Another AI success story

[duncan_bayne]

This screenshot was taken by a family member today, who was searching for guidelines to interpret Bell helicopter model numbers.



It astounds me that people are using LLMs unsupervised, for any purpose. And yet they keep doing it; just today I saw a company (PocketOS) whose operations were destroyed when an unsupervised LLM destroyed their production database and backups.

[row13please]

To err is human.
To really mess up takes a computer.

[Bravo73]

This is genius: “For example, the Bell 412 is known for having 4 engineers, 1 problem,
and 2 days to fix it.” 🤣🤣🤣

Nearly as good as an engineer telling me in the past that the 109 is called that because it requires 9 hours of maintenance for every 1 hour of flight time. 🫣

[helispotter]

Originally Posted by duncan_bayne

This screenshot was taken by a family member today, who was searching for guidelines to interpret Bell helicopter model numbers.



...

So once the 4 engineers have spent 2 days fixing the one problem on the Bell 412, is it trouble free for the rest if it's life? Hope the LLM can enlighten me!

[212man]

Interesting, I just asked ChatGPT where the H225 gets its designation from:

🔢 So what does “225” mean?

The “225” in EC225 → H225 is essentially an internal model number, not a direct reference to:

• passenger capacity
• weight
• engine power

However, there is some logic behind it:

• “2” → part of Eurocopter’s naming family for medium-to-heavy twin-engine helicopters
• “25” → indicates a specific variant/evolution level within that family, following earlier models like:
  • AS332 (earlier generation)
  • EC225 (next major upgrade)

So the number reflects where it sits in the product lineage, not a spec sheet value.

Totally incorrect, of course.

Now, to be a total anorak, I'll explain the real reason.
At that time (Eurocopter and then Airbus) all models had three digit numbers:

• First number: 1 (Civil) or 6 (military)
• Second number: weight category in tonnes
• Third number: 0 (one engine) or 5 (two engines)

So, EC120 - single engine, two tonne civil aircraft. EC635, twin engined, 3 tonne military helicopter.

EC1125. Oh, Can't have four digits, so:

1
12
5
-----
225

Same for the 725

I'll get my coat.......

(For some reason the 5 won't stay where I put it, aligned right!)

[TasmanAir]

It helps to think of an LLM as a text-generator which is coincidentally correct about answers to questions which are commonly associated in its training data but no capacity to actually validate information presented beyond these associations.

It is also trained to be confident in its presentation and to give a certain level of content or detail even if the information requested does not exist in its data. This leads to the lying and hallucinations that LLMs are notorious for and it’s extremely present in topics that require a high level of niche knowledge such as aviation.

Any sufficiently advanced science is indistinguishable from magic and everyone who promotes LLMs as “intelligent” is lying to you for money.

[row13please]

Originally Posted by TasmanAir

everyone who promotes LLMs as “intelligent” is lying

It's the obvious nonsequiturs that amaze me, such as consecutive sentences that are completely the opposite. An intelligent 3-year-old would detect that even with no knowledge. But "AI" despite its resources clearly has no actual intelligence. If it were a human, it would have been written off as insane long ago and never talked to again for fear of catching whatever it has.

So glad it's making hiring decisions, driving cars into other cars, etc. But maybe it could get an order right at a burger drive-through. No human can.

[Bravo73]

Originally Posted by 212man

At that time (Eurocopter…

I always felt that it was a shame that when they rebranded as ‘Airbus Helicopters’, the company abolished this naming convention. 🤓

The Airbus designators have about as much logic to them as the Agusta/Leonardo ones do. 🤦🏻‍♂️

[avtech23]

Originally Posted by duncan_bayne

It astounds me that people are using LLMs unsupervised, for any purpose.

I've seen someone use ChatGPT on multiple occasions to make airworthiness based decisions.

For example, asking it if two standards are the same to determine if compliance with one satisfies the other (these are paywalled national standards, so it's doubtful that the LLM can access beyond the title of the standards and a few sketchy blog posts related to it).

Or asking it whether or not a procedure is compliant with Part 42 regulations.

Or asking it if a particular generic part is compatible with a particular aircraft.

And on and on.

Using an LLM as a crutch to hide behind when you are in a position beyond your capability is unconscionable, yet these people walk amongst us.

[DIBO]

Originally Posted by 212man

1
12
----5
------
225

(For some reason the 5 won't stay where I put it, aligned right!)

Fixed that for you

[212man]

Originally Posted by DIBO

Fixed that for you

thanks! I guess I could have asked ChatGPT to help! 🤣🤣

[[email protected]]

AI is definitely not intelligent - it is simply very fast data scraping - hence the size of the server farms needed to run it.

[retoocs]

I wonder if it pulled it from this thread?
Bell Helicopters naming

[212man]

Originally Posted by [email protected]

AI is definitely not intelligent - it is simply very fast data scraping - hence the size of the server farms needed to run it.

The power consumption is eye watering. Not to mention water usage. But, it is the future and I certainly use it daily. It’s great for document gap analysis, data extraction and tabulation and finding regulations, to name a few. No intelligence just huge processing power and data access.

I recently stuck in my wife’s employee manual snd the collective labour agreement - it spat out the comparison in a second or two.

[TimdeBoer]

Originally Posted by retoocs

I wonder if it pulled it from this thread?
Bell Helicopters naming

Hahaha 🤣

Oh yeah it definitely did!

The problem with a lot of LLM's is that they always want to give an answer... 🙄

These models should have good programming to be able to say:
"I don't know the answer to what you are asking"
when the model truly has no idea.

Instead often some ridiculous response is just made up.

[row13please]

Originally Posted by TimdeBoer

These models should have good programming to be able to say:
"I don't know the answer to what you are asking"
when the model truly has no idea.

Instead often some ridiculous response is just made up.

Maybe its creators make it tend to do that because they've invested so much time, money and effort, and don't want people to think AI is stupid. Like in school we were taught to always answer the question, at least make an intelligent guess. And in some cultures, confidence is a necessity.

But when we don't want it to do that, are you aware that the AI can be conditioned? For at least some, there are settings and an optional configuration file where you can indicate preferences, or you can just tell it your preferences in the session.

i.e. "Please only give answers if you have reliable and specific source material that is at least 70% likely to be correct. Otherwise please just tell me that you don't know the answer. But you can suggest where to look."

I'm a major AI skeptic and would vote to have it banned, but some of it is quite impressive and can be useful.

[albatross]

One very funny instructor many years ago defined the whole aviation flight planning process as “ Making ill informed, life threatening decisions” and actual flight control movements as “a continuous process of correcting for the errors you have just made!” ( very true in hovering a helicopter ) .
He often pointed to me as the living proof of his theory. He was a joy to fly with and a great instructor.

I postulate that reliance on AI flight planning may lead to even more “ Ill informed life-threatening decisions.”

[First_Principal]

I have been researching the use of LLMs in a variety of settings for the past two or three years. Note that these are locally-hosted systems, not Saas, thus ensuring sensitive data is kept in-house and experiments can be, er, extensive.

It's my view that such systems can be useful in instances where their reach is bounded. By that I mean that if you have a specific query, or specify a text that you want to analyse, and you also specify a set of source documents to use with said analysis or query, and you are very precise in your prompting, then they can produce some quite timely and useful results. In other tests I have found them to have mixed outcomes with general coding, and with analysis of handwritten documents, and generally positive or amusing outcomes when asked to write a story (eg. "write a 10 page story about Darcy and Elizabeth's great-grandchildren's first motor vehicle crash"). The latter is assisted if you give some background documents on early English motoring...

But, as the LLM sphere is at present, I would most definitely NOT use them for a general, unbounded, query engine! I'm entirely unsurprised by the OP's relative's result, amused by some of the responses here, and heartened by others, so perhaps not all is lost, however I have seen posts on this site where people have presented the results of a stochastic parrot as authoritative - which I find disturbing, particularly when they come from 'senior' members.

Takeaway: if you want to know the answer to a general knowledge question, try going to the fundamental source, then Wikipedia perhaps where at least there's usually some peer-review process. @TasmanAir is particularly on the point about lack of validation, and perhaps that will come in time (despite sounding negative I'm not, the year-on-year advancements are astounding - see the 2025 and 2026 Stanford AI index report for example), but they're definitely not there yet.

FP.

[MechEngr]

Originally Posted by TimdeBoer

Hahaha 🤣

Oh yeah it definitely did!

The problem with a lot of LLM's is that they always want to give an answer... 🙄

These models should have good programming to be able to say:
"I don't know the answer to what you are asking"
when the model truly has no idea.

Instead often some ridiculous response is just made up.

They always have an answer because they are building a large probability table and will select the most probable response they can make. Even if that result is nonsense or wrong, it's what they have to work with. They don't store how some probable connection was created - they don't recall that it came from a specific paper or standard; no traceability exists in the probability table. There is no place in the LLM for an administrator to make a point fix. If somehow the model is fed information where a person's name is consistently spelled incorrectly, there isn't a way to search-and-replace to fix it.

Contrast with Wikipedia where, for the most part, community guidelines require sources and almost every individual sentence in many.articles has one or more listed/footnoted sources.

Mary Shelly's Frankenstein was a work of fiction. AI now is the reality of what happens when someone puts a bunch of random parts together. The makers rob people of intellectual property, just as D. Frankenstein was robbing graves. I'll have to consider more about the parallels, but I cannot help but wonder when the Creation will turn against the Creator.

[Lonewolf_50]

Originally Posted by First_Principal

I have been researching the use of LLMs in a variety of settings for the past two or three years. Note that these are locally-hosted systems, not Saas, thus ensuring sensitive data is kept in-house and experiments can be, er, extensive.

It's my view that such systems can be useful in instances where their reach is bounded.

Very much agree. I have had to take a few training modules lately on use of AI (our company training folks put them out) and how not to use it.

however I have seen posts on this site where people have presented the results of a stochastic parrot as authoritative - which I find disturbing, particularly when they come from 'senior' members.

Yeah, and I agree with both your diagnosis and recommended treatment, doctor.