Actually using "AI"

DaleB said:
My wife's neurosurgeon is one of the best in the field. He told us that they are using AI to evaluate MRI and MRA images and evaluate the risk of strokes and rupturing aneurysms, and that AI is doing a fantastic job of it.
Click to expand...
Yep. Some of that work goes back to research done 20-25 years ago by the national cancer institute for early detection of breast cancer by looking at MRIs taken over a period of years. The NCI is part of NIH and is in the process of getting its budget cut dramatically.
 
schmookeeg said:
I use it daily. Treat it like a "Junior level employee" that you can delegate scudwork to and you won't be disappointed. Ask the Junior to do Senior work and... you'll get about the same result as it enters a rabbit-hole of its own making and you get to fish it out :)
Click to expand...
This is exactly what we realized a couple years ago. For software development, you basically treat it like a junior employee and tell it exactly what you want. It's usually at 90% correct, or better.

Until the LLMs can map from the business speak to technical needs, mid or senior level engineers are going to be necessary. And even then, how far do you trust what an purely AI generated application creates or does?

I'm pretty sure I'll be retired by the time it reaches the level some people believe is imminent where it can do it all. But I do believe it will eventually get to a point where a large percentage of the software industry will be an AI babysitter and troubleshooter outside of specialized niches.
 
I put in a request for route planning and got super basic information (check atis!)

Is there some trick to it? Should I be reading some site which explains/trains ?

tia
 
I have been playing with Ai for a while, in its current stage I think it’s vastly overrated
 
mark.staggs said:
Until the LLMs can map from the business speak to technical needs, mid or senior level engineers are going to be necessary. And even then, how far do you trust what an purely AI generated application creates or does?
Click to expand...
Yeah, people underestimate how much of a programmers job (much less a designer, architect or product owner) is translating the business need into the application. AI is, currently, super effective at taking precise english instructions and converting them into python/c++/sql/javascript.

But it can't even write a compelling short story (or, if it does, it's by dumb luck every 10,000th throw of the dice), much less take into account the vagaries of any company large enough to need an application written.

With current solutions, I can't even tell an AI something as simple as, "write an iphone app that accepts the FAA data stream and displays current traffic on a zoomable vfr sectional".

mark.staggs said:
I'm pretty sure I'll be retired by the time it reaches the level some people believe is imminent where it can do it all. But I do believe it will eventually get to a point where a large percentage of the software industry will be an AI babysitter and troubleshooter outside of specialized niches.
Click to expand...
Me too. I was telling someone earlier that I have a gray hair for every memory leak I've dealt with in C/C++ over the decades (I have a lot of gray hair). AI should be pretty effective at solving those problems. But that's going to be a programmer aid, rather than a programmer replacement, for a long time.
 
Go take a look at the multi-agent and autonomous AI demos... You can have a bunch of specialized agents (with specific LLMs) that talk to each other to solve a problem. This is in its infancy still, but where things are going before we reach AGI (artificial general intelligence)
 
Bob Noel said:
let me know when an AI bot gets an MBA...
Click to expand...

This isn’t 2000s anymore

The MBA is becoming a high debt low return degree

Doubt many AIs could pay for housing and electrical for their power supply after servicing their monthly “education” debt

At this point “AI” is still a slightly upgraded search engine
 
Bob Noel said:
let me know when an AI bot gets an MBA...
Click to expand...
You'll know. It'll go around randomly firing really good people because "they're too expensive".

Actually... Never mind, it's already happened.

172andyou said:
Yeah, people underestimate how much of a programmers job (much less a designer, architect or product owner) is translating the business need into the application. AI is, currently, super effective at taking precise english instructions and converting them into python/c++/sql/javascript.
Click to expand...
That's just the thing - That's what software is. Creating precise instructions.

Now, it may translate them into whatever language, but it still needs the instructions. When I'm developing software, it isn't writing the syntax that's a challenge (unless I'm working with a new language). It's the "How do we get these requirements down to the precise, detailed instructions necessary to make this work the way it's supposed to."

172andyou said:
With current solutions, I can't even tell an AI something as simple as, "write an iphone app that accepts the FAA data stream and displays current traffic on a zoomable vfr sectional".
Click to expand...
Someone else needs to do each of the pieces first for it to learn from... Or you need to be more detailed in your description.
 
yup - my "let me know when an AI bot gets an MBA..." really was intended as a straight line as well as a jab at the MBA
 
jesse said:
I sure hope that's just someone speaking a little out of context. ChatGPT is a great tool to act as an assistant to a skilled engineer. It's a terrible idea for someone that knows zero about programming to start slinging production code with ChatGPT.
Click to expand...
So as I understand the story, an employee with zero programming experience was able to get gpt to write a piece of code to automate some menial task they were doing. I don't know enough about programming to understand what exactly they were doing, but he mentioned "300 lines of code". I presume it's taking data from one software package and plugging it into another. I believe one of our IT guys was watching over their shoulder. The story was that it took three iterations; the first code did something wrong, which they told the AI, and it fixed it. That happened again and the third try worked correctly. We are strictly limiting the number of people who are empowered to use it at this time, but several thousand man hours have been saved already. The nature of the grain elevator business is that there's LOTS of contracts and transactions that have to be reconciled, which is pretty low hanging fruit for this kind of stuff. One of the topics of discussion at the meeting was implementing an official AI policy.

OTOH, last night my brother and I realized we should do something for our dad's 80th birthday this weekend. He asked me to pick a restaurant about halfway between us. I thought "this is a perfect application", so I asked Grok for restaurant recommendations in Decatur. I assumed it would know where I was and search Decatur, IL, but it went with Decatur, GA. Correcting that, it gave me some recommendations, and I asked about price. It said it couldn't find menus, but gave me a guess (which turned out to be reasonably accurate). I then asked which of the three I picked were open on Sunday. It said all three were. I then asked for the websites, which it did give me successfully. By visiting their websites, I quickly determined all three are closed on Sunday, and found their menus quickly. So I just went to google to finish my research.
 
Jim K said:
I then asked which of the three I picked were open on Sunday. It said all three were. I then asked for the websites, which it did give me successfully. By visiting their websites, I quickly determined all three are closed on Sunday, and found their menus quickly. So I just went to google to finish my research.
Click to expand...
GPT, please let me introduce Mr. Clemens...

1740755831801.png
 
Jim K said:
So as I understand the story, an employee with zero programming experience was able to get gpt to write a piece of code to automate some menial task they were doing. I don't know enough about programming to understand what exactly they were doing, but he mentioned "300 lines of code". I presume it's taking data from one software package and plugging it into another. I believe one of our IT guys was watching over their shoulder. The story was that it took three iterations; the first code did something wrong, which they told the AI, and it fixed it. That happened again and the third try worked correctly. We are strictly limiting the number of people who are empowered to use it at this time, but several thousand man hours have been saved already. The nature of the grain elevator business is that there's LOTS of contracts and transactions that have to be reconciled, which is pretty low hanging fruit for this kind of stuff. One of the topics of discussion at the meeting was implementing an official AI policy.
Click to expand...
It's excellent that a practical use case was discovered and has tangible output and is making life easier.

But my takeaway from this is perhaps a little more harsh. Primarily that the IT guys who supervised development should perhaps consider a new line of work, or perhaps rolling their sleeves up and getting more involved with the workers and understanding areas for workflow improvement.

I say that because any automation task that is achievable by only a handful of prompt iterations by a non-techie, which already has produced a tangible ROI of thousands of man hours saved, is an automation task those IT guys almost certainly should have already done.
 
ArrowFlyer86 said:
It's excellent that a practical use case was discovered and has tangible output and is making life easier.

But my takeaway from this is perhaps a little more harsh. Primarily that the IT guys who supervised development should perhaps consider a new line of work, or perhaps rolling their sleeves up and getting more involved with the workers and understanding areas for workflow improvement.

I say that because any automation task that is achievable by only a handful of prompt iterations by a non-techie, which already has produced a tangible ROI of thousands of man hours saved, is an automation task those IT guys almost certainly should have already done.
Click to expand...
Most of us "IT guys" don't spend our time automating non-IT business tasks just because. If the business identifies a need and asks for it, then sure, we'll get involved - but that's typically a project.

Lots of drudge work doesn't get automated, because nobody bothers to ask for it. IT staff, whether programmers or infrastructure or ops, have their own work to do. If jobs are going to be automated, those jobs will typically be infrastructure and ops/SRE first. If somebody's sitting at their desk playing spreadsheet copy/paste half the day, well, that's the job. If you want it automated and dno't know how to do it yourself, then you can ask, but typically nobody is going to come around once a month and ask whether you have any odd jobs that need to be done. We know better.
 
DaleB said:
Most of us "IT guys" don't spend our time automating non-IT business tasks just because. If the business identifies a need and asks for it, then sure, we'll get involved - but that's typically a project.

Lots of drudge work doesn't get automated, because nobody bothers to ask for it. IT staff, whether programmers or infrastructure or ops, have their own work to do. If jobs are going to be automated, those jobs will typically be infrastructure and ops/SRE first. If somebody's sitting at their desk playing spreadsheet copy/paste half the day, well, that's the job. If you want it automated and dno't know how to do it yourself, then you can ask, but typically nobody is going to come around once a month and ask whether you have any odd jobs that need to be done. We know better.
Click to expand...
In the firms I've worked at and where I've managed it's been much different. Your boss is going to ask where you're spending your time.
If you say "I'm moving data from A to B manually" or "I'm reformatting data A to format B manually", they're going to say "that's a really bad use of time". And within a few days you're going to have IT on the phone finding a way to automate it for you (that's assuming you didn't contact them under your own initiative already). At that point you just have to hope you have other things on your plate to keep you busy, or you're out of a job.
 
ArrowFlyer86 said:
In the firms I've worked at and where I've managed it's been much different. Your boss is going to ask where you're spending your time.
If you say "I'm moving data from A to B manually" or "I'm reformatting data A to format B manually", they're going to say "that's a really bad use of time". And within a few days you're going to have IT on the phone finding a way to automate it for you (that's assuming you didn't contact them under your own initiative already). At that point you just have to hope you have other things on your plate to keep you busy, or you're out of a job.
Click to expand...
Right, but the key here is that the worker or the manager is going to engage IT. If nobody does that and the person is sitting there doing easily automated drudge work, the blame is not on IT for not knowing about it.
 
DaleB said:
Right, but the key here is that the worker or the manager is going to engage IT. If nobody does that and the person is sitting there doing easily automated drudge work, the blame is not on IT for not knowing about it.
Click to expand...
That's totally fair. I guess if the culture doesn't really encourage it, you can't blame them for not just automatically knowing :)
 
ArrowFlyer86 said:
At that point you just have to hope you have other things on your plate to keep you busy, or you're out of a job.
Click to expand...
Do you happen to move jobs often? I’ve seen far more buisnesses with that attitude in management go out of business than ones that can even keep their doors open. It might work to show a couple quarters of increased earnings and get you a good performance bonus, but I think it’s a terrible strategy for the buisness on a two year projection. Ya, you can trim some fat, but when your workforce all see it your good workers will immediately start looking for work at different companies, and find it.
 
Flatiowa said:
Do you happen to move jobs often? I’ve seen far more buisnesses with that attitude in management go out of business than ones that can even keep their doors open. It might work to show a couple quarters of increased earnings and get you a good performance bonus, but I think it’s a terrible strategy for the buisness on a two year projection. Ya, you can trim some fat, but when your workforce all see it your good workers will immediately start looking for work at different companies, and find it.
Click to expand...
I don't switch too often.
But this boils down to knowing where your workers spend their time. If you don't know, and you don't know what resources are available to boost your workers productivity, then it's difficult to be an effective manager.

And to clarify -- if you free up someone's schedule by automating repetitive work, it's not like you just automatically fire the person lol. You help them find other things to work on that would be valuable.
 
I have zero knowledge of AI and can barely operate a keyboard; keep that in mind with my request.

I have a 20 page form to complete (ie many questions on each page, some with checkboxes) and within the form are requirements to generate, de novo, other related documents - and I find the terms and verbiage mind-boggling. (I have a university degree, too)
I cannot find any of these (completed, or 'example') forms or documents on the internet.

Is it likely that AI is going to be able to help?
Where would I begin?
Can I put the names of the form or additional documents on an AI website and hope to receive any help?
Or, provide and AI service with a pdf of the form?

It's all aviation-related.
 
Let'sgoflying! said:
I have zero knowledge of AI and can barely operate a keyboard; keep that in mind with my request.

I have a 20 page form to complete (ie many questions on each page, some with checkboxes) and within the form are requirements to generate, de novo, other related documents - and I find the terms and verbiage mind-boggling. (I have a university degree, too)
I cannot find any of these (completed, or 'example') forms or documents on the internet.

Is it likely that AI is going to be able to help?
Where would I begin?
Can I put the names of the form or additional documents on an AI website and hope to receive any help?
Or, provide and AI service with a pdf of the form?

It's all aviation-related.
Click to expand...
There's a good chance that AI can assist in explaining what is being requested. I'd recommend using ChatGPT.

For starters: you could try pasting in the contents of the first page of the document, and ask it to explain what each item being requested is. That's called "giving it a prompt". Give it a prompt like: "There is a document questionnaire requested by my insurance company for my aircraft. Some of the wording is nuanced and I'm not sure what it means. For the more complex questions can you provide a concise, simplified explanation of what it is asking for? The following is the contents from the document page: " (then paste contents in).

Alternatively if you don't need help on every question, you could just copy and paste the more mind-boggling questions that you see on the form and repetitively ask (one-by-one) "I'm not sure what XYZ question means, can you help explain it to me. If any additional documents are being requested in the question, can you help explain which ones those are? Here is the question: " (then paste the question in)

By letting it know it's dealing with aviation/insurance, aviation/regulatory, etc -- that should help at least get started in pointing you in the right direction.

EDIT: For the documents you can't find on Google, try asking GPT for it. There's a reasonable chance it knows of it, or can describe it. It's recall ability is quite good. I just wouldn't trust it to fill out the form for you :)
 
My experience is underwhelming. I spent time in Excel learning how to setup linear regressions properly. Now, I'd like to fill the multiple equation unknowns (x1, x3, x3, xn for each row) automatically from a list of example sample models. Nope - it didn't know how to fill the matrix right. I spent some time thinking this through, and decided to give it a square matrix to start with so we had a potentially invertible determinate. Nope, it got confused with the placement of the determinate(Y axis). So, I flipped the determinate the other way: Y = Ax1 + Bx2 + Cx3... and it got lost.

By the time I worked through the process I'd spent ~2 hours playing with it.

Then I tried some relatively simple integrals. Hopelessly lost! I set limits for integration steps, and it crashed. I removed the limits for steps, and it crashed. I set bounds for the taper(min, max, span) values and it crashed. I'm sure the guys doing protein modeling have spent HOURS setting up the mesh params, and they are able to do much more with the better version of AI I was using(free).

For stuff like interpreting what someone wants from a free language source it's pretty good. For using basic well-known math functions I was pulling my hair out. By the time I would get it resolved, I could simply copy the matrix of unknowns right into the matrix I set up in Excel, and in a few seconds, or a minute, Bam! I had my offsets, my slope, and the individual cell multiplier as well as my RsQ value.

If I had spent another 4-5 hours fighting through it, I likely would have been able to set it up. Or - I could HIRE someone who is skilled with AI to set up the deal. The right human teacher might even get it to integrate will some skill. Maybe, after N hours of work.


Oh, MBA students out there? Don't use AI to write papers. We use it to discover AI assisted fumblings from grad students:


justdone.ai

AI Plagiarism Checker | JustDone.ai

Detect plagiarism with JustDone. Try our AI-powered online plagiarism checker to scan your content and ensure originality in your writing.
justdone.ai justdone.ai

I don't teach subjects which would lend themselves to AI cheating, but colleagues do use it often. It's apparently pretty good success.
 
I tried using ChatGPT awhile ago and was forcefully reminded that it isn't a search engine, just a language model. It may have improved since, but yeah, you'd better already know what you're doing. I'm a scientist and I've been working on a mineral called dolomite. I actually just did it for fun, so I wasn't trying to seriously do research. I asked ChatGPT to tell me about dolomite and right away spotted errors. When I pointed them out, it apologized (charming) and spit out more errors. My favorite thing was I asked it to give me references and, while most of the journal names were legit, everything else (authors, publication year, article title, volume and page numbers) were all bogus.

On a more serious note, I wanted to see if it could generate a usable image that I could use to illustrate some of my findings on researching the ecology of the Navajo Sandstone (a Jurassic desert deposit). No matter how detailed my instructions were, it insisted on putting pterosaurs in the picture, even if I specifically instructed it not to include pterosaurs (none are know from the Navajo Sandstone). It would put some correct plants in, but with pterosaurs, so I would ask it to keep those plants and get rid of the pterosaurs, but every time it generated a new image, it would change the plants and add pterosaurs. The images it did generate were pretty nice, but unusable. I hear what @flyingcheesehead is saying about having to be very specific, but I don't know how much more specific "keep the cycads and omit pterosaurs" can be.
 
I had the same experience. About a year ago or so I asked ChatGPT a couple of questions for which I already knew the answers. All of the answers are easy enough to find from publicly available sources, and have been there since well before 2020. The answers it gave were wrong in every possible way. I pointed out the errors, got an apology, and a fresh set of different but equally incorrect answers other than the one fact that I had corrected. I repeated that cycle a few times, never getting even close to the correct information.
 
Yesterday I asked chatgpt how gene hackman and his wife died. It said they were both very much alive. I replied saying no they are both dead. It asked me to cite a source. I replied “literally any source on the planet”. It came back with the full overview of everything that was known up until that point about their death. I asked ‘can you clarify why you said they were alive when they are in fact dead?” It said it was using the info it had up until I pushed it to take another look. Really weird.
 
Prompt engineering is a science in itself. One has to know how to use AI. If you don't, you won't get good results. Sh** in, sh** out very much applies to AI.
 
judypilot said:
but yeah, you'd better already know what you're doing. I'm a scientist and I've been working on a mineral called dolomite. I actually just did it for fun, so I wasn't trying to seriously do research. I asked ChatGPT to tell me about dolomite and right away spotted errors.
Click to expand...
Absolutely. There's guys like my buddy who ask it a random question about quantum mechanics at dinner, then show us all the screen filled with long descriptions and mathematical formulas to "prove" that it has a PhD level understanding of quantum physics. Doesn't matter than he has zero chance of ever validating that the answer is correct. But some people really are convinced by just the presence of an intelligent sounding/looking answer. Dangerous!

judypilot said:
No matter how detailed my instructions were, it insisted on putting pterosaurs in the picture, even if I specifically instructed it not to include pterosaurs (none are know from the Navajo Sandstone). It would put some correct plants in, but with pterosaurs, so I would ask it to keep those plants and get rid of the pterosaurs, but every time it generated a new image, it would change the plants and add pterosaurs. The images it did generate were pretty nice, but unusable
Click to expand...
This is a problem with textual response too, though it presents itself more obviously in images. It has trouble disassociating certain words from particular outputs.
Even using negative prompting and explicitly telling it not to do certain things is not enough. There's some times you need to find the particular phrase that is derailing things and remove it. And some items must be so correlated in its training data that it can't do things like remove one item but keep the other (that's my speculation).

Chrisgoesflying said:
Prompt engineering is a science in itself. One has to know how to use AI. If you don't, you won't get good results. Sh** in, sh** out very much applies to AI.
Click to expand...
It's true that good prompts are key to getting the results. But there are plenty of cases where no amount of prompt engineering is going to get you there because the system is not capable.
 
I asked chatgpt and grok3 to tell me about myself. It was interesting how different the results were. Neither were wrong, but I was really surprised how different they were.

chatgpt went into almost a google mode. It gave me sentences and paragraphs describing my channel, but also quickly switched to links to my various socials. It was also pretty broad. It had links to a startup I did in Mexico about a decade ago and displayed that as the first result, though it definitely hasn't been relevant for a very long time. On the whole, it mostly had the facts right, but not the priorities and read very much like it was machine generated.

grok3 started by noting that I had a YT channel that seemed centered around the midwest while it also correctly found me on LinkedIn with an Austin, TX location. It thought that those two finds were probably the same person, but explained the reason to think otherwise as well. It then wrote a page or so bio that was very accurate and more human sounding than I can ever remember getting out of a chatgpt response.

So, I think I'll stick with chatgpt for tech work. Mostly out of habit. But also because it hasn't steered me wrong yet.

But I'm going to seriously consider grok3 for prose. My first experiment was that much better.
 
eman1200 said:
Yesterday I asked chatgpt how gene hackman and his wife died. It said they were both very much alive. I replied saying no they are both dead. It asked me to cite a source. I replied “literally any source on the planet”. It came back with the full overview of everything that was known up until that point about their death. I asked ‘can you clarify why you said they were alive when they are in fact dead?” It said it was using the info it had up until I pushed it to take another look. Really weird.
Click to expand...

I've actually gotten replies much the same. It spits out something obviously wrong and when it is pointed out with a reference it replies, "thanks for the update" ...
 
Daleandee said:
I've actually gotten replies much the same. It spits out something obviously wrong and when it is pointed out with a reference it replies, "thanks for the update" ...
Click to expand...
Yeah, this generation of technology is not a replacement for google when it comes to current events. These models take something like a week to a month to train. The actual number seems to be a corporate secret that hasn't been replaced. I also don't know how often the models are retrained as that also seems to be a secret. It's definitely massively expensive to do the training, so I wouldn't be surprised if the retraining was as infrequent as they can get away with.
 
DaleB said:
I had the same experience. About a year ago or so I asked ChatGPT a couple of questions for which I already knew the answers. All of the answers are easy enough to find from publicly available sources, and have been there since well before 2020. The answers it gave were wrong in every possible way. I pointed out the errors, got an apology, and a fresh set of different but equally incorrect answers other than the one fact that I had corrected. I repeated that cycle a few times, never getting even close to the correct information.
Click to expand...
One of the major issues that needs to be worked out with these models is that they don't seem to be able to simply say "I don't know". They are currently designed so that they have to give an answer, with the confidence of a politician or mid-level corporate manager, regardless of its accuracy.
 
Baked Potato said:
One of the major issues that needs to be worked out with these models is that they don't seem to be able to simply say "I don't know". They are currently designed so that they have to give an answer, with the confidence of a politician or mid-level corporate manager, regardless of its accuracy.
Click to expand...
If I'm taking my mind off things, sometimes I'll go down a prompt rabbit hole and see how deep it goes before getting nonsensical results. I thought this one was amusing because it's one that a human artist could easily do a good job with, but the AI didn't seem able to combine patterns across knowledge domains. So, instead of saying, "I don't know", it came up with this drawing to represent "a tiger-striped, giraffe-necked, antlered hippo reacting to wheat flour allergy with sneezing, watery eyes, and a slightly swollen snout."

1740947427704.png
 
We had a hallucination problem at work last year. We were using AI to track students through online classroom activities and predict what class would be useful to do next based on their previous experience. If the most useful class didn't actually exist, it didn't stop the AI from hallucinating it, adding it to the course offerings and suggesting it to the students.
 
rwellner98 said:
If I'm taking my mind off things, sometimes I'll go down a prompt rabbit hole and see how deep it goes before getting nonsensical results. I thought this one was amusing because it's one that a human artist could easily do a good job with, but the AI didn't seem able to combine patterns across knowledge domains. So, instead of saying, "I don't know", it came up with this drawing to represent "a tiger-striped, giraffe-necked, antlered hippo reacting to wheat flour allergy with sneezing, watery eyes, and a slightly swollen snout."

View attachment 138637
Click to expand...

I think it’s pretty spot on for a first attempt
 
You must log in or register to reply here.
Back
Top