Developer AI Token Costs Could Exceed Their Salaries in Two Years (infoworld.com) 127
"Enterprises may soon be paying as much for their developers' AI token usage as they do for their salaries," writes InfoWorld:
According to Gartner, these costs will meet, or even exceed, the typical software engineer's monthly salary within the next two years. This is not only because developers are increasingly adopting generative AI and agentic tools, it reflects a trend toward consumption-based licensing models as vendors balance infrastructure investments with profitability...
Gartner senior principal analyst Nitish Tyagi explained that it's important to note that Gartner's prediction is based on a global average salary of $2,000 per month; it doesn't mean AI token usage will exceed all salaries. For instance, in the US, yearly pay rates can be six digits or more. However, that kind of spend is not out of the realm of possibility, Tyagi emphasized. "I have heard scary numbers like 'My developer consumed $20K last month,' or 'A business user consumed $32K'."
If these amounts sound shocking, that's the point. "The goal is to alarm the industry about the impact of token cost if it is not governed and controlled," he said... AI coding vendors have yet to deliver "mature, built-in cost optimization capabilities," Tyagi said, and prices will likely only continue to rise as vendors further build out their models while at the same time trying to remain profitable. Thus, enterprises struggle to forecast and control costs, and, because AI is moving so fast, many organizations lack the "maturity and frameworks" to determine ROI, he noted. Agent-driven workflows are difficult to govern, context windows become bloated, budgets are wiped out earlier than anticipated, and token spend becomes hard to justify....
"Without a governed engineering operating model, costs can escalate faster than the productivity gains these tools are designed to deliver," Tyagi said.
Gartner senior principal analyst Nitish Tyagi explained that it's important to note that Gartner's prediction is based on a global average salary of $2,000 per month; it doesn't mean AI token usage will exceed all salaries. For instance, in the US, yearly pay rates can be six digits or more. However, that kind of spend is not out of the realm of possibility, Tyagi emphasized. "I have heard scary numbers like 'My developer consumed $20K last month,' or 'A business user consumed $32K'."
If these amounts sound shocking, that's the point. "The goal is to alarm the industry about the impact of token cost if it is not governed and controlled," he said... AI coding vendors have yet to deliver "mature, built-in cost optimization capabilities," Tyagi said, and prices will likely only continue to rise as vendors further build out their models while at the same time trying to remain profitable. Thus, enterprises struggle to forecast and control costs, and, because AI is moving so fast, many organizations lack the "maturity and frameworks" to determine ROI, he noted. Agent-driven workflows are difficult to govern, context windows become bloated, budgets are wiped out earlier than anticipated, and token spend becomes hard to justify....
"Without a governed engineering operating model, costs can escalate faster than the productivity gains these tools are designed to deliver," Tyagi said.
software engineer's $2,000 monthly salary (Score:5, Insightful)
Re: (Score:2)
Re:software engineer's $2,000 monthly salary (Score:5, Informative)
This is a GLOBAL average. For France the pay grade of 2 000 €/month (30 k€/year) would Junior level for a developer (look here for current job offers: as "developer", typical is 40 k€/year before taxes, remove 25% for taxes, then divide by 12: https://www.apec.fr/candidat/r... [www.apec.fr] ). For Southern Europe 2000 € is considered a good salary.
Re: (Score:2)
But they won't get a developer below 60k.
Job adverts like this are "the company is showing presence" only. This are not real jobs.
No one would apply for such a job, the money you get would not pay the rent for a flat.
For Southern Europe 2000 â is considered a good salary.
a) that is not southern Europe
b) it is not a good salary
c) it is a joke
Do you have mental problems?
Re: (Score:2)
Next year I will apply for such a job if in a location I like.
At the moment I making more but I plan to move towards my retirement place, then 2kE will be fine for me...
Re: (Score:2)
As per statistics (and my argument), entry-level developer gets you 2 600 € per month in France https://resoforces.fr/salaire-... [resoforces.fr]
For Southern Europe 2000 â is considered a good salary.
a) that is not southern Europe
b) it is not a good salary
c) it is a joke
It is an overall mid-career salary for Spain (Southern Europe).
Re: (Score:1)
France is not Spain.
Get a damn map.
Your link is not about software development entry jobs.
Learn to read.
No one is going into software development below 6k netto (after taxes and pension and health insurance).
You are simply misinformed or your French sucks.
Re: (Score:3)
Re: (Score:2)
One small detail: in the entire EU, they have national healthcare systems. There's no need to spend hundreds a month for "insurance".
Re: (Score:2)
Re: (Score:2)
almost everywhere. ofc you won't get a posh flat in st. germain, and you might have a hard time saving but ... living? sure. and mostly with good quality of life.
to almost every european the cost of living in the us appears just ludicrous. you guys are used to it, hence the big paychecks. your expats are just amazed, most of them ain't coming back.
ofc things might change in the not so distant future ...
Re: (Score:3)
In France (and of course elsewhere), the cost for an employee is not the net income of the employee. It is the gross income plus whatever taxes and benefits the employer pays. My bet is an employer in France for a 30k€/year salary is paying closer to 45k (or nearly 4000 per month, not 2000). This is what an employer would care about.
Re: (Score:2)
Re: (Score:2)
LMAO....Gartner senior principal analyst Nitish Tyagi explained that it's important to note that Gartner's prediction is based on a global average salary of $2,000 per month; What was he smoking to come up with that lowball # ???
AI isn't replacing some senior engineers in your cushy little western office. It's replacing shitting meaningless tasks that you outsource to the lowest bidder. What's he smoking? He's smoking India. Senior engineer is earning at the high end 30LPA. that's 3,000,000 rupees. That is 32,000 USD / year or ~ $2500 per month.
That is a high end estimate of the most senior engineers. A mid level engineer is earning more like $1000 per month, and entry level engineer as little as $250.
One of my employees has the sa
Re: (Score:2)
Re: software engineer's $2,000 monthly salary (Score:2)
It isn't just bullshit on that end, it's bullshit on the expense end. Most businesses are on $20/user/month plans, some individuals maybe on $100/$200
Where the fuck is $2000/month/user coming from? FTA:
I have heard scary numbers
Uh huh. Where did those come from. Ticking off every AI upgrade option from every SaaS product you rent, AI summaries of AI summaries of AI code reviews everywhere? I don't think THAT even gets you to $2000 a user. Can you burn $2000 if you try, absolutely, is there a story there... well if we don't know the
Re: software engineer's $2,000 monthly salary (Score:2)
Pretty much. We use Claude where I work. We are on some kind of budget enterprise plan. So we are judicious on what we use it on. If we hit the limit for the month then we wait or request an increase with justification.
Also good because we keep using our brains instead of outsourcing everything to the robots, and that is going to be critical going forward imo.
Re: (Score:2)
You do realize that the AI companies are taking a loss on those plans?
We used to be on some fixed-rate-burn-all-the-forests-you-want-to-power-the-stack-of-GPUs plan, but then the vendor switched to charging per token. Costs went up 10x overnight. We switched to another fixed-rate plan elsewhere but they will have to switch to charging per token before long or go bankrupt.
And the more that companies sack humans and replace them with AI, the more locked in they become to AI companies who can increase fees at
Re: software engineer's $2,000 monthly salary (Score:2)
Re: (Score:2)
The global average income is around $13-15k USD per year. That includes the trillionaire to the person working the paddy fields for 10 cents a day.
That's the problem with averages.
But it also includes high-tech workers in India, who may make half that, as well as plenty of entry lev
Re: (Score:2)
Humanity really isn't set up for the current situation. Our brains don't work like that. The problem is that we evolved to function in small tribes fighting over water and herds of animals. We need to be functioning like a global tribe where everybody raises everybody else's wages so that the ruling elite that ran Epstein Island can't play us against each other like crabs in a bucket.
Re: (Score:2)
Did you miss the *global* part? How much do you think an indian Software engineer makes a month?
Yep! 2k per month (Score:2)
pay peanuts get monkeys (Score:2)
I guess all you are going to get is a vibe coder if you're only paying $2k
Re: pay peanuts get monkeys (Score:2)
Tooling exceeds Machinist Cost (Score:1)
Airplane crews are Shocked
Machinists are Shocked
Radiologists are Shocked
Tower Crane Operators are Shocked.
Trading Firm Analysts are Shocked.
Garbage Men are Shocked.
Concreate Drivers are Shocked.
Re: (Score:3)
AI is opex, not capex though. You get to pay that cost every year. You are renting time on a machine not buying it. A well taken care of lathe lasts 70+ years, so the $50k you spend on it now pays dividends for as long as you keep it, and you can sell it used when you upgrade. The money on AI tooling is just gone. No machine shop would rent that lathe at $25k/yr
Re:Tooling exceeds Machinist Cost (Score:5, Informative)
Yeah, somebody let slip that out larger organization is blowing $300k/mo on AI a while back when somebody else complained that our smaller section bought a nice server with a bunch of GPU for $50k. We get along just fine with our one-time purchase and really don't see the need to rent servers.
We ran into something similar where we were looking at the cloud compute cost for a project that we could buy a better server every month for the price we were paying and had a bunch of folks trying to keep us in the cloud. At the end of the day it always boils down to that they don't want to have employees to maintain the physical infrastructure and it's mind-blowing how much people are willing to blow on renting this stuff.
Re: (Score:2)
You can generally tell when the true operational cost, including cost of capital, significantly exceeds employee cost by looking at whether they pay people to work in the middle of the night.
There are quite a lot of places where that happens. Just using the OP's list:
Airplanes: there are fewer flights at night, but that's when a lot of required maintenance happens. The Internet tells me the average lease on a Boeing 737 is around a few hundred thousand USD per month. Bigger planes would have even bigger dif
Re: (Score:2)
That's a big "if". Most studies show a 50-100% increase in code output at best, and code produced with AI assistance requires more rework. Measuring end-to-end, AI increased developer productivity by 10-30% or so. Sure, the tools are still improving, and developers are still learning how to use them to best advantage, but we're a long long way off from a x20 increase. That will likely require a breakthrough rather than incremental improvement of AI
Re: (Score:2)
The most important difference in all your examples is things like tooling and equipment are either strictly necessary to complete the job, or produce so much value in terms of productivity that they are worth the cost. Also, most tooling and equipment lasts longer than an AI token so the cost tends to get spread our over several jobs...
Using AI coding agents has not proven to increase quality or productivity in any meaningful way - increased volume of code does not mean productivity unless you're a middle m
Re: (Score:2)
If you the same programmer can be x20 more productive while efficient
If wishes were fishes.
Maybe you can be 20x more productive, but your code is unmaintainable and big pieces of it will have to be thrown away later because it's faster than fixing it. That would be shocking to people building airplanes.
Everything new is old again (Score:2)
Seems like we're circling back to The Feeling of Power [wikipedia.org].
Re: (Score:2)
"humans live in a computer-aided society and have forgotten the fundamentals of mathematics, including even the rudimentary skill of counting."
Can relate. Must read. Forgot how though. I'll have my computer phone read it to me.
Peddlers of doomsday machines (Score:2)
The peddlers of doomsday machines and their cronies are making random pseudo-predictions to market how scaaaary and all-powerful their doomsday machines are!
P.S.: send more venture capital to throw into the furnace plxplx
I'm shocked! (Score:2)
Shocked, I tell you! /sarc
It seems that at some point, some parasites end up feeding on others of their own kind.
I wonder what will happen to all the parasites when the non-parasite host at the bottom of the food chain is ultimately sucked dry...
Subsidies can't last forever (Score:4, Insightful)
Investor funds and market-share-fights have been subsidizing the cost of AI for end-users. Even if the predictions in the article are exaggerated, the Day(s) of Reckoning will eventually arrive and there will be lots of WTFs.
Re:Subsidies can't last forever (Score:5, Interesting)
The Wall Street Journal had a story a weekend ago about how OpenAI and Anthropic are getting pressure to cut their prices, that will certainly ding their path to profitability and thus their stock prices (were it to come to pass). Apparently there are some Open Source AI thingies out there and the Chinese are always ready with a cheaper product. Some companies are starting to use them.
I suppose we'll get another Idiot-Gram from Project 2025, foghorned into the blathersphere by la Presidenta about how using Chinese AI could turn the users into Chinese eating Chinese food or Ohio pets, or using Communist AI that doesn't have an upfront price. Those sneaky Antifa operatives will stop at nothing to screw White America. They've already destroyed the reflecting pool, they'll be coming by your homes to pee in your pools as well.
Re: (Score:1)
> and the Chinese are always ready with a cheaper product.
Chinese gov't is likely subsidizing those also to gain market-share. So it's in the same boat, perhaps at a different pace though, as they don't have to care what Wallstreet wants
> Open Source AI thingies
Running on in-premises hardware? If cloud-based, the subsidization time-bomb may still play out.
Re: (Score:2)
China builds tools that work instead of chasing bullshit. So their stuff is actually viable without trillionaires.
"The above statistics include all user requests from web, APP, and API. If all tokens were billed at DeepSeek-R1’s pricing (*), the total daily revenue would be $562,027, with a cost profit margin of 545%."
https://github.com/deepseek-ai... [github.com]
Re: (Score:2)
Yes and their current five year plan involves open sourcing models. That's to your advantage and causes the large companies headaches.
I think if things go according to Anthropic's vision we'll heading toward the cyberpunk dystopia where you rely on the mega corporations AI to be able to participate in society and usual people get just enough credits err tokens to survive their usual life without having the option to try something different.
I guess reality is never that dramatic, nevertheless I think China i
Re: (Score:2)
You know the cost (plus profit) of a server farm at scale pretty well as companies like AWS are offering them for longer than most people know what (current) AI is. The hourly price is expensive if your load is low, but the point in a profitable AI-API business is to make sure utilization is high. The challenge is managing the request queues such that a) The never get empty, so you can batch requests instead of processing batches before they are full so you pay hourly rates without processing requests b) do
and Gartner didn't even mention subsidies (Score:2)
They focused entirely on the cost to the AI consumers - which is not unreasonable since that's who Gartner's customers are.
But it's still a noteworthy omission for anyone with the slightest interest in the trajectory of the LLM industry as a whole. If OpenAI / Anthropic / XAI are providing their services at a loss, and their customers are consuming those services at a loss, then the GPU and RAM vendors are the only winners.
Local LLMs (Score:2)
So just run your own inference infrastructure then?
A traditional "developer" laptop generally costs around $2000-3000 for high end engineers. An extremely capable inference laptop, otherwise doing the rest of the same jobs, is around $5000-6000. One time cost. Everything runs locally, which also solves the "do I trust LLM vendor?" issue.
Re: Local LLMs (Score:4, Informative)
You need a huge model to even come close to what you get from claude type tools. Then you also need a huge context window so a tool like claude can even work when you direct it to your local llm. That means your developer laptop needs like 128gb of ram and most of that needs to be allocated to vram from your apu. Unless you are doing some experimental setup where you can split between vram and normal ram without taking a perf hit. Either way though, you are going to be generating tokens quite slowly. Enpecially with the power tdp of a laptop
A laptop with that much ram and a good apu (newest amd strix) will be like over 5k these days i bet.
You can't just throw some tiny 12b sized model from ollama on and think it's equiv to claude. It very much wont be even close. You need something like the 100+ billion parameter models to even get a little close and still have memory to have a very large context window ( which will almost definitely be smaller than claude). Good luck doing that on a laptop.
There's a reason why ai providers are losing money ... The real world cost of actually usable and productive ai is high... Not just in hardware but energy too. Every less is basically a toy, and wont be usable for serious work beyond maybe the simplest tasks.
Re: (Score:3)
You don't put 128 GB RAM in your laptop. You put 256 GB VRAM in your GPU server and let your developers access it. The more it is used, the better it can utilize batching, so it becomes more efficient than every developer running AI on their one machine.
Re: (Score:2)
You do realize these top-tier models require way more than 256GB? I can run Qwen and gpt-oss models pretty well in 128GB but there's a lack of open models between there and 512GB to make it worth upgrading to 256.
Re: (Score:2)
The large gpt-oss is merely 120B. People run it highly quantized on their desktop computer (okay, with quite "interesting" builds with multiple GPUs ...). Most models that need more than 256GB are not available anyway with Kimi being a notable exception. And we were talking about replacing what developers may run on their laptops ... and there a highly quantized ~100 model is surely the upper limit and I'm not sure if the quality and speed is sufficient for a productivity gain. The productive "going local"
Re: (Score:2)
Most models that need more than 256GB are not available anyway
Yes, the most powerful models are not available anyway, so the best you can do on your own hardware is second or third-best. Great, you can be an also-ran!
The productive "going local" is (currently) not the GPU in the workstation. In five years it may be
You mean when the models are even bigger? Sigh.
Re: (Score:2)
Do you need to even bigger models? Currently hobbyists seem to agree that the Qwen 27B model beats several older three digit models. So why not having in five years a let's say 120B model that runs on your then current GPU with the capabilities of a two year old previous Frontier model, which is just "good enough" for your purposes? I think chasing newer models will sooner or later be like chasing the latest programming language or javascript framework. You can do it and they certainly have their advantages
Re: (Score:2)
Do you need to even bigger models? Currently hobbyists seem to agree that the Qwen 27B model beats several older three digit models.
Oh good, it beats older models! You can be out of date with just the RAM you have already!
I'm developiong C++ using superpowers locally... (Score:2)
OpenAI et al are doomed when enough companies realize they are being ripped off.
Re: (Score:2)
I tried (Score:5, Interesting)
I tried hard in the last 2 months: I bought Claude, came up with a project I thought seemed reasonable: JS-based rich-text editor with plugin system. Existing solutions (CKEditor, TinyMCE, Quill, etc etc) are old, unwieldy, sometimes proprietary, and modern browsers have many newly supported features... My goal was an HTML web component of /
I tried to carefully prompt. Before coding I used Claude to help research the issues involved (dead standards, browsers handling edges cases differently, generated HTML questions, etc.). Claude was thoughtful and reassuring. I knew it would be more complex than Claude kept insisting, but that's OK. As usual with LLMs, at first I was more than impressed, I was blown away.
Still, bugs. That's expected. Fixes were easy and it was amazing how Claude understood the issues. But the more I tested, the more the bugs proliferated. Some issues activated Claude to rewrite whole architectural parts of the codebase, which broke dependencies. Fixing the dependent stuff introduced new bugs. I slowly had to learn more and more about the implementation specifics. More and more I had to audit the code changes, revert, try again. Soon I found myself questioning Claude's approaches in what seemed to be subtle ways. At times I was forced to really dig in, and the code – which looked so clean and organized – was a true spaghetti mess. Out-of-date comments. Repeated blocks of functionality with small differences. Convoluted back-and-forth paths across files, functions, classes. Each plugin had drifted to requiring its own long list of specialized one-off supporting worlds of code. Basic browser functions got overwritten with convoluted bespoke mish-mash slop with long interruptions of exceptions work-arounds and crazy shit.
Maybe the thing works. But the bugs are brutal! Everything is delicate! I've lost track of what the hell is going on.
But all of this was very familiar! It all looked like what USED TO happen to me before I got experience. What happened when I instructed programming newbies to take a crack without supervision. What happened when someone paid $5 to Upwork for something the boss thought would be easy.
We are not there yet. Not even close. It is 1998 and we are using for layout with the "100% td width" work-around.
Re: (Score:2)
Claude should be your companion coder. Not THE coder. You're using the tool wrong. Let Claude do the mundane stuff. Your output should be at least 50% higher if you use the tool right.
Re: (Score:2)
I'm not saying this because I like it, I'm saying it because that's what's going on out there right now.
Riding the overhang for a while (Score:2)
The big failure is that you have to build a large number of narrow purpose AI agents to check things.
The list of things to check grows seemingly without bound, necessitating too many AI agents.
The larger issue is that the AI just links every part of the code base to every other part of the code base given enough time.
Ditto: have to throw away 1/2 of what Claude does (Score:5, Interesting)
The only reason I am glad it's there is that when I prompt it something...it gives a wrong answer that usually leads me to the correct one. It's pretty useless for the languages I know well. It's too unreliable to save me time. The only benefit is for languages and platforms I know nothing about. I will admit, when it speeds the process along greatly. Although even then, like Regex, which I used like 4x a year....it'll write a TERRIBLE RegEx that's 100 characters long, but it works and it jogs my memory well enough to fix it down to a 20 character RegEx like a how someone who isn't a moron would have written it. It also allows me to be braver in technology I don't work with...for better or worse. As they say...nothing is more dangerous than a guy with "a little" knowledge...and access to really sharp tools!
Re: (Score:2)
I've given up trusting it to write unit tests.
Claude is exceptionally good at inferring test scenarios, but in terms of implementation IMHO it writes tests that are far too much low level and focused on implementation details.
In my existing projects I almost always have my own test patterns and utilities that effectively provides a mini-framework at the abstraction level I consider appropriate. I can tell Claude to use these as basis for new tests and it helps immensely in getting the kind of test implementations I want.
Note though that the reason most
Re: (Score:2)
Just wait, it gets better! .md file based on the PRD file created by product (also using Claude). You polish this plan and start Claude executing it. Unfortunately Claude Code needs to be running on your laptop in order to orchestrat
The near future, probably in a month or two if your company is following everyone else's trajectory:
You are mandated to use Claude and a set of skills called the SDLC that agentify (not a word) the work you used to do. You use Claude to generate a mult-repo, mult-agent spec/plan
Re: (Score:2)
I tried hard in the last 2 months: I bought Claude, came up with a project I thought seemed reasonable: JS-based rich-text editor with plugin system.
There's your problem: Javascript.
I'm actually serious. LLMs in their current incarnation need a lot of guardrails, and I think they do much better with a very strict, statically-typed language.
If I did have to write Javascript and I wanted to use an LLM I would focus hardcore on extremely thorough and extensive unit tests, because that's the only way to provide the necessary guardrails. And I would closely scrutinize all test changes made... or maybe just mark the test code as read-only so the AI can
Re: (Score:2)
The model you use (and the level of effort you select) makes a big difference in the outcome. If you use a better model, you get a better result, but it also costs more tokens.
No developer at my company has written any code in many months, it's all 100% Claude, and it's actually working out pretty well for that.
Cheap solutions get replaced by cheap solutions (Score:3)
India should definitely be worried that their entire "outsourced economy" is about to turn belly up.
Re:Cheap solutions get replaced by cheap solutions (Score:4, Interesting)
I think the point being made here is that the "cheap solution" remains India. When you're paying entry level engineers $250US / month, and senior experienced engineers $2000 USD / month the whole point of this story is that AI can't compete.
Re: (Score:2)
Exactly this!
Sadly, I just lost my mod points. They don't last very long. :(
Re: (Score:2)
You might want to check your napkin math there because it will take the person in India 10x longer to do the same work.
Re: (Score:2)
You might want to check your napkin math there because it will take the person in India 10x longer to do the same work.
That's not a problem though. That's a business decision. I have been literally told that "It doesn't matter if our Indian office delivers a project incorrectly 5 times, it's still cheaper to this there."
The funny part is our Indian office outsource it to an Indian subcontractor to save costs, so while my Indian counterpart makes 1/6th of the salary I do (literally, I have an employee on my team with the same grade and same industry experience), the actual project costs are about 1/10th or 1/15th of what it
Cloak of Invisibility Wanted? (Score:2)
This morning's weird thought was that Adam Smith's great mistake was to make the invisible hand visible. It was a flawed idea in the first place, basically a claim that things would work out "for the best" without knowing the details, but it was really a justification for social Darwinism before Darwin. In a sense bankruptcy serves as the equivalent of death in evolutionary systems in the biological world.
Once the invisible hand had lost its cloak of invisibility all the scoundrels in the world leapt to the
So the point isn't to save money (Score:3)
AI breaks capitalism.
Try to imagine you're a trillionaire. You are incredibly rich and powerful. And it's never ever ever enough.
Now right now someone out there is furiously typing how Elon Musk doesn't really have a trillion dollars because it's based on the valuation of his companies. It's based on how many goods and services he can sell to people.
Now stop. Ask yourself. Do you think he hasn't realized that?
What do you think Elon Musk thinks about the basic reality that he is entirely dependent on a Schlub like you for his wealth and power and status? Do you think he is content with that situation?
I'll spell it out for the 12-year-olds here. No he's not. He is extremely and bitterly angry that a god king like himself requires you to consent to his godhood. Right now you're thinking I could just boycott him and he'd be brought down to my level. Never mind that that isn't really true let's imagine for a second that it is. You have power over him. He doesn't like that.
The trillionaires will spend any amount of money to break that dependency. They do not care if they have to pay 10 times the money for an AI programmer as long as they don't have to be nice to the human programmer in order to make them make their products or get them to buy their products.
And the problem is I don't know an easy way to explain this to people who think like 12-year-olds and that's 60% of this country based on reading levels.
We gave so much money and so much power to the people at the top that they do not care about profits anymore. They don't have to worry about it. No amount of loss has any effect on their standard of living because it's not money anymore it's power and we gave them unlimited power. Never mind for a second why we did that we still did it.
So they can use AI to impoverish all of us and even though it breaks capitalism they don't care anymore because they are trying to create a post capitalism society without socialism. And holy fuck that is a sentence way too complicated for the people who need to get the message and I do not know how to dumb it down.
Re: So the point isn't to save money (Score:2)
You sound angry and jealous.
And at that point (Score:2)
It would be more beneficial to hire a human if you do a cost benefit analysis. But the business types need to help investor money in the pot so they gotta say something
Just another reminder of the upcoming auctions (Score:3)
There's no way to interpret these costs, that nobody is ever going to be willing to pay, as a reminder that soon these companies are going to be bankrupt.
Every time I see an AI story like this, it makes me realize I really have no idea what the AI bubble hardware is actually like, and how it might be used after auction.
A few months from now you might find yourself at an auction where 4TB of faster-than-anything-you-have RAM might be for sale for $80, but of course it won't be in the usual DIMMs that any of your existing mobos can use, will it? What will it be, and how do we best exploit it?
Re: (Score:2)
As I've said before, governments will introduce Cash for Clankers programs so all that weapons-grade RAM and GPUs don't get into the hands of the proles.
And to bail out their friends, but mostly to ensure we can't buy it cheap.
Re: (Score:2)
it makes me realize I really have no idea what the AI bubble hardware is actually like
It's mostly just basically big GPUs with no video connectors, that form factor is the cheapest way to deliver the tech. As such it's really not useful for anything other than GPGPU. You could use the hardware to build render farms or something, but it's fundamentally only useful for similar projects. No doubt if you wanted to do nuclear blast modeling, it would be really handy, but it's not going to help with gaming.
Re: (Score:2)
Except that (Score:3)
As of today frontier models are heavily subsidised to get you hooked and dependent. Be smart. Use the current relatively low cost of tokens to build what you cannot and get the low complexity grunt work done by local models so in the next 3-5 years you won't need those frontier models for most things and when you do, for specific well-scoped work, the cost will be roughly the same or at least tolerable.
If everything you do and know depends on a frontier AI model be prepared to spend. We got a taste with Fable. DOUBLE the cost of the already expensive Opus.
Currently to run a 1M context window AI model in roughly 30B range you need two RTX6000 Blackwells. That's £20k - so when you think about your £180/m subscription consider this as the teaser rate. Get the value out of it ASAP while it's cheap.
But then again (Score:2)
Buying into the manager's dreams is expensive (Score:3)
Unless the upper management understands that AI is just a tool, not the objective, expenses will spin out of control.
And I'm just like the kid with Bruce Willis : I see incompetent engineers using AI everywhere.
Better idea: (Score:2)
Re: (Score:2)
Good luck with that suggestion. You might as well suggest that people stop using...cars.
Re: (Score:2)
Re: (Score:2)
Maybe you haven't tried AI recently.
In November, I decided I needed to convert my old jQuery genealogy website to Vue.js, which I didn't know at all. I did this entirely with GitHub Copilot and Claude. In the process, I learned in depth not only how to get what I want from AI, but also how Vue.js works. I can now write Vue.js on my own, but I don't need to do the typing because AI does such a good job. Yes, I have to review everything it does to make sure it's right, and yes, it makes lots of mistakes. But
Re: (Score:2)
Now, if you take someone like me, who is constantly creating new things or new ways of doing things (and no, I won’t describe them because of trade secrets), AI just gets in
Re: (Score:2)
Again, you are making it clear that you haven't actually used AI in any significant way. AI is not in fact limited to copying things other people have already done. This is a surmise that many have aimed at AI, without justification.
Yes, AI has its shortcomings, and you certainly can accidentally introduce security risks if you don't watch for this possibility. But that doesn't make it "not useful."
Creating "new ways of doing things" can be overrated. APIs and data entry forms are "predictable" because...pr
Re: (Score:2)
Re: (Score:2)
Time has already told. The slop you reference is real, there are many areas where AI has not yet reached a level of quality rivaling human output. But in code generation, for the more popular languages and frameworks at least, it has. In fact, the quality of AI-written code has actually surpassed that of many programmers. It's well-structured, well-written, readable, with good, meaningful names, and following good programming practices. Is it perfect? No, not at all. That's not the standard. The standard is
Re: (Score:2)
Re: (Score:2)
When specifically did you last try one of these experiments? If it was more than a handful of months ago, the landscape has changed significantly. Specifically, GitHub Copilot and Claude Code have improved dramatically, month by month, in recent months.
Tokenmaxxing (Score:2)
Some companies are pushing employees to use as much AI as they can, so they do exactly that.
https://en.wikipedia.org/wiki/... [wikipedia.org]
"or more"? (Score:2)
in the US, yearly pay rates can be six digits or more
Wait, what?
Take gun in both hand, aim at foot, shoot (Score:3)
At what point do companies decide that it's cheaper to simply hire people than pay for chatbots? Esp. since 90% of all companies that brought them in find no value - 5%-10% at most, but wasting more for people to fix the hallucinations.
Already does... (Score:2)
Who wants to be the cheapest developer on the team (Score:2)
Tokens are cheap (Score:2)
I expect that when laptops with 128GB RAM and a decent NPU happen, I'll start burning closer to 100 million tokens a day.
I don't think I could spend $2000 a year on tokens if I tried... Even a laptop is a 5 year purchase, so that can't cost m
Re: (Score:2)
This is really fucking stupid.
Re: (Score:3)
go woke, go broke
People keep saying this but is it actually true? What company actually went broke because of woke and specifically because of woke?
A week of bad PR maybe but beyond that? I think this is made up and there is evidence of the opposite direction as well but enough either way to make this more than just feel good rhetoric.
Re: (Score:2, Insightful)
Re: DEI costs MORE than tokens. Call them EMPTY TO (Score:2)
"There are lots of extremely successful companies who treat their people right"
Name one.
Re: (Score:2)
Costco
Re: (Score:2)
My company does this sort of theatre: partnering with non-profits that provide the look and feel of being a progressive and caring company, but ultimately they are beholden to capital and not to the employees in any way.
Is your company employee-owned, or do you believe the investors and management are actually good people who are motivated by more than increasing shareholder value?
"support community service by their employees" - that's called good will for free. You have to work against goals/deadlines/OKRs