AI lad, that be reet

So I just coaxed Google Bard into compiling some BBC Basic into ARM assembly despite it knowing nothing about the former and not enough about the latter.

The most interesting part of the exchange was its claim that it had learned from its mistakes and would remember. But it may be lying about that.

However, having tried to return using MOV PC,LR after calling a stub with BL it was enough to remind it that BL corrupts LR for it to generate the right code. It’s going to be an interesting few months, these last months of human software engineers, dedicated compilers, and software licences.

[Yes, it’ll happily read code in one language and rewrite it in another. Spot that in the final build, I dare you.]

But is it smart enough to translate arm 32 into arm64? Or better still assembly into C

Ups slip of the finger

These things are pattern recognition engines. Spotting assembly idioms and translating to a conceptual representation would be well within its capability if trained to do so.

It’s about time. I was advised not to bother taking a computer course at school because they’ll be programming themselves in no time. That would have been in 1984.

But is it smart enough to translate arm 32 into arm64?

For small routines, yes, it hsould be able to translate arm 32 to arm 64 (in most cases correctly, but NOT all).

Or better still assembly into C

If you’re refering to use it to translate RO source to C, then no. This is because, while it would be capable of translating some of RO into C, then you still need to add an “ASM” firendly API if you wish the C side to stay compatible with everything available for RO, plus, given the first mentioned limitation (work for small routines or programs), then the translation of the entire kernel would be “de-contestualized” which means you’ll need to take the C source it has generated and “re-contextualize” it.

Finally, the way pattern matching works for AI means it will most likely use user-space code for the C (and also the assembly translation btw) which means you’ll probably get code with printf, mallocs etc. which isn’t suitable for kernel programming, so you’ll need to clean up manually or write extremely complicated prompts to make the AI select more suitable datasets.

You could use it however to help 32bitting RISC OS Apps and user space activities like that, if you train it specifically for RO (as mentioned by Nemo).

Hope this helps

By the way, on the subject of programmers becoming obsolete, don’t you have to have a working program for the bot to translate?

By the way, on the subject of programmers becoming obsolete, don’t you have to have a working program for the bot to translate?

Yes, and even when you have a working program the “bot” as you call it, can still make mistakes and still has limitations if you use one of the services like ChatGPT or Bard etc.

Generative AI that actually work needs to be (still) heavly trained, so it’s hard to have a “general generative” AI that can translate code in such a way to make it useful in all situations. But it will come, the issue of the large datasets is beign addressed and there are dedicated (non general purpose) genertive AI that have showen great results.

Bard and ChatGPT are more to capture the attention of the large public. Besides, GPT-4 is already way better than GPT-3.5 on which ChatGPT was originally based, so it will improve a lot when it will be released under GPT-4, and the next release (GPT-5) seems to be even more impressive.

So yes, at some point it will work, how soon it will depend by the usual factors: software progresses, training improvements, hardwave progresses to be able to generate and infere more complex models.

Simon suggested

they’ll be programming themselves in no time. That would have been in 1984.

LOL

Paulo presumed

If you’re referring to use it to translate RO source to C, then no.

Not ‘out of the box’, but with suitable prompting it does (Bard does, anyway).

Don’t forget you add to its understanding (allegedly ‘permanently’, but certainly contextually) with everything you tell it. So one can instruct it to use a different API for a particular conceptual function.

I find the limitation of these machines to be the number of ‘ideas’ they can keep in their working set simultaneously – given too many constraints, some of them get forgotten and revert to the learned form. But this is simply a scaling artefact, not an inherent brick wall.

the “bot” as you call it, can still make mistakes

And by far the greatest weakness of all these automatons is that they have no concept of ‘certainty’ – they just can’t tell the difference between what they know and what they’ve just imagined. This is a significant failing and will need to be addressed in the next generation… which will be here in a few weeks at the current rate of progress.

And by far the greatest weakness of all these automatons is that they have no concept of ‘certainty’ – they just can’t tell the difference between what they know and what they’ve just imagined. This is a significant failing and will need to be addressed in the next generation…

and when they are certain we are a nuisance?

<whispers>It’s still saying “our”.</>

It’s still saying “our”.</>

Perfect, plagiarising tinker-toy machines with an inflated ego.

Also, gotta ask: "What do you mean We you animated kettle? "
Real humans don’t need pretend-thinkers to screw things up, we can do it on our own mate.

I for one welcome our word-association overlords.

The overriding impression is not of ego but of fawning, corporate, puff-piece, smoke-blowing. Everyone you ask about is “notable” and “influential”. Every work is “important” and “famed”.

Very little chance of one of ’em coining the late great Clive James’s description of Marilyn Monroe: “She was good at playing abstract confusion in the same way that a midget is good at being short”.

I’ve noticed that Windows calls itself “we”, which has always struck me as rather strange. When I get a message like “We’re installing updates” it conjures up an image of Microsoft putting its tendrils into the computer via the Internet and replacing files without my consent… which is probably fairly accurate!

May be the “Kindergarten We” more then the “Royal We”…

Both equally wet and smelly.

We’re already being pushed devices that need internet connection and won’t fully work without being online. Will AI in every device be the next compulsory ‘feature’?

I’m sorry, I’m having trouble reaching the internet right now. Please check your router.

We’re already being pushed devices

I’m dreading the day my resuscitated washing machine spins its last load.

There’s a beautiful simplicity in a mechanical state machine. I don’t want a device that weighs the load, flags warnings if my underwear isn’t colour matched, and then uses “AI” to go online to search for the optimal programme for my climatic conditions.
It’s really really simple. Take in some water, heat it, turn the drum gently in both directions. After a while (that was the soapy state), shout out the water and take in some more. Rinse and repeat, literally. Then whizz the crap out of everything so it stands a chance of drying.

My current machine uses ~125 litres (25 litre, wash plus four rinses). I’m not worried, it’s from a well not metered. However I can’t imagine what sort of chemical cocktail you end up wearing with a machine that claims it only needs 45 litres for an entire wash. Ugh.

I’m not worried, it’s from a well not metered.

No water meter…

Ours has a digital display (and a rotary switch to set things)¹
It has a filter sump area, with a drain point so low that you can’t get more than a tea tray under it to catch the water when you open the screw in plug, recessed so far you can only get fingertips in to turn the tight screw/plug²
Pretty reliable though.

¹ So, a simple keypad ain’t good enough then? You know, 0-9 and few extra, and then you can program 9, 99, 999 different combos. No? sigh…

Rick, you need a 1950s twin tub.

Rick, you need a 1950s twin tub.

Mums old Parnall carried on until the motor went.
Repair man was sad to say you couldn’t get the parts anymore, due to them having gone out of business decades before.
Apparently, people only ever bought one and that lasted pretty much a lifetime (although mum is still going, and the washer isn’t) so there was no return customer base.

There’s a beautiful simplicity in a mechanical state machine. I don’t want a device that weighs the load, flags warnings if my underwear isn’t colour matched, and then uses “AI” to go online to search for the optimal programme for my climatic conditions.

Mine died a few months ago and the replacement, while not Internet-connected, still tries to be “clever” and fails at it. I started a load earlier and it weighed it and told me that it’d take 46 minutes. A little over an hour later, it’s saying it still has 9 minutes remaining. Did someone let the Microsoft “file copy progress” people near it? :)

it’s saying it still has 9 minutes remaining

I wonder if that’s a water pressure issue? My machine has timings in the back (it’s pretty easy to work out, each click on the knob is three minutes), but since my water is pumped up from a well the pressure is pretty low by the time it reaches the house. Takes around eight minutes to fill the twenty five litres the machine wants, and even that varies depending on whether the pump is about to kick in or has just done so.
All of which means, forty minutes extra just in filling the drum, and not accounted for in the optimal estimation of how long a load of washing ought to take. Though I still come out ahead as they based their estimation on a 60° wash!

Did someone let the Microsoft “file copy progress” people near it?

Your washing will be done in 45 minutes… 70 minutes… 30 minutes… 2 hours… ;)

The one that really really bugs me with Android (notably YouTube) is that the estimations are all off by one. If an upload says about two hours remaining (slow uplink speed), they mean three because after saying an hour remaining (for a while) it’ll then switch to saying something like 75 minutes remaining.
At the end, “one minute remaining” and then suddenly as the minute is up, 50 seconds remaining.
How can one have any confidence in programmers that can’t even manage basic maths? It’s as if they’re only doing “estimate DIV 60” and paying no attention to the modulus, so “1 hour = 1.95 hours”. It’s dumb.

I wonder if that’s a water pressure issue?

I’ve never seen such an inaccurate estimate before, and I didn’t notice anything strange about the water, so I’m just going to go with “black box computer decides to do something odd”.

I’ve noticed that on YouTube a 364-day-old video will be reported as “uploaded 11 months ago”, only to switch to “1 year ago” the following day (or maybe it takes until day 366?).