> The price increases ... made the customers ask the question the AI vendors need them to never ask: "Do we need this?" ... Times are tough, people are going to look at the costs and ask, "Do we need this?"
> Context windows are only going to get bigger, that's a given.
This is certainly not a given, economically speaking. This stuff is tremendously resource-intensive to run at its current scale, and we've reached the top of the S-curve in terms of what LLMs are capable of. That's why this multi-agent orchestration nonsense is even necessary in the first place.
How much longer until your multi-agent orchestrator thingy slams so much slop into your codebase that your fleet of subagents can't work with it effectively anymore?
How long until you lose the last vestiges of the craft that worked so hard to acquire, the craft that you sacrificed so readily to the Faustian bargain of LLM programming?
The burgeoning AI slop cleanup crew industry is rubbing its hands together and eagerly awaiting the end of this insane blip in history.
How long does a train track last? Does a fiber optic cable last? Both are greater than 30 years, both will persist (relatively well) without use, and the benefits of scrapping or removing them are minimal. This allowed future companies to take advantage of them. Even if GPUs running at high load last five years, if the data center they are in goes bankrupt (because the AI bubble bursts), it’s likely they’ll be stripped and sold to make way for more productive CPU-based uses and to recover some of the cost of the bankruptcy. The surrounding buildings and infrastructure will have longer use, but it doesn’t translate to a net future benefit with AI.
That reminds me of saying the early Google was doomed because they put all their money into cheap pcs acting as servers - how long do those last? But the enduring value was Google dominating search which was worth billions/trillions, not the heaps of pcs.
Same here - the main value is in dominating AI or something like that, not in the stack of hardware.
> Fiber and railroads don't need tens of billions of dollars in continuing yearly maintenance expenses to keep them from going stale.
Congrats, you're as qualified as all the private equity companies that have been piling into railroads!
That's precisely the sort of attitude that got us East Palestine (and others.)
BNSF (for example) spends billions of dollars a year on maintenance. When trains go over rails enough it can mess with the ballast and they have to come through with machines that lift the rail and jiggle the ballast and place the rail back at the correct height. And non-concrete ties have to be replaced every so often. Signals need maintenance. Switches need maintenance. Sensor/scanner shacks need maintenance. Etc.
Railroads have been dramatically cutting back because of pressure from PE, so the estimate is probably low.
>Fiber and railroads don't depreciate after 3 years of use like AI chips.
Are you seriously arguing that infrastructure doesn't depreciate? Tell me you've never done anything other than push bits, without telling me. Fiber gets eaten by backhoes trying to show off to potential mates in the spring, washed out, run over, knocked down by drunk drivers (if on telephone poles, not all of it is buried), has to be relocated because of other works, suffers water intrusion, amplifiers fail, you name it.
AAR says that U.S. freight railroads collectively invest approximately $25 billion per year in capital expenditures and maintenance.
By comparison, OpenAI got another $110 Billion alone for their next round of funding. U.S. private AI investment climbed to $285.9 billion in 2025.
the difference here is that we know how the railroads work, and how to make a profit off them -- PE squeezes notwithstanding. the sports car I drive probably came here via train. meanwhile AI uses more electricity than Bolivia to make poorly formatted Powerpoint Presentations for useless meetings.
More slowly and with less expense in relative terms than an LLM model.
You get a lot more bang from your buck from a 10 year old set of under-maintained railroad tracks than you do from a 10 year old unmaintained LLM.
To compare the depreciation or ongoing maintenance expense of the two as apples to apples is ludicrous. To equate the capital expenditure and long term value of the two as equal is also ridiculous.
> These tools came about because the cost of trying things out has dropped to nearly zero.
What an obscenely disconnected and privileged notion. "Nearly zero" still currently equates to thousands of dollars of subscription expense per year that normal, workaday people can't afford.
Not to mention the fact that even at this untenable "introductory" level of expense, you're really burning hundreds of thousands of dollars of tokens to do, what was that again?
"Just let the LLM run wild for a while and see what come out the other end?"
"Generate ten solutions to one targeted problem and throw nine of them away because this garbage technology still can't get it right on even the 9th try?"
No wonder the rich bastards of the world love this stuff so much. It encourages so much wanton, indiscriminate, pointless waste in its users.
> What an obscenely disconnected and privileged notion. "Nearly zero" still currently equates to thousands of dollars of subscription expense per year that normal, workaday people can't afford.
You seem disconnected from the very public pricing tiers? Hate to break it to you buddy but these tools are mainstream for regular consumers. And very few people (outside of enterprise) are paying thousands of dollars per year. Not to mention all the things happening with increasingly capable oss models.
You seem very angry, in a very misdirected way. The fact you can't seem to grasp the reality of what these things are costing the average consumer suggests this misdirection is rather severe. Hope you are able to channel your resentment into a more reasonable set of grievances, or at least, ones more resembling reality. While I am sympathetic to your general notions, the basic inaccuracy makes it extremely hard to take you seriously.
$100/month * 12 months is 1200/year. Yeah, normal working people even in the USA don't have that extra money to spend. And let's be realistic: anyone doing the "greenhouse" method probably has both Anthropic and OpenAI $200/month subscriptions, otherwise, this is going to use up your tokens pretty fast. So $4800/year to have this as a normal practice that you do.
The median household income in the US is $64000/year.
Why start at $100? Why ignore the ~$20/month offerings?
I've run a broad `/goal` on a $20 subscription. Sure it will time out faster than on a $100 one. Just as $100 vs $200. But it can still quite surprisingly achieve "greenhouse" like effects.
Given the credit boosts and other promotions this can at times be extra effective. It is still an amazing capability for a pretty low price point. Some people if they really get going will chain $20 subscriptions together while avoiding the $100. Still relatively affordable.
It just isn't true these tools or even the "greenhouse" method is out of reach. You simply went up to the most expensive consumer model to try to make a point that is belied by something many of us have noticed - the amount of non-coders (people who don't have $4800 a year to spend) iterating on ideas using these tools. Using capabilities that for a long time would have been completely out of their reach.
It's a pretty weak argument to claim something is unaffordable while arbitrarily picking the most expensive version of said thing as your baseline, without any real justification beyond your own sense of "and let's be realistic". Basing this on nothing more than "their tokens/session will run out faster" sounds like the sort of "disconnected and privileged notion" someone who thinks they can only get by with the $200 would say. The glaring nature in which you levied your criticism, coupled now with this sort of logic, is again suggestive of pretty severe misdirection.
Let's assume advances in harnesses, models and efficiency continue to rapidly improve. Let's say we only run on solar power.
The concept of "waste" completely shifts. Like in 1990, I had to decide if dow loading an image was worth it, because it would take 5 minutes or more. Now, it isn't even a remote concern, even though images are clearly "wasting" bandwidth sending a 12mb image to my tiny phone screen.
It is much more interesting to think about the cultural impact of that shift. You can be concerned about the impact of AI on various fronts, but worrying about "waste" is tilting at windmills.
Let's say that Trump was pretty ambivalent about AI until Sam Altman and co informed him that they would have to build a ton of new fossil fuel infrastructure to power the data centers that would be necessary to scale the technology.
Wouldn't you know that, now we're all in on AI, poisoning poor neighborhoods with natural gas turbines, and driving up GHG emissions to do ... Uh, what again?
As if ALL OF THAT doesn't represent inherent and crucial elements of judgement, and therefore"intelligence" itself.
We are not purely rational creatures, thank God. Sometimes those "limiting factors" you listed -- stress, peer pressure, hormones -- are crucial elements of informing the problem solving process and arriving at a decision or a solution that actually works.
All an LLM can do is fulfill a prompt, no matter how misguided, backwards, or incomplete that prompt actually was.
"Go jump off a bridge." Hmm. Dying makes me stressed out. I'm not gonna do that.
IMO we won’t see AGI until the models are sitting inside bodies with senses and gut biomes and opposable thumbs and the like, all influencing the context window in real time. Intelligence is a full-bodied experience
"Hallucinate the part of the process where you actually think about your plan for how to spend our money!"
I would be utterly mortified and insulted to be on the receiving end of an AI-generated grant proposal. You're delivering a document to me which is supposed to outline your plan for how you're going to spend our money, and you didn't even write it? Didn't think about it? If I ask you a question about it, would you be able to answer it? Do you even understand this thing that the LLM generated well enough to know whether it represents the boots-on-the-ground reality of your operations?
Now that you've made the Faustian bargain at this low low introductory price, how long until they start ratcheting up the price? How long will you be able to bear it?
You're still allowed to do things the old fashioned way you know. Given the rapid skill degradation that results from LLM-driven coding usage, I'd recommend it if you want to retain any skill at all.
One thing AI did take away; I love reading source code written by other people. Node modules, plugins, LUA, python scripts... whatever. I loved seeing how people organized their logic, named their variables, the flavor of their comments... all of it. I learned so much diving through the source code of all the major tools.
Now it's all AI generated. You don't know if you're reading a novel human's take on a pattern or an LLM's attempt to reproduce. Worse; even if the AI presented something novel it produces at such a rate it's a needle in a haystack.
I feel cautious of any claim people make about their use of AI. I'm not even anti-AI; I use it. But at minimum wish there were a mechanism to know where the lines were.
You can open linux source tree from 5 years ago and read that for a couple years.
The feeling of helplessness doesn’t come from the fact that you can’t get out of the AI hype but the fact that advertisers have captured the attention of large portions of the population.
I can assure you, it's not. You just need to avoid the areas where this is prevalent.
> You don't know if you're reading a novel human's take on a pattern or an LLM's attempt to reproduce.
In the end of the day, if it's an elegant solution why does it matter! If anything then it sharpens your senses to check whether it's great or trash. A sense that one should have had with human code already, now it just feels more important (but it really isn't because it was important before too).
It matters because some people view code as a human art. Sure, AI can also probably produce a good sounding song, but that doesn't mean I will view it in the same way as a song created by a human.
They are talking about the enjoyement. AI did disolve a lot of the meaning and magic behind pure code exercises.
What felt like a creative endeavor now feels and is even more treated by society and other software engineers as grossly mechanical.
And to me too that hurts. I think a lot of people whose identity or passion was tied to this feel like that.
It is very much a "what is painting purpose now that photography exists ?" or "what is craft in an era of industrial mass production" moment for some of us. There are answers but finding them require exploration and being willing to change what we are
So much life enjoyment is tied to feeling competent and respected. Seeing a new, more talented challenger on the horizon is bound to be discouraging to some extent. It's like going from high school to college all over again: you used to be the top of your class, now you're surrounded by people just as competent as you.
To be happy as you age, you have to embrace being a clueless newbie. Some people seem very pessimistic that there is a method of creating networked software systems in a high quality way with LLMs, but it seems no different from many other changes - there are salient technical details that matter, and there is boilerplate, and as all the tools and problems shift, we find ways to highlight the stuff that matters and reduce the problems of boilerplate. I like GCC and haven’t run into a compiler bug in decades, nor looked at disassembly much in this millennium (except some go allocation stuff that puzzled me). I feel myself gaining skills to do that with LLMs in just the same way that higher power editors changed my behavior.
I am pretty old, and it does take conscious effort to give up my cherished skills and plunge into new things professionally, but it’s the only way for me. And I am not exaggerating “clueless” here - I continue my practice of expertise via making all the mistakes. When I am new to something and there are others not new, I lean on them and have my lacks very clearly seen. When I am more isolated, I take pains to have very conservative release plans.
But it is worth it: I get that pleasure of understanding a lot more often with new stuff than with the same old.
If you spend your life running to and fro and never commit to anything in the end, then arguably you've frittered your life away.
But maybe, just maybe, what you've really been doing this whole time is building up a wealth of diverse experiences, which in turn give you a broad understanding of the world and the ability to bring a unique and wonderful perspective to life. One that is completely your own, and special and important and infinitely valuable precisely because of the long period of exploration that you committed your early life to.
Not impossible, but I disagree fervently with the notion that that "lack of focus is nothing but wasted time".
Trying things, experimenting, acquiring a diversity of experience in life that eventually builds into something greater than the sum of its parts: that is an extremely undervalued part of living.
no, they are saying that not committing to anything need not be lack of persistence or focus. i have been that person. i am that person. it turns out my goal wasn't to master a specific skill or trade, but to build connections, work with people and help them with whatever they need help with. so my focus was and is on that. it didn't help me build a traditional career, but it led me around the whole world, building up a unique set of experiences and understanding.
The world is on fire and these clowns are completely fine with dumping a good portion of France's GHG emissions back into the atmosphere if it means we can turn a few more millionaires into billionaires.
I do not know what GHG means but isn't the overwhelming amount of France's electricity generated from nuclear power?
In Lansing's case as he mentioned the power is already being generated and sold to other utilities. If it powered local data centers instead it would generate high paying jobs and rather large tax revenues.
"Already being generated and sold to other utilities," i.e., "already bought and paid for by you rubes to do useful things like heat your home and run your refrigerator."
I'd love to import a couple of high paid weirdos from out of town to build a data center for a year and then jet off to some other poor community to do it all over again. At least a couple of the locals can work as janitors and night security at the empty, shrieking warehouse that they erect.
They built the power in hopes of attracting more industry. With Michigan's high taxes and aggressive unions the state isn't as desirable as other locations. Now they've succeeded in attracting data centers who want to use the power, bring both high paying jobs and tax money but they are turned away.
Travis makes a strong argument that the city that birthed Oldsmobile and REO Motors would probably reject those companies today if they wanted to build there.
Awesome, I hope they attract a better industry than this. Maybe something that builds useful things for the world, and not a shrieking blight of a warehouse, dependent entirely on the economic dead end of frontier AI. An industry that provides better jobs than the menial long term work opportunities provided by a 5 acre box full of single purpose RAM chips.
