BBC BASIC Program Generated by ChatGPT

It looks correct but fails. The following program is from ChatGPT which searches the decimal digits of PI to find the “999999” sequence. Mathematicians amongst you will know it is the Feynman point. Not being a mathmatician I’m stuggling to understand the so called spigot algorithm in lines 90-190. Can anyone explain in simple terms and if you can make the program work that would be a bonus. I’ve found quite a few examples that fail – even mixing BBC BASIC with keywords that don’t exist.

10 REM Feynman Point Finder in BBC BASIC
20 REM Spigot Algorithm for Digits of Pi
30 DIM a(1000), pi$(10000)
40 N = 1000
50 FOR i = 0 TO N
60 a(i) = 2
70 NEXT
80
90 FOR i = 1 TO N
100 REM Perform the division step
110 k = 0
120 FOR j = N TO 0 STEP -1
130 k = k * j + 10000 * a(j)
140 a(j) = k MOD (2 * j + 1)
150 k = k / (2 * j + 1)
160 NEXT
170 REM Extract the digit
180 pi$(i) = MID$(STR$(k / 10000), 3, 1)
190 NEXT
200
210 REM Search for Feynman Point
220 FOR i = 1 TO LEN – 5
230 IF MID$(pi$, i, 6) = “999999” THEN
240 PRINT "Feynman Point starts at position "; i
250 END
260 ENDIF
270 NEXT
280
290 PRINT "Feynman Point not found in first "; N; " digits of PI."
300 END

It might help if you posted that as a code block, so that textile doesn’t “improve” the contents…

prefix the first line with bc..<space> like this

bc.. 10 REM Feynman Point Finder in BBC BASIC

to create a multi-line code block.

If you wanted to follow on with plain text again, you could prefix the first non-code line with p.<space>.

Does that code even have something that calculates pi?

Change it to simply print pi. I’d be surprised if it was better than 22/7.

It’s very rough, and I’ve not debugged it, but going by this I suspect this is what the code is trying to do.

digits% = 4
columns% = 10 * digits% / 3

PRINT "Calculating ";digits%;" digits, using ";columns%;" columns..."

DIM a%(columns% - 1)
a%() = 2

FOR digit% = 1 TO digits%
  a%() = a%() * 10

  carry% = 0

  FOR column% = columns% - 1 TO 0 STEP -1
    a%(column%) += carry%

    REM A special case for column zero, as we need to take the 10s as the next
    REM digit of Pi, and send on the units.
    IF column% > 0 THEN
      column_numerator% = column%
      column_denominator% = 2 * column% + 1
    ELSE
      column_numerator% = 1
      column_denominator% = 10
    ENDIF

    PRINT "Column ";column%," denominator is ";column_denominator%," value is ";a%(column%);

    carry% = (a%(column%) DIV column_denominator%) * column_numerator%
    PRINT ," carrying ";carry%;

    a%(column%) = a%(column%) MOD column_denominator%
    PRINT ," saving back ";a%(column%)
  NEXT column%

  PRINT "Extracted digit ";digit%;" of Pi: ";carry%
NEXT digit%

We know AI is shit, so please don’t post any of it’s output here.

We know AI is shit, so please don’t post any of it’s output here.

Hopefully that wasn’t aimed at the code which preceded it…

It’s an interesting demonstration of just why AI isn’t there quite yet, especially for languages like BBC BASIC. Richard’s code is almost like BASIC written by someone who thinks it’s just another dialect of Python, or something. The Feynman part appears to “functional” aside from completely ignoring the fact that strings are limited to 255 characters, for instance, and apparently trying some kind of slicing or array indexing on them – very C# or Pythonesque.

On the other hand, the Spigot algorithm contains

a(j) = k MOD (2 * j + 1)

which is close to what I wrote, but then

k = k / (2 * j + 1)

which is clearly no substitute for

carry% = (a%(column%) DIV (2 * column% + 1)) * column%

if only because it hasn’t realised that reals aren’t integers…

It’s a statistical word guesser, it will never understand what is correct code and what isn’t. Let’s keep this platform free of it’s pollution.

I will point you towards https://heyrick.eu/blog/index.php?diary=20230920

It seems to me that tools such as these are useful for explaining concepts when you don’t quite grok an idea; but there’s a vast gulf between concepts and functional code.

I thought I’d written up about the mess it made when asked to make a raycaster in BASIC (wrong keywords, code that was gibberish to the point I gave up trying to understand it) but I guess I didn’t… Maybe I’ll do it again soon and make an article out of it.

I’m too lazy to look closely at what the generated code does. But I think I recognized a few small errors when I took a quick look at it.
I quickly adapted my code (also Spigot) and fiddled with “Feynman”. Not pretty but it works for me.

Download

Note: It will be very slow with bigger digits of Pi. Any days for 10^6 ;-)

I didn’t think it would generate so much interest but thanks for all your replies. The other thing (I really shouldn’t be surprised) is how dogmatic and uncompromising (I love this word) some people are. Thanks to Steve and Raik who put effort into providing a solution and again to Steve for the tip about posting inside code block. I’ll have a look at your examples – the spigot algorithm is not simple to understand but maybe going through your ideas might help. Don’t worry I won’t be posting anything from ChatGPT again as I find it ends up rendering code that bamboozles me even more – was it hallucinating I wonder?

how dogmatic and uncompromising (I love this word) some people are

I think some of us see very very strong parallels with the dot com boom all those years ago.

Honestly I don’t care either way, but I’d like to see three things:

1, An actual viable use case for LLMs, and not everybody falling over themselves to shove it in to everything they can because they can. They are interesting toys, but I sure as hell wouldn’t trust one to review a police bodycam and write a report summarising what happened (yes, really…).

2, Proper attention to data hygiene around the use of LLMs. If one, for example, is helping you get fitter by creating custom routines according to your time and effort, is this information being shared with any third parties? If so, who and why? (and I’m counting so-called anonymised information here)

3, A proper resolution to the question of why ten seconds of a song in the background of a video of a copyright strike, yet these AI companies can grab a copy of everything in sight, mash it up with everything else, and make loads of money claiming it’s something shiny and new rather than a regurgitation of everything that they, essentially, stole (if one is to apply as strict an interpretation of copy/moral rights as the content companies wish to apply to us).

I’m still working my way through your examples. Raik you’ll be pleased to know that both your examples work really well with the second example showing precisely how many decimal digits must be counted before the “999999” sequence starts which mathematicians will know as 762. Both examples seem reasonably fast on a R-Pi4 machine with processing 10,000 iterations taking approx. 2min. I think 10^6 would be pushing it! I wonder what other interesting sequences exist in Pi? I did hear of a 12345678 sequence (might have a 9 in it but can’t remember)

3 good points Rick – difficult to argue with them!

The other thing (I really shouldn’t be surprised) is how dogmatic and uncompromising (I love this word) some people are.

Dogmatic and uncompromising ensures:-

• Airbus aircraft don’t fall out of the sky
• Interbank gateways don’t lose your money
• British fighter jets are maintained correctly
• Various missile systems hit what they are supposed to
• Formula One, Indycar, Nascar and FE cars don’t stop working during a race
• Millions of video meetings work without a glitch (not Teams obviously)
• Payment terminals do what they are supposed to do
• Filecore disks can be repaired

AI will never be involved in the writing of high integrity software as long as there is breath in my body.

That’s fine as a personal opinion but any RISCOS related subject should be up for discussion on this forum; and for some people, that includes AI.

Formula One, Indycar, Nascar and FE cars don’t stop working during a race

Carbon fibre shell – meet Alex; Alex Albon, licensed to destroy. :)

I did hear of a 12345678 sequence

On position 17.387.594.880 should be 0123456789 but I will not try it ;-)
You can try to find 999999 also on 193.034 (should work) or 888888 on 222.299 and 000000 on 1.699.927 ;-)

You can try a C version without Feynman-search. Is much faster.

Download

In case it interests anybody, a long meandering article that skewers “AI”. Includes dodgy C code, dodgy BASIC code, dodgy attempts at making humans, and a summary of me that is so off-the-wall it has to be seen to be believed. And a bowl of linguine¹ which is what I’m about to go and cook…

Have fun! https://heyrick.eu/blog/index.php?diary=20241214

¹ Makes sense in context, but my regular readers will be used to this sort of thing.

I tried getting version 3 to generate BBC BASIC and the result was abysmal, much worse than above.

It occurred to me at the time its highly likely that there would not be much training material about given that its generally a tokenised language, and the bulk of the published source was well before the internet.

I was also amused that it thought the alias “THE RAM and BUFFER” for you local boozer was from Douglas Adams and not The Micro User at all.

What do you expect from an “artificial intelligence” that was created by a “natural intelligence” that proves its lack of intelligence every day?

I have upload a small app from my Pistuff.
Calc if you need (the first Million digits are ready inside but you can play with) and search feynman or what ever…

Download

the spigot algorithm is not simple to understand

My App is based on this (? – found the author in my remarks but the code is older) publication and the first I programmed in MSX-Basic.