ChatGPT: ASCII to decimal conversion

[BigEd]

All these systems are stochastic - they don't produce the same text, even for the same inputs. And they are text generators, not lookup machines, so you're not getting information, you're getting something which plausibly resembles information. That resemblance can be very good, but there's still a difference. Even when they provide links as if they are citations, those links may not support the statements made - they are merely plausible links. They might even be to defunct websites or websites which have since been domain-squatted.

Although these systems are getting better, the only way to use them reliably is to check everything they output. Which of course you did.

[BigDumbDinosaur]

Martin_H wrote:

I am porting code from the 6502 to 65816, and I needed a routine to print a 16-bit binary number as decimal...In Edge I type "65816 assembly to print number as decimal number" and at the top is panel with AI generated code. But the code doesn't look right. It has repeated blocks of code and do nothing blocks.

I’m a bit amazed that AI would know enough about the 65C816 and its assembly language to produce anything of note, never mind whether it would work. 

In the same vein of relying on artificial “intelligence,” there was this news story about autonomous automobiles, aka “self-crashing cars.”  The revelation that that autonomy isn’t quite what it seems does a lot to bolster my confidence in the ability of machines to make timely decisions that, if wrong, could result in large hospital or mortician bills.

Quote:

But here's the weird part. I redo the search out of curiosity, and it generates a completely different block of code. Initially it looks better, but as I read it, I realize its likewise bonkers, so I gave it the thumbs down feedback.

As Ed more or less suggests, if you redo your search ten times, you will likely get ten different chunks of code, all equally dysfunctional. 

Quote:

I'm not wasting my time trying this, but I thought I would share it. I added a few comments in caps where it's off the wall.

I didn’t dissect the AI-generated code, but it seems to be a lot of code for what is supposed to get done.

Quote:

Go home AI you're drunk.

No kidding!

[Martin_H]

@BDD, thanks for the link to the Waymo story. I've been telling my brother-in-law that they had human rescue drivers and now I have a source.

[BigDumbDinosaur]

Martin_H wrote:

@BDD, thanks for the link to the Waymo story. I've been telling my brother-in-law that they had human rescue drivers and now I have a source.

Too bad the “rescue drivers” are a 14-hour plane ride away from the locations of the Waymo cars when they go whammo into an obstacle.

BTW, you should know that my automobile also has a “rescue driver”: me! 

[barnacle]

Why does MS claim that 'AI' produces 30% of its code? Because after they spent so many billions on it, they can't afford programmers any more...

Except in very specialised applications, 'AI' is delivering the distilled and averaged intelligence of the internet. Ask yourself which part of the population posts loudest and most often... and rejoice that now this idiocy can also include its cohort's idiocy in its own training material. Eventually, the answer to any question asked of generative 'AI' will be "Eh?"

Fifteen years ago, I wrote a Master's dissertation on the subject of autocorrection for OCR'd documents of book length - 100k to 1M words. The software I wrote for it included knowledge of English spelling, frequency of letters and partial word forms, common errors, detailed noise models for particular OCR programs. It generated correction lists of words which were not in normal spelling directories; eventually it either decided that a word was correct, or offered a _list_ of alternatives in descending probability. It was very very good, and I used that software to correct a lot of scanned books. But I would never allow it to correct something on its own, as people trust 'AI' to do...

Neil ([/rant])

[BigDumbDinosaur]

barnacle wrote:

Why does MS claim that 'AI' produces 30% of its code? Because after they spent so many billions on it, they can't afford programmers any more...

That’s why they are exhorting you to install Windows 11.  They need the money to meet the payroll and hire more offshore talent.

Quote:

Except in very specialised applications, 'AI' is delivering the distilled and averaged intelligence of the internet. Ask yourself which part of the population posts loudest and most often... and rejoice that now this idiocy can also include its cohort's idiocy in its own training material. Eventually, the answer to any question asked of generative 'AI' will be "Eh?"

AI is the classic “monkeys on typewriters” scenario.  In the past, it was postulated that given enough monkeys on enough typewriters, the Immortal Bard’s complete works could be recreated.  The Internet version of monkeys-on-typewriters would have a hard time getting the first page in print.

Quote:

Fifteen years ago, I wrote a Master's dissertation on the subject of autocorrection for OCR'd documents of book length - 100k to 1M words.

I will have you know that I have read your thesis (as we Yanks would call it) and found it quite interesting.  The problems illustrated and the solutions offered truly illustrate the size of the chasm between human intellect and the capabilities of even the most powerful computers.

[BigEd]

While microsoft will have strong business reasons (as they see them) to do what they do, there's no shortage of cash: a quick search says $86 billion.

[Paganini]

AI has come a long way in the nearly 3 years since I started this thread. I've primarily been using Claude (Sonnet 4.5) to assist me with PUNIX. It does still make mistakes from time to time (I would say with somewhat lower frequency that myself), but is able to accept instruction and make corrections. It also will correct my mistakes in sensible ways. Basically, it's like working with a collaborator. It's not an infallible coding god, but it also no longer just emits nonsense. The code it generates typically will assemble and run with expected results, and the mistakes it makes are semantically full - for example, it would like there to be an (indirect,x),y addressing mode (don't we all), and will sometimes wishfully employ it. When this is pointed out by me Claude will then acknowledge the correction, and output the appropriate fix (which, for my architecture, means using a zero page "pseudoregister" as an intermediate storage target for the indirection).

[Martin_H]

Interesting video on the failure rate of LLMs in real world usage:
https://youtu.be/z3kaLM8Oj4o?si=8n_iTOBFc_7_QO3r

[Dr Jefyll]

Paganini wrote:

it would like there to be an (indirect,x),y addressing mode (don't we all)

Paganini wrote:

[...] the appropriate fix (which, for my architecture, means using a zero page "pseudoregister" as an intermediate storage target for the indirection).

Back in 1988 I dreamed up a lot of new instructions and addressing modes for the 65C02, and -- yes -- (ind,x),y mode is one of them. As with Paganini's fix, it uses a zero page pseudoregister as intermediate storage. (The KK Computer is real, working hardware, built before the advent of CPLDs and FPGAs.)

Usage is pretty simple. Imagine that X selects a z-pg pair which points to a multi-byte array. The 65C02 can very easily fetch the 0th byte of the array by using...

Code: Select all

LDA   (0,X)

... but then there's a speed bump (ie, having to do a 16-bit inc of the z-pg pair) before the 1th byte of the array can be fetched. Instead, here's the fix.

KK recognizes undefined opcode $CB as an alias, which gets translated to $A1 -- LDA (ind,X) -- before the 'C02 sees it. This means $CB is equally apt at speedily fetching the 0th byte of the array. But! $CB also secretly snoops the bus as the LDA (ind,X) executes; and, during cycles 4 and 5 when the z-pg pair is fetched, it copies the pair into a 16-bit register (a pair of 'HC574s) known as W.

Because W is readable at a fixed location in z-pg, we don't need X to access it. Thus it becomes fast and easy to use LDA (W),Y to access the remaining bytes in the array. In one very commonly used snippet (in which the array is simply a 16-bit word), performance almost doubles! (an 89% speedup)

-- Jeff

[barnacle]

I have to say that my experiments in FAT32 are leading me to wonder about a 32-bit 6502. But I fear it would have so many extra instructions: I find it annoying, for example, that the inc/dec pair weren't included in the (zp) instruction. I'd also like to be able to use x and y as sources and destinations for the ALU operations. But then, we could add perhaps sixteen 32-bit register pairs and we've reinvented the ARM. Hmmm.

Neil

[BigDumbDinosaur]

barnacle wrote:

I have to say that my experiments in FAT32 are leading me to wonder about a 32-bit 6502...

Yes, and Harland and Wolff’s senior management probably wondered about whether using a higher grade of steel for hull plates and rivets would have benefited Titanic.

You are in a somewhat similar boat (I just had to get that in ).  You’ve invested a lot of time in developing a FAT32 filesystem driver running on the 65C02 and are seeing some fruition from your work.  Things are getting harder and the 65C02, while capable of doing the job, could benefit from 32-bit registers and better arithmetic.  Such a critter doesn’t exist.  So you work with what you have.

While the 65C02 has excellent throughput with handling load/store instructions and register operations, it’s mathematically weak.  Aside from efficient disk accesses, the filesystem driver needs a strong arithmetic library.  It’s the nature of the filesystem beast, something that came clear to me years ago when I was working with the Xetec Lt. Kernal.  Filesystem stuff is not easy, even with a 32-bit MPU with built-in multiply and divide instructions.

I say look at what you’ve accomplished to date and keep pushing forward.  Just watch out for icebergs.

[barnacle]

Encouragement is always appreciated. I feel at this point, I have all the major algorithms in place, mostly working (but not f_alloc(), which is sulking). It's a question of sorting this in minimal assembly; it's about 4k at the moment, though with a lot of unnecessary debugging code in place.

I feel I could - with the knowledge I now have - probably write a C program to handle the major arcana in a week. Assembly, a little longer perhaps. C is, of course, assembly language with structures and 32-bit arithmetic.

I'm writing up a small screed to summarise the thread as I go along. It's currently around 20 pages...

Neil

[BigEd]

I think it's a good adventure. We all know (surely) that any reasonable CPU can do anything that any other reasonable CPU can do, given memory and time. So FAT32 on the '02 is surely going to be feasible, it's just that it's somewhat tricky. It might be that some of the trickiness would be just as challenging on another CPU with more and wider registers. It might be that suitable macros would hide some of the details of doing 32 bits on an 8 bit micro. But we've been doing that ever since we ran a sensible Basic on our 6502 computers - most of us never even thought to consider the lack of an FPU, because an 8 bit micro can do anything. Indeed, that's part of the joy!

[BigEd]

I do hope it's been an enjoyable adventure, even if challenging.