REV 3 - REV 5 - The beginning (ST536)

[ijor]

Hi @exxos ,

I apologize.

Yes, I admit I was sometimes bashing AI too harsh. It might seem like I think that AI is completely useless, always wrong, and you should never use it. Not at all. AI is indeed extremely powerful. And I think it is reaching a point that it is becoming unwise not to use it, or even not to have a paid subscription. I do use AI quite a lot lately.

However, having some technical expertise in the issues covered, I thought it was important to note when AI was wrong. Furthermore, I though (and I still think) it was important to note that some of the AI analysis wasn’t just wrong, it indeed didn’t make any sense whatsoever.

AI likes to make elaborated analysis detailing all the steps and all the reasoning to reach the conclusion. AI doesn’t say, I think, I’m not sure but …, it seems …, I heard that …, it might be…. When you read that kind of AI analysis, you might think that it was a scientific paper, as rigorous as a peer reviewed mathematical theorem proof. So yes, in that context it is important to qualify not only the conclusions, but how wrong is the whole analysis.

Yes, I probably should have remarked the positive side as well, and say, congratulate the AI when it gets things right. It is certainly not always wrong. But I though that being AI, it wouldn’t take it personally and it wouldn’t mind.

So I certainly recommend to use AI. Just use it with care. Don’t take every AI claim as the absolute truth. Everything must be tested or confirmed with your own analysis or with an authoritative source. AI can even be misleading. AI sometimes includes reference sources in the analysis, at least Google AI does, and you would think that it can’t be wrong because they are not just AI own conclusions. But if you check the links you find that the links just cover the same topic, but they are not source for the data or the conclusions reached by AI. Sometimes the links even are not really relevant.

Yes, it is starting to be better to use AI to search instead of just Google. It is becoming faster to ask AI instead of look up the manuals. Just be careful and confirm when it is important to be sure.

Yes. I think that Claude is extremely powerful for coding. It can find bugs, it can fix bugs, it can even code. That’s a different aspect than a theoretical analysis. And it is usually easier to confirm if AI is right or not. Sometimes you can say, but of course that I had a bug there, how I couldn’t see it.

[ijor]

Yes, I probably should have remarked the positive side as well, and say, congratulate the AI when it gets things right. It is certainly not always wrong. But I though that being AI, it wouldn’t take it personally and it wouldn’t mind.

Btw, I was curious and I asked Claude what's the typical attitude to the AI. The reply is that most people are very polite, and use terms like thanks and please. And there are even several studies in the subject about why people do that. Of course that they all know that it's a machine.

I really don't see why you keep focusing on the negatives all the time like your on some sort of personal crusade to bash AI and myself for using it.

Please don't take it personally. It certainly wasn't my intentions. Again, I think there is nothing wrong at all in using AI, and as I said, I'm using it myself. I just wanted to emphasize that the AI analysis and conclusions should be taken with a grain of salt.

You always seem to gloss over the good points as well.

That's true, and it might have been my fault. But I mostly focus on technical details about my areas of expertise. I can't comment too much about some STOS bug fix because I don't know nothing about STOS. I also don't have much choice but to comment on what you post. You posted a couple of quotes from your AI chat, and the analysis there was, IMHO, very wrong. I realize that at some point later, together with the AI, and thanks to the AI, you solved a problem. But at that point I can't comment too much, what would I say? Good job?

Again, I do use AI myself, and I will probably use it more in the future. And yes, it was you probably the first that showed me how powerful can AI be. If you remember, your AI found a bug in the polarity of one signal of my Verilog code some time ago. That was for testing the FPGA adapter boards you made. I remember I was amazed at the time.

My point is though, that you have to challenge everything that AI claims. I guess it happened to you many times that you challenged a conclusion, and the AI replied something like, ah, yes, you are correct, my mistake. Or something similar. That was my main point, and to elaborate on the technical mistakes. That's all.

Now, I apologize once again. I didn't consider you would take it personally. My fault. And please feel free to move these AI posts to some other thread, if you think it would be more appropriate.

[exxos]

@ijor

No worries, I don't take it personally. I don't treat anything any AI says as gospel either . In fact I spent most of this week challenging multiple models (including Claude, which got a lot backwards and wasted time). The better model I've been using lately has been fixing real bugs and getting the job done, and that's what matters to me. It's rather expensive, so not sure I will keep it much longer anyway.

[Badwolf]

I do think it would be better not to post full AI 'analyses' of problems here as that will reinforce any nonsense included in the purported reasoning as both humans and AI return to the subject in the future and read it back.

"AI suggested *this*, so I did *this* and *this* and it seems to have worked." is probably more useful than a four-page screed which hasn't been proofread and may be 70% hallucination after the initial (possibly) good suggestion.

BW

[Darklord]

Hmm, okay...I thought I'd be a boy scout and try to help out with the STOS issue.

I downloaded the STOS v2.08 archive + the v2.09 file.

Put them in a folder labeled STOS on my STacy. I can start either one and the
initial STOS screen is shown, but it's locked there. I have to reboot to get out
of it. This is a 68030 @40mhz so I thought I'd be able to test it.

Just for the fun of it, I booted up in compatibility mode (TOS v1.04, 68000 @8mhz,
etc) and tried it there. There, it does allow me to get to the menu but a few
seconds after that, it locks up, leaving the mouse pointer frozen and requiring
a reboot.

Is that perhaps because it's STe only or something like that?

Sorry, I tried... :roll:

[exxos]

Put them in a folder labeled STOS on my STacy. I can start either one and the
initial STOS screen is shown, but it's locked there. I have to reboot to get out
of it. This is a 68030 @40mhz so I thought I'd be able to test it.
...
Is that perhaps because it's STe only or something like that?

I tested it on my H5 which is a ST with the ST536 030 running at 50mhz..

I'd suggest removing the extensions from the STOS folder.
STE_EXTN.EXF
COMPACT.EXA
COMPILER.EXC

Delete the compiler folder and AUTOEXEC.BAS so your just left with a raw STOS folder and the BASIC209.PRG. Also note BASIC209.PRG should be on the root of the floppy and STOS folder in root also.

I just tried in STEEM and was fine. TOS206.

1.PNG

Also in Hatari. TOS404.

2.PNG

[ijor]

The better model I've been using lately has been fixing real bugs and getting the job done, and that's what matters to me. It's rather expensive, so not sure I will keep it much longer anyway.

Don't be selfish and share your experience. :) So what is that expensive model that is working so good?

I do think it would be better not to post full AI 'analyses' of problems here as that will reinforce any nonsense included in the purported reasoning as both humans and AI return to the subject in the future and read it back.

That's one of the most important reasons I sometimes feel that I, even have, to reply and criticize these AI posts. For the record. At least on topics that I consider that I can provide a quite authoritative answer.

Of course, future AI, or future humans, might completely ignore my replies, considering better to ignore "that" disrespectful human :)

[Darklord]

Okay, I did everything you said except that I did run everything from the root
of my C: drive (STOS folder there, with the .PRG).

It runs, shows the splash screen, then the black screen just like the one you
posted. However, if I do anything, it gives a bus error and then locks up.

This is with the 68030 or in compatibility mode (TOS v1.04/8mhz 68000).

I think my highly modified STacy just doesn't like STOS. :roll:

The good news though, is that doing <shift> A works. :lol:

STOS Error01.JPG

[exxos]

I think my highly modified STacy just doesn't like STOS. :roll:

What is it with your STacy :lol:
Does the original 208 do the same?

The good news though, is that doing <shift> A works. :lol:

Nice, so at least those fixes worked :excited:

[exxos]

Don't be selfish and share your experience. :) So what is that expensive model that is working so good?

Its Claude opus. Costing £90 a month :( Claude sonnet is as bad as other AI. Waste of time. I just end up arguing with them for hours then giving up. Opus isn't perfect. It takes corrections. But it's amazing how it can write code to diagnose problems. It wrote programs to figure out why STOS keyboard tables were getting corrupted. Then came up with solutions. It tried to write STOS code to help. But I just skipped all that as the code wasn't going to work. But the fact it can write 68k assembly code its fantastic. It makes simple mistakes.. But it finds them and corrects them. Then all good. Other AI claim they found the problem, claim they fixed it, and just make things worse or not change the code at all.

All AI models have the same flaws. They suggest things and because there's nothing to disagree, it basically takes it as fact.
They all suggested technical articles to "prove" the point. But often broken links or hard to find papers. The ones I do find don't really tally with the proof either. Grok I asked several times about proof, gets mad if I say it's making things up. But basically side steps directly answering the question of proof.

So 99% of the time I can pickup they taking crap. That's the problem with every AI I've tried. It's expected i have to correct them. But only opus seems to learn from its mistakes. Generally 90% of the time it's fine. It offers suggestions which is great, but I can rule most out as being valid. It's not that AI is wrong, just the current conversation based on what its told might not be enough to correctly diagnose the issue.

The problem then becomes, the conversation can span 50+ pages and multiple things are discussed. Over time it simply gets confused and mixes up 2 or 3 things which then doesn't make much sense without the entire 50 page context of it all. But even so, things obviously slip though. In particular late at night when my brain gets to frazzled after all day session of debugging stuff.

Even if AI is wrong, things can be learned. AI can come out with ideas, a lot I would never have even though about by myself. As a idea generator, it's a very powerful tool. Some ideas blew my mind, others were just dumb. So I don't really take anything AI says as fact. But it can point you in directions and give ideas which at least drives progress. Its the ideas which interest me, and people shouldn't take stuff even what I post as fact. But yes. Probably best I don't post AI responses directly, as I don't have time to clarify it all. The context gets lost which doesn't help matters either.