LLM Discussion

[Niya]

dday9 edit:
The start of this post is very odd if you don't have the context of where they were spun off from, here: https://www.vbforums.com/showthread....asic-Criticism

Basically, let's talk LLMs. Now everything below this was Niya's original post that was spun off.

-----------------

To be honest, I don't see how any programmer in the future could NOT make AI a part of their workflow. It would be insane not to take advantage of it. AI is becoming way too good to ignore any longer.

I think we are in the final years where we could safely ignore AI. 2026 and onward, we cannot. We'd be outcompeted way too easily by anyone using it.

[fafalone]

AI can be helpful with a lot of things but it would have to improve dramatically to be of use for programming something complicated and novel.

As an illustration, when I was making a Media Foundation player recently. There's documentation for interfaces to display subtitles, but MS doesn't explicitly document how to actually obtain it. Every single AI did nothing but waste my time with hallucinations that *looked* right but failed, varying from 'good code for getting some MF interfaces but not subtitles' to 'made up methods that don't exist' as I pushed for alternative solutions after failures.

Eventually I found a single extant project that answered my question; must not have been in the training material.

Another funny one was asking for help writing a driver. "VB6 can't do that because..." "Yes it can" "No, it can't" "Really it can" "Ok, here's an error filled routine for calling a driver, the opposite of what you asked!"
AI just isn't there for advanced programming with topics that don't perfectly correlate to code well represented in training data.

And there's good reason to believe we're very close to the limits of the current methodology... ChatGPT for instance has gotten substantially *worse* to the point I don't ask it for help with any vb6/tB code. It's not helping solve exotic edge cases deep in compiler codegen.

[wqweto]

To be honest, I don't see how any programmer in the future could NOT make AI a part of their workflow. It would be insane not to take advantage of it. AI is becoming way too good to ignore any longer.

I think we are in the final years where we could safely ignore AI. 2026 and onward, we cannot. We'd be outcompeted way too easily by anyone using it.

We might see Year of the Linux Desktop meme come true about this time too. . .

cheers,
</wqw>

[Niya]

AI can be helpful with a lot of things but it would have to improve dramatically to be of use for programming something complicated and novel.

As an illustration, when I was making a Media Foundation player recently. There's documentation for interfaces to display subtitles, but MS doesn't explicitly document how to actually obtain it. Every single AI did nothing but waste my time with hallucinations that *looked* right but failed, varying from 'good code for getting some MF interfaces but not subtitles' to 'made up methods that don't exist' as I pushed for alternative solutions after failures.

Eventually I found a single extant project that answered my question; must not have been in the training material.

Another funny one was asking for help writing a driver. "VB6 can't do that because..." "Yes it can" "No, it can't" "Really it can" "Ok, here's an error filled routine for calling a driver, the opposite of what you asked!"
AI just isn't there for advanced programming with topics that don't perfectly correlate to code well represented in training data.

And there's good reason to believe we're very close to the limits of the current methodology... ChatGPT for instance has gotten substantially *worse* to the point I don't ask it for help with any vb6/tB code. It's not helping solve exotic edge cases deep in compiler codegen.

Two points.

One, I'm looking at the rapid pace with which AI is advancing. Sure, it might seem that it's stalled a bit as per your experience with ChatGPT, but I don't believe for a second that it will remain so inevitably. I have every confidence that breakthroughs will be made in the coming years that will take AI to levels we can barely imagine right now. There is no doubt in my mind.

Secondly, I think you might be using it wrong. Right now, it's not nearly good enough for us to offload our cognitive tasks but it is really really good at following instructions and executing them way faster than any human can even dream of doing. Current AI models perform far better when they have to think less. The best relationship to have with current models is one where you're doing all the thinking, planning, research and design. The AI just a low-skilled laborer that does all the heavy lifting to execute your vision, something it could do far better and faster than any human.

[fafalone]

Thinking we're close to a radical shift in AI is like thinking routine space travel to other star systems was on the horizon after we landed on the moon. There's a ceiling for the current underlying technology... it will advance, but not to the point of revolution.

And I'm not saying I don't use it or it's not useful. It's great for simple boilerplate, routine basic coding an amateur could do... but you have to know the limits or you spend more time rewording prompts and fixing mistakes than it would have taken to just do it yourself.

Another example with just the last thing I did:

> I need to get the total amount of video RAM for each adapter, from a 32bit application, where the total can exceed 4GB. On Windows, with C++. How can I do that?

Since it has no actual reasoning ability, even an amateur human would catch the issue here... a SIZE_T or UINT32 can't store numbers above 4GB. But since a language model is a technology without any actual understanding, it gives me the standard methods: DXGI_ADAPTER_DESC (SIZE_T, wrong), and WMI (uint_32, wrong).

Asking for corrections only gets the same or worse results; it changed to using DXGI_ADAPTER_DESC1, falsely claiming that worked when it's the same size_t. It doesn't even understand the same problem exists. Another solution actually does support the proper 8-byte types, but only provides information about available to the current app/allocated to the current app memory (which is more useful, but not the right answer).

And this is hardly compiler development.

That's not a problem that will be solved by throwing more compute and increasingly contaminated data at it. It's a fundamental limitation like burning hydrocarbon-based propulsion never getting us to FTL.

[Niya]

Thinking we're close to a radical shift in AI is like thinking routine space travel to other star systems was on the horizon after we landed on the moon. There's a ceiling for the current underlying technology... it will advance, but not to the point of revolution.

I see what you're trying to say but there is something you didn't consider: we have no non-theoretical models for advanced space travel. We don't know of anything that could surpass the cosmic speed limit. All we have are theories, most of which tell us that there is a ceiling. However, we do have numerous examples of what neural networks are capable of all around us. The ants in your backyard are constantly solving problems of ventilation, defense, construction, acquiring energy, and reproduction, all of which are powered by neural networks that could fit in the palm of your hand. The neural network in a dragonfly’s head controls an exceptionally advanced flight system, coordinating four independently moving wings while simultaneously processing complex data on air pressure, currents, angles of attack, speed, distance, and timing. The end result is a hunting success rate of 90%, the highest in the animal kingdom. We haven't even mentioned humans yet.

My point is, nature has already given us countless examples, not even counting humans, of what neural nets are capable of. We haven't tapped into that yet, but we know it is possible. We also have 2 major advantages over nature: energy and storage capacity. The human brain can only hold so many neurons, and the body can only provide so much energy. Our artificial neural nets have no such constraints. Imagine how potent AI will become when we do unlock the remaining secrets of neural networks.

Another example with just the last thing I did:

> I need to get the total amount of video RAM for each adapter, from a 32bit application, where the total can exceed 4GB. On Windows, with C++. How can I do that?

Since it has no actual reasoning ability, even an amateur human would catch the issue here... a SIZE_T or UINT32 can't store numbers above 4GB. But since a language model is a technology without any actual understanding, it gives me the standard methods: DXGI_ADAPTER_DESC (SIZE_T, wrong), and WMI (uint_32, wrong).

Asking for corrections only gets the same or worse results; it changed to using DXGI_ADAPTER_DESC1, falsely claiming that worked when it's the same size_t. It doesn't even understand the same problem exists. Another solution actually does support the proper 8-byte types, but only provides information about available to the current app/allocated to the current app memory (which is more useful, but not the right answer).

Well, like I said, our current models aren't ready for this level of cognition, but I wouldn't bet on this being the case indefinitely, after all, the neural net in your head can do it. I see no reason why we can't build an artificial one to do the same. The only valid point of contention I can see in all this is time. When will it happen? Two years from now or thirty? But I am sure it will happen.

[fafalone]

I see what you're trying to say but there is something you didn't consider: we have no non-theoretical models for advanced space travel. We don't know of anything that could surpass the cosmic speed limit. All we have are theories, most of which tell us that there is a ceiling. However, we do have numerous examples of what neural networks are capable of all around us. The ants in your backyard are constantly solving problems of ventilation, defense, construction, acquiring energy, and reproduction, all of which are powered by neural networks that could fit in the palm of your hand. The neural network in a dragonfly’s head controls an exceptionally advanced flight system, coordinating four independently moving wings while simultaneously processing complex data on air pressure, currents, angles of attack, speed, distance, and timing. The end result is a hunting success rate of 90%, the highest in the animal kingdom. We haven't even mentioned humans yet.

My point is, nature has already given us countless examples, not even counting humans, of what neural nets are capable of. We haven't tapped into that yet, but we know it is possible. We also have 2 major advantages over nature: energy and storage capacity. The human brain can only hold so many neurons, and the body can only provide so much energy. Our artificial neural nets have no such constraints. Imagine how potent AI will become when we do unlock the remaining secrets of neural networks.
.

LLMs are not an accurate representation of how our brains work and the fundamental problem is we have no idea how they do. That's what I'm trying to say. Just because we've made one discovery that gets us a little way of the way there, doesn't mean that exact technology can take us to the next level, and we frankly have absolutely no idea how to do it. You're right that it's different in that we know intelligence is possible because we can observe the human brain, but that doesn't mean we can get there just by throwing more conventional processing power and memory at LLMs. Because the technology has a ceiling we're nearly at. Going much further would require a revolutionary new approach that we're as far if not further from understanding than propulsion methods that would get us to other planets (which isn't ruled out; a relativistic ship accelerating at a constant 1g would get us to a whole other galaxy in under 30y... for the occupants).

[yereverluvinuncleber]

Come on, It's not rocket science...

[dday9]

Let's try to stick to the topic of the thread, which is tB criticism.

I love the discussions on LLMs, but we have a pretty solid practice of sticking to one topic per thread. I'd encourage that y'all bring this discussion to General Developer Forum and if you want, I can move some of these posts to that thread. I'd not suggest bringing it to the Chit-Chat for fear of it being coopted by... well the Chit-Chat

[Niya]

and if you want, I can move some of these posts to that thread.

I appreciate that. I wasn't planning on saying much more than I already have. If faf wants to continue I have no objections. I love speculating about the future of AI and there is more I can say.

[Shaggy Hiker]

Two points.

One, I'm looking at the rapid pace with which AI is advancing. Sure, it might seem that it's stalled a bit as per your experience with ChatGPT, but I don't believe for a second that it will remain so inevitably.

I'm not nearly as sanguine as you are about this. I hear more and more about how this is a case of smoke and mirrors. They are both running out of training material, and perhaps more importantly, NONE of the LLM AI models appears capable of making money. ChatGPT loses money on every request. The newer models appear to be able to lose money faster.

[dday9]

The chief architect at my old job stood up his own instance on his own machine that didn’t “phone home” at all, I.e., he unplugged his Ethernet and ripped out his network card but it still worked.

Do these not cost as much to run, they just aren’t as accurate?

[Niya]

I'm not nearly as sanguine as you are about this. I hear more and more about how this is a case of smoke and mirrors. They are both running out of training material, and perhaps more importantly, NONE of the LLM AI models appears capable of making money. ChatGPT loses money on every request. The newer models appear to be able to lose money faster.

Growing pains. That's what it is. AI is still new, so there will be growing pains. I think everyone investing in AI understands that. AI is already changing the very foundation of how modern societies operate. It has the potential to revolutionize how everything is done in every facet of life. Losing a little money along the way is hardly anything to be concerned about in my opinion. AI is here to stay, whether we like it or not. There is no way this genie goes back into the bottle just because a couple of companies are losing money in the present.

[Niya]

One other thing: there are a lot of people making serious money off AI right now, so even if ChatGPT is bleeding money, it's certainly not an indictment of the technology, it's more an indictment of the business model. The market has already proven beyond all doubt that AI is highly profitable when used competently.

[Shaggy Hiker]

Growing pains. That's what it is. AI is still new, so there will be growing pains. I think everyone investing in AI understands that. AI is already changing the very foundation of how modern societies operate. It has the potential to revolutionize how everything is done in every facet of life. Losing a little money along the way is hardly anything to be concerned about in my opinion. AI is here to stay, whether we like it or not. There is no way this genie goes back into the bottle just because a couple of companies are losing money in the present.

That's possible, but if they can't figure out any means by which to make money, then who is going to be paying for this? Right now, it appears to be people pouring money into it hoping that there will eventually be a gold mine here, but so far, it's been a losing proposition. If there is hope in sight, then that's one thing, but if that hope never arrives, then there has to be some business model other than, "talk positively and hope for a miracle."

The tools are quite impressive. If the price of them is ruinous, then they...will lead to ruin. There HAS to be a profitable business model, eventually. That model does not yet exist, nor has anybody suggested where it might be found.

[Shaggy Hiker]

One other thing: there are a lot of people making serious money off AI right now, so even if ChatGPT is bleeding money, it's certainly not an indictment of the technology, it's more an indictment of the business model. The market has already proven beyond all doubt that AI is highly profitable when used competently.

Who?

Who is making money, and are they doing so solely because they don't have to bear the cost? Heck, ANYBODY can make serious money if they are given hundred dollar bills (for free) and told to sell them for twenty bucks a piece.

[Niya]

Who?

Who? Everybody. It's like every day, something new is popping up. There are companies selling AI girlfriends and companions. There are AI influencers now. People making short films and concept trailers. There are also about a dozen or so Chat Music clones milking YouTube. These are only a handful of examples. AI is everywhere now.

Who is making money, and are they doing so solely because they don't have to bear the cost? Heck, ANYBODY can make serious money if they are given hundred dollar bills (for free) and told to sell them for twenty bucks a piece.

Errr. I think you might be looking at it from the wrong POV. I mean, the more money I make off AI, the more money you are going to make off AI because I am paying you for use of it. No one is getting anything for free. If the money I'm paying you to provide me with AI is not covering your costs, then either you need to charge me more which I'm happy to pay as my business is very profitable or you can invest in looking for ways to reduce your costs. You can even do both.

Right now, it appears to be people pouring money into it hoping that there will eventually be a gold mine here

No one wants to go back to a world without AI. We're all addicted now. If that is not a gold mine, I don't know what is.

There HAS to be a profitable business model, eventually. That model does not yet exist, nor has anybody suggested where it might be found.

There will be. I have to stress again, this genie isn't going back into the bottle. The world is hungry for AI. They are going to make this work no matter what it costs because it will be worth it in the end.

[Shaggy Hiker]

Errr. I think you might be looking at it from the wrong POV. I mean, the more money I make off AI, the more money you are going to make off AI because I am paying you for use of it. No one is getting anything for free. If the money I'm paying you to provide me with AI is not covering your costs, then either you need to charge me more which I'm happy to pay as my business is very profitable or you can invest in looking for ways to reduce your costs. You can even do both.

No, we're looking at it the same way, and it's quite reasonable. We can't say much about whether or not it will get cheaper. Technology changes in unpredictable ways, so it may well be the case that it gets cheaper. Currently, though, none of the LLM are even close to being profitable, which means they aren't charging anywhere near the break even point. Why is that? I see speculation, but I don't know why that is. I think it is because the use of LLM would plummet if the true cost was passed along to consumers in the current iteration.

I got looking for articles on the subject. There were too many, and I wandered off reading about them, eventually ending up on a page about a murder. In any case, they aren't currently profitable. Will they be? Perhaps. Might they not be? Also, perhaps. Not much point in arguing about it, though. That's the future. It will get here.

[wes4dbt]

which means they aren't charging anywhere near the break even point. Why is that?

Maybe they're using an old drug dealers trick. "Here, try it, it's free" lol

It's an interesting point AI is currently at. As of now it's a useful tool for some but still very limited. But as more and more data is being continually gathered AI should get better.

[Arnoutdv]

I use “AI” or LLM often. Sometimes for a skeleton for a algorithm, but often also to discuss things as you do with coworkers.
Last week I wanted to reorganize some data structure on the fly before processing it. All my tested methods were to slow. So I asked chatGPT for alternatives, it did come with the methods I already tried. So I explained in more detail why I didn’t want to use the proposed methods. While doing this I thought of some other method which does work much faster.
Why not discussing this with a coworker, because I hardly have any, working from home most days of the week

[Niya]

We can't say much about whether or not it will get cheaper. Technology changes in unpredictable ways, so it may well be the case that it gets cheaper. Currently, though, none of the LLM are even close to being profitable, which means they aren't charging anywhere near the break even point.

Ah. I see now where we we actually disagree. What is going to happen here I think is very predictable because it has happened several times. Whale oil made way for petroleum, the horse and carriage made way for the combustion engine, ENIAC evolved into the laptop, cellphone and desktop. The electrical infrastructure took us out of the dark. Powered flight gave humans an entirely new dimension of travel. The postal system was later supplemented by the telegraph system, which eventually went extinct due to the telephone system. Then all of these things were eventually dwarfed by the internet, email, and WhatsApp.

Growing pains and exorbitant costs were always part of the deal for technological progress. AI will be no different. Just like electricity, the internet, and the combustion engine, AI is a revolutionary technology that quickly ensnared us into a level of dependency that guaranteed its success, and like all of these previous developments, it will succeed and change the very foundations of how modern societies operate.

Now we can contrast this with something like fusion energy. Fusion energy is perpetually 30 years away, and the reason it's having trouble succeeding is because we already have a perfectly functional petroleum-based energy infrastructure. EVs fall into this category as well. The demand for EVs is entirely dependent on artificial demand driven by how many people buy into climate change hysteria. We don't actually need them which is why despite being around for nearly 2 decades, they have failed to dominate the market. Unlike the internet, aircraft, electricity and AI, EVs and fusion energy don't give us anything new. They are just different ways of doing things we have already mastered.