[HN Gopher] AI capex is so big that it's affecting economic stat...
___________________________________________________________________
AI capex is so big that it's affecting economic statistics
Author : throw0101c
Score : 163 points
Date : 2025-07-18 19:59 UTC (3 hours ago)
(HTM) web link (paulkedrosky.com)
(TXT) w3m dump (paulkedrosky.com)
| oytis wrote:
| I just hope when (if) the hype is over, we can repurpose the
| capacities for something useful (e.g. drug discovery etc.)
| te_chris wrote:
| As the author says: this capex isn't a railroad, it's a very
| expensive immediately depreciating asset.
|
| We should be so lucky
| midnightclubbed wrote:
| Maybe there will be a benefit from the additional power
| generation and infrastructure that will still be available if
| the data centers don't need it.
|
| Can't believe I have to state the obvious and say that is
| only a potential gain if the power/cooling is from renewable
| sources. But I do
| intended wrote:
| That's a mighty big if.
| ch33zer wrote:
| We won't. Well lay off engineers to balance the books and
| destroy unneeded capacity.
| toomuchtodo wrote:
| Very Grapes of Wrath.
| teaearlgraycold wrote:
| A competent government might rent these company's GPUs to cure
| major diseases once they're sitting idle.
| Bluestein wrote:
| (or they'll end up mining crypto as the only viable
| alternative, crashing the whole tent )
| dgfitz wrote:
| Bitcoin still needs a-minin'.
| Retr0id wrote:
| GPUs haven't been useful for bitcoin mining for a long time.
| apwell23 wrote:
| yearly AI capex is more than NASA and NIH combined
| actionfromafar wrote:
| So, not very much, all things considered.
| lenerdenator wrote:
| Tell that to Congress.
| esseph wrote:
| That would be $75 billion combined for 2025. A drop in the
| bucket.
|
| -- > From 2013 to 2020, cloud infrastructure capex rose
| methodically--from $32 billion to $119 billion. That's
| significant, but manageable. Post-2020? The curve steepens.
| By 2024, we hit $285 billion. And in 2025 alone, the top 11
| cloud providers are forecasted to deploy a staggering $392
| billion--MORE than the entire previous two years combined.
|
| https://www.wisdomtree.com/investments/blog/2025/05/21/this-.
| ..
| armchairhacker wrote:
| Very cheap game consoles and VR headsets. Unironically that
| could really help world peace and QOL: less news and
| doomscrolling, and people would have an outlet for stress,
| anger, and boredom.
| lenerdenator wrote:
| I've never seen a group more easily angered than gamers.
|
| "What do you mean the women in this game have proportions
| roughly equivalent to what's actually possible in nature?!?!"
| drdaeman wrote:
| That's a _subset_ of gamers, specifically gamers who
| actively want a hypersexualized characters and still haven
| 't figured out how to ignore the irrelevant products and
| still chose to be angered if some product doesn't fit their
| requirements (despite no shortage of games that would fit
| the bill).
| intended wrote:
| The most trust and safety research I've seen - on things
| like recidivism, ban success, pro social behavior - comes
| from gaming.
|
| Dota, league, he'll - Roblox, twitch, discord - have some
| of the most data on how angry humans are when they play
| vidya.
| lm28469 wrote:
| Is this the technosolutionist version of "just one more lane
| and we'll fix traffic jams for good"?
| smokel wrote:
| You may be interested in alternative approaches such as
| playing chess, running, singing, dancing, or reading
| literature.
| amelius wrote:
| This is one of the more hopeful explanations of the Fermi
| paradox.
|
| https://en.wikipedia.org/wiki/Fermi_paradox#Hypothetical_exp.
| ..
|
| "Alien species may isolate themselves in virtual worlds"
| etherlord wrote:
| Have you read Ready Player 1 or watched the Matrix?
| 827a wrote:
| If you have a GPU which you've used for AI training, but that's
| no longer valuable, you could sell that GPU; but then you'd
| incur taxable revenue.
|
| If you destroy the GPU, you can write it off as a loss, which
| reduces your taxable income.
|
| Its possible you could come out ahead by selling everything
| off, but then you'd have to pay expensive people to manage the
| sell off, logistics, etc. What a mess. Easier to just destroy
| everything and take the write-off.
| phil21 wrote:
| Capital equipment is depreciated over time. By the time
| you're selling it off, it's pennies on the dollar and a small
| recoupment of cost. Paying 30% (or whatever) taxes on that
| small amount of income and having 70% remaining is still
| better than zero dollars and zero taxes.
|
| It is a giant pain to sell off this gear if you are using in-
| house folks to do so. Usually not worth it, and why things
| end up trashed as you state. If I have a dozen 10 year old
| servers to get rid of - it's usually not worth anyone's time
| or energy to list them for $200 on ebay and figure out
| shipping logistics.
|
| However, at scale the situation and numbers change - you can
| call in an equipment liquidator who can wheel out 500 racks
| full of gear at a time and you get paid for their disposal on
| top of it. Usually a win/win situation since you no longer
| have expensive people trying to figure out who to call to get
| rid of it, how to do data destruction properly, etc. This
| usually is a help to the bottom line in almost all cases I've
| seen, on top of it saving internal man-hours.
|
| If you're in "failed startup being liquidated for asset
| value" territory, then the receiver/those in charge typically
| have a fiduciary duty to find the best reasonable outcome for
| the investors. It's rarely throwing gear with residual value
| in the trash. See: used Aeron chair market.
| triceratops wrote:
| > that's no longer valuable, you could sell that GPU; but
| then you'd incur taxable revenue
|
| Unless GPUs are like post-Covid used cars you're going to
| sell them at a loss which can be written off. Write-offs
| don't have to involve destroying the asset. I don't know
| where you got that idea.
| barbazoo wrote:
| On the other hand, drug discovery sounds like it's a candidate
| for really benefitting from AI. To fuel AI model development,
| there maybe has to be all the garbage that comes with AI.
| lm28469 wrote:
| What drugs? The leading causes of death are routine induced,
| 80% of cancers are caused by your environment, &c.
|
| We have much better things to do with these billions
| ribosometronome wrote:
| Drugs to cure the diseases caused by your environment. It's
| not as if people are suddenly going to be making perfect
| decisions (e.g. never getting a sunburn, not eating meat,
| avoiding sugary foods).
| lm28469 wrote:
| So all these things that skyrocketed in the span of 75
| years are immutable facts of life, but magic drugs are
| somehow in the realm of possibilities?
|
| What's easier, educate your people and feed them well to
| build a strong and healthy nation OR let them rot and
| shovel billions to pharma corps in the hope of finding a
| magic cure?
| astrange wrote:
| Second one's easier. Technological improvements are
| always easier than social change.
|
| > shovel billions to pharma corps in the hope of finding
| a magic cure?
|
| What do you mean finding? We already found it (GLP-1
| inhibitors). Ozempic is even owned by a nonprofit (Novo
| Nordisk). See, everything's fine.
| ribosometronome wrote:
| >skyrocketed in the span of 75 years are immutable facts
| of life
|
| A number of them seem to have skyrocketed with quality of
| life and personal wealth. I suspect my ancestors were
| skinny not because they were educated on eating well but
| because they lacked the same access to food we have in
| modern society, especially super caloric ones. I don't
| super want to go back to an ice cream scarce world.
| Things like meat consumption are linked to colon cancer
| and most folk are unwilling to give that up or practice
| meat-light diets. People generally like smoking!
| Education campaigns got that down briefly but it was
| generally not because people didn't want to smoke, it's
| because they didn't want cancer. Vaping is clearly
| popular nowadays. Alcohol, too! The WHO says there is no
| safe amount of alcohol consumption and attributes lots of
| cancer to even light drinking. I suspect people would
| enjoy being able to regularly have a glass of wine or
| beer and not have it cost them their life.
| logicchains wrote:
| >Drugs to cure the diseases caused by your environment
|
| Humans have so far completely failed to develop any drug
| with minimal side effects to cure lifestyle diseases;
| it's magical to think AI can definitely do it.
| astrange wrote:
| Everything has side effects. In this case we have three
| pretty good interventions, Ozempic, FMT, and telling
| people to drink Coke Zero. The worst "side effect" is
| just that the first two are expensive.
|
| Oh, in this case GP seems to be including sunscreen as a
| treatment for lifestyle diseases. Pretty sure those don't
| have side effects, but Americans don't get the good ones.
| nerevarthelame wrote:
| Is your objection just over the word "cure"? Because
| hypertension, depression, arthritis, asthma are a few in
| an absurdly long list of lifestyle diseases that use
| drugs as a primary method of treatment.
| schmidtleonard wrote:
| Bullshit. Heart disease and cancer (and a long tail of
| medical problems) cook up with age and kill ~everyone
| inside ~100 years. If you think that environment and
| exercise can fix this, show me the person who is 200 years
| old.
|
| We would have to 100x medical research spending before it
| was clearly overdone.
| lm28469 wrote:
| You can 1000x the research if you want, a 50kg overweight
| person who doesn't exercise, drink alcohol and lives next
| to a highway is statistically fucked no matter what.
| You'd need straight up magic to undo the damages
|
| You're not going to fix lifestyle diseases with drugs,
| and lifestyle diseases are the leading cause of death
| conception wrote:
| Yes. Let's ignore 20% of people with cancer. That's cool.
| alphazard wrote:
| The rest of the world has not caught up to current LLM
| capabilities. If it all stopped tomorrow and we couldn't build
| anything more intelligent than what we have now: there would be
| years of work automating away toil across various industries.
| Terr_ wrote:
| Creating oodles of new jobs in internally QAing LLM results,
| or finding and suing companies for reckless outcomes. :p
| alphazard wrote:
| As long as liability is clearly assigned, it doesn't have
| an economic impact. The ambiguity of liability is what
| creates negative economic impact. Once it's assigned
| initially through law, then it can be reassigned via
| contract in exchange for cash to ensure the most productive
| outcome.
|
| e.g. if OpenAI is responsible for any damages caused by
| ChatGPT then the service shuts down until you waive
| liability and then it's back up. Similarly if companies are
| responsible for the chat bots they deploy then they can buy
| insurance or put up guard rails around the chat bot, or not
| use it.
| Terr_ wrote:
| > As long as liability is clearly assigned, it doesn't
| have an economic impact
|
| In a reality with perfect knowledge, complete laws always
| applied, and populated by un-bankrupt-able immortals with
| infinite lines of credit, yes. :P
| sterlind wrote:
| my experience using LLM-powered tools (e.g. copilot in agent
| mode) has been underwhelming. like, shockingly so. like not
| cd-ing to the wrong dir where a script is located, and
| getting lost, disregarding my instructions to run ./tests.ps1
| and running `dotnet test`, writing syntactically incorrect
| scripts and failing to correct them, particularly being
| overwhelmed by verbose logs. sometimes it even fails to
| understand the semantic meaning of my prompts.
|
| whereas my experience describing my problem and actually
| _asking_ the AI is much, much smoother.
|
| I'm not convinced the "LLM+scaffolding" paradigm will work
| all that well. sanity degrades with context length, and even
| the models with huge context windows don't seem to use it all
| that effectively. RAG searches often give lackluster results.
| the models fundamentally seem to do poorly with using
| commands to accomplish tasks.
|
| I think fundamental model advances are needed to make most
| things more than superficially automatable: better
| planning/goal-directed behavior, a more organic connection to
| RAG context, automatic gym synthesis, and RL-based fine
| tuning (that holds up to distribution shift.)
|
| I think that will come, but I think _if_ LLMs plateau here
| they won 't have much more impact than Google Search did in
| the '90s.
| break_the_bank wrote:
| I'm curious which was the model you used when you ran into
| the cd-ing bug?
|
| I'd give building with sonnet 4 a fair shot. It's really
| good, not accurate all the time but pretty good.
| baxtr wrote:
| Re hype: Why is it that so many people are completely obsessed
| with replacing all developers and any other white-collar job?
| They seem to be totally convinced that this will happen. 100%
|
| To me, this all sounds like an "end-of-the-world" nihilistic
| wet dream, and I don't buy the hype.
|
| Is it's just me?
| ToucanLoucan wrote:
| > Re hype: Why is it that so many people are completely
| obsessed with replacing all developers and any other white-
| collar job? They seem to be totally convinced that this will
| happen. 100%
|
| Because the only thing that gets the executive class hornier
| than new iPhone-tier products is getting to layoff tons of
| staff. It sends the stock price through the roof.
|
| It follows from there that an iPhone-tier product that also
| lets them layoff tons of staff would be like fucking catnip
| to them.
| astrange wrote:
| The job description of a developer is to replace themselves
| by automating themselves so they can get promoted/find a
| new more relevant role. That's the point of compilers and
| new programming languages.
|
| There's no such thing as taking people's jobs, nobody and
| nothing is going to take your job except for Jay Powell,
| and productivity improvements cause employment to increase
| not decrease.
| ivape wrote:
| People really hate late stage capitalism.
| noitpmeder wrote:
| Because developer and other white-collar job salaries are the
| top expense of most companies.
| kulahan wrote:
| Oh boy I can't wait until we get rid of the highest paying
| jobs!
|
| Your response doesn't explain why so many people are hyped
| about it, just why CEOs are.
| fnimick wrote:
| There's a lot of non-engineering people who are very
| happy to see someone else get unemployed by automation
| for a change. The people who formerly were automating
| others out of a job are getting a taste of their own
| medicine.
| linotype wrote:
| Of course that would cause a tremendous drop in demand
| for the services the schaudenfreuden folks provide,
| hurting them as well.
| JumpCrisscross wrote:
| > _that would cause a tremendous drop in demand for the
| services the schaudenfreuden folks provide, hurting them
| as well_
|
| You're correct. But it doesn't matter. Remember the San
| Francisco protests against tech? People will kill a
| golden goose if it's shinier than their own.
| oytis wrote:
| If this goose is also pricing others out of housing
| market it's not entirely unreasonable
| JumpCrisscross wrote:
| > _If this goose is also pricing others out of housing
| market it 's not entirely unreasonable_
|
| It's self-defeating but predictable. (Hence why the
| protests were tolerated to backed by NIMBY interests.)
|
| My point is the same nonsense can be applied to someone
| not earning a tech wage celebrating tech workers getting
| replaced by AI. It makes them poorer, _ceteris paribus_.
| But they may not understand that. And the few that do may
| not care (or may have a way to profit off it, directly or
| indirectly, such that it 's acceptable).
| oytis wrote:
| I don't quite follow. What exactly have non-tech people
| of San Francisco got from all the tech people working
| there? How did they become richer (ok, apart from
| landlords) or how would they become poorer if they lose
| their jobs?
| daedrdev wrote:
| The services will get cheaper, since most companies make
| much profit and the moat of high salaries will be gone.
| satyrun wrote:
| I am not an engineer and I expect my white collar job to
| be automated.
|
| The reason to be excited economically for this is if it
| happens it will be massively deflationary. Pretending
| CEOs are just going to pocket the money is economically
| stupid.
|
| Being able to use a super intelligence has been a long
| time dream too.
|
| What is depressing is the amount of tech workers who have
| no interest in technological advancement.
| overgard wrote:
| I'm not sure exactly what you mean by deflationary, but
| in general deflation in an economy is a very bad thing.
| The most significant period of economic deflation in the
| US was 1930-1933, ie, the great depression, and the most
| recent period was the great recession.
|
| And since when do business executives NOT pocket the
| money? Pretty much the only exception is when they
| reinvest the savings into the business, for more growth,
| but that reinvestment and growth usually is only
| something the rest of us care about if it involves
| hiring..
| citrin_ru wrote:
| White collar jobs as a white - yes, but even in software
| companies it is not uncommon when sales and marketing cost
| more than engineering.
| rpcope1 wrote:
| Notice how it wasn't and isn't a big deal when it's not in
| your own back yard (i.e. destroying blue collar professions);
| our chickens have just come home to roost. It's amazing the
| number of gullible, naive nerds out there that can't or won't
| see the forests for the trees. The number of ancap-lite
| libertarians precipitously drops when it's their own
| livelihood getting its shit kicked in.
| phil21 wrote:
| One of the more annoying parts about being in the tech
| community was listening to the tone-deafness of these sorts
| of folks. Coming from a very blue collar background and
| family.
|
| It's difficult to have much empathy for the "learn to code"
| crowd who seemingly almost got a sense of joy out of
| watching those jobs and lifestyles get destroyed. Almost
| some form of childhood high school revenge fantasy style
| stuff - the nerd finally gets one up on the prom king.
| Otherwise I'm not sure where the vitriol came from. Way too
| many private conversations and overheard discussion in the
| office to make me think these were isolated opinions.
|
| That said, it's not everyone in tech. Just a much larger
| percentage than I ever thought, which is depressing to
| think about.
|
| It's certainly been interesting to watch some folks who a
| decade ago were all about "only skills matter, if you can
| be outcompeted by a robot you deserve to lose your job"
| make a 180 on the whole topic.
| rightbyte wrote:
| No it is not just you. If true it would be beyond dystopic
| due to current political immaturity.
| ModernMech wrote:
| The dream of many business owners is running their business
| with no products, no employees, and no customers, where they
| can just collect money. AI promises to fulfil this dream. AIs
| selling NFTs to other AIs paying in crypto is the final boss
| of capitalism.
| eastbound wrote:
| > completely obsessed with replacing all developers
|
| I'm paid about 16x an electronics engineer. Salaries in IT
| are completely unrelated to the person's effort compared to
| other white collar jobs. It would take an entire career to
| some manager to reach what I made after 5 years. I may be
| 140IQ but I'm also a dumbass in social terms!
| ivape wrote:
| Jealousy is absolutely a thing.
| bcrosby95 wrote:
| That's cool. It sounds like most of us aren't making what
| you make. I don't make 16x what someone paid minimum wage
| makes, much less an electrical engineer.
| MyOutfitIsVague wrote:
| The median wage for programmers is about 5x the average
| minimum wage. You're paid significantly higher than the
| average programmer.
| phil21 wrote:
| Same reason so many people got excited in the early Internet
| days of how much work and effort could be saved by
| interconnecting everyone. So many jobs lost to history due to
| such an invention. Stock trading floors no longer exist, call
| centers drastically minimized, retail shopping completely
| changing, etc.
|
| I had the same thought you did back then. If I could build a
| company with 3 people pulling a couple million of revenue per
| year, what did that mean to society when the average before
| that was maybe a couple dozen folks?
|
| Technology concentrates gains to those that can deploy it -
| either through knowledge, skill, or pure brute force
| deployment of capital.
| 2944015603 wrote:
| > Why is it that so many people are completely obsessed with
| replacing all developers and any other white-collar job?
|
| For the same reason people are obsessed with replacing all
| blue-collar jobs. Every cent that a company doesn't have to
| spend on its employees is another cent that can enrich the
| company's owners.
| oytis wrote:
| The general view in my bubble was that blue collar jobs are
| seen as dumb, physically demanding and dangerous, so we
| kind of replacing them for their own good, so that they can
| do something intellectual (aka learn coding). Whereas
| intellectual labour is kind of what humans exist for, so
| making intellectual work redundant is truly the end of the
| world.
|
| Maybe it's my post-communist background though and not
| relevant for the rest of the world
| detourdog wrote:
| I have no problem with replacing jobs with automation. The
| question to ask is the investment justified by the
| efficiency gains?
|
| I'm skeptical that this is a good use of resources or
| energy consumption.
| overgard wrote:
| I find it weird and uncomfortable too. Like, why are there a
| bunch of people excited about mass unemployment?
| Ekaros wrote:
| It is not about mass unemployment. It is about efficiency.
| Producing the same thing for cheaper. Leading to cheaper
| products to buy.
|
| That is what allowed our current lifestyles. It is good
| thing. Now it is just coming to next area.
| rightbyte wrote:
| In a capitalist society?
| Ekaros wrote:
| It is the one thing I believe capitalism at some level
| works for. Invest capital to build something that gives
| competitive advantage over other capital. Buying newer
| bigger factory that allows producing more for cheaper to
| compete with others.
|
| With AI it is white collar work.
| rpdillon wrote:
| This is exactly it. I was talking to my wife about this
| this morning. She's a sociologist researcher, and a lot
| of people that work in her organization are responsible
| for reading through interviews and doing something called
| coding, where you look for particular themes and then tag
| them with a particular name associated with that theme.
| And this is something that people spend a lot of hours
| on, along with interview transcription, also done by
| hand.
|
| And I was explaining that I work in tech, so I live in
| the future to some degree, but that ultimately, even with
| HIPAA and other regulations, there's too much of a gain
| here for it not to be deployed eventually, And those
| people in their time are going to be used differently
| when that happens. I was speculating that it could be
| used for interviews as well, but I think I'm less
| confident there.
| LinXitoW wrote:
| It's allowed our lifestyle for a short period of time, to
| small amount of people. It's not given to poor people, to
| people in places we've bombed, or extracted resources
| from, or people in the future since it's destroying the
| planet.
|
| We're all far closer to poor than we are to having enough
| capital to live off of efficiency increases. AI is the
| last thing the capitalist class requires to finally throw
| of the shackles of humanity, of keeping around the filthy
| masses for their labor.
| overgard wrote:
| Ok, but who benefits from these efficiencies? Hint: not
| the people losing their jobs. The main people that stand
| to benefit from this don't even need to work.
|
| Producing things cheaper sounds great, but just because
| its produced cheaper doesn't mean it is cheaper for
| people to buy.
|
| And it doesn't matter if things are cheap if a massive
| number of people don't have incomes at all (or even a
| reasonable way to find an income - what exactly are white
| collar professionals supposed to do when their profession
| is automated away, if all the other professions are also
| being automated away?)
|
| Sidenote btw, but I do think it's funny that the investor
| class doesn't think AI will come for their role..
|
| To me the silver lining is that I don't think most of
| this comes to pass, because I don't think current
| approaches to AGI are good enough. But it sure shows some
| massive structural issues we will eventually face
| rightbyte wrote:
| In some sense I think it is related to how pyromaniacs like
| to lit and watch houses burn. Some sort of fetisch?
| oytis wrote:
| Nobody wants to be unemployed, but people generally love
| the idea of getting what they want without having to
| interact with not to say pay to other people.
|
| Like you have a brilliant idea, but unfortunately don't
| have any hard skills. Now you don't have to pay enormous
| sums of money to geeks and have to suffer them to make it
| come true. Truly a dream!
| jahewson wrote:
| Because perceived existential risks capture people's
| attention and imagination. If you're a white collar worker
| then it's a big deal to you.
| olalonde wrote:
| Most people prefer not having to work.
| jdietrich wrote:
| I'm afraid that this might sound flippant, but the answer to
| your question comes through another question - why were early
| 19th century industrialists obsessed with replacing textile
| workers? Replacing workers with machines is not a new
| phenomenon and we have gone through countless waves of social
| upheaval as a result of it. The debate we're currently having
| about AI has been rehearsed many, many times and there are
| few truly novel points being made.
|
| If you want to understand our current moment, I would urge
| you to study that history.
|
| https://en.wikipedia.org/wiki/Swing_Riots
| noncoml wrote:
| Programers are going to be replaced by AI in the same way
| accountants got replaced by VisiCalc and engineers by CAD
| and mathematicians by calculators and software like
| Mathematica
| detourdog wrote:
| We could produce more cloth in safer working conditions.
| blibble wrote:
| there was nothing safe about those early machines
| crop_rotation wrote:
| > Why is it that so many people are completely obsessed with
| replacing all developers and any other white-collar job?
|
| > They seem to be totally convinced that this will happen.
|
| The two groups of people are not same. I for example belong
| to the 2nd but not the 1st. If you have used the current gen
| LLM coding tools you will realize they have gotten they are
| scary good.
| throw310822 wrote:
| Nobody is obsessed with it. People are afraid of it. And yet,
| what will you do? Will you renounce adopting a tool that can
| make your work or someone else's work faster, easier, better?
| It's a trap: once you've seen the possibilities you can't go
| back; and if you do, you'll have to compete with those who
| keep using the new tools. Even if you know perfectly well
| that in a few years the tools will make your own job useless.
|
| Personally, however, I would find it possibly even more
| depressing to spend my day doing a job that has economic
| value only because some regulation prevents it being done
| more efficiently. At that point I'd rather get the money
| anyway and spend the day at the beach.
| whydoyoucare wrote:
| Aren't advancements in AI actually helping drug discovery?
| jahewson wrote:
| This implies that a human life is valuable but what they spend
| it doing is not.
| charleshn wrote:
| I'm always surprised by the number of people posting here that
| are dismissive of AI and the obvious unstoppable progress.
|
| Just looking at what happened with chess, go, strategy games,
| protein folding etc, it's obvious that pretty much any
| field/problem that can be formalised and cheaply verified -
| e.g. mathematics, algorithms etc - will be solved, and that
| it's only a matter of time before we have domain-specific ASI.
|
| I strongly encourage everyone to read about the bitter lesson
| [0] and verifier's law [1].
|
| [0] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
|
| [1] https://www.jasonwei.net/blog/asymmetry-of-verification-
| and-...
| oytis wrote:
| It's very different from chess etc. If we could formalise and
| "solve" software engineering precisely, it would be really
| cool, and probably indeed just lift programming to a new
| level of abstraction.
|
| I don't mind if software jobs move from writing software to
| verifying software either if it makes the whole process more
| efficient and the software becomes better as a result. Again,
| not what is happening here.
|
| What is happening, at least in AI optimist CEO minds is
| "disruption". Drop the quality while cutting costs
| dramatically.
| charleshn wrote:
| I mentioned algorithms, not software engineering, precisely
| for that reason.
|
| But the next step is obviously increased formalism via
| formal methods, deterministic simulators etc, basically so
| that one could define an environment for a RL agent.
| bigyabai wrote:
| I'll bet you $1,000*10^32 that AI never formalizes a
| novel FFT algorithm worth more than a dime.
| puchatek wrote:
| It's unlikely that LLMs are gonna get us there though.
| They ingested all relevant data at this point at the net
| effect might very well kill future sources of quality
| data. How is e.g. stackoverflow gonna stay alive if the
| next generation of programmers relies mainly on copilot
| and vibe coding? And what will the LLMs scrape once it's
| gone?
| bigyabai wrote:
| People assume (rightly so) that the progress in AI should be
| self-evident. If the whole thing is really working that
| great, we should expect to see real advances in these fields.
| Protein-folding AI should lower the prices of drugs and
| create competitive new treatments at an unprecedented rate.
| Photo and video AI should be enabling film directors and game
| directors to release higher-quality content faster than ever
| before. Text AI should be spitting out Shakespeare-toppling
| opuses on a monthly basis.
|
| So... where's the kaboom? Where's the giant, earth-shattering
| kaboom? There _are_ solid applications for AI in computer
| vision and sentiment analysis right now, but even these are
| fallible and have limited effectiveness when you do deploy
| them. The grander ambitions, even for pared-back "ASI"
| definitions, is just kicking the can further down the road.
| TheBicPen wrote:
| The kaboom already happened on user-generated media
| platforms. YouTube, Facebook, tiktok, and so on are flooded
| with AI-generated videos, photos, sounds, and so on. The
| sheer volume of this low-quality slop is because AI lowered
| the barrier of entry for creating content. In this space
| the progress is not happening through pushing the upper
| bound of quality higher but by reducing the cost for
| minimal quality to down to near-0.
| overgard wrote:
| We need to stop calling what we have AI. LLMs can't reliably
| reason. Until they can the progress is far from unstoppable.
| mvieira38 wrote:
| Your examples are not LLMs, though, and don't really behave
| like them at all. If we take the chess analogy and design an
| "LLM-like chess engine", it would behave like an average 1400
| London spammer, not like Stockfish, because it would try to
| play like the average human plays in it's database.
|
| It isn't entirely clear what problem LLMs are solving and
| what they are optimizing towards... They sound humanlike and
| give some good solutions to stuff, but there are so many
| glaring holes. How are we so many years and billions of
| dollars in and I can't reliably play a coherent game of chess
| with ChatGPT, let alone have it be useful?
| tootie wrote:
| Cloud gaming 2.0
| jandrewrogers wrote:
| During the 1990s dotcom boom we massively overbuilt fiber
| networks. It was indiscriminate and most of the fiber was never
| lit.
|
| After the dotcom crash, much of this infrastructure became
| distressed assets that could be picked up for peanuts. This
| fueled a large number of new startups in the aftermath that
| built business models figuring out how to effectively leverage
| all of this dead fiber when you don't have to pay the huge
| capital costs of building it out yourself. At the time, you
| could essentially build a nationwide fiber network for a few
| million dollars if you were clever, and people did.
|
| These new data centers will find a use, even if it ends up
| being by some startup who picks it up for nothing after a
| crash. This has been a pattern in US tech for a long time. The
| carcass of the previous boom's whale becomes cheap fuel for the
| next generation of companies.
| segmondy wrote:
| I can barely get 50Mbps up/down and I only have xfinity in
| this area. No fiber, I will pay for it, but here we are. 2025
| in good ol USA. In an urban area too.
| datadrivenangel wrote:
| Now is this AI CapEx or data and IT CapEx? Because everyone and
| their mother are labeling regular data centers as AI data
| centers.
| hatthew wrote:
| It seems like it's specifically based on nvidia's sales, which
| I assume almost entirely for deep learning. "Regular" data
| centers don't need many GPUs.
| gazpacho wrote:
| I just wish we forced every new data center to be built with
| renewables or something. The marginal cost over a conventional
| data center can't be that big compared to the total cost, and
| these companies can afford it. Maybe it can help advance the next
| generation of small modular nuclear reactors or something.
| jraph wrote:
| I wish the hardware itself were renewable too.
| loeg wrote:
| Data centers want firm power to avoid underutilizing expensive
| assets. Solar and wind are intermittent. New gas has a years-
| long lead out. 12+ hours of batteries (solar in winter) are, in
| fact, not free or de minimis.
| bpodgursky wrote:
| These companies would all LOVE to use small modular nuclear
| reactors in their datacenters... if the NRC ever got around to
| approving a license for one.
| phil21 wrote:
| Many of these companies are very interested in small nuclear
| tech as a means to power these facilities. The major bottleneck
| now for most of them is finding sites where grid capacity
| exists to power them.
|
| Talking to anyone in the space for more than 30 minutes and
| nuclear with come up.
|
| I very much hope the hype cycle lasts long enough for some of
| this capital raining down from the sky to get these reactors
| deployed in the field, because those will be a lasting positive
| from this hype cycle - much like laying railroad infrastructure
| and fiber optic cables came from other hype cycles.
|
| I've often said that the robber barons sucked, but at least
| they left massive amounts of physical infrastructure for the
| average person to benefit from. The hype cycles of late leave
| very little in comparison.
| nocsi wrote:
| Why renewables? These guys don't have an energy problem. The
| states have made sure give them energy credits and other
| assurances on that. The problem they do have though is
| siphoning the water from everyone around them.
|
| Microsoft actually has a design for mini datacenters that stay
| cool in the ocean and collect tidal energy. But it's way more
| fun to have states trying to court you into building
| datacenters cause it'll bring some jobs.
| jjk166 wrote:
| At least in the US, it doesn't need to be forced. In 2024 94%
| of new US powerplant capacity in the US were renewables or
| battery storage, we're on track for 93% for 2025, and based on
| announced plans the next few years will see very similar
| numbers. What few fossil fuel power plants are being built are
| exclusively natural gas, and a decent number of them are
| conversions of former coal plants to natural gas. Planned
| additional natural gas plant capacity is lower than at any
| point since the shale revolution started. Renewables have won.
| Voloskaya wrote:
| > I just wish we forced every new data center to be built with
| renewables or something.
|
| Already the case in Europe. And in the US, most of the biggest
| player are doing this: Google, Microsoft, Meta, AWS. By now
| those 4 are the largest buyers of renewable purchase agreement
| in the entire world. MS alone invested something like 20B in
| renewable purchase.
|
| But the issue is that installation of renewable in the US is
| not bottlenecked by lack of demand, it's bottlenecked by
| permitting, zoning issue etc. The queue for power deployment
| right now is something like 100GW (i.e. how much production is
| paid to be built, but not yet built), that is around 10% of the
| current total US power capacity. So it's not really clear to me
| if buying more renewables helps making it's deployment faster
| through economies of scale, or if the purchase order is just
| sitting in a queue for years and years.
|
| One notable exception is xAI/Grok, who has one of the biggest
| cluster, is powering it 100% with gas and afaik did not offset
| it by buying the equivalent in renewable. Having built the
| cluster in a what was a washing machine factory that does not
| have adequate power supply or cooling tech, they have been
| rolling in 35 mobile gas turbines (large trailer trucks that
| you connect to gas pipes) and 50+ refrigeration trucks. IMHO,
| it should be illegal to build such an energy consuming system
| with such a poor efficiency, but well.
| stefan_ wrote:
| Apparently the telecom boost was 2020? What am I missing?
|
| Using sus statistics to draw weird conclusions.
| betaby wrote:
| There were some major network upgrades during COVID times dueto
| traffic growth induced by work from home.
| jsnell wrote:
| I don't see how you can make the argument that a large portion of
| funds used for AI capex were diverted from other investments (and
| starving other industries), while simultaneously applying the
| economic multiplier to the whole sum when going from the
| investments to the GDP impact.
|
| Surely you only get one of the two, because for diverted
| investments the multiplier applies equally on both sides of the
| equation.
| arthurofbabylon wrote:
| Hence the title, "Honey, AI Capex is Eating the Economy."
| jsnell wrote:
| That seems like a non sequitur.
|
| He is making two arguments. One is that AI capex is starving
| other industries. And the other that AI capex is causing
| major GDP growth, attributed to both the direct investments
| themselves, as well as the multiplier effects.
|
| One of those could be true. But I assert that both cannot be
| true at the same time. If these direct investments were going
| to happen elsewhere if they weren't happening for AI
| infrastructure _then that counterfactual spending would show
| up in the GDP instead_ , as would the multiplier effect from
| that spending.
| x86x87 wrote:
| I hear AI data centers are consuming more power than the entire
| country of Argentina /s
|
| But I don't hear anyone worried about the massive power
| consumption without a clear indication if this is a net positive
| for our society.
| criddell wrote:
| > if this is a net positive for our society
|
| Maybe that's something that can only be determined looking
| back. There are so many unknown _unknowns_.
| zegax wrote:
| Roughly 5% of energy in the US is dedicated towards AI
| datacenters. The current usage will double to 70-90 TWh/year by
| 2026-2027. For software heavy tech businesses it makes sense to
| host your own AI data center so that you could train it on your
| codebase and have developers build faster. Not sure if this
| benefits humanity that much..
| mattmanser wrote:
| Will it? Google originally started trying to sell search
| servers you'd install on site. And on prem (even on your
| cloud) is not a model any of the big boys will want to
| follow.
|
| We're yet to see if it's going to be a winner takes all
| market or whether there will end up a Linux equivalent pop up
| that destroys all the investment from the big players because
| programmers are too tight to pay for software.
| woah wrote:
| Is Argentina a net positive for our society? There's the
| grilled beef, but every country has some kind of barbecue.
| There's a few soccer players I guess? Is Argentina worth the
| energy expenditure though?
| asdev wrote:
| how many signs of a bubble do we need?
| noman-land wrote:
| The only sign you'll ever get is the pop, no?
| B1FF_PSUVM wrote:
| And they're not even making paperclips, as in
| https://www.google.com/search?q=ai+paperclipping
|
| The 1880s 6% on railroads is an interesting number, I didn't know
| it was that much.
| jonas21 wrote:
| I don't know... 1.2% of GDP just doesn't seem that extreme to me.
| Certainly nowhere near "eating the economy" level compared to
| other transformative technologies or programs like:
|
| - Apollo program: 4%
|
| - Railroads: 6% (mentioned by the author)
|
| - Covid stimulus: 27%
|
| - WW2 defense: 40%
| zachlatta wrote:
| We are only 2 years in! 1.2% of GDP is enormous! The fact that
| we can even make any of these comparisons is stunning.
| dude250711 wrote:
| It's hard for me to tell what is a bigger misspending of
| money - LLMs or Apollo... At least I have a direct access to
| LLMs. Not sure I would need a direct access to moon rocks
| though.
| AlotOfReading wrote:
| 1.2% is larger than either agriculture or mining (including
| O&G).
| EA-3167 wrote:
| - The birth of the space age, and more realistically the birth
| of the ICBM and satellite age. Both key to national security,
| and in the context of a cold war.
|
| - 40% of long-distance ton miles travel by rail in the US. This
| represents a VAST part of the economic activity within the
| country.
|
| - A literal plague, and the cessation of much economic
| activity, with the goal of avoiding a total collapse.
|
| - ...Come on.
|
| So we're comparing these earth-shaking changes and reactions to
| crisis with "AI"? Other than the people selling pickaxes and
| hookers to the prospectors, who is getting rich here exactly?
| What essential economic activity is AI crucial to? What war is
| it fighting? It mostly seems to be a toy that costs FAR more
| than it could ever hope to make, subsidized by some obscenely
| wealthy speculators, executives fantasizing about savings that
| don't materialize, and a product looking for a purpose
| commensurate to the resources it eats.
| ToucanLoucan wrote:
| > What essential economic activity is AI crucial to?
|
| The continued devaluing of skilled labor and making smaller
| pools of workers able to produce at higher levels, if not
| their automation entirely.
|
| And yeah AI generated code blows. It's verbose and
| inefficient. So what? The state of mainstream platform web
| development has been a complete shit show since roughly 2010.
| Websites for a decade plus just... don't load sometimes.
| Links don't load right, you get in a never-ending spinning
| loading wheel, stuff just doesn't render or it crashes the
| browser tab entirely. That's been status quo for Facebook,
| YouTube, Instagram, fuck knows how many desktop apps which
| are just web wrappers around websites, for just.. like I
| said, over a decade at this point. Nobody even bats an eye.
|
| I don't see how ChatGPT generating all the code is going to
| make anything substantively worse than hundreds of junior
| devs educated at StackOverflow university with zero oversight
| already have.
| detourdog wrote:
| I agree we don't have an actual ROI on AI yet. There is a ton
| of activity but the progress for society is speculation at
| this point. I don't think we will have an idea on societal
| progression for at least 10 years and maybe 30.
| kulahan wrote:
| As a % of GDP it doesn't seem very large, but that's because
| our GDP is so massive. This is still an entire Norway's worth
| of GDP.
|
| Like 1.2% isn't a big percentage, but neither is 3.4% - our
| total military expenditures this year.
| pinkmuffinere wrote:
| You're right that it's large in an absolute sense, but any
| sector of the US economy is going to be large in an absolute
| sense. It's not a very meaningful statement. Using
| percentages allows comparison to other items, which for some
| purposes gives a more useful sense of size. For instance,
| based on your numbers, AI expenditure is about 1/3 the total
| military expenditure. I tend to agree that this is less than
| I expected, and generally makes me feel a bit better about
| the (imo excessive) hype.
| noosphr wrote:
| The entire population of Norway fits in Queens and Brooklyn.
| If everyone there decided to whittle spoons we'd be midly
| concerned about just what got in their water, but it won't be
| an existential crisis for the rest of us.
|
| I will never understand people who use tiny European
| countries as meaningful comparisons to continent sized ones.
| raincole wrote:
| Yeah that's my first reaction to. 1.2% doesn't sound much. It's
| just people making headlines out of thin air. If it lists the
| water and energy consumption I might be more concerned.
|
| Slightly off-topic, but ~9% of GDP is generated by "financial
| services" in the US. Personally I think it's a more alarming
| data point.
| linotype wrote:
| https://youtu.be/HA1YKg_OLBw
|
| Financial services makes the unrealistic consumption of rich
| countries possible. That's worth 9%.
| anticorporate wrote:
| Wait, so we could end unrealistic consumption in rich
| countries and get 9% of our economy back to doing something
| useful? Sounds like a win-win.
| rogerrogerr wrote:
| Yes, and you could go back to agrarian life too! Win-win-
| win!
|
| The finance industry's ability to teleport value across
| time and space is a massive boon for quality of life
| across the world.
| goda90 wrote:
| I can't help but wonder if there's a middle ground
| between people not being able to obtain credit to pursue
| new enterprises, and entire productive enterprises being
| swallowed up in the pursuit of short-term rent seeking.
| Ntrails wrote:
| Could such a middle ground exist? Sure. Could someone
| design a system where that middle ground was a natural
| equilibrium? Unsure. I don't see how you incentivise the
| goldilocks behaviour (but I am not the smartest bear so
| maybe someone else can)
| wizzwizz4 wrote:
| There's a good book on this topic by a Scottish
| philosopher: _An Inquiry into the Nature and Causes of
| the Wealth of Nations_ by Adam Smith, LL. D. and F. R. S,
| formerly Professor of Moral Philosophy in the University
| of Glasgow.
| lukas099 wrote:
| The economy is not a fixed pie that you can just take
| slices from one sector and give them to another.
| Financial services provide liquidity that supports every
| other sector; getting rid of them would cause contraction
| in every other sector.
| puchatek wrote:
| That's true to a degree. But giving them free reign
| inventivizes the kind of behavior that gave us the 2008
| financial crisis. So the commenter can be forgiven for
| wanting to limit that sector.
| UltraSane wrote:
| How do they do that?
| afiodorov wrote:
| Nice clip yet it does not make it clear why 9% is the good
| value of GDP. Why not 7%?
| arthurofbabylon wrote:
| You're looking at dots on a graph when you should be looking at
| lines and curves (and slopes of curves). The author makes this
| argument:
|
| * Movement of capital from other fields to "AI". * Duration of
| asset value (eg, AI in months/years vs railroad in
| decades/centuries). * "Without AI datacenter investment, Q1 GDP
| contraction could have been closer to -2.1%".
| ViscountPenguin wrote:
| Covid stimulus isn't directly comparable to the other three
| programs you listed. It was more like tax legislation than an
| industry.
| dyauspitr wrote:
| 1/100th of the _US economy_ sounds massive to me.
| afiodorov wrote:
| The crux of the article is asking whether such a large
| investment is justified; downplaying the article saying it's
| only X% of the GDP compared to Y doesn't address the issue.
| mikewarot wrote:
| I'm waiting for the shoe to drop when someone comes out with an
| FPGA optimized for reconfigurable computing and lowers the cost
| of llm compute by 90% or better.
| pdhborges wrote:
| Where is that improvement coming from? Hardware is already here
| to compute gemm as fast as possible.
| leakyfilter wrote:
| Raw gemm computation was never the real bottleneck,
| especially on the newer GPUs. Feeding the matmuls i.e memory
| bandwidth is where it's at, especially in the newer GPUs.
| trhway wrote:
| We're already seeing it with DeepSeek's and other optimizations
| - like that law with highways - the wider the highway the even
| more usage of it. Dropping by 90% would open even more use
| cases.
|
| For white-collar jobs replacement - we can always evolve up the
| knowledge/skills/value chain. It is the blue-collar jobs where
| bloodbath with all the robotics is coming.
| crystal_revenge wrote:
| This is where I do wish we had more people working on the
| theoretical CS side of things in this space.
|
| Once you recognize that all ML techniques, including LLMs, are
| fundamentally compression techniques you should be able to come
| up with some estimates of the minimum feasible size of an LLM
| based on: information that can be encoded in a given parameter
| size, relationship between loss of information and model
| performance, and information contained in the original data
| set.
|
| I simultaneously believe LLMs are bigger than the need to be,
| but suspect they _need_ to be larger than most people think
| given that you are trying to store a fantastically large amount
| of information. Even given lossy compression (which ironically
| is what makes LLMs "generalize"), we're still talking about an
| enormous corpus of data we're trying to represent.
| sfpotter wrote:
| Getting theoretical results along these lines that can be
| operationalized meaningfully is... really hard.
| charleshn wrote:
| We do already have ASICs, see Google's TPU to get some cost
| estimates.
|
| HBM is also very expensive.
| d-moon wrote:
| I wouldn't wait. fpgas weren't design to serve this model
| architecture. yes they are very power efficient but the
| layout/p+r overhead, the memory requirement (very few on-the-
| market fpgas have hbm), slower clock speed, and just an
| unpleasant developer experience makes it a hard sell.
| wagwang wrote:
| My chip friends constantly complain about Qualcomm owning a
| bunch of FPGA related patents that stifle any meaningful FPGA
| progress.
| padjo wrote:
| > the scale and pace of capital deployment into a rapidly
| depreciating technology is remarkable
|
| That's an interesting perspective. It does feel a bit like we're
| setting money on fire.
| siavosh wrote:
| What if anything would it take to actually change the markets
| perception that expectations may not be met in a significant way?
| tptacek wrote:
| Can you annualize Nvidia's Q1 results simply by multiplying them
| by 4?
| jsnell wrote:
| For a first approximation, yes? Their earnings have been
| growing steadily for years ($4-5B growth in each of the last 8
| quarters), with no seasonal effects.
|
| Q1*4 is highly likely to be a better estimate of their eventual
| 2025 calendar revenue than their current trailing 12 months
| revenue would be. Probably still a bit conservative, but easier
| to justify than projecting that growth continues at exactly the
| same pace.
| lucbocahut wrote:
| The main argument builds on the assumption that the economy is a
| zero sum game when it clearly is not. Just because we invest
| these ressources in AI does not mean we could mobilize the same
| capital for other pursuits.
|
| Precisely AI is being built out today because the value returned
| is expected to be massive. I would argue this value will be far
| bigger than railroads ever could be.
|
| Overspending will happen, for sure, in certain geographies or for
| specialty hardware, maybe even capacity will outpace demand for a
| while, but I don't think the author makes a good case that we are
| there yet.
| arthurofbabylon wrote:
| > "... builds on the assumption that the economy is a zero sum
| game when it clearly is not"
|
| Be cautious making assessments as to compounding effects; while
| it remains the critical attribute, the compounding nature of a
| system is not always obvious. For example, the author is
| correct that financing for AI CapEx is starving other fields of
| investment at least in the short term.
| schimmy_changa wrote:
| In addition, the massive investment may be overbuilt, but then
| likely will end up being useful long-term. This is just as the
| initial buildout of the internet to power pets.com was a bust
| at the time, but then led to Amazon, YouTube, Zoom, etc and in
| fact led to us being able to weather Covid better than you'd
| have expected.
|
| The modern internet came from folks getting connected over-
| exuberantly based on near-term returns (with a lot of investors
| losing their shirts) but then humans figured out what the
| _actual_ best use of the technology.
|
| Highly recommend this book for more, Carlota Perez is very
| insightful:
| https://en.wikipedia.org/wiki/Technological_Revolutions_and_...
| constantcrying wrote:
| The premise of AI and certainly what a large subset of executives
| and investors believe is that AI will provide a significant
| productivity increase to a significant part of the work force.
|
| Of 30% of the work is done 10% faster that leaves a 3% gain for
| other economic activities. If that is true the CapEx is
| justified.
___________________________________________________________________
(page generated 2025-07-18 23:00 UTC)