The big hype "anything can be done by ChatGPT" was never going to last long.
The next step will be embedding these models in other products as practical tools (for example automatically summarising things) and they become part of day to day usage. That's already happening.
Remember when tesla told us their cars could drive themselves?
Then, like five years later, they still weren’t really driving themselves. (=
AI feels sort of like that.
ChatGPT is fun, but gets as much wrong as it gets right.
The only things I consistently use it for are rewriting my emails in different tones and to summarize large documents / emails that I don’t feel like reading.
But ChatGPT is already useful if you use it correctly. I don't try to make it write my code. I use it as an expert knowledge system. If I need a specific BASH command or I want to figure out how to do something in some badly documented but widely used API, I just ask it and it gives me good hints.
This reflects my experience with it so far but I see so many claims of multiplicative "productivity" boosts but there's always a lack of concrete examples of what this means. Avoiding a quick google/stack overflow search is not my definition of productivity boosts.
ChatGPT is useful for generating text that look okay at a distance. I can see the use for creating low quality content to drive traffic to a website, something like medium.
For coding its not useful for any language you are proficient in. There is just so much hype around it but I just can't find the usecase for it.
It's been great as I've been trying out Nix, exactly because I'm not familiar with Nix. It doesn't just magically solve all my problems, but it's been super helpful in showing me general aporoaches that I wasn't aware of, and also gets the answer right most of the time.
Every HN thread has plenty of examples, refactoring/extending/explain8 code, structuring text, knowledge summary and so on. Where do you guys live that you are still looking for the plethora usecases that people are using this tech for?!
This is my opinion as a professional content creator: of course, the boom is over, if by 'boom' you mean the frenzy generated on the internet. But, the ONLY reason why this is true is because nowadays the internet functions on trends. It was always true that people followed trends, but the difference between attention to any one trend between when it first becomes trendy and mere days later is far greater than it's ever been.
On the other hand, the economic and social impact of AI has only just begun. Now it will move more quietly to slowly integrate itself in society like a virus, making us more and more dependent on AI.
This integration will push us to becoming more like automatons, needing people less, and further concentrating the wealth of the world towards Silicon Valley, because it will create a worldwide addiction and dependency on something we never needed in the first place.
> slowly integrate itself in society like a virus, making us more and more dependent on AI
so you could say the same for electricity. And yet i dont see you calling it a virus!
> needing people less
no, needing people to do less isn't the outcome. The same amount of people can now produce _even_ more. Just like having tractors pushed out farmers that used to hire lots of manual labour.
It's a good outcome in the long term. It sucks for those who get displaced - but the train of technological progress runs over tracks made of the bones of those it displaces. this has always been the case in history, and will continue to be until the day we reach post-scarcity.
> so you could say the same for electricity. And yet i dont see you calling it a virus!
It is like a virus in a way.
> no, needing people to do less isn't the outcome. The same amount of people can now produce _even_ more. Just like having tractors pushed out farmers that used to hire lots of manual labour.
We have reached the point of diminishing returns when it comes to technological development. We do NOT need to produce more things more efficiently. That only leads to more resource usage, which this planet can no longer accommodate.
Instead, we need to concentrate on sustainability rather than productivity.
> It's a good outcome in the long term. It sucks for those who get displaced - but the train of technological progress runs over tracks made of the bones of those it displaces. this has always been the case in history, and will continue to be until the day we reach post-scarcity.
If we go down that route, we will use up too many resources. We should not aim for post-scarcity, but rather post-productivity. Technology is no longer making the world better, it's making it worse. If you live in a city, simply look out the window. We're destroying our planet all in the name of endless economic growth, which is only making us slaves to the technological machine.
Like what? The lack of indoor plumbing? Or perhaps the high infant mortality rates? The only way to believe this is to live in total ignorance of history.
@dlkf since I can't reply further.... I think many people regard the current system of life unthinkable. I would be more than happy to live in a society with tech from the 90s. To be honest, the world has not gotten better, and the small ways it has would be a fair trade for less technology from the past.
It’s getting started. Serious use cases never have the glamour of hype. But I am starting to see generative AI cover more and more ground into business utility.
Saying it’s over is like saying that the internet is over after the dotcom bubble burst.
Nah. This is a revolutionizing foundational tech. Bigger than the internet even IMO. More like computerization of business, or the steam engine.
Completely agree. These foundational models are more akin to the invention of transistors or the internet than anything else. It's a little disheartening from a startup perspective to see already large "incumbents" in the form of Google, Open-AI but I think most of the value created, in the future, from generative AI, will not be captured by these players.
But models like GPT-4 depend on good quality training data, and that resource is limited, means maybe we aren't so long away from peak at least for LLMs.
There is no unlimited growth, and after the low hanging fruits each step in progress will be exponentially harder.
Llama 2 is very close to GPT 3, same with the new code Llama.
That said, while it scores high, it's not as "generally" competent on most issues as GPT is, however, it's an exponentially smaller model and for that it does, it's really impressive.
I don’t particularly use ChatGPT in any serious capacity or for work. But, for my amusement and intellectual assisting needs, Llam2-uncensored and codewhisper models run using Ollama locally on my Mac has been a pleasure to use.
Like most things, if you steer clear of all the hype there is a golden core of utility to be found. It sure is not just a party truck as the article daringly claims/questions it to be. The ones that make clever use of the foundational tool will stand to reap the benefits. Just like the ones that benefited from the spread of the internet.
Depends on what you want to do. For writing and RPG style interactions some of the LLama 2 based models are pretty good. Beauty is, even the 13B models work in those cases. So it works reasonably well on consumer grade hardware.
It is not really a fair comparison, because chatgpt is not just a model. There is clever output sampling, scoring, sub model selection based on prompt and more. While open source LLMs are "just a model" with the most basic (greedy) output sampling.
To have real LLM comparison we would need access to pure chatgpt model's input/outputs. I wouldn't be surprised if certain open source LLMs like falcon-40b-instruct or llama2 have already surpassed chatgpt 3.5 models in quality.
But as far as I know no one so far has taken a bunch of open source models and wrapped them in a service like chatgpt 4. Why? Because the main cost of running such service using open source code (assuming the service would be open source too) would be the cost of renting hardware. Consequently the barrier to entry for competition would be very low so no one would sink their life savings in a company like that.
However, once it becomes possible to run multiple models like falcon-40B and llama2 (without quantization to 4bits) on a typical "high end pc" that will change. We will see open source projects that want to achieve a "chatgpt service" offline on consumer hardware. Why only then? Same reason why Linux was written once 386PC's became available. Could someone like Linus Torvalds write Linux 20 years before on their "university" or "company" computer? Sure, but except certain exceptions (gnu) most people engage in writing open source software they themselves want to run on their own hardware.
Big companies know this and they know very well such developments will out compete them shortly (like Linux out competed unix) so they are already doing all they can to stop or slow it down. By talking about "the dangers of AI", by not releasing powerfull ai accelerators on the market (google's tpu). And so on. But it will happen. Give it 10 years. I
> This is a revolutionizing foundational tech. Bigger than the internet even IMO.
Oh is it? Maybe you can tell us then how you'd get the models home without the internet? On a DVD? Or interact with a server that runs the model for you? Or how we'd even be having this discussion?
People said similar delusions about Second Life back in the day.
Something can depend on another thing and still be "bigger" than it. (as woolly as that definition is)
I don't actually agree with the point you're contesting with but your reasoning here is flawed. Also - you might want to check your tone for future posts. It's slighty combative.
You're making a logical connection betweeen x depending on y and x being (bigger/more important/more significant) than y.
That implies that nothing can be more important than all of the things that were neccesary for it to come to pass. And there are a huge number of neccesary preconditions for most things. Your definition would result in a very counterintuitive definition of importance.
I understood your criticism of my post. What I was asking for as how you would have expressed the criticism of the post I was criticising, since you mentioned that you also don't agreee with it, just like me.
People miss the "hidden bit of the iceberg" when dismissing gen AI, or making claims that the hype is over. It's just getting started, indeed.
Anyone who did data work in big, boring enterprises knows the staggering amount of dark, unstructured data they have. Free-form text input, chat/phone conversations, but also plain old documents (pdf scans, PowerPoint files, anything).
Have a LLM interpret what is there and you turn this unstructured mess into data with a clear schema that can go into a relational database for further processing or summarization.
LLMs are groundbreaking for this kind of automated data structuring. And many use cases don't need perfect accuracy; detecting trends or summarizing at a high level is good enough for inputs to downstream process mining tasks.
This is such a naive take. I’m using gpt4 daily for anything coding, sql, directions, text editing, research etc. It’s absolutely invaluable and if it was $1k a month I’d seriously consider getting it.
Computer just had their GUI moment again and we’re just getting started.
For many companies, it's already hard to get a human support person on the line. "It's all on our website, just look there!" In the future this will be even harder.
I use it as an extension of my mind towards getting answers for grey areas. A grey area such as "write a unit test skeleton for this method signature", where the method in my head is either too abstract or has complexity that increase my short term memory span.
I find it a helpful extra pair of eyes that seems to know a lot more than can be expected by a human.
It is getting blended into life - so you won’t realise AI is being used. Companies are too focussed on efficiency.
I have bought a perplexity pro subscription which is excellent - it creates useful summaries of search results across various sources but also has generative AI with chatgpt 3.5 FT, 4, Claude 2 along with llama 2.
I personally have difficulty trusting generative AI - too many hallucinations- so to see sources is helpful.
BTW, people get excited with "chatgpt for enterprise", but bing is doing something somewhat similar. For example if you go to bing.com from the internal network of one company I work for you get a "work" menu item(in addition to "news" "images" etc). If you search there it's not only searching all the company wikis, sharepoint sites, and any shared documents you might have access to, but it also answers questions "a la chatgpt" in company's context.
I'm not sure how long this has been available. I noticed it last month, but it is pretty useful to search for documentation or if you don't know internal procedures. You can just just describe what you want to achieve and it is reasonably good in finding the right documents. The AI boom is far from over. It hasn't even started properly yet.
To me, the hype is about big flashy use cases. The usage is like google search - scattered throughout my work week. Generate a power BI function (I don’t know the syntax at all), extract info from a strange document in english (not my mother tongue), write boring copy for marketing (I don’t do it often, I can fix the facts myself but ChatGPT’s english is better).
Some use cases are enabled just the last few weeks by company-internal models.
This will accelerate and evolve. Generative AI is a new appliance. We just got started.
An acquaintance of mine wrote commercial content for a living. Pretty repetitive work, which he performs remotely. He recently quit because of AI. This will happen more and more and the people concerned will often have a negative bias on AI.
To wrap AI technology into a chatbot is a deceiving, strategic move. It is a wolf in sheep's clothes aiming for people to believe the technology is there to help them. As such it is a solution looking for the problem, and now a year later there are not enough problems that a chat interface can solve.
In reality the genie is out of the box, but not to help people rather to replace them. Reduction of ChatGPT traffic does not tell us what is really ahead of us.
Well, yeah. The use cases are fundamentally limited by its approach. Statistical techniques underperform compared to symbolic rules no matter how much hardware you throw at the problem. This has been known from the start.
Anyway, all known AI techniques struggle with common sense reasoning. The majority of ambiguities in natural language are distinct and uncorrelated. Combining known techniques is not going to solve the problem. Either we discover new theories, or revise our natural languages to be less ambiguous (likely a deeply unpopular social change). Both will probably occur in this century.
As for ChatGPT and Bing? I expect these brands will be dead long before we get to the next wave of AI.
I feel like a passable reasoning will be achieved by an equally dumb / brute force approach. Basic reasoning isn't different from pattern matching. When one calculates an integral, he matches the function against known patterns until something sticks and applies the transformation. Similarly, logical reasoning is often just noticing patterns (a or not b) = (b -> a).
Several other LLMs score higher than 95.0 as well.
I suspect some upcoming multimodal LFMs will be very close to human level since their training data would include video, audio, and possibly even physical interaction input.
Moreover, top LLMs already surpass the human baseline on SuperGLUE, a suite of many language tasks.
This is my experience when using GPT-4 as well. It’s like an extremely well-read intern who follows instructions much better than an average person on the street. It is just often a little drunk or sleepy at work ¯\_(ツ)_/¯, ……for now.
Like the Turing test, Winograd is not a very good test and more of a philosophical question than a metric for engineering.
Sure there's nothing wrong with a model that at some point in time passes 100% and is even more consistent than humans, but it's not possible to make an exhaustive list of what humans can disambiguate that these models can't.
It's a moving target. We're just a cultural shift, scientific discovery, or political regime change away from the models needing to be updated all over again. For some applications these lagging inaccuracies are not acceptable. For yet still more applications, it's not possible to allow any amount of time to train the models ahead of seeing novel ambiguity.
There's a lot of work left for AI research and the threat of authoritarian-induced cultural stagnation just to serve the AI gods and the convenience of governance is absolutely terrifying. i.e. You can't speak that way because it confuses the AI, you can't look or act that way because it confuses the AI, etc. AI in its current form totally sucks and is unfit for much beyond low risk tasks.
I was addressing your statements regarding the capabilities of neural AI:
> Anyway, all known AI techniques struggle with common sense reasoning. The majority of ambiguities in natural language are distinct and uncorrelated. Combining known techniques is not going to solve the problem.
> Statistical techniques underperform compared to symbolic rules no matter how much hardware you throw at the problem.
The best empirical machine learning techniques to measure a system's capabilities are good enough for many practical purposes. AI's commonsense and language understanding differs from that of humans but it's not necessarily inferior to the average human's understanding, even in its current state. AI's understanding is shallower but more interconnected to more diverse domains of knowledge, which is very useful for practical purposes.
We're not going to change our language to avoid confusing AI. If anything, it can already communicate, using our natural language, better than an average human in some practical contexts.
I do agree though that AI research in language understanding is far from completed.
This article feels like it’s written by someone who wants the boom to be over. Just spending a few hours with AutoGPT made me realize it’s very far dram over. It just started.
I don't think the technology is going anywhere, but the hype is cooling down, as limitations are starting to be better understood, and the 'Dude! Don't you get it! At this pace[1] we'll have a 1e304 parameter GPT17 model and weaponized AGI by noon on Thursday'-bros have quieted down a bit.
If anything it's underhyped, in that the technology will be foundational to how we do work and sort through tremendous data volumes in the future. The pushback is simply that these giant companies haven't found any clear way to monetize primary LLMs as a service-- and they won't, because open-source models are improving at a breakneck pace and hardware requirements (and thus cost) is dropping due to community development.
Ultimately I don't see those primary LLMaaS offerings being particularly profitable. Secondary LLM applications are a completely different matter. For example, LLMs acting as medical scribes, sitting in your doctor's office and filling out your chart, handling a job that formerly required a skilled human. That sort of thing will make tons of money. It just probably won't use Azure/Amazon/Google LLMs to do it.
71 comments
[ 5.3 ms ] story [ 337 ms ] threadThe next step will be embedding these models in other products as practical tools (for example automatically summarising things) and they become part of day to day usage. That's already happening.
Then, like five years later, they still weren’t really driving themselves. (=
AI feels sort of like that.
ChatGPT is fun, but gets as much wrong as it gets right.
The only things I consistently use it for are rewriting my emails in different tones and to summarize large documents / emails that I don’t feel like reading.
ChatGPT is useful for generating text that look okay at a distance. I can see the use for creating low quality content to drive traffic to a website, something like medium.
For coding its not useful for any language you are proficient in. There is just so much hype around it but I just can't find the usecase for it.
Evidently: for coding in languages you are not proficient in! What could go wrong?
You're being facetious, but I agree unironically.
On the other hand, the economic and social impact of AI has only just begun. Now it will move more quietly to slowly integrate itself in society like a virus, making us more and more dependent on AI.
This integration will push us to becoming more like automatons, needing people less, and further concentrating the wealth of the world towards Silicon Valley, because it will create a worldwide addiction and dependency on something we never needed in the first place.
so you could say the same for electricity. And yet i dont see you calling it a virus!
> needing people less
no, needing people to do less isn't the outcome. The same amount of people can now produce _even_ more. Just like having tractors pushed out farmers that used to hire lots of manual labour.
It's a good outcome in the long term. It sucks for those who get displaced - but the train of technological progress runs over tracks made of the bones of those it displaces. this has always been the case in history, and will continue to be until the day we reach post-scarcity.
It is like a virus in a way.
> no, needing people to do less isn't the outcome. The same amount of people can now produce _even_ more. Just like having tractors pushed out farmers that used to hire lots of manual labour.
We have reached the point of diminishing returns when it comes to technological development. We do NOT need to produce more things more efficiently. That only leads to more resource usage, which this planet can no longer accommodate.
Instead, we need to concentrate on sustainability rather than productivity.
> It's a good outcome in the long term. It sucks for those who get displaced - but the train of technological progress runs over tracks made of the bones of those it displaces. this has always been the case in history, and will continue to be until the day we reach post-scarcity.
If we go down that route, we will use up too many resources. We should not aim for post-scarcity, but rather post-productivity. Technology is no longer making the world better, it's making it worse. If you live in a city, simply look out the window. We're destroying our planet all in the name of endless economic growth, which is only making us slaves to the technological machine.
Every generation says this, while at exactly the same time regarding previous ways of life completely unthinkable.
The world was MUCH better without all this junk.
There is going to be a ton of losers with the incumbent big tech companies ending up as the winners.
Saying it’s over is like saying that the internet is over after the dotcom bubble burst.
Nah. This is a revolutionizing foundational tech. Bigger than the internet even IMO. More like computerization of business, or the steam engine.
Look at what GPT-4 can achieve from only receiving text and a primitive form of image training.
There is no unlimited growth, and after the low hanging fruits each step in progress will be exponentially harder.
Most data has not been incorporated into a model yet. Most data is not in text format. We don't know what is possible once that changes.
Hopefully more and more tinkerers are going to fuel this fire even further, making it possible to run AI on-device with custom native chips.
That said, while it scores high, it's not as "generally" competent on most issues as GPT is, however, it's an exponentially smaller model and for that it does, it's really impressive.
Like most things, if you steer clear of all the hype there is a golden core of utility to be found. It sure is not just a party truck as the article daringly claims/questions it to be. The ones that make clever use of the foundational tool will stand to reap the benefits. Just like the ones that benefited from the spread of the internet.
To have real LLM comparison we would need access to pure chatgpt model's input/outputs. I wouldn't be surprised if certain open source LLMs like falcon-40b-instruct or llama2 have already surpassed chatgpt 3.5 models in quality.
But as far as I know no one so far has taken a bunch of open source models and wrapped them in a service like chatgpt 4. Why? Because the main cost of running such service using open source code (assuming the service would be open source too) would be the cost of renting hardware. Consequently the barrier to entry for competition would be very low so no one would sink their life savings in a company like that.
However, once it becomes possible to run multiple models like falcon-40B and llama2 (without quantization to 4bits) on a typical "high end pc" that will change. We will see open source projects that want to achieve a "chatgpt service" offline on consumer hardware. Why only then? Same reason why Linux was written once 386PC's became available. Could someone like Linus Torvalds write Linux 20 years before on their "university" or "company" computer? Sure, but except certain exceptions (gnu) most people engage in writing open source software they themselves want to run on their own hardware.
Big companies know this and they know very well such developments will out compete them shortly (like Linux out competed unix) so they are already doing all they can to stop or slow it down. By talking about "the dangers of AI", by not releasing powerfull ai accelerators on the market (google's tpu). And so on. But it will happen. Give it 10 years. I
I hang around a lot of solopreneurs and there is a serious cohort of founders/hustlers who are using AI tools to greatly improve their productivity.
Not sure how much of that can carry forth to a major corporation though.
Oh is it? Maybe you can tell us then how you'd get the models home without the internet? On a DVD? Or interact with a server that runs the model for you? Or how we'd even be having this discussion?
People said similar delusions about Second Life back in the day.
I don't actually agree with the point you're contesting with but your reasoning here is flawed. Also - you might want to check your tone for future posts. It's slighty combative.
How would you express your angle regarding the contested point?
That implies that nothing can be more important than all of the things that were neccesary for it to come to pass. And there are a huge number of neccesary preconditions for most things. Your definition would result in a very counterintuitive definition of importance.
GitHub copilot and sit.ilar will become defacto standard.
Image generation is already here.
School systems already have to act on this due to pupil using it for essays etc.
AI literally creates a new ara for us: it's the first time we structure or learn on data to have a knowledge agent.
Anyone who did data work in big, boring enterprises knows the staggering amount of dark, unstructured data they have. Free-form text input, chat/phone conversations, but also plain old documents (pdf scans, PowerPoint files, anything).
Have a LLM interpret what is there and you turn this unstructured mess into data with a clear schema that can go into a relational database for further processing or summarization.
LLMs are groundbreaking for this kind of automated data structuring. And many use cases don't need perfect accuracy; detecting trends or summarizing at a high level is good enough for inputs to downstream process mining tasks.
Computer just had their GUI moment again and we’re just getting started.
I use it as an extension of my mind towards getting answers for grey areas. A grey area such as "write a unit test skeleton for this method signature", where the method in my head is either too abstract or has complexity that increase my short term memory span.
I find it a helpful extra pair of eyes that seems to know a lot more than can be expected by a human.
I have bought a perplexity pro subscription which is excellent - it creates useful summaries of search results across various sources but also has generative AI with chatgpt 3.5 FT, 4, Claude 2 along with llama 2.
I personally have difficulty trusting generative AI - too many hallucinations- so to see sources is helpful.
I'm not sure how long this has been available. I noticed it last month, but it is pretty useful to search for documentation or if you don't know internal procedures. You can just just describe what you want to achieve and it is reasonably good in finding the right documents. The AI boom is far from over. It hasn't even started properly yet.
Some use cases are enabled just the last few weeks by company-internal models.
This will accelerate and evolve. Generative AI is a new appliance. We just got started.
To wrap AI technology into a chatbot is a deceiving, strategic move. It is a wolf in sheep's clothes aiming for people to believe the technology is there to help them. As such it is a solution looking for the problem, and now a year later there are not enough problems that a chat interface can solve.
In reality the genie is out of the box, but not to help people rather to replace them. Reduction of ChatGPT traffic does not tell us what is really ahead of us.
Anyway, all known AI techniques struggle with common sense reasoning. The majority of ambiguities in natural language are distinct and uncorrelated. Combining known techniques is not going to solve the problem. Either we discover new theories, or revise our natural languages to be less ambiguous (likely a deeply unpopular social change). Both will probably occur in this century.
As for ChatGPT and Bing? I expect these brands will be dead long before we get to the next wave of AI.
Additional reading for the curious: https://en.wikipedia.org/wiki/Winograd_schema_challenge
https://github.com/LudwigStumpp/llm-leaderboard/blob/main/RE...
The current top model, Vega v2, is already at 98.6 (compared to human’s 100.0). https://super.gluebenchmark.com/leaderboard
Several other LLMs score higher than 95.0 as well.
I suspect some upcoming multimodal LFMs will be very close to human level since their training data would include video, audio, and possibly even physical interaction input.
Moreover, top LLMs already surpass the human baseline on SuperGLUE, a suite of many language tasks.
This is my experience when using GPT-4 as well. It’s like an extremely well-read intern who follows instructions much better than an average person on the street. It is just often a little drunk or sleepy at work ¯\_(ツ)_/¯, ……for now.
Like the Turing test, Winograd is not a very good test and more of a philosophical question than a metric for engineering.
Sure there's nothing wrong with a model that at some point in time passes 100% and is even more consistent than humans, but it's not possible to make an exhaustive list of what humans can disambiguate that these models can't.
It's a moving target. We're just a cultural shift, scientific discovery, or political regime change away from the models needing to be updated all over again. For some applications these lagging inaccuracies are not acceptable. For yet still more applications, it's not possible to allow any amount of time to train the models ahead of seeing novel ambiguity.
There's a lot of work left for AI research and the threat of authoritarian-induced cultural stagnation just to serve the AI gods and the convenience of governance is absolutely terrifying. i.e. You can't speak that way because it confuses the AI, you can't look or act that way because it confuses the AI, etc. AI in its current form totally sucks and is unfit for much beyond low risk tasks.
> Anyway, all known AI techniques struggle with common sense reasoning. The majority of ambiguities in natural language are distinct and uncorrelated. Combining known techniques is not going to solve the problem.
> Statistical techniques underperform compared to symbolic rules no matter how much hardware you throw at the problem.
The best empirical machine learning techniques to measure a system's capabilities are good enough for many practical purposes. AI's commonsense and language understanding differs from that of humans but it's not necessarily inferior to the average human's understanding, even in its current state. AI's understanding is shallower but more interconnected to more diverse domains of knowledge, which is very useful for practical purposes.
We're not going to change our language to avoid confusing AI. If anything, it can already communicate, using our natural language, better than an average human in some practical contexts.
I do agree though that AI research in language understanding is far from completed.
[1] https://xkcd.com/605/
Ultimately I don't see those primary LLMaaS offerings being particularly profitable. Secondary LLM applications are a completely different matter. For example, LLMs acting as medical scribes, sitting in your doctor's office and filling out your chart, handling a job that formerly required a skilled human. That sort of thing will make tons of money. It just probably won't use Azure/Amazon/Google LLMs to do it.
[1] https://news.ycombinator.com/item?id=37362181