### Bookstore

 Computer Math and Games in Pascal (preview) Lazarus Handbook

### Author Topic: Generative AI - opinion  (Read 4451 times)

#### ccrause

• Hero Member
• Posts: 933
##### Re: Generative AI - opinion
« Reply #15 on: July 15, 2024, 08:02:55 am »
two others:
temperature and color
temperature and distance
As stated the relationships above are underdetermined, similar to my original prompt.  Adding context can make it solvable, but then the context should be provided, or if assumed, clearly listed.

For instance temperature and colour can be related via the black body radiation theory.  But this theory is not directly applicable AFAIK to the colour of light emitted by an LED.  Also, one needs to distinguish between emitted and reflected light.

Quote
This particular - the steam - equation is called the Antione equation and AI came up with the correct answer
You did not state your specific prompt that lead to this.  Of course even when asking for a relationship between temperature and pressure for water specifically, one needs to understand the limitations of the suggested solution:
* The Antoine equation has a small range of accuracy
* It is applicable only for the saturated vapour/liquid equilibrium region
* It is based on simplifying assumptions of the underlying thermodynamics

#### Eugene Loza

• Hero Member
• Posts: 729
##### Re: Generative AI - opinion
« Reply #16 on: July 15, 2024, 08:24:47 am »
Pascal, Kelvin and Celsius walk into a Bar...
....Because the original prompt mentioned bar the AI model used the conversion from Pa to bar because it was (I assume) statistically likely to be relevant.
I'm almost sure it decided to convert Pascals to Bars because you've asked for that in Pascal language and it got confused between relevant keywords.
My FOSS games in FreePascal&CastleGameEngine: https://decoherence.itch.io/ (Sources: https://gitlab.com/EugeneLoza)

#### MarkMLl

• Hero Member
• Posts: 7493
##### Re: Generative AI - opinion
« Reply #17 on: July 15, 2024, 08:35:39 am »
I'm not sure where the specific 100000 number came from, but as the steam table link points out "Steam and temperature posess a reliable realationship whereby if the pressure of the steam is known, it's temperature can be predicted (and vice versa)."  The appropriate response, I suppose, would be to put the question to ChatGPT

In that case read the OP:

Quote
It also provided the following explanation:
...
2. **KelvinToBar function**: Converts a temperature from Kelvin to bar. The conversion factor used here is 1 bar = 100000 Pascal.

The so-called AI has latched onto the fact that Pascal was mentioned in what it was being asked to do, and made the invalid assumption that units of pressure (Pascals) were somehow relevant and should be worked into the program.

That is why elsewhere we are assuming that an LLM which lacks domain-specific training in NBG for document preparation. The example I gave was "seeking backwards in a file" being confused with "abusing a metalworking file by drawing it backwards, but this one is equally good... perhaps better, since actually tested.

MarkMLl
MT+86 & Turbo Pascal v1 on CCP/M-86, multitasking with LAN & graphics in 128Kb.
Logitech, TopSpeed & FTL Modula-2 on bare metal (Z80, '286 protected mode).
Pet hate: people who boast about the size and sophistication of their computer.
GitHub repositories: https://github.com/MarkMLl?tab=repositories

#### TRon

• Hero Member
• Posts: 3141
##### Re: Generative AI - opinion
« Reply #18 on: July 15, 2024, 08:36:28 am »
@ccrause
I waited with a response on purpose because I had a hunch the first post was intended to be more or less sarcastic (and please don't get me wrong, I get it).

If one blames the prompt (which /is/ a fair thing to do) then that also indicates that in case you are not familiar with the subject  (or programming language) you can get lied to and be deceived by answers that are inaccurate or plain wrong. But if the user providing the prompt is not aware then you run into mayhem pretty fast.

Besides that, even if you are aware the presented results are wrong and try to correct that a gai often gets confused pretty fast when drawing attention to these errors and more often than not results in hallucination (because the AI doesn't know what to do either so starts spiffing nonsense).

Especially larger blocks of code or more complicated solutions often contain minor errors which get bigger and bigger the moment you instruct to rectify those. And something that I have seen often is after specifying the error that the AI apologizes and tells that it will rectify things to end up presenting exactly the same faulty code. Sometimes it will only add comments to match the original question (which I interpret as an additional error because it now also explicitly lies about the mistake it made somehow believing such comment will actually fix the original error).

In short: nothing you should use to start learning a programming language and also not use for real practical examples (besides all the other obvious obstacles such as copyright)
All software is open source (as long as you can read assembler)

#### ccrause

• Hero Member
• Posts: 933
##### Re: Generative AI - opinion
« Reply #19 on: July 15, 2024, 09:52:52 am »
@ccrause
I waited with a response on purpose because I had a hunch the first post was intended to be more or less sarcastic (and please don't get me wrong, I get it).

If one blames the prompt (which /is/ a fair thing to do) then that also indicates that in case you are not familiar with the subject  (or programming language) you can get lied to and be deceived by answers that are inaccurate or plain wrong. But if the user providing the prompt is not aware then you run into mayhem pretty fast.

Besides that, even if you are aware the presented results are wrong and try to correct that a gai often gets confused pretty fast when drawing attention to these errors and more often than not results in hallucination (because the AI doesn't know what to do either so starts spiffing nonsense).

Especially larger blocks of code or more complicated solutions often contain minor errors which get bigger and bigger the moment you instruct to rectify those. And something that I have seen often is after specifying the error that the AI apologizes and tells that it will rectify things to end up presenting exactly the same faulty code. Sometimes it will only add comments to match the original question (which I interpret as an additional error because it now also explicitly lies about the mistake it made somehow believing such comment will actually fix the original error).

In short: nothing you should use to start learning a programming language and also not use for real practical examples (besides all the other obvious obstacles such as copyright)

Yep, agreed!

#### gidesa

• Full Member
• Posts: 102
##### Re: Generative AI - opinion
« Reply #20 on: July 15, 2024, 01:11:57 pm »
Pascal, Kelvin and Celsius walk into a Bar...
I'm almost sure it decided to convert Pascals to Bars because you've asked for that in Pascal language and it got confused between relevant keywords.

Yes, nice idea

#### Curt Carpenter

• Hero Member
• Posts: 503
##### Re: Generative AI - opinion
« Reply #21 on: July 15, 2024, 04:13:22 pm »
...

Besides that, even if you are aware the presented results are wrong and try to correct that a gai often gets confused pretty fast when drawing attention to these errors and more often than not results in hallucination (because the AI doesn't know what to do either so starts spiffing nonsense).
...

Sounds remarkably like many human responses to error to me

#### TRon

• Hero Member
• Posts: 3141
##### Re: Generative AI - opinion
« Reply #22 on: July 15, 2024, 04:17:46 pm »
Sounds remarkably like many human responses to error to me
Thank you for the ROTFLOL as it made my day (but sadly there is truth in that).
All software is open source (as long as you can read assembler)

#### Curt Carpenter

• Hero Member
• Posts: 503
##### Re: Generative AI - opinion
« Reply #23 on: July 15, 2024, 04:23:25 pm »

That is why elsewhere we are assuming that an LLM which lacks domain-specific training in NBG for document preparation...

Fair enough.  What is "NBG"?

#### MarkMLl

• Hero Member
• Posts: 7493
##### Re: Generative AI - opinion
« Reply #24 on: July 15, 2024, 04:28:59 pm »

That is why elsewhere we are assuming that an LLM which lacks domain-specific training in NBG for document preparation...

Fair enough.  What is "NBG"?

No Bloomin' Good.

Should obviously have read "is NBG".

MarkMLl
MT+86 & Turbo Pascal v1 on CCP/M-86, multitasking with LAN & graphics in 128Kb.
Logitech, TopSpeed & FTL Modula-2 on bare metal (Z80, '286 protected mode).
Pet hate: people who boast about the size and sophistication of their computer.
GitHub repositories: https://github.com/MarkMLl?tab=repositories

#### Weiss

• Full Member
• Posts: 168
##### Re: Generative AI - opinion
« Reply #25 on: July 17, 2024, 04:49:56 am »
Pascal, Kelvin and Celsius walk into a Bar...
....Because the original prompt mentioned bar the AI model used the conversion from Pa to bar because it was (I assume) statistically likely to be relevant.
I'm almost sure it decided to convert Pascals to Bars because you've asked for that in Pascal language and it got confused between relevant keywords.

If there was a confusion, wasn't this thing designed to say so, and ask questions? Instead, the thing "invented" conversion.

#### Curt Carpenter

• Hero Member
• Posts: 503
##### Re: Generative AI - opinion
« Reply #26 on: July 17, 2024, 05:04:26 am »
Pascals to Bars is a real conversion, not invented by the thing.  I agree that it won't really match our human abilities until it is able to say "what?" (and lie)  though.

#### TRon

• Hero Member
• Posts: 3141
##### Re: Generative AI - opinion
« Reply #27 on: July 17, 2024, 05:27:36 am »
I agree that it won't really match our human abilities until it is able to say "what?" (and lie)  though.
I believe the point there is that there is no incentive to deny a request. In the end the idea behind all this is to engage the end-user as long as possible.
All software is open source (as long as you can read assembler)

#### Weiss

• Full Member
• Posts: 168
##### Re: Generative AI - opinion
« Reply #28 on: July 17, 2024, 05:42:05 am »
Pascals to Bars is a real conversion, not invented by the thing.  I agree that it won't really match our human abilities until it is able to say "what?" (and lie)  though.

I am engineer, steam tables were my bread and butter at some point. Pressure and temperature are connected of course (P*V:=T). But you cannot convert directly, without reference to the process being isochoric, and definitely there is no formula like the one invented by the thing, anywhere. That little thing could not have used a reference from any source to convert Kelvins to bars.

if it was confused, the thing would have used well known and working conversion formula. Like, "sorry I mixed up". But question was non-ambiguous. Convert units of temperature to units of pressure. To me, question of this sort would either make me laugh or shake my head and doubt the sanity of inquirer. The thing did not flinch, proceeded straight to boosheating its way through hoping no one would know the difference.

This is an emerging intelligence, folks. We just need to invent a vice to pinch its balls every time it is cought lying like that.

#### TRon

• Hero Member
• Posts: 3141
##### Re: Generative AI - opinion
« Reply #29 on: July 17, 2024, 05:47:15 am »
This is an emerging intelligence, folks. We just need to invent a vice to pinch its balls every time it is cought lying like that.
And with that you have also effectively killed all new ootb idea's that aren't established (by vice) yet
All software is open source (as long as you can read assembler)