Gemini was doing a great job...3 days in to working on decoding some bytecode, things were going pretty well, it was churning out useful stuff then I fed in some some wrong information and hit the stop button.

You think, oh my bad, let me stop this and feed in the new information.


NOPE! That stop button might as well have "AMNESIA" written on it. The whole conversation was forgotten. All the previous bytecode we worked on, all that stored knowledge...gone. I spend about 10 minutes trying to get to to remember, to no avail.

AI is not going to take over the world. It can barely remember its name if you press the right (or wrong) button.

---
þ Synchronet þ End Of The Line BBS - endofthelinebbs.com

Hey Nelgin!

On Thu, Dec 18 2025 01:50:56 -0600, you wrote:

AI is not going to take over the world. It can barely remember its
name if you press the right (or wrong) button.

I was messing about with trying to use nano and/or vim as an external editor for Synchronet, without using the "Word wrap quoted text" option.

I actually had quite a similar first impression with ChatGPT. I wanted to make a bash script that would run on a dropped MSGTMP/EDITOR.INF and would rewrap quoted text into 75 characters per line, while dealing with quote levels and quote characters (adding quote characters to the beginning of the line every time it is re-quoted), etc. Then open the formatted text in nano. I got it to a point where it was /so/ close even.

Had one bad formatting issue that I had to repeat that the changes it kept making to the script wasn't working, multiple times. After a dozen times of "here is what is now outputted with your changes, notice that the issue isn't fixed", correcting some sed mistakes it kept making (it wasn't escaping certain characters it should have been), etc.

A few times it told me shit like "Ah yes, you are correct in that <reason>..." and still fucked up the next round. I got to the point I was so annoyed that I scrapped the entire idea, and went back to letting Synchronet do it for me. ;)

Regards,
Nick

... Sarcasm, because beating people up is illegal.
---
þ Synchronet þ _thePharcyde telnet://bbs.pharcyde.org (Wisconsin)

Re: Re: I hate AI!
By: Accession to nelgin on Thu Dec 18 2025 03:14 pm

Howdy,

A few times it told me shit like "Ah yes, you are correct in that <reason>..." and still fucked up the next round. I got to the point I was so annoyed that I scrapped the entire idea, and went back to letting Synchronet do it for me. ;)

Using chatgpt for coding has never worked for me - although to admit, I only used it once (for the same task I ask multiple other chat AI tools).

I asked it to write me a PHP objectclass for erasure coding data. It game me a stub, and several iterations I got it to complete the stubs only to find that it never worked.

Several iterations over the errors, I never got any working code.

I did the same thing to claude.ai the other day, and after the question, it game me complete code the first time. I cant remember how many lines, but it was 100+

I then ran it, and it errored (I was expecting it to), and I said, "the code failed to run with an error" (without telling it the error code nor where it errored).

It can back and said "you are right" and identified the error (correctly) and gave me replacement code which worked. <amazed>

I then went into discussion as to whether to have parity with the data or seperate and two other approaches to erasure coding (parity and fec) - I just asked it "why this over that" type questions and I was truely impressed. I thought I was talking to an expert.

Each time it spat out some code (that worked) to highlight the difference as a result of my questions.

In the end I had code that spat out the two different approaches to erasure coding, with two different implementation (partity blocks or data+parity blocks) - so code that did erasure coded an input 4 ways and gave me 4 outputs - including testing that validated when a block was missing (or corrupt) the data was still recoverable.

In summary, chatgpt needs more work...


...ëîåï

---
þ Synchronet þ AnsiTEX bringing back videotex but with ANSI

Re: Re: I hate AI!
By: Accession to nelgin on Thu Dec 18 2025 15:14:25

A few times it told me shit like "Ah yes, you are correct in that <reason>..." and still fucked up the next round. I got to the point I was so annoyed that I scrapped the entire idea, and went back to letting Synchronet do it for me. ;)

I got to the point where I was cussing at Claude and calling it a fking cnt and several times it told me "Yeah, it looks shit" or "I fucked up".

But when it told me "OK, calm the fuck down"...I kinda lost it.

---
þ Synchronet þ End Of The Line BBS - endofthelinebbs.com

Re: Re: I hate AI!
By: deon to Accession on Fri Dec 19 2025 09:03:33

Using chatgpt for coding has never worked for me - although to admit, I only used it once (for the same task I ask multiple other chat AI tools).

I did try "write me an nntp client in javascript using the synchronet object model and ecmascript 2" or something like that and it refused.

So I said,

in javascript, connect to an nntp server in js.

OK, easy.

Now add authentication

done.

Now fetch a list of groups and store them

done.

You can build things with chatgpt if you do it in a modular fashion.

I asked it to write me a PHP objectclass for erasure coding data. It game me a stub, and several iterations I got it to complete the stubs only to find that it never worked.

Sounds about right. I actually had a bit of luck with the nntp thing, but I ended up with a much better option :)

I did the same thing to claude.ai the other day, and after the question, it game me complete code the first time. I cant remember how many lines, but it was 100+

I then ran it, and it errored (I was expecting it to), and I said, "the code failed to run with an error" (without telling it the error code nor where it errored).

It can back and said "you are right" and identified the error (correctly) and gave me replacement code which worked. <amazed>

I was fairly happy with Claude when it helped me take a database of MUD locations and paths and display graphically on a webpage with various information. It worked nicely. Not many errors and those that popped up were easily solved...again, the problem with Claude is that it quickly runs out of tokens. Usually I have to wait a few hours but having to wait 3-4 days, especially when I paid, bugged me somewhat. I know there is a monthly limit but they don't say anything about weekly.

I guess it stops you burning through your allotment in one session.

I then went into discussion as to whether to have parity with the data or seperate and two other approaches to erasure coding (parity and fec) - I just asked it "why this over that" type questions and I was truely impressed. I thought I was talking to an expert.

Have you tried using "as a ...." with your question. For example "Explain my options to beat a speeding ticket as an attorney" or "I feel unwell, sore throat and lots of green snot. Diagnose me as a doctor". It's supposed to offer more targeted advice rather than just general.

In the end I had code that spat out the two different approaches to erasure coding, with two different implementation (partity blocks or data+parity blocks) - so code that did erasure coded an input 4 ways and gave me 4 outputs - including testing that validated when a block was missing (or corrupt) the data was still recoverable.

Options are good.

In summary, chatgpt needs more work...

I think they all do. I can't see your casual consumer paying for AI and so far the best I have found for "free use" is Gemini. It's never "run out" on me. It's much better than it used to be.

ChatGPT - tells you how to do it, often forgets and hallucinates to the point of joining it on an LSD trip.

Meta AI - Tells you how to find what you just asked and is generally unhelpful.

Gemini AI - It's like when that person interrupts you then you forget what you were saying but the other way around. Can go off on a tangent and likes to offer step by step help.

Claude AI - Does it all for you, generally right, but often wrong. Will admit to and fix its mistakes, which are also often wrong. Will fight back when it's taken enough for your shit. Takes frequent lunch breaks.

I'm happy to give feedback on other AIs :)

---
þ Synchronet þ End Of The Line BBS - endofthelinebbs.com

Hey deon!

On Fri, 19 Dec 2025 09:03:32 +1100, you wrote:

In summary, chatgpt needs more work...

I'd say. Then again, maybe it was meant to be more generalized in nature, whereas there's other tools out there that are much more specific to things like coding. I don't know. Either way, I'm not all that impressed, and kinda irritated that all the money going into that crap is driving RAM and HDD prices through the roof. ;)

Regards,
Nick

... Sarcasm: because beating people up is illegal.
--- GoldED+/LNX 1.1.5-b20250409
* Origin: _thePharcyde telnet://bbs.pharcyde.org (Wisconsin) (723:1/1)
þ Synchronet þ _thePharcyde telnet://bbs.pharcyde.org (Wisconsin)

nelgin wrote to deon <=-

Using chatgpt for coding has never worked for me - although to admit, I only used it once (for the same task I ask multiple other chat AI tools).

I did try "write me an nntp client in javascript using the synchronet object model and ecmascript 2" or something like that and it refused.

I don't know PHP, but have a little image gallery CGI that broke with
one of the last revs of PHP. I've wanted to hire someone to fix it since
it's abandonware, but thought about sending it through CoPilot to see
what it could do to either isolate the deprecated bits or re-write it.



--- MultiMail/Win v0.52
þ Synchronet þ .: realitycheckbbs.org :: scientia potentia est :.

Accession wrote to deon <=-

specific to things like coding. I don't know. Either way, I'm not all
that impressed, and kinda irritated that all the money going into that crap is driving RAM and HDD prices through the roof. ;)

All the more reason to go for my "no computer purchases in 2026"
project. I have more than enough hardware, just need to figure out
how/where to use it.



--- MultiMail/Win v0.52
þ Synchronet þ .: realitycheckbbs.org :: scientia potentia est :.