[NA]The quintessential AI thread

Let'sgoflying!

Let'sgoflying!

Touchdown! Greaser!
Joined
Feb 23, 2005
Messages
21,872
Location
west Texas
Display Name
Dave Taylor
Home Airfield
KMRF
Since the beginning, I was told to be leery of Chatgpt (I removed it) ; Gemini (an option in google searches I still play with) ; and Claude (more of a coder’s tool, I think) Copilot, plus others.

I was leery, and I quickly learned why this was essential advice.

AI has spewed out such impossible fabrications (then cheerfully offered corrections when the error is pointed out) that it’s laughably unusable.

But, not all people understand this; I had a client tell me last week how it is the greatest thing since sliced bread and he was oblivious to its propensity to produce hallucinations; his blind acceptance frankly scares me.

Repositories of legal hallucinations: https://www.halupedia.com/

and

naturalandartificiallaw.com

AI Hallucination Cases Tracker

AI hallucination cases are no longer rare. This tracker lists decisions worldwide, each linked but verify yourself. Updated regularly.
naturalandartificiallaw.com naturalandartificiallaw.com

hmm

Halupedia - Wikipedia

en.wikipedia.org en.wikipedia.org



Might be careful about investing in any AI ventures just yet




I’m hoping we have all had a similar experience and are now cautious about AI:

Recently I asked one of them the specs for an airplane. It confidently told me. in addition to many other numbers that the cruise speed was the same as its vne. I asked about it and again, a cheerful correction. I cannot believe any of the specs it provided now.



If these things are so smart I don’t get why they are not programmed to listen to human challenges to incorrect info, and then ‘look into it’ so the next person gets better info. (I asked once if it could update the wrong info so the next person is not lied to and it basically refused).

I think Wikipedia has gotten better over the years due to human challenges(?).

Are all these billions of $$ going into data centers mostly to help people/business/governments answer questions better? That seems to me they hope to monetize this soon (I mean betting what they do now for coders etc; the average public user might have to pay for it)

I don’t get the feeling of imminent danger to the human race like the media is saying, but I could easily foresee a rogue agent getting ahold of military assets, so…

Another side topic:
so far I only mentioned a narrow slice of AI activities. Photos, videos, voices (ie fraudulent manipulation of someone’s voice for nefarious purposes), probably things like such market manipulation, banking & coin theft, interference by & of governments are easily foreseeable. It’s just so powerful.

Edit:
As usual, corrections and additions welcome.
 
Last edited:
Let'sgoflying! said:
If these things are so smart I don’t get why they are not programmed to listen to human challenges to incorrect info, and then ‘look into it’ so the next person gets better info.
Click to expand...
They do. This is part of the reason they are so much better than they were five years ago.

But you want it to be "right damn now!" and that's unreasonable for both technological and practical reasons.

From a practical standpoint, they do get human feedback, in a manner somewhat similar to wikipedia, via human editors. This is called data annotation and every AI company hires teams of humans that are specialized in certain skills to correct data and tune which sources are more or less trusted. They also allow you to down vote bad answers and this is used as part of the next round of training. This all takes time to accumulate and be validated.

From a technological standpoint, it takes two to six months to train a new frontier model (all of your examples were frontier models). So even if you were viewed by ChatGPT as the single authority to end all authorities on a subject, a correction wouldn't be in the next model for two to six months.
 
Claude does real well in the M&A world though. Good at crunching numbers for acquisitions and evaluating fit with current distribution. I'm also automating some accounting stuff with it.

But yesterday I had to point out to Gemini that it did math wrong.
 
Better (paid) models produce better results.

Better prompting produces better results.

The models improve every quarter.

But yes, ask the LLM to bring receipts or you will get lied to because it is rather too based on humanity -- it will lie, cheat, and steal if you let it and is one lazy mofo when prompted to work on something difficult.
 
schmookeeg said:
But yes, ask the LLM to bring receipts or you will get lied to because it is rather too based on humanity -- it will lie, cheat, and steal if you let it and is one lazy mofo when prompted to work on something difficult.
Click to expand...
:yeahthat:

"Do this thing for me."
"Ok!"
"Cool, looks good! Now do the same thing on this list of 2000 inputs."
"Woah! That looks hard. NFW."
"Fine, be that way. Write a program that prompts an LLM to do this thing..."

... and then I run the GD script in a loop against the same GD underlying LLM and get the answers I need :D
 
schmookeeg said:
Better (paid) models produce better results.

Better prompting produces better results.

The models improve every quarter.

But yes, ask the LLM to bring receipts or you will get lied to because it is rather too based on humanity -- it will lie, cheat, and steal if you let it and is one lazy mofo when prompted to work on something difficult.
Click to expand...
This. The newer models are pretty impressive. They need less complex prompts, but are not perfect yet.
I use a system prompt to cut the crap, always list authorized references and give me a tldr that is short and blunt. This made the experience a lot better.
 
MacFlier said:
This. The newer models are pretty impressive. They need less complex prompts, but are not perfect yet.
I use a system prompt to cut the crap, always list authorized references and give me a tldr that is short and blunt. This made the experience a lot better.
Click to expand...
This is the load-bearing way to do it :D
 
Let'sgoflying! said:
I don’t get the feeling of imminent danger to the human race like the media is saying, but I could easily foresee a rogue agent getting ahold of military assets, so…

Another side topic:
so far I only mentioned a narrow slice of AI activities. Photos, videos, voices (ie fraudulent manipulation of someone’s voice for nefarious purposes), probably things like such market manipulation, banking & coin theft, interference by & of governments are easily foreseeable. It’s just so powerful.
Click to expand...

The fear comes down to 'agentic' AI. That style of AI is capable of 'doing things.' Things such as opening the browser on your computer, logging into your bank (as you), and performing any action (benign or malevolent) as instructed or not instructed.

If you hear a news story about an AI doing something (like ruining someone's reputation or compromising computer security) it was likely agentic AI.

What Is Agentic AI? Definition, 6 Levels & Examples (2026)

Agentic AI acts on its own — planning, using tools, and adapting until a task is done. Learn how it differs from chatbots, the 6 levels of autonomy, and real tools to try.
agentic.ai agentic.ai

mitsloan.mit.edu

Agentic AI, explained | MIT Sloan

The age of agentic AI — systems that are semi- or fully autonomous and can act on their own — has arrived. Here’s what you need to know, according to MIT experts.
mitsloan.mit.edu mitsloan.mit.edu
 
2-Bit Speed said:
The fear comes down to 'agentic' AI. That style of AI is capable of 'doing things.' Things such as opening the browser on your computer, logging into your bank (as you), and performing any action (benign or malevolent) as instructed or not instructed.

If you hear a news story about an AI doing something (like ruining someone's reputation or compromising computer security) it was likely agentic AI.

What Is Agentic AI? Definition, 6 Levels & Examples (2026)

Agentic AI acts on its own — planning, using tools, and adapting until a task is done. Learn how it differs from chatbots, the 6 levels of autonomy, and real tools to try.
agentic.ai agentic.ai

mitsloan.mit.edu

Agentic AI, explained | MIT Sloan

The age of agentic AI — systems that are semi- or fully autonomous and can act on their own — has arrived. Here’s what you need to know, according to MIT experts.
mitsloan.mit.edu mitsloan.mit.edu
Click to expand...
No agentic for me yet. Mine is all manual prompts. Last week I taught it how to build my monthly analysis spreadsheets for me. That will be nice not to have to do that copy paste waste of time.

What I'd like to build is an agent where the sales team can ask the agent about inventory levels from their phone. Still trying to decide which way is cheapest since getting the data to the agent requires either licenses or tokens.
 
I use AI every day for work (Claude Code).
In fact, my job is pretty much only to build FinTech products that embed AI (Anthropic/OpenAI/Gemini models).

... And I have never been more bearish/pessimistic about AI's ability to actually deliver results (outside of coding agents).
 
ArrowFlyer86 said:
I use AI every day for work (Claude Code).
In fact, my job is pretty much only to build FinTech products that embed AI (Anthropic/OpenAI/Gemini models).

... And I have never been more bearish/pessimistic about AI's ability to actually deliver results (outside of coding agents).
Click to expand...
I use speech to text models running locally on a daily basis to help create notes and action items with a fidelity vastly better than I was ever able to consistently achieve manually.

I have also written a teleprompter iOS app, backed by the same speech to text model, that displays my scripts, listens to what I'm saying and paces the teleprompter to match the speed at which I'm presenting.
 
rwellner98 said:
I use speech to text models running locally on a daily basis to help create notes and action items with a fidelity vastly better than I was ever able to consistently achieve manually.

I have also written a teleprompter iOS app, backed by the same speech to text model, that displays my scripts, listens to what I'm saying and paces the teleprompter to match the speed at which I'm presenting.
Click to expand...
Not far from part of our utility. Since early 2024 we built out a speech to text for financial earnings calls, industry conferences, investor presentations, etc with pretty broad coverage -- and have that going in near real time so people don't have to wait for other vendors to put out a published transcript. It's paired with an app front end that shows the real time transcription so people can follow along multiple calls -- and it can listen for specific topics/guidance/etc, log when they occur and then fire off alerts as needed. Then write summaries/takeaways across multiple calls.
 
ArrowFlyer86 said:
I use AI every day for work (Claude Code).
In fact, my job is pretty much only to build FinTech products that embed AI (Anthropic/OpenAI/Gemini models).

... And I have never been more bearish/pessimistic about AI's ability to actually deliver results (outside of coding agents).
Click to expand...
I'm more bearish than I was a few months ago, but more due to human factors than tech.
 
I am finding two things happening at once in my own circles:

1. Nobody wants a pure engineer anymore because anyone with an Anthropic subby can do what I can do with 30+ years' experience.

2. As such, I am being asked to blend engineering and X where X was previously some midgrade knucklehead who worked in some unrelated part of the business.

(Case in point: my current contract is literally as a FDE or Forward Deployed Engineer -- a sneaky way to say Engy and Sales)

Per the above, I can bleed engineering into those other X roles handily and efficiently. Today I asked Claude to make some visualizations of a dataset I recently unearthed. I'd have made some cute line graphs in Excel. Claude made lovely, eye-catching heatmaps and graph types I have never heard of -- and it's dazzling.

MY role? continues.

Former X roles for this: (Data Analyst? Accounting Assistant? Some wog who knows excel? whatever we used to call "Sales Engineers" that never touch code) -- they're toast.

K-shaped? yeah confirmed. And I'm sweating THAT part. Not unemployed or underpaid Engys -- wah. I'm worried about all of the BS connective-tissue jobs that an LLM can spread into and make redundant. In the orgs I usually play in, there are a LOT of those weird people and not enough jump to conclusion mat ideas to keep them employed.
 
I did the same as the LGF (the OP) and removed every AI agent I could find on my phone. I have all the browsers I use configured NOT to give me AI responses. Unfortunately, AI cannot be avoided. Many websites silently use it for content. I wish all AI generated content was marked.

However, I am considering trying one of the AI agents to generate my next book cover--Rio Legacy. All my other covers have been collaborative projects of me and my cover girl. Any suggestions on which AI agent to use would be appreciated.
 
Domenick said:
However, I am considering trying one of the AI agents to generate my next book cover--Rio Legacy. All my other covers have been collaborative projects of me and my cover girl. Any suggestions on which AI agent to use would be appreciated.
Click to expand...

You don't want an 'agent.' You want generative AI. "Agent" has a very specific meaning in the context of AI.
 
I still have all the voice stuff off on my phone and car, but I am using AI more. AI search is about the only way to get any useful results these days. Used to be a search for "m3 pan head slotted zinc plated" would work, now all you get is "m3 button head, philips, stainless" at least with the AI I can go "no, I really do mean slotted pan head"... of course I ended up ordering from McMaster anyway.
For personal and work adjacent stuff I have a setup with local AI that I can get to write passable code and make images. It is somewhat nice to say "Hey, I know I asked for a native MCP23017 implementation, but let's add an on-board gray code decoder so I don't have to sample over 5Hz." And then just come back later and flash it and test it, which I can actually have it automate if I trusted it with hardware access, which I don't.
 
ArrowFlyer86 said:
There are some human factors that compound my bearishness. Which ones are you most pessimistic on?
Click to expand...

Where to start?

Well, you know, this is just like, uh, my opinion, man... but I saw the best minds of my generation destroyed by madness.

AI psychosis is a thing. The denial and withdrawal symptoms look a lot like those of a serious substance abuser. Pretty soon, there will be billboards that say "Struggling with AI addiction? We can help", the same way there already are for gambling. It's gonna be a thing for a whole slice of the population.

From my side, I see: Decrease in quality of work product (and/or an increase in the variance of the quality). Loss of situational awareness. Loss of critical thinking ability. Loss of ability to communicate.

Then, what happens to the team dynamic when you start mixing in personalities like that? It kills credibility and trust. It's dehumanizing and sad and divisive and horrible for morale.
 
We are already seeing graduates who can't read or write or do simple math. Colleges have already instituted remedial courses for large numbers of their students. Think reliance on AI will improve things when the user has no basis to judge the veracity of the AI response?

AI reliance is another step toward human ignorance and generations of sheep.
 
GeorgeC said:
Where to start?

Well, you know, this is just like, uh, my opinion, man... but I saw the best minds of my generation destroyed by madness.

AI psychosis is a thing. The denial and withdrawal symptoms look a lot like those of a serious substance abuser. Pretty soon, there will be billboards that say "Struggling with AI addiction? We can help", the same way there already are for gambling. It's gonna be a thing for a whole slice of the population.

From my side, I see: Decrease in quality of work product (and/or an increase in the variance of the quality). Loss of situational awareness. Loss of critical thinking ability. Loss of ability to communicate.

Then, what happens to the team dynamic when you start mixing in personalities like that? It kills credibility and trust. It's dehumanizing and sad and divisive and horrible for morale.
Click to expand...
Those are good examples IMHO and align with my own.

From my perspective I'm worried about critical thinking ability, an impaired willingness to try/challenge AI (aka: "cognitive surrender"), and a reduced capacity to retain knowledge.

I'm not exaggerating when I say that the engineers I know, all of whom use Claude/Codex/Cursor -- their professional brains have atrophied significantly over the last 2 years. Not only in the ability to write code, which would be fine since the mechanical process of writing lines of code is all but dead now and forever, but also in terms of their ability to think through problems, appreciate nuance, and come up with solutions.

What started as a neat text editor that could autocomplete small functions for you, morphed into a system that could accept more generalized instructions (neat!). Which morphed into a system that could accept sloppier, lazily written requests and then fill in the gaps for you (cool-ish). Which then morphed into a system where you can have a potato for a brain and just make some vague statements about a problem and it will come up with some passable solution [<-- WE ARE HERE].

That evolution is wonderful if you've never built anything before and you're slapping together some low effort, low risk application as a hobby.
That's not good when it's the professional norm and quickly becoming the standard mode of operation for professionals.

In my experience, even the highest paid professionals are outsourcing their thinking to Claude, simply because it's easier. The solutions aren't nearly as good or importantly -- DIFFERENTIATED -- as what they'd come up with if they thought through it, but the solution is available NOW and it costs them nothing to do it. So that's what they do. It's like digital fast food. Cheap, easy and available everywhere. You know it's not great, but it's satisfying enough to be acceptable.

And this is just a software example, but I know similar things are playing out in consulting, finance and law.
 
^Also, whatever damage this is doing to experienced professionals? The damage is 100x on students/kids, because they'll never develop the skills to begin with.

Put an [EASY] button in front of every kid and make it available for every problem: writing a history paper, solving math homework, planning an organic synthesis... Done! You don't even need to give it parameters anymore. Just snap a photo of the assignment instructions and send it to GPT Sol, and wait while it does all your thinking for you.

Why worry about building silly skills like focusing and conveying thoughts in writing? Just vomit up some jumbled heap of thoughts into AI and it will compose it for you!
 
Last edited:
John Henry was a coding man,
Had a keyboard in his hand.
Said, “I’ve shipped through nights and pager storms,
A machine won’t take my stand.”

The boss came down and told the team,
“We’ve got a brand-new plan:
This agent writes the tickets now,
And codes them while it can.”

They gave John Henry half the repo,
The agent got the rest.
It opened branches by the score
And boasted it knew best.

John Henry traced the ancient bugs,
The kind no chart can show.
The agent learned from every fix,
Then claimed the final glow.

Now the agent ships a thousand lines
Before a man writes ten.
But when it says, “The work is done,”
You’d better check again.
Click to expand...
 
Time to do a search for "I Robot" on your various streaming options. Probably pretty timely again. :eek: :cool:
 
schmookeeg said:
Now the agent ships a thousand lines
Before a man writes ten.
But when it says, “The work is done,”
You’d better check again.
Click to expand...
That's my life now- sorting through the wreckage of all the "progress" of the last couple months, as many things have now been contaminated by the premature ejaculation of AI.

ArrowFlyer86 said:
The damage is 100x on students/kids, because they'll never develop the skills to begin with.
Click to expand...
Each generation has its curses, I suppose. Mine had above ground nuclear testing, leaded gasoline, indoor smoking, and the omnipresent threat of annihilation.

Kids and parents have no choice but to navigate the world and establish what they think are appropriate boundaries. As with any interesting/disruptive tech, it will be dual use, and there will be winners and losers. It will probably end badly for the kids that end up with an AI in their face as soon as they can hold a tablet in their stroller.

There was just a $17B wrist slap for social media addiction, so I expect the techbros will be laughing all the way to the bank for the next 20 years until the AI settlement.
 
How are the
AI blockers/filters working out?
AI Content Shield, ByeAI.
I really need one for YT, FB.
 
Let'sgoflying! said:
I really need one for YT, FB.
Click to expand...
OH I would love a blocker for YouTube!
I was thinking about it just yesterday... I notice quite a few channels now that may use AI for their thumbnails so good chance that a filter would snag those too
Personally the thumbnail is not something I care about so that would be unfortunate, but I'd take it..... boy I hate those AI cartoons that look real!

I'm not anti-AI for everything though. I have enjoyed using it as a research tool. The way I've used it, it's really just a conversational interface with Google or whatever search engine, it seems. Certainly not perfect though, it spouts a lot of bogus info that I have to filter through, but it does help in research at the same time....
 
I was just thinking of another twist that might be hard to filter.... there are videos that seem to be legit/real, but they may use AI generated visual aids. Such as a real person discussing some flight mishap or situation...an accident debrief or whatever.... it's a real person talking head, but they are using AI generated cartoons to "show" what happened....
In my opinion that's not something I'd necessarily want to block.
 
I was behind on a new cooling tower project the other day so I handed 30 year old submittals to my new best friend, AI, and received back a plausible summary of existing equipment, changes in design/materials over the thirty years, etc. All in about 15 minutes of prompting.

Yay I’m not behind any more!

5 minutes later: Hey randomskylane, here’s another new project…

Ugh.

The (you have ai, do the work of three era) is ramping up (yes I should use ai to clean up this line, I hope your ai readers properly summarize for you)
 
ArrowFlyer86 said:
Not far from part of our utility. Since early 2024 we built out a speech to text for financial earnings calls, industry conferences, investor presentations, etc with pretty broad coverage -- and have that going in near real time so people don't have to wait for other vendors to put out a published transcript. It's paired with an app front end that shows the real time transcription so people can follow along multiple calls -- and it can listen for specific topics/guidance/etc, log when they occur and then fire off alerts as needed. Then write summaries/takeaways across multiple calls.
Click to expand...
I think there are some important differences. In your case, you're trying to do realtime transcription to do something useful and digestible. None of the models small enough to run in realtime on a phone were anywhere near accurate enough for that. Tons of errors.

But for my app, I didn't need accuracy. I was able to use a very lightweight model and some heuristics to let the transcription be 'close enough' to the known source of truth, the script, that it could scroll along.

I've done a bunch of other things, with bigger models on better hardware and/or without the realtime requirements, and it's kind of amazing what AI speech to text is capable of compared to older approaches from just a few years ago. But running on a phone is still a big constraint as of 2026.
 
rwellner98 said:
I've done a bunch of other things, with bigger models on better hardware and/or without the realtime requirements, and it's kind of amazing what AI speech to text is capable of compared to older approaches from just a few years ago. But running on a phone is still a big constraint as of 2026.
Click to expand...
That reminds me of a type of AI I'd like that I haven't seen yet: One that can listen to a song and then break out all of the musical parts for it. Maybe "music" just isn't a language that an LLM can understand yet, but I've only found one site (since forgotten, sorry) that claimed to do something like that, and it just failed miserably. Utterly useless.
 
flyingcheesehead said:
That reminds me of a type of AI I'd like that I haven't seen yet: One that can listen to a song and then break out all of the musical parts for it. Maybe "music" just isn't a language that an LLM can understand yet, but I've only found one site (since forgotten, sorry) that claimed to do something like that, and it just failed miserably. Utterly useless.
Click to expand...
Have you tried AudioStrip? It's been useful for the stuff I wanted to do.
 
I found a decent everyday use for AI last week.
My business insurance doubled for for no reason so I shopped it with two other composites.
So I had three quotes (including the original company) in the form of many-page pdfs.
All with limits, deductibles, exclusions for many different coverages. Would have taken hours to sort through.
Uploaded the pdfs to Gemini, it was able to tell me in chart form, the differences.
I verified, made it easy to choose. (saved $5000, too)
 
flyingcheesehead said:
That reminds me of a type of AI I'd like that I haven't seen yet: One that can listen to a song and then break out all of the musical parts for it. Maybe "music" just isn't a language that an LLM can understand yet, but I've only found one site (since forgotten, sorry) that claimed to do something like that, and it just failed miserably. Utterly useless.
Click to expand...
Now that Finale shut down (the company admitted the code was too old and too convoluted to maintain and forget about enhancements) look into Sibelius.
 
Last edited:
rwellner98 said:
Have you tried AudioStrip? It's been useful for the stuff I wanted to do.
Click to expand...
It looks like that gets halfway there by splitting things into parts. Now, if it could just output notation...

murphey said:
Now that Finale shut down (the company admitted the code was too old and too convoluted to maintain and forget about enhancements) look into Sibelius.
Click to expand...
I'm not looking for plain old music notation software - I already have MuseScore for that. The pros can have their Sibelius and Dorico, both are way too expensive for what I do, which is mainly arranging stuff for my tuba quartet. I just want something to make it easier by doing the listening and transcription parts for me.

M20K said:
Hmmm... Reminds you of an AI you haven't seen yet?????? What's wrong with this picture?? :D
Click to expand...
Well, it reminded me of an AI that I had the idea to create a while back but I have no free time to do so because kids, so I'm waiting for someone else to do it. :rofl:
 
You must log in or register to reply here.
Back
Top