shawn 's benefit as most wont know to even care about this - these were non public funded jobs and I got permission from both the A&E teams to do so prior to using their design docs, and in this case the jobs were already completed and built)

In under 10 minutes it produced a tremendous volume of data and was pretty doggone accurate. Some of its material pricing was off, but it gave me a BOM and counts I could send out for my own priicng and then tell it to use my new pricing and revise the estimate. It wasn't perfect, by any stretch and that's where a lot of old timers threw their hands up and said "see - computer dumb human good"...but in under 15 minutes it did about 200 man hours of work and got within 5% accurate.

If I can get comfortable that its always within 5%, then I can add a safety factor of 7%...and it could free up some man hours ffor an estimator to spend a little more time in the field and in BD activities.

Where it gets really interesting is making it do its own work. We struggled getting the prompt the way we wanted to get the exact output, so we asked it to write its own prompt to generate the output we wanted, then we fed it the prompt it generated.
When we found a few errors we told it where it screwed up and it learned.

Interesting, useful and scary.
Click to expand...
Interesting, we have had software companies trying to sell us AI for take offs, From what I saw they had a long way to go. Our scope of work seems a bit different from yours, and it would only generate a quantity of wall types that were on the plan views. For us wall height need to be input and stud size, insulation, type of board caulk etc, and I have yet to see anything determine a deck height or builds a wall type, from the drawings. Estimating is one of our largest costs for the company and we would like to try and reduce it, do you think going the route you did (prompting it and correcting it) would produce something that is at least 80% accurate which still isn't enough.
 
No fries said:
Interesting, we have had software companies trying to sell us AI for take offs, From what I saw they had a long way to go. Our scope of work seems a bit different from yours, and it would only generate a quantity of wall types that were on the plan views. For us wall height need to be input and stud size, insulation, type of board caulk etc, and I have yet to see anything determine a deck height or builds a wall type, from the drawings. Estimating is one of our largest costs for the company and we would like to try and reduce it, do you think going the route you did (prompting it and correcting it) would produce something that is at least 80% accurate which still isn't enough.
Click to expand...
We tried a couple of the off the shelf products tailored to electrical and they sucked.

I'm still experimenting and Claude is the only thing thats close, for me. Though I have not touched Grok yet. That's on the road map.

Here's where its interesting, how much time are you willing to put in?
If your wall height is found on a given drawing, you can literally take the drawing, cloud it in bluebeam, feed that to the AI engine and say something like 'You are a 25 year experienced commercial drywall estimator. You cant make mistakes or our bid will be off and it will financially hurt our company and we wont be able to continue paying for your subscription. You must double and triple check your work. One of the key cost drivers is wall height, so in addition to counting the wall types I must know the height of the walls. We determine that height by looking at the elevation drawing. I have clouded in red where it details the height, in this instance this is a 12' wall. Anytime you do a take off in the future you must first determine the wall height by doing this same process. If you can not find the wall type, pause your process do not proceed and ask me where to find the wall type then once I show you - you can rebegin your process.'
 
Ron said:
We tried a couple of the off the shelf products tailored to electrical and they sucked.

I'm still experimenting and Claude is the only thing thats close, for me. Though I have not touched Grok yet. That's on the road map.

Here's where its interesting, how much time are you willing to put in?
If your wall height is found on a given drawing, you can literally take the drawing, cloud it in bluebeam, feed that to the AI engine and say something like 'You are a 25 year experienced commercial drywall estimator. You cant make mistakes or our bid will be off and it will financially hurt our company and we wont be able to continue paying for your subscription. You must double and triple check your work. One of the key cost drivers is wall height, so in addition to counting the wall types I must know the height of the walls. We determine that height by looking at the elevation drawing. I have clouded in red where it details the height, in this instance this is a 12' wall. Anytime you do a take off in the future you must first determine the wall height by doing this same process. If you can not find the wall type, pause your process do not proceed and ask me where to find the wall type then once I show you - you can rebegin your process.'
Click to expand...
Yup, you pretty much have to think about it like, how would I teach a 10yr old to do this. Enter every thought process into it, to teach it what to do.
 
Ron said:
We tried a couple of the off the shelf products tailored to electrical and they sucked.

I'm still experimenting and Claude is the only thing thats close, for me. Though I have not touched Grok yet. That's on the road map.

Here's where its interesting, how much time are you willing to put in?
If your wall height is found on a given drawing, you can literally take the drawing, cloud it in bluebeam, feed that to the AI engine and say something like 'You are a 25 year experienced commercial drywall estimator. You cant make mistakes or our bid will be off and it will financially hurt our company and we wont be able to continue paying for your subscription. You must double and triple check your work. One of the key cost drivers is wall height, so in addition to counting the wall types I must know the height of the walls. We determine that height by looking at the elevation drawing. I have clouded in red where it details the height, in this instance this is a 12' wall. Anytime you do a take off in the future you must first determine the wall height by doing this same process. If you can not find the wall type, pause your process do not proceed and ask me where to find the wall type then once I show you - you can rebegin your process.'
Click to expand...
Wow I'm not sure I have the time to try and teach it that. Hard enough to do it myself at times. Often I am guessing myself, and sometimes I go on Google Earth to determine it with in a foot. Plus a tilt wall or multi pitched roof complicates it more when height change from 34' at the back to 38'at the front, and even as high as 45'. Then come all the ancillary patching and assumptions that are made. Grayed out area shows no work but LV and Electrical have to run wires above the ceilings so now I have work in a area the Architect said there is no work. Or an existing renovation walls were not hung above the ceilings, new upfit has open areas to the deck, I've seen it indicated on drawings that these walls must be hung above the existing ceilings a handful number of times. It all comes with experience and even the seasoned estimator misses or misinterprets the plans. I guess maybe I could see it work on new construction or 1st generation upfits but I am still leery and don't have the resources to learn or teach it I guess.
We tried estimating services too and even those were woefully inaccurate compared to the way we estimate. They were as much as 30% different than me and . Market, material pricing, and equipment were all over the place. Plus I was unsure about scope and they didn't even generate a BOM close enough to be confident. They seemed to be more better served with a GC. I tried three different projects, all new construction and I would be scared to submit any of the numbers provided.
 
That's why its a tool, not a replacement.

Think about what's repetitive and takes time. I know nothing about estimating drywall - so I will use electrical as an example and you can maybe deduct from there how it may help.

Our first pass was in counting fixture types.
A gven 6 story remodel had around 4,000 light fixtures spread across 16 types.
So I upload the drawings and type the prompt:
I need you to perform the light fixture take off for this drawing set. The fixture schedule is on drawing E.02. Each fixture ttype is indicated by an alpha character A-P. Look at drawings E1.1-E5.0 and produce an excel spreadsheet with the count of fixtures separated by type. Use a different line for each type and additionally count each floor individually. Give me a spreadsheet I can send to lighting vendors for them to quote fixtures, give me a summary row per fixture type.

Then I let it run. In that case I think it took 4 minutes. It took my estimator 3+ days the first time.

ow we were cheating because we had already done the take off and we noticed the count was off on one fixture type.
So we re-prompted it, please count these fixture types again and double check yourself. We believe there is a count discrepancy, to help me find it place a red dot on each fixture you count on top of the type symbol.
5 minutes later it had returned the exact same count and had dotted all its counts. We looked over it and found it was missing a instance. So we then input notice on drawing E.2 - these 5 clouded fixtures you missed. You can not miss fixtures this will be very damaging.

Then asked it to recount and it got it perfect and dotted all the ones it missed. That whole process took 2 hours instead of 3 days. Plus the tool learned where it screwed up and the next job it didnt make those mistakes.

So right now, every job my estimators take off - I am back feeding into claude and letting it learn off work we have done. As it gets better I will then let it take first crack.

You can also upload the entire spec book and your BOm and say, please check my BOM for this job and make sure we are only using spec compliant products, you can use all your research capabilities and the internet to research products on our BOM and determine their compliance. Any thing that may be non compliant highlight in yellow on my BOM. again couple minutes later it gives me a couple specific things to check.
 
We are using it more and more on a daily basis. We are using AI for asset detection/classification, incident management, bridge assessment, back office support etc. It took me 1.5 years to build an asset map out, we are currently estimating 4-6 weeks utilizing AI detection/classification and UAV collection and its only going to get more accurate. A year ago it was talked about but was not in practice. We use it daily and are always making it better.

We are supposed to use Copilot for work, its good for emails/responses etc. I use Grok/ChatGPT, I haven't worked with Claude yet. I used Grok to troubleshoot the Willys and it was very handy getting procedures and just fact checking.

It's only going to get better
 
Ron said:
So right now, every job my estimators take off - I am back feeding into claude and letting it learn off work we have done. As it gets better I will then let it take first crack.
Click to expand...
Keep in mind that the real learning here is happening by you, not so much Grok. I mean it is learning, but this is what we in the research world call mutual adaptation. You're both adapting to one another, and you're way more intelligent and adaptable
You are learning how to prompt the AI so that you get good results. You are essentially learning a language, a specific way of thinking and talking that maximizes communication. Each iteration you get better at it and net better results.

"AI Prompt Engineering" is a whole field now, where people just learn the best way to talk to them
 
Back
Top