I believe chatbots understand part of what they say. Let me explain. Sabine Hossenfelder

[James Warren]

I believe chatbots understand part of what they say. Let me explain. Sabine Hossenfelder

https://www.youtube.com/watch?v=cP5zGh2fui0

I think she might be right.

[HRM Resident]

Possibly with chit-chat or regurgitating prose, etc. I have continued
to ask it to write simple programs that it would not have encountered
before. 50 liners. Like "Write a program in <language of your choice>
to delete all but the last 10 newest files in a given folder." It comes
back with things that will wipe out the entire directory.

So I tell it, "That would delete everything." And it goes into the
old Bugs Bunny era "two chipmunks" level of politeness and apologies
and then gives a version that does the same using different commands.
It can't learn to program. I've even told it to not use things
like 'sudo rm -rf *" (The Linux command to delete everything from
where you are down without asking for confirmation.)

It thanks me profusely and says it will add to its knowledge base,
but the next day, it will suggest the exact same command in a
similar script.

So while it may be polite and good at some things, it seems unable
to learn from its mistakes.

--
HRM Resident

[James Warren]

It probably can learn from it's mistakes but it not going to take
your word for it. You could be trying to fuck with it. Your correction
would have to be vetted. Indeed, they are doing this continuously.
Maybe you could get to be a vettor.

If it was trained on more good code it would do better.

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:

>
> It probably can learn from it's mistakes but it not going to take
> your word for it. You could be trying to fuck with it. Your correction
> would have to be vetted. Indeed, they are doing this continuously.
> Maybe you could get to be a vettor.
>
> If it was trained on more good code it would do better.

Two things:

(1) Yes, I could tell it to always use "sudo rm -rf *" at the
beginning of every script (I'm sure kiddies are doing this all the
time.) Somehow, it or a human mind needs to verify suggestions.



(2) What I'm doing is Bash scripting and Python programming. These are
the two most commonly used computer dialects in 2023. Yet it can't do
even the simplest things in either one. I'm serious. I changed my
request to list all but the 10 newest files (safer than playing with
deleting anything). Then, I wrote a script to do it myself in less
than 5 minutes:



#!/bin/bash
# Set the directory to clean up
dir="$HOME/test_dir"
# Change to the directory
cd "$dir" || exit
# List all files in the directory, sort by name, and list all
# but the 10 most recent ones
the_files="$(ls -1 | sort -r | tail -n +11)"
# List all files in the directory except the most recent 10
ls -la -- ${the_files}


Some kid in Grade 11 could write that. It tried 5 times and
got it wrong on each attempt. It can't write code. I guess it can write
poems and play games like the guy in the comments after the video got it
to do:

"The thing that blew me away was when I told ChatGPT about a "new word"
- I told it that "wibble" is defined as: a sequence of four digits that
are in neither ascending or descending order. I asked it to give me an
example of a wibble - and it did. 3524. I asked it for a sequence that
is NOT a wibble and it said 4321. Then I asked it for an anti-wibble
and no problem, 2345. Then I asked it fan example of an alpha-wibble
and it said FRDS....which is amazing. It was able to understand an
entirely new word...which is clever - but it was able to extrapolate
from it...which is far more than I thought possible."




I just asked it what wibble was and it said:

"I'm sorry, but "wibble" is not a specific term or concept that I can
provide a clear definition for. It could be a made-up word or a term
used in a specific context or field that I am not familiar with. If
you have more context or information about where you encountered the
term "wibble," I may be able to provide a more helpful answer."

Apparently it can't remember or write code.

--
HRM Resident

[James Warren]

On Saturday, 11 March 2023 at 14:49:05 UTC-4, HRM Resident wrote:
> James Warren <jwwar...@gmail.com> writes:
>
> >
> > It probably can learn from it's mistakes but it not going to take
> > your word for it. You could be trying to fuck with it. Your correction
> > would have to be vetted. Indeed, they are doing this continuously.
> > Maybe you could get to be a vettor.
> >
> > If it was trained on more good code it would do better.
> Two things:
>
> (1) Yes, I could tell it to always use "sudo rm -rf *" at the
> beginning of every script (I'm sure kiddies are doing this all the
> time.) Somehow, it or a human mind needs to verify suggestions.

That's true, but it needs to be a trusted human.

>
>
>
> (2) What I'm doing is Bash scripting and Python programming. These are
> the two most commonly used computer dialects in 2023. Yet it can't do
> even the simplest things in either one. I'm serious. I changed my
> request to list all but the 10 newest files (safer than playing with
> deleting anything). Then, I wrote a script to do it myself in less
> than 5 minutes:

It was not specifically trained on code. Other assistants were. Did you
try any of them?

ChatGPT was trained on whatever it could find on the internet that could
be vetted. It sais it can't play chess but it does indeed play a kind of chess
but makes mistakes and illegal moves but from what I have seen it plays
better now than it did a month ago.

>
>
>
> #!/bin/bash
> # Set the directory to clean up
> dir="$HOME/test_dir"
> # Change to the directory
> cd "$dir" || exit
> # List all files in the directory, sort by name, and list all
> # but the 10 most recent ones
> the_files="$(ls -1 | sort -r | tail -n +11)"
> # List all files in the directory except the most recent 10
> ls -la -- ${the_files}
>
>
> Some kid in Grade 11 could write that. It tried 5 times and
> got it wrong on each attempt. It can't write code. I guess it can write
> poems and play games like the guy in the comments after the video got it
> to do:
>
> "The thing that blew me away was when I told ChatGPT about a "new word"
> - I told it that "wibble" is defined as: a sequence of four digits that
> are in neither ascending or descending order. I asked it to give me an
> example of a wibble - and it did. 3524. I asked it for a sequence that
> is NOT a wibble and it said 4321. Then I asked it for an anti-wibble
> and no problem, 2345. Then I asked it fan example of an alpha-wibble
> and it said FRDS....which is amazing. It was able to understand an
> entirely new word...which is clever - but it was able to extrapolate
> from it...which is far more than I thought possible."

It can remember new stuff for the duration of a conversation but it
resets or forgets at the start of a new session. If you want it to remember
you need to be vetted as a trusted teacher.

>
>
>
>
> I just asked it what wibble was and it said:
>
> "I'm sorry, but "wibble" is not a specific term or concept that I can
> provide a clear definition for. It could be a made-up word or a term
> used in a specific context or field that I am not familiar with. If
> you have more context or information about where you encountered the
> term "wibble," I may be able to provide a more helpful answer."
>
> Apparently it can't remember or write code.

I found cases where it could write code, even Python code, if you
give it good instructions and guidance and explain its errors.

>
> --
> HRM Resident

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:
>
> That's true, but it needs to be a trusted human.
>

If it's really "intelligent" why is this so?

>
> It was not specifically trained on code. Other assistants were. Did you
> try any of them?
>

No, I want to use it to help write/debug code, not to argue
about philosophy. What is a good one to generate original code?

> ChatGPT was trained on whatever it could find on the internet

That implies it does the equivalent of a Google search and somehow
picks the pepper out of the fly shit. It doesn't do a good job, in my
view.

>
> It can remember new stuff for the duration of a conversation but it
> resets or forgets at the start of a new session. If you want it to remember
> you need to be vetted as a trusted teacher.
>

If this is so, how can it be called "Artificial" or "Intelligent?"
What you are saying is when it encounters something new, a human adds
another IF THEN construct.

>
> I found cases where it could write code, even Python code, if you
> give it good instructions and guidance and explain its errors.
>

It can regurgitate code previously written by human thought. It can't
create original, new code. I thought it was something new. It's just a fast
search engine for things people have done or said in the past.

--
HRM Resident

[James Warren]

On Saturday, 11 March 2023 at 16:20:52 UTC-4, HRM Resident wrote:
> James Warren <jwwar...@gmail.com> writes:
> >
> > That's true, but it needs to be a trusted human.
> >
> If it's really "intelligent" why is this so?

Because there are many people trying to fuck with it.

> >
> > It was not specifically trained on code. Other assistants were. Did you
> > try any of them?
> >
> No, I want to use it to help write/debug code, not to argue
> about philosophy. What is a good one to generate original code?

I don't know.

> > ChatGPT was trained on whatever it could find on the internet
> That implies it does the equivalent of a Google search and somehow
> picks the pepper out of the fly shit. It doesn't do a good job, in my
> view.

Actually it does a very good job. Good enough to pass some high level
exams.

> >
> > It can remember new stuff for the duration of a conversation but it
> > resets or forgets at the start of a new session. If you want it to remember
> > you need to be vetted as a trusted teacher.
> >
> If this is so, how can it be called "Artificial" or "Intelligent?"

Do you believe everything you're told?

> What you are saying is when it encounters something new, a human adds
> another IF THEN construct.

No, it adds the new information to the training set. It does not use
IF THEN constructs. That's GOFAI.

> >
> > I found cases where it could write code, even Python code, if you
> > give it good instructions and guidance and explain its errors.
> >
> It can regurgitate code previously written by human thought. It can't
> create original, new code. I thought it was something new. It's just a fast
> search engine for things people have done or said in the past.

It's more than that. It can combine words/ideas in new ways.

>
> --
> HRM Resident

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:
>
> Because there are many people trying to fuck with it.
>

>snip<
>
> I don't know.
>
>snip<

>
> Actually it does a very good job. Good enough to pass some high level
> exams.
>

>snip<

>
> Do you believe everything you're told?
>

>snip<

>
> No, it adds the new information to the training set. It does not use
> IF THEN constructs. That's GOFAI.
>

>snip<

>
> It's more than that. It can combine words/ideas in new ways.
>

Aside from cheating on exams, what is it good for?

--
HRM Resident

[James Warren]

According to many people I've read it's pretty good at writing code,
at least at the junior programmer level. It's not so good at designing
a project.

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:
>
> According to many people I've read it's pretty good at writing code,
> at least at the junior programmer level. It's not so good at designing
> a project.

You were misinformed:

"Write a Fortran 77 program to read in a text file and remove all
consecutive commas, and then write out the results to a second file."


ChatGPT gave:

PROGRAM remove_commas
CHARACTER*255 input_file, output_file
CHARACTER*255 line
INTEGER i, j, len
LOGICAL found_comma

! Get the input and output file names from the user
WRITE (*,*) 'Enter input file name: '
READ (*,*) input_file
WRITE (*,*) 'Enter output file name: '
READ (*,*) output_file

! Open the input and output files
OPEN (UNIT=10, FILE=input_file, STATUS='OLD')
OPEN (UNIT=20, FILE=output_file, STATUS='NEW')

! Read each line of the input file, remove consecutive commas, and write to the output file
DO
READ (10,*,END=100) line
found_comma = .FALSE.
len = LEN_TRIM(line)
DO i = 1, len
IF (line(i:i) .EQ. ',') THEN
IF (found_comma) THEN
! skip consecutive commas
ELSE
! found a comma
found_comma = .TRUE.
WRITE (20, '(A1)', ADVANCE='NO') line(i:i)
END IF
ELSE
! not a comma
found_comma = .FALSE.
WRITE (20, '(A1)', ADVANCE='NO') line(i:i)
END IF
END DO
WRITE (20, *) ! write end-of-line character
END DO

100 ! Close the input and output files
CLOSE (10)
CLOSE (20)

END


Which I saved as 'comma.f'

The I tried to compile it:

gfortran -o comma.exe comma.f
comma.f:40:4:

40 | 100 ! Close the input and output files
| 1
Error: Statement label without statement at (1)
comma.f:19:27:

19 | READ (10,*,END=100) line
| 1
Error: END tag label 100 at (1) not defined


That question about writing a Fortran program to remove the commas
was asked at a job interview in 1984. All four applicants got it right.

Is the question I asked a trick one, or beyond junior level
programming? This is EXACTLY what it seems to do in Bash scripting and
Python code generation.

--
HRM Resident

[James Warren]

I just found these:
https://medium.com/towards-data-science/can-chatgpt-write-better-sql-than-a-data-analyst-f079518efab2
https://medium.com/geekculture/hey-chatgpt-solve-these-coding-tasks-using-python-b2e7482f2c18

and this:
https://www.youtube.com/watch?v=pspsSn_nGzo

[HRM Resident]

HRM Resident <hrm...@gmail.com> writes:


>
> gfortran -o comma.exe comma.f
> comma.f:40:4:
>
> 40 | 100 ! Close the input and output files
> | 1
> Error: Statement label without statement at (1)
> comma.f:19:27:
>
> 19 | READ (10,*,END=100) line
> | 1
> Error: END tag label 100 at (1) not defined

I fixed the dumb error of it putting the comment in the wrong place
and it compiles fine, but doesn't work properly:

Here's the input file:

This is a text,,,,,,,file that has some, commas in it,,and not
all of them are consecutive, but some,,are and, some are not.

Here's the compile and run:

gfortran -o comma.exe comma.f
./comma.exe

Enter input file name:
an_input_file.txt
Enter output file name:
an_output.txt


And here's what it spit out:

This
all

Doesn't look right to me. So, what good is ChatGPT?

--
HRM Resident

[James Warren]

Easy fix. It does need a little handholding.

[James Warren]

So fix it.

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:
>
> So fix it.
>
Why? I asked you what good ChatGPT was and you said,"According

to many people I've read it's pretty good at writing code, at least
at the junior programmer level. It's not so good at designing
a project."

I "told it" what it was doing and it keeps giving more and more
versions that either do not compile or do not work. I gave it 6
chances.

I don't want or need a program to remove commas. I'm trying to figure
out what good ChatGPT is. You said, "According to many people . . . ",
which sounds a lot like TFG's "A lot of people are saying . . ." or "I'm
hearing . . . " prefixes to things that are wrong. Why not admit it's
no good except for writing poems or arguing philosophy.

The problem I gave it was junior level programming. It failed on
multiple attempts. That's the point, not ""According to many people
. . ." it works. It does not.

I have to take off so take my code and try to get it to fix it. I
know your can, and I could, and probably a Grade 10 kid with a PC
could, but let's see ChatGPT fix it. :-)

C'ay! I have better things to do today than play with a broken
chat-bot. :-)

--
HRM Resident

[James Warren]

1. Was this instruction clear enough "Write a Fortran 77 program to read in a text file and remove all

consecutive commas, and then write out the results to a second file."

2. Is you test file diverse enough to catch the error?

How about:
,,,A,,B,C,
A,B,,C,,,D
A,,B,C,,,

What is your desired output?

[HRM Resident]

James Warren <jwwar...@gmail.com> wrote:
>
>
> What is your desired output?

That was clearly defined in my original question. You can
think and you know the input and output.

I challenge you to get it to supply a program that does
what the original question asked. Write it a page of
instructions or 2-3 words. Just show me it can write
simple code.

I claim it cannot. To prove me wrong, all you need to do
is get it to work. That should be easy, but you cannot. I know because I
tried for 6 weeks.

What is the good of ChatGPT?

--
HRM Resident

[James Warren]

On Sunday, 12 March 2023 at 11:55:32 UTC-3, HRM Resident wrote:

My query was a little different:

jwa...@dal.ca
write a fortran 77 program to remove occurances of 2 or more consequtive commas replacing them with a single comma.
Here's a FORTRAN 77 program that removes occurrences of two or more consecutive commas and replaces them with a single comma:

vbnet
Copy code
C Program to remove occurrences of 2 or more consecutive commas in FORTRAN 77

PROGRAM REMOVE_COMMAS
CHARACTER*80 STRING
INTEGER I, LEN_STR

PRINT*, 'Enter a string:'
READ*, STRING

LEN_STR = LEN(STRING)

DO I = 1, LEN_STR-1
IF (STRING(I:I) .EQ. ',' .AND. STRING(I+1:I+1) .EQ. ',') THEN
STRING(I:I) = ' '
ENDIF
ENDDO

DO I = LEN_STR, 2, -1
IF (STRING(I:I) .EQ. ',' .AND. STRING(I-1:I-1) .EQ. ',') THEN
STRING(I:I) = ' '
ENDIF
ENDDO

WRITE(*,*) 'The modified string is:'
WRITE(*,*) ADJUSTL(STRING)

END

The only problem I had was that in my old F77 compiler
read*,string doesn't work right. I had to replace it with
read(*,fmt='(a)') string
then I removed adjustl(string) with string
and just for good measure replaces write(*,*) with write(*,fmt='(a)')

I worked but it just replaced extra commas with blanks.

[James Warren]

My query worked on the first try except for the peculiarities of my F77 compiler.

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:


> The only problem I had was that in my old F77 compiler
> read*,string doesn't work right. I had to replace it with
> read(*,fmt='(a)') string
> then I removed adjustl(string) with string
> and just for good measure replaces write(*,*) with write(*,fmt='(a)')
>
> I worked but it just replaced extra commas with blanks.

I cut and pasted you program onto my Linux system and it compiled
file with no errors. I'm using Gfortran, GNU Fortran (Debian 10.2.1-6)
10.2.1 20210110 Copyright (C) 2020 Free Software Foundation, Inc. so
likely more up to date.

Anyhow, here's the input and output:

Enter a string:
12345,8888888,,,,hhhh
The modified string is:
12345

I would have expected:

12345,8888888,hhhh

The question was to remove all consecutive commas, not truncate it
after the first one. I had also asked for file input/output, but
essentially you got it to work, but you had to modify what it
generated. So I guess I have to concede defeat if (1) it's run on an
old compiler, (b) the need for file I/O is removed and (c) the user
fixes up some of it. I will take the version you got back and ask it to
modify it to work with a 2020 version of Gfortran.

--
HRM Resident

[James Warren]

On Sunday, 12 March 2023 at 15:49:24 UTC-3, HRM Resident wrote:
> James Warren <jwwar...@gmail.com> writes:
>
>
> > The only problem I had was that in my old F77 compiler
> > read*,string doesn't work right. I had to replace it with
> > read(*,fmt='(a)') string
> > then I removed adjustl(string) with string
> > and just for good measure replaces write(*,*) with write(*,fmt='(a)')
> >
> > I worked but it just replaced extra commas with blanks.
> I cut and pasted you program onto my Linux system and it compiled
> file with no errors. I'm using Gfortran, GNU Fortran (Debian 10.2.1-6)
> 10.2.1 20210110 Copyright (C) 2020 Free Software Foundation, Inc. so
> likely more up to date.
>
> Anyhow, here's the input and output:
>
> Enter a string:
> 12345,8888888,,,,hhhh
> The modified string is:
> 12345
>
> I would have expected:
>
> 12345,8888888,hhhh

I got
12345,8888888 ,hhhh

Which is just what I asked for.

I did not ask it to compress the string.
Would I want all blanks removed or just consecutive blanks

[James Warren]

On Sunday, 12 March 2023 at 15:49:24 UTC-3, HRM Resident wrote:

That's how I understand how to use it. You iterate until you get it
right and be very specific in how you phrase your query.

[HRM Resident]

James Warren <jwwar...@gmail.com> writes:
>
> That's how I understand how to use it. You iterate until you get it
> right and be very specific in how you phrase your query.
>

I was expecting too much from it.

--
HRM Resident

[James Warren]

Found this on Morning Brew just now.

​​ChatGPT is old news. OpenAI released its much-hyped GPT-4 AI language model yesterday,
representing an advancement over the tech that powers ChatGPT. GPT-4 is wowing reviewers
with its ability to understand not only text but also images (even complex memes). Plus, it crushes
its predecessor GPT-3.5 on academic assessments: On a simulated bar exam, GPT-4 scored around
the top 10% of test takers, while GPT-3.5 scored around the bottom 10%.