I am not going to comment on the entire article, but this bit is funny to me > $1,199.79 for Anthropic Claude Code > $980.37 for OpenAI Codex This sounds like my mother when she explains how it...
I am not going to comment on the entire article, but this bit is funny to me
I just ran the ccusage tool on my laptop to get an estimate of how much I would have spent if I were to pay for API tokens in the past 30 days and got:
> $1,199.79 for Anthropic Claude Code
> $980.37 for OpenAI Codex
That’s $2,180.16 worth of tokens for $200—not bad at all
This sounds like my mother when she explains how it was such a great deal to buy the baby clothes at Kohl's for no specific use because it is such a good deal that those clothes could have cost $80 but I got them for $20.
–––– (imagine a line here because I am bad at markdown on my phone)
I hope that the increasing prices do force people to stop throwing every single query at them.
Funnily enough, my office did get a memo from one of the higher ups asking how the Dev team could figure out how to lower our LLM usage because they are seeing all of these articles about price jumps. None of us are using them, we have no LLM licensing costs. The top brass just assume we all are.
We are allowed to if we wanted, but we are a small team in a relatively small company and the hard part of writing our code and fixing bugs are just cases where "this piece of data will never be...
We are allowed to if we wanted, but we are a small team in a relatively small company and the hard part of writing our code and fixing bugs are just cases where "this piece of data will never be X" ends up being "X" and we need to update the business rules to handle it. Often my PRs are in the realm of a dozen lines. It would regularly take me more effort to explain the fix to the LLM than it would take to just change the code myself.
Small company, small problems, all business rule kind of code issues rather than generating pipelines where it would have more benefit to me specifically.
Finding something useful as a tool in some aspects of your work is different from being forced to use something in all aspects of your work even where it does not fit. I can see why someone...
Finding something useful as a tool in some aspects of your work is different from being forced to use something in all aspects of your work even where it does not fit.
I can see why someone working in an environment where it is the latter rather works in one where there is none of that.
Even good things can be a bad thing in sufficient quantities (see: tribbles). Though, much like tribbles, I'm not even sure LLMs are a good thing used sparingly (I know not everyone will agree...
Even good things can be a bad thing in sufficient quantities (see: tribbles). Though, much like tribbles, I'm not even sure LLMs are a good thing used sparingly (I know not everyone will agree with me here; just try to accept my peculiar form of digital veganism, I guess).
I’m not sure how that follows since Willison really is using LLM’s a lot and has legitimate reason to do so. Besides being a programmer, he’s become a prominent influencer (for lack of a better...
I’m not sure how that follows since Willison really is using LLM’s a lot and has legitimate reason to do so. Besides being a programmer, he’s become a prominent influencer (for lack of a better word) who reports LLM-related news.
Just the entire concept of a business comparing their own rates against themselves to show you what a great deal you got. I have no opinion on how this person actually uses LLMs.
Just the entire concept of a business comparing their own rates against themselves to show you what a great deal you got. I have no opinion on how this person actually uses LLMs.
That’s the “API price” which is the standard rate, and it’s quite high for Anthropic. The subscriptions are heavily discounted, but have restrictions, like you have to use Claude Code. I’m...
That’s the “API price” which is the standard rate, and it’s quite high for Anthropic. The subscriptions are heavily discounted, but have restrictions, like you have to use Claude Code.
I’m actually paying the standard rate via an intermediary, which is why I’ve been trying out Chinese competitors recently, GLM and DeepSeek.
And all it took was SpaceX selling them compute as a significant discount! Sure is nice of them to do that for exactly two months to help juice the quarter. No colluding or circular finance at all!
Anthropic are strongly rumored to be about to have their first profitable quarter.
Apparently this year the world’s managers have gone mad and decided it’s very important to spend, spend, spend on AI. It’s fascinating technology but I don’t really see what the hurry is? Can it...
Apparently this year the world’s managers have gone mad and decided it’s very important to spend, spend, spend on AI. It’s fascinating technology but I don’t really see what the hurry is?
Can it last? Hell if I know. When I joined Google I wondered if buying Internet advertising was a bubble since I never clicked on the ads myself. Turns out, not so much.
I think it will be difficult to predict the end of the bubble if that’s what it is. It seems more likely we’re somewhere in the middle of it.
This shift towards B2B was certainly predictable. While API costs can be burdensome, they are likely still lower than the cost of hiring a developer. Following this logic, a company needs an...
This shift towards B2B was certainly predictable.
While API costs can be burdensome, they are likely still lower than the cost of hiring a developer. Following this logic, a company needs an architect who directs an AI to write the code, rather than an entire team.
Based on my company’s current usage, they’re projected to spend 50x my yearly gross salary on AI (and I’m paid very well) this year. Or 100x more junior employees’ doing the same job. We’ve only...
Based on my company’s current usage, they’re projected to spend 50x my yearly gross salary on AI (and I’m paid very well) this year. Or 100x more junior employees’ doing the same job. We’ve only been using it short of three months.
I can’t see how this is sustainable, considering that the prices hikes have only just begun. From where I’m sitting, I don’t see the return on investment. Frankly what I’m seeing instead are hits to morale, more frustrated & anxious colleagues, and lower overall productivity.
Salary is not the only number you have to consider. Time is also important. An AI can make in seconds what a human can make in days. I may pay an AI 50x a developer, but I probably get 1000x...
Salary is not the only number you have to consider.
Time is also important.
An AI can make in seconds what a human can make in days.
I may pay an AI 50x a developer, but I probably get 1000x outcomes.
So if we look at the result, I am saving money.
I think this is what concerns me the most about rising AI use. The numbers do seem to make a lot of business sense in the near term. Thinking about what it takes to fill out an entire dev team,...
I think this is what concerns me the most about rising AI use. The numbers do seem to make a lot of business sense in the near term. Thinking about what it takes to fill out an entire dev team, there are a lot of steps, and a lot of potential for failure. There's selecting engineers with good fit, bringing them up to speed with the codebase, and retention. You also need seniors or leads to code review and drive good practices. It's a lot.
Compare that to the example of a single (or smaller group) or architects that are guiding LLMs. The architects are likely doing a lot of the same tasks that the senior/lead engineers are doing in the previous scenario, and the "developers" are much faster, and their "salary" comes with none of the overhead of humans (ramp up time, benefits, culture fit, etc).
If we're only looking at this snapshot it makes a ton of sense to shift to AI. But I think it all falls apart when you consider the background of the architect. The architect is able to suggest good patterns and detect problems in generated code in large part because they spent years as a junior learning those hard lessons. But in our second scenario, the junior developer is largely replaced by LLMs. So fast forward 10/20/30 years when the architect is ready to retire. Where is the source of new architects to replace him?
This is an interesting problem for sure, and I’ve spent a lot of time thinking about it. On the other hand, we used to say that about assembly language programmers.
This is an interesting problem for sure, and I’ve spent a lot of time thinking about it. On the other hand, we used to say that about assembly language programmers.
Yes, I'm trying very hard not to fall in the trap of blocking progress due to fear. Maybe this is similar to the evolution of compilers and programming languages, and in the end we just use LLMs...
Yes, I'm trying very hard not to fall in the trap of blocking progress due to fear. Maybe this is similar to the evolution of compilers and programming languages, and in the end we just use LLMs all the way down and the human is just the product owner that says what they want and observes the results with no need to dig into the code. Java/C++ developers likely never need to dig into machine code after all. Maybe someone could make the argument that the compiler is making less efficient use of the machine code (similar to how a WYSIWYG HTML editor produces spaghetti HTML under the hood). But at the end of the day it's worth it for the time savings and the ability to work on higher level tasks. It's hard to say for sure if my concerns are really warranted, or if it's just a sign of progress.
People used to be able to make an argument that compilers were less efficient than hand tuned machine code, but nowadays, I’m pretty confident that all in all but the most unusual circumstances an...
People used to be able to make an argument that compilers were less efficient than hand tuned machine code, but nowadays, I’m pretty confident that all in all but the most unusual circumstances an optimizing compiler will far outperform a human.
What this means I think is that our best efforts from our brightest minds in the machine language space are now spent on optimizing compilers instead of writing bespoke machine code to solve individual problems.
This is what I see in the future: we will still have high level programming experts and mathematicians who know C or Rust or high dimensional mathematics in a deep way, but their efforts will be largely focused on improving coding agents who then in turn interface with “regular” developers, like compilers or interpreted languages do today.
Crucial differences are that compilers are deterministic, understandable and limited in their scope. LLMs are tools to generate next likeliest token based on their training data. There is no...
Crucial differences are that compilers are deterministic, understandable and limited in their scope.
LLMs are tools to generate next likeliest token based on their training data. There is no guarantee or accuracy or anything else and it will not fundamentally change with the current technology.
Personally I was only able to effectively utilize llms when I was already proficient in the domain, the quality of the result didn't matter or it was possible to trivially check the output. The problem is also the scope. Lllms can be used, badly, for everything.
I've been referring to this at work as "don't multiply by zero". Meaning, if you don't have some base level of skill, OR you don't put in a base level of effort, the powerful AI will multiple your...
Personally I was only able to effectively utilize llms when I was already proficient in the domain
I've been referring to this at work as "don't multiply by zero". Meaning, if you don't have some base level of skill, OR you don't put in a base level of effort, the powerful AI will multiple your zero by 1000 and you will get zero back out (slop).
My experience was more along the lines of large efficiecy gains in rare situations, eg setting up a project. The rest of the time the llm is somewhat helpful, decreased a lot by the need to...
My experience was more along the lines of large efficiecy gains in rare situations, eg setting up a project. The rest of the time the llm is somewhat helpful, decreased a lot by the need to recheck all of output thoroughly which takes time and is mentally taxing.
Now imagine: you are the CEO of a software company, or one whose core business depends on internally developed software, and you are about 40 years old. Do you care about what will happen in 30...
Now imagine: you are the CEO of a software company, or one whose core business depends on internally developed software, and you are about 40 years old.
Do you care about what will happen in 30 years?
Probably not. You are likely trying to make as much money as possible over the next 20 years so you can retire on a yacht and not give a damn about the rest of the world.
Regarding the concept of an architect, I have my own personal opinion here, and I may be wrong, but: someone with no computer science knowledge, but with a logical mind, could direct an AI to write code for them.
I believe having a simple foundation of how software communicates is enough to get an AI to understand what it needs to write.
I believe I am a concrete example: I have no computer science knowledge and a completely different background. Yet, I managed to get ChatGPT (OpenAI), the free version, to write pieces of code for me to modify certain WordPress functionalities.
I am not referring to HTML. I am referring to code, interaction between plugins (with each other and with WordPress), PHP functions, etcetera.
And if I succeeded with NO IT background, imagine what someone who has studied programming could get an AI to do.
My company has already started imposing token limits and sending out memos to reduce usage. And because our private equity enshittifier overlords said we didn't make enough money for them last...
My company has already started imposing token limits and sending out memos to reduce usage. And because our private equity enshittifier overlords said we didn't make enough money for them last year, nobody got raises and at this rate I highly doubt we're getting raises next year either - which will probably mean the collapse of the company. I'm already looking for the way out, just a matter of time now...
We've probably had a 15% turnover from the no raises stance already. There's been no replacements - they went all in on "AI will make up the differences". Nobody is happy except the one guy who's been harping about how amazing AI has been.
I can’t tell, we have a couple thousand users, and I don’t have access to granular data. Suppose the developers still all become “10x engineers” (which I am skeptical of) - in that case the...
I can’t tell, we have a couple thousand users, and I don’t have access to granular data. Suppose the developers still all become “10x engineers” (which I am skeptical of) - in that case the bottleneck simply shifts somewhere else. Project management, sales, design, testing, market research, HR, and on and on. There’s plenty of support functions.
Fine, all of them become 10x. Then the bottleneck will be the customer, and who’s going to buy all this software? My expendable income did not 10x, and this is how it is for most of everyone. And even if it did, I for one am not terribly excited about paying for such software.
Is it just going to be corporations paying each other then? Then I could argue that’s unsustainable from another point of view.
Maybe there's another way to look at it: there are people who control spending decisions, either because it's their own money (investors) or they control a budget (legislators, managers, pension...
Maybe there's another way to look at it: there are people who control spending decisions, either because it's their own money (investors) or they control a budget (legislators, managers, pension funds, etc). If they save money on labor, what do they spend it on?
Currently it seems like investors want to build data centers and managers want to pay for the computing, and I wouldn't have predicted so much enthusiasm. But what comes after that?
I don't see teaching and healthcare becoming less important.
If businesses don’t need labor & end up selling only to each other, then what are you & I going to eat? Are we both going to become teachers or healthcare professionals? What happens if I don’t...
If businesses don’t need labor & end up selling only to each other, then what are you & I going to eat? Are we both going to become teachers or healthcare professionals? What happens if I don’t have any desire or inclination towards either of those professions?
I don't know what's going to happen, but it seems like a mini-economy where businesses sell only to each other is basically a bubble. It could only work while investors are propping it up....
I don't know what's going to happen, but it seems like a mini-economy where businesses sell only to each other is basically a bubble. It could only work while investors are propping it up. Eventually it runs out of new sources of funding.
Eliminating all labor is a far-fetched hypothetical. Maybe putting together a list of all the kinds of labor that can't be automated any time soon would make that clearer?
Probably the wrong thread, but I have been tossing an idea around in my head. Suppose current LLM performance matches claims, and scales as claimed. Observe the same trend in hardware performance....
Probably the wrong thread, but I have been tossing an idea around in my head.
Suppose current LLM performance matches claims, and scales as claimed. Observe the same trend in hardware performance.
I am reminded of myths around early CGI studios:
After months of negotiations.... Alvy ran the numbers very seriously for the first time. He discovered to his dismay that Moore's Law still had not proceeded far enough to make computation of a movie cost effective, and so backed out of the deal, just as Pixar was spinning out from Lucasfilm. Alvy Pixar Myth 3
A few observations:
LLMs become better/faster than any team of human developers or older LLM
hardware continues to improve
running LLMs or employing developers has costs
Why not keep a product idea to oneself, and risk waiting a few years? Avoid wasting money in the meanwhile on current LLMs and humans.
Depending on one's forecast, it would be faster to implement, with better results, later. Faster presumably translates into cheaper.
Some ideas would be at risk of immediate copycats though. Only non-software restrictions like legislation, startup capital, access to customers or prerequisite data could feasibly block a new entrant. Some ideas might be differentiated by data quality also.
But it stands to reason that anything without those "moats" have no protection. And depending on how far LLM scaling is projected to go, even those may not matter.
In a market where you expect costs to drop later, it makes some sense to wait, but you do miss out in the meantime. So, you might want to buy something cheaper that will do for now? For a laptop,...
In a market where you expect costs to drop later, it makes some sense to wait, but you do miss out in the meantime. So, you might want to buy something cheaper that will do for now?
For a laptop, this might mean sticking with what you have or buying something reasonably priced like a Macbook Neo.
Similarly, I don’t see why businesses are spending money on AI like a drunken sailor now when AI’s are likely to get cheaper and better, but it still makes sense to use it more conservatively in the meantime.
But as you say, perhaps some companies need to move faster because there’s a market opportunity and competition for it. What opportunities would those be?
The water consumption for data centers is tiny compared to that of agriculture and generally a tiny percentage compared to what's available. See here or here from r/askengineers.
The water consumption for data centers is tiny compared to that of agriculture and generally a tiny percentage compared to what's available. See here or here from r/askengineers.
Proximity to power plants and data networks are both important factors in site location. The largest data centers won't be brownfield development though. We're not talking about dense urban...
Proximity to power plants and data networks are both important factors in site location. The largest data centers won't be brownfield development though. We're not talking about dense urban neighborhoods being displaced by data centers.
The issue of water being wasted has been widely misunderstood. Water is arguably the only resource on the planet that has had, still has, and will always have a continuous cycle. Water used to...
The issue of water being wasted has been widely misunderstood.
Water is arguably the only resource on the planet that has had, still has, and will always have a continuous cycle.
Water used to cool a data center will be discharged into a pipe, which ends up in a river, then eventually in some ocean, or is otherwise absorbed into the ground. By following its course, it will return to hydrate the soil or a water reservoir in the form of rain, because that is where it ends up regardless.
Water will only start to run out in a few thousand years, when the planet temperature including the atmosphere has risen so much that condensation at high altitudes is no longer possible, thus preventing the natural cycle. And no, I am not talking about global warming; I am talking about the natural evolution of the solar system. The Sun is expanding and it is actually in an explosive phase and sooner or later, in about 4.5 billion years, it will turn into a supernova and burn the entire solar system. But that is not our problem; we will be extinct.
I have digressed. In short, I meant to say that water is not the problem.
I think you might misunderstand why people are worried about water use. It is not that people believe that water will be used up, that we are going to drain our oceans, empty our lakes, that is...
I think you might misunderstand why people are worried about water use. It is not that people believe that water will be used up, that we are going to drain our oceans, empty our lakes, that is absurd. They are worried about local water reserves. Because moving water is not trivial. And cleaning water is not trivial. And desalinating water is not trivial. All are possible, yes, but not without effort, often on industrial scale.
So, when people complain that so and so industry is using a lot of water, it is not that they are stupid and think that the water will just go away. It is that the local limited water will be used, and that water will be locally more expensive, scarce, maybe even dirty to the point of being undrinkable.
Also, the planet heating up to the point that clouds can't form in a thousand years? What? The sun won't be running out of hydrogen in its core for billions of years, and only after will it start expanding at any pase that would lead to such drastic temeperature increases over such a short time. The current expansion slow. As in water-won't-be-a-problem-for-millions-of-years slow. Your understanding of our suns evolution and its effect on our water is simply not correct.
But the overwhelming majority of wasted water is agricultural. Data centers are a non-factor in every serious report I've read. Focusing on their water usage is a distraction. Their power...
But the overwhelming majority of wasted water is agricultural. Data centers are a non-factor in every serious report I've read.
Focusing on their water usage is a distraction. Their power consumption should arguably be the biggest point of contention.
I don't know whether that is true and not, I tgink it is, but it is not relevant to my comment. My reply was mainly to LostInBerlin's seeming misunderstanding of the water cyvle, the sun, peoples...
I don't know whether that is true and not, I tgink it is, but it is not relevant to my comment. My reply was mainly to LostInBerlin's seeming misunderstanding of the water cyvle, the sun, peoples genaral knowkedge on the water c^vle and why people might be aggrieved when they think their water is threatened.
But now that you brought it up: what does it matter that there is another bigger waste of water? That does not make concerns of water waste on smaller scale less justified. Yes, we should really really get around to reducing overwatering (and all the other absurd pravtices) that plague the agricultural system, but we should also ensure that datacenters are build where water isn't scarce and ensure that the water they might leak out isn't contaminaited.
Now, I d(n't believe that dataventers using water is an (and far from the) argument against building them. But they should be held to high standards, not just in their capacity as a datacebter, but as their part in the local and global encironment. Gotta build sustainably, not just effeciently.
Haha, English is my second language - the internet was my third parent growing up. And also, the keyboard on my phone is non-standard; each of the buttons have 5 modes, tapping and swiping into...
Haha, English is my second language - the internet was my third parent growing up. And also, the keyboard on my phone is non-standard; each of the buttons have 5 modes, tapping and swiping into one if the 4 corners. Letters are typed by tapping the keys, numbers btyswiping up and to the left on the upper row, symbols by usually by swiping down and to the right or left, with som few outliers. You'd probably be able to determine the layout of the swipeable characters by looking at the word with an errant character, and factoring in that I type with two thumbs, usually.
This is not true. The Sun will never go supernova. We can say this with extreme certainty. The only supernova it could trigger is one that requires a binary star system. The Sun will grow to a red...
Water will only start to run out in a few thousand years, when the planet temperature including the atmosphere has risen so much that condensation at high altitudes is no longer possible, thus preventing the natural cycle. And no, I am not talking about global warming; I am talking about the natural evolution of the solar system. The Sun is expanding and it is actually in an explosive phase and sooner or later, in about 4.5 billion years, it will turn into a supernova and burn the entire solar system. But that is not our problem; we will be extinct.
This is not true. The Sun will never go supernova. We can say this with extreme certainty. The only supernova it could trigger is one that requires a binary star system. The Sun will grow to a red giant approximately the size of the Earth's orbit. But that won't be for 4.5 billion years, as you noted.
It will not permanently lock away the world's water in "a few thousand years."
What is currently considered likely is that in ~250 million years, the continents will all merge into one supercontinent and the Sun's heat will be 2.5% more intense on Earth. This combination will lead to temperatures exceeding those that mammals can currently survive in, plus a huge drought in the enormous continental interior, leading to conditions that will make basically all land inhospitable.
But mammals haven't even existed for 250 million years, so who knows what will be going on then?
I am not going to comment on the entire article, but this bit is funny to me
This sounds like my mother when she explains how it was such a great deal to buy the baby clothes at Kohl's for no specific use because it is such a good deal that those clothes could have cost $80 but I got them for $20.
–––– (imagine a line here because I am bad at markdown on my phone)
I hope that the increasing prices do force people to stop throwing every single query at them.
Funnily enough, my office did get a memo from one of the higher ups asking how the Dev team could figure out how to lower our LLM usage because they are seeing all of these articles about price jumps. None of us are using them, we have no LLM licensing costs. The top brass just assume we all are.
No LLM usage? You wouldn't be hiring, are you?
We are allowed to if we wanted, but we are a small team in a relatively small company and the hard part of writing our code and fixing bugs are just cases where "this piece of data will never be X" ends up being "X" and we need to update the business rules to handle it. Often my PRs are in the realm of a dozen lines. It would regularly take me more effort to explain the fix to the LLM than it would take to just change the code myself.
Small company, small problems, all business rule kind of code issues rather than generating pipelines where it would have more benefit to me specifically.
I'm surprised. Pretty much everyone I know finds them useful, even if they're not going crazy running up huge bills.
Finding something useful as a tool in some aspects of your work is different from being forced to use something in all aspects of your work even where it does not fit.
I can see why someone working in an environment where it is the latter rather works in one where there is none of that.
Even good things can be a bad thing in sufficient quantities (see: tribbles). Though, much like tribbles, I'm not even sure LLMs are a good thing used sparingly (I know not everyone will agree with me here; just try to accept my peculiar form of digital veganism, I guess).
I’m not sure how that follows since Willison really is using LLM’s a lot and has legitimate reason to do so. Besides being a programmer, he’s become a prominent influencer (for lack of a better word) who reports LLM-related news.
Just the entire concept of a business comparing their own rates against themselves to show you what a great deal you got. I have no opinion on how this person actually uses LLMs.
That’s the “API price” which is the standard rate, and it’s quite high for Anthropic. The subscriptions are heavily discounted, but have restrictions, like you have to use Claude Code.
I’m actually paying the standard rate via an intermediary, which is why I’ve been trying out Chinese competitors recently, GLM and DeepSeek.
If you're on your phone, why did you specifically use en dashes? Four hyphens will do it:
I intended to use hyphens and apparently my phone keyboard has en dashes on that key
You can also use asterisks or underscores.
Good to know! I thought it was just hyphens and it doesn't come up often enough for me to have researched
And all it took was SpaceX selling them compute as a significant discount! Sure is nice of them to do that for exactly two months to help juice the quarter. No colluding or circular finance at all!
And Amazon's tokensmaxxing!
If that revenue trend doesn’t reverse, they’re not going to need a discount.
Sure, but it will. They are all just racing to IPO and cashing in before it all falls apart.
I don’t share your confidence that it will be over that quickly.
Apparently this year the world’s managers have gone mad and decided it’s very important to spend, spend, spend on AI. It’s fascinating technology but I don’t really see what the hurry is?
Can it last? Hell if I know. When I joined Google I wondered if buying Internet advertising was a bubble since I never clicked on the ads myself. Turns out, not so much.
I think it will be difficult to predict the end of the bubble if that’s what it is. It seems more likely we’re somewhere in the middle of it.
This shift towards B2B was certainly predictable.
While API costs can be burdensome, they are likely still lower than the cost of hiring a developer. Following this logic, a company needs an architect who directs an AI to write the code, rather than an entire team.
Based on my company’s current usage, they’re projected to spend 50x my yearly gross salary on AI (and I’m paid very well) this year. Or 100x more junior employees’ doing the same job. We’ve only been using it short of three months.
I can’t see how this is sustainable, considering that the prices hikes have only just begun. From where I’m sitting, I don’t see the return on investment. Frankly what I’m seeing instead are hits to morale, more frustrated & anxious colleagues, and lower overall productivity.
Salary is not the only number you have to consider.
Time is also important.
An AI can make in seconds what a human can make in days.
I may pay an AI 50x a developer, but I probably get 1000x outcomes.
So if we look at the result, I am saving money.
I think this is what concerns me the most about rising AI use. The numbers do seem to make a lot of business sense in the near term. Thinking about what it takes to fill out an entire dev team, there are a lot of steps, and a lot of potential for failure. There's selecting engineers with good fit, bringing them up to speed with the codebase, and retention. You also need seniors or leads to code review and drive good practices. It's a lot.
Compare that to the example of a single (or smaller group) or architects that are guiding LLMs. The architects are likely doing a lot of the same tasks that the senior/lead engineers are doing in the previous scenario, and the "developers" are much faster, and their "salary" comes with none of the overhead of humans (ramp up time, benefits, culture fit, etc).
If we're only looking at this snapshot it makes a ton of sense to shift to AI. But I think it all falls apart when you consider the background of the architect. The architect is able to suggest good patterns and detect problems in generated code in large part because they spent years as a junior learning those hard lessons. But in our second scenario, the junior developer is largely replaced by LLMs. So fast forward 10/20/30 years when the architect is ready to retire. Where is the source of new architects to replace him?
edit:typo
This is an interesting problem for sure, and I’ve spent a lot of time thinking about it. On the other hand, we used to say that about assembly language programmers.
Yes, I'm trying very hard not to fall in the trap of blocking progress due to fear. Maybe this is similar to the evolution of compilers and programming languages, and in the end we just use LLMs all the way down and the human is just the product owner that says what they want and observes the results with no need to dig into the code. Java/C++ developers likely never need to dig into machine code after all. Maybe someone could make the argument that the compiler is making less efficient use of the machine code (similar to how a WYSIWYG HTML editor produces spaghetti HTML under the hood). But at the end of the day it's worth it for the time savings and the ability to work on higher level tasks. It's hard to say for sure if my concerns are really warranted, or if it's just a sign of progress.
People used to be able to make an argument that compilers were less efficient than hand tuned machine code, but nowadays, I’m pretty confident that all in all but the most unusual circumstances an optimizing compiler will far outperform a human.
What this means I think is that our best efforts from our brightest minds in the machine language space are now spent on optimizing compilers instead of writing bespoke machine code to solve individual problems.
This is what I see in the future: we will still have high level programming experts and mathematicians who know C or Rust or high dimensional mathematics in a deep way, but their efforts will be largely focused on improving coding agents who then in turn interface with “regular” developers, like compilers or interpreted languages do today.
Crucial differences are that compilers are deterministic, understandable and limited in their scope.
LLMs are tools to generate next likeliest token based on their training data. There is no guarantee or accuracy or anything else and it will not fundamentally change with the current technology.
Personally I was only able to effectively utilize llms when I was already proficient in the domain, the quality of the result didn't matter or it was possible to trivially check the output. The problem is also the scope. Lllms can be used, badly, for everything.
I've been referring to this at work as "don't multiply by zero". Meaning, if you don't have some base level of skill, OR you don't put in a base level of effort, the powerful AI will multiple your zero by 1000 and you will get zero back out (slop).
My experience was more along the lines of large efficiecy gains in rare situations, eg setting up a project. The rest of the time the llm is somewhat helpful, decreased a lot by the need to recheck all of output thoroughly which takes time and is mentally taxing.
Now imagine: you are the CEO of a software company, or one whose core business depends on internally developed software, and you are about 40 years old.
Do you care about what will happen in 30 years?
Probably not. You are likely trying to make as much money as possible over the next 20 years so you can retire on a yacht and not give a damn about the rest of the world.
Regarding the concept of an architect, I have my own personal opinion here, and I may be wrong, but: someone with no computer science knowledge, but with a logical mind, could direct an AI to write code for them.
I believe having a simple foundation of how software communicates is enough to get an AI to understand what it needs to write.
I believe I am a concrete example: I have no computer science knowledge and a completely different background. Yet, I managed to get ChatGPT (OpenAI), the free version, to write pieces of code for me to modify certain WordPress functionalities.
I am not referring to HTML. I am referring to code, interaction between plugins (with each other and with WordPress), PHP functions, etcetera.
And if I succeeded with NO IT background, imagine what someone who has studied programming could get an AI to do.
I am not a developer, and based on my experience so far: I’ll agree to disagree.
My company has already started imposing token limits and sending out memos to reduce usage. And because our private equity enshittifier overlords said we didn't make enough money for them last year, nobody got raises and at this rate I highly doubt we're getting raises next year either - which will probably mean the collapse of the company. I'm already looking for the way out, just a matter of time now...
We've probably had a 15% turnover from the no raises stance already. There's been no replacements - they went all in on "AI will make up the differences". Nobody is happy except the one guy who's been harping about how amazing AI has been.
How much is that per developer?
I can’t tell, we have a couple thousand users, and I don’t have access to granular data. Suppose the developers still all become “10x engineers” (which I am skeptical of) - in that case the bottleneck simply shifts somewhere else. Project management, sales, design, testing, market research, HR, and on and on. There’s plenty of support functions.
Fine, all of them become 10x. Then the bottleneck will be the customer, and who’s going to buy all this software? My expendable income did not 10x, and this is how it is for most of everyone. And even if it did, I for one am not terribly excited about paying for such software.
Is it just going to be corporations paying each other then? Then I could argue that’s unsustainable from another point of view.
Maybe there's another way to look at it: there are people who control spending decisions, either because it's their own money (investors) or they control a budget (legislators, managers, pension funds, etc). If they save money on labor, what do they spend it on?
Currently it seems like investors want to build data centers and managers want to pay for the computing, and I wouldn't have predicted so much enthusiasm. But what comes after that?
I don't see teaching and healthcare becoming less important.
If businesses don’t need labor & end up selling only to each other, then what are you & I going to eat? Are we both going to become teachers or healthcare professionals? What happens if I don’t have any desire or inclination towards either of those professions?
I don't know what's going to happen, but it seems like a mini-economy where businesses sell only to each other is basically a bubble. It could only work while investors are propping it up. Eventually it runs out of new sources of funding.
Eliminating all labor is a far-fetched hypothetical. Maybe putting together a list of all the kinds of labor that can't be automated any time soon would make that clearer?
Probably the wrong thread, but I have been tossing an idea around in my head.
Suppose current LLM performance matches claims, and scales as claimed. Observe the same trend in hardware performance.
I am reminded of myths around early CGI studios:
A few observations:
Why not keep a product idea to oneself, and risk waiting a few years? Avoid wasting money in the meanwhile on current LLMs and humans.
Depending on one's forecast, it would be faster to implement, with better results, later. Faster presumably translates into cheaper.
Some ideas would be at risk of immediate copycats though. Only non-software restrictions like legislation, startup capital, access to customers or prerequisite data could feasibly block a new entrant. Some ideas might be differentiated by data quality also.
But it stands to reason that anything without those "moats" have no protection. And depending on how far LLM scaling is projected to go, even those may not matter.
In a market where you expect costs to drop later, it makes some sense to wait, but you do miss out in the meantime. So, you might want to buy something cheaper that will do for now?
For a laptop, this might mean sticking with what you have or buying something reasonably priced like a Macbook Neo.
Similarly, I don’t see why businesses are spending money on AI like a drunken sailor now when AI’s are likely to get cheaper and better, but it still makes sense to use it more conservatively in the meantime.
But as you say, perhaps some companies need to move faster because there’s a market opportunity and competition for it. What opportunities would those be?
So when we run out of clean water and green space for data centers how will ai fit into the market then?
The water consumption for data centers is tiny compared to that of agriculture and generally a tiny percentage compared to what's available. See here or here from r/askengineers.
We're not running out of space for data centers. There's plenty of empty space.
Right! That's why theyre going for the space in peoples neighborhoods and game lands!
Proximity to power plants and data networks are both important factors in site location. The largest data centers won't be brownfield development though. We're not talking about dense urban neighborhoods being displaced by data centers.
The issue of water being wasted has been widely misunderstood.
Water is arguably the only resource on the planet that has had, still has, and will always have a continuous cycle.
Water used to cool a data center will be discharged into a pipe, which ends up in a river, then eventually in some ocean, or is otherwise absorbed into the ground. By following its course, it will return to hydrate the soil or a water reservoir in the form of rain, because that is where it ends up regardless.
Water will only start to run out in a few thousand years, when the planet temperature including the atmosphere has risen so much that condensation at high altitudes is no longer possible, thus preventing the natural cycle. And no, I am not talking about global warming; I am talking about the natural evolution of the solar system. The Sun is expanding and it is actually in an explosive phase and sooner or later, in about 4.5 billion years, it will turn into a supernova and burn the entire solar system. But that is not our problem; we will be extinct.
I have digressed. In short, I meant to say that water is not the problem.
I would be more concerned about other effects.
I think you might misunderstand why people are worried about water use. It is not that people believe that water will be used up, that we are going to drain our oceans, empty our lakes, that is absurd. They are worried about local water reserves. Because moving water is not trivial. And cleaning water is not trivial. And desalinating water is not trivial. All are possible, yes, but not without effort, often on industrial scale.
So, when people complain that so and so industry is using a lot of water, it is not that they are stupid and think that the water will just go away. It is that the local limited water will be used, and that water will be locally more expensive, scarce, maybe even dirty to the point of being undrinkable.
Also, the planet heating up to the point that clouds can't form in a thousand years? What? The sun won't be running out of hydrogen in its core for billions of years, and only after will it start expanding at any pase that would lead to such drastic temeperature increases over such a short time. The current expansion slow. As in water-won't-be-a-problem-for-millions-of-years slow. Your understanding of our suns evolution and its effect on our water is simply not correct.
But the overwhelming majority of wasted water is agricultural. Data centers are a non-factor in every serious report I've read.
Focusing on their water usage is a distraction. Their power consumption should arguably be the biggest point of contention.
I don't know whether that is true and not, I tgink it is, but it is not relevant to my comment. My reply was mainly to LostInBerlin's seeming misunderstanding of the water cyvle, the sun, peoples genaral knowkedge on the water c^vle and why people might be aggrieved when they think their water is threatened.
But now that you brought it up: what does it matter that there is another bigger waste of water? That does not make concerns of water waste on smaller scale less justified. Yes, we should really really get around to reducing overwatering (and all the other absurd pravtices) that plague the agricultural system, but we should also ensure that datacenters are build where water isn't scarce and ensure that the water they might leak out isn't contaminaited.
Now, I d(n't believe that dataventers using water is an (and far from the) argument against building them. But they should be held to high standards, not just in their capacity as a datacebter, but as their part in the local and global encironment. Gotta build sustainably, not just effeciently.
Off topic, why is your spelling the way it is?
Haha, English is my second language - the internet was my third parent growing up. And also, the keyboard on my phone is non-standard; each of the buttons have 5 modes, tapping and swiping into one if the 4 corners. Letters are typed by tapping the keys, numbers btyswiping up and to the left on the upper row, symbols by usually by swiping down and to the right or left, with som few outliers. You'd probably be able to determine the layout of the swipeable characters by looking at the word with an errant character, and factoring in that I type with two thumbs, usually.
This is not true. The Sun will never go supernova. We can say this with extreme certainty. The only supernova it could trigger is one that requires a binary star system. The Sun will grow to a red giant approximately the size of the Earth's orbit. But that won't be for 4.5 billion years, as you noted.
It will not permanently lock away the world's water in "a few thousand years."
What is currently considered likely is that in ~250 million years, the continents will all merge into one supercontinent and the Sun's heat will be 2.5% more intense on Earth. This combination will lead to temperatures exceeding those that mammals can currently survive in, plus a huge drought in the enormous continental interior, leading to conditions that will make basically all land inhospitable.
But mammals haven't even existed for 250 million years, so who knows what will be going on then?