Merged Artificial Intelligence

[Roger Ramjets]

Yes, we're talking about microcontrollers to demonstrate how wildly off-base your post was.

It's changing the goalposts.

Using an LLM does not imply incorporating an LLM into your codebase or your OS, the code they write is not bloated at all, and they can easily target MCUs with far fewer resources than the antique computers you were talking about.

Again, irrelevant. An MCU with only enough memory for 512 instructions should be easy for AI generate to code for - unless it has to do something tricky. The problem comes when projects get larger.

But hey, as AI gets more powerful it will get better, right? Meanwhile humans are relegated to curating it and lose their coding skills. This is why many developers are going back to doing their own coding. AI is making them dumber at the same time as it produces code that's hard to maintain. When AI gets involved the slop multiplies to the point where humans don't have the time to check it. The end-game here isn't good.

What's the relevance? 17K would still be significantly less than 1M.

You seem to know nothing about the difference between them. One is running bare metal code perhaps with a tiny RTOS, the other a full-fledged multitasking OS with GUI, and the code must work on several different OS versions and hardware.

And coding models can write code for them, too, so why bring it up, unless you want to be even wronger?

I didn't bring it up.

You don't understand how LLMs work at all, and it's amusing that you're attempting to position yourself as anti-trillionaire when there's only ever been one of those in the world and you worship him.

Don't I? How do you know?

I don't worship Musk either. Show me one post where I worship him.

 

[Roger Ramjets]

But this isn't an AI song. It's a song made using AI.

By that definition nothing is an AI creation.

 

[mumblethrax]

It's changing the goalposts.

No, it's a simple illustration.

Again, irrelevant. An MCU with only enough memory for 512 instructions should be easy for AI generate to code for - unless it has to do something tricky. The problem comes when projects get larger.

It's easy for them to code for even if you ask them to do something tricky.

For example, how would you go about generating and storing in volatile memory the first thousand prime numbers on an MCU with only 1K of SRAM?

But hey, as AI gets more powerful it will get better, right? Meanwhile humans are relegated to curating it and lose their coding skills. This is why many developers are going back to doing their own coding. AI is making them dumber at the same time as it produces code that's hard to maintain. When AI gets involved the slop multiplies to the point where humans don't have the time to check it. The end-game here isn't good.

That all might be true, but it has no bearing on how they work, how they are used or what they are capable of.

You seem to know nothing about the difference between them. One is running bare metal code perhaps with a tiny RTOS, the other a full-fledged multitasking OS with GUI, and the code must work on several different OS versions and hardware.

I am fully aware of the differences between them. Those differences are not relevant to the point being illustrated--that LLMs can write efficient code even for target microcontrollers with extremely limited resources.

I didn't bring it up.

You did. You mistook the point of that comment about MCUs for some kind of weird pissing contest, and thus brought up an ultra-low RAM microcontroller...failing to appreciate that it in no way answers my objection to the idea that using an LLM produces inefficient, bloated code.

Don't I? How do you know?

Because of your posts in this thread where you demonstrate fundamental misunderstandings about how they work.

I don't worship Musk either. Show me one post where I worship him.

For example, the recent one where you attempt to derive genius from incompetence. But it's a glaring pattern that I won't be the only one to have noticed.

 

[Frank Newgent]

And you know this how?

AI doesn't know country don't have flutes.

And I had the album thirty years ago.

 

[W.D.Clinger]

The problem comes when projects get larger.

What would you consider to be a larger project?

For example: Would you consider 800,000 commented lines of code to be a larger project?

 

[Roger Ramjets]

What would you consider to be a larger project?

For example: Would you consider 800,000 commented lines of code to be a larger project?

Certainly. The program I am working on only has 102,612 lines of 68k assembly language (commented on almost every line).

 

[W.D.Clinger]

Again, irrelevant. An MCU with only enough memory for 512 instructions should be easy for AI generate to code for - unless it has to do something tricky. The problem comes when projects get larger.

What would you consider to be a larger project?

For example: Would you consider 800,000 commented lines of code to be a larger project?

Certainly. The program I am working on only has 102,612 lines of 68k assembly language (commented on almost every line).

It looks to me as though you think a "problem comes" when the developers of that 800,000-line project begin to write code with the help of AI tools.

Can you please explain what sort of problem you have in mind? Because I'm not seeing it.

 

[Beelzebuddy]

By that definition nothing is an AI creation.

Correct. It's an important distinction that gets lost in the rush to vilify and/or sanctify AI's use. All these stories of AI did this or AI did that all have people behind the AIs knowingly directing them. It has no will of its own. Remember the Riddle of Steel: what use is a tool without the hand that wields it?

It looks to me as though you think a "problem comes" when the developers of that 800,000-line project begin to write code with the help of AI tools.

Can you please explain what sort of problem you have in mind? Because I'm not seeing it.

Same problems as every large project has, and the entire field of software engineering exists to help solve. Whether a spicy rock or a flesh sack, eventually there's just too much code for a single developer to manage. 800k lines simply will not fit in the context windows of coding LLMs. It has to leave some parts out to work on the rest, which means you have to figure out how to make sure it has loaded the parts it needs, which means you need a solid architecture and plans in place before things get squirrely. Even without that, performance degrades as the context fills. If you've crammed 900k tokens into a million-long context window, little statements like "don't do [stupid thing]" at the beginning of the context tend to get lost in the noise.

 

[Dr.Sid]

Correct. It's an important distinction that gets lost in the rush to vilify and/or sanctify AI's use. All these stories of AI did this or AI did that all have people behind the AIs knowingly directing them. It has no will of its own. Remember the Riddle of Steel: what use is a tool without the hand that wields it?

If you don't know what the AI will do, it does has will on its own. It's not because AI is too smart, it's because you are too stupid in commanding it.

 

[steenkh]

It's an important distinction that gets lost in the rush to vilify and/or sanctify AI's use. All these stories of AI did this or AI did that all have people behind the AIs knowingly directing them. It has no will of its own. Remember the Riddle of Steel: what use is a tool without the hand that wields it?

Or what happens if the mind that wields the hand that wields it tells the AI to do whatever it pleases?

 

[Wudang]

Or what happens if the mind that wields the hand that wields it tells the AI to do whatever it pleases?

At this point we're coming close to the old philosophical question: a mechanic with an artificial arm is working on a car engine. Is the arm part of the organism struggling with the environment or part of the environment the organism is struggling with? <lights up a Gauloise and coughs horribly>

 

[Beelzebuddy]

If you don't know what the AI will do, it does has will on its own. It's not because AI is too smart, it's because you are too stupid in commanding it.

If you're using AI and thinking it has a will of its own because you're too stupid to understand how you are commanding it you're probably listening to top40 country music anyway, so that tracks.

Or what happens if the mind that wields the hand that wields it tells the AI to do whatever it pleases?

You get the stochastic continuation of the baseline, i.e. the slop that people refer to. But don't take my word for it, try it yourself. Load up a chatbot, ask it for a story, and just tell it to keep going when it stops for more prompt. Let me know when it craps out Shakespeare on its own.

 

[W.D.Clinger]

Same problems as every large project has, and the entire field of software engineering exists to help solve. Whether a spicy rock or a flesh sack, eventually there's just too much code for a single developer to manage. 800k lines simply will not fit in the context windows of coding LLMs. It has to leave some parts out to work on the rest, which means you have to figure out how to make sure it has loaded the parts it needs, which means you need a solid architecture and plans in place before things get squirrely. Even without that, performance degrades as the context fills. If you've crammed 900k tokens into a million-long context window, little statements like "don't do [stupid thing]" at the beginning of the context tend to get lost in the noise.

If the problems are the same as in every large project, regardless of whether it uses AI tools, then why would someone blame those problems on the use of AI tools?

It seems to me that one important attribute of a solid software architecture is that it has been designed so you don't have to keep all 800,000 lines of code in your head as you're working on one localized module. That attribute is important for software designed and maintained by humans. Why wouldn't it be equally important for software designed and maintained by humans aided by AI?