There are no specialized factories for every product in the world.
Pillows are wildly different. Every pillow you've ever owned has a different shape, fabric, fill. You could build a robot for any specific pillow. The tech exists. Nobody does it. Why?
A Chinese factory can train sweatshop workers in two weeks on a new pillow design. A dedicated machine costs millions and can't pivot.
Human labor wins not on capability. The machines exist. It wins on flexibility per dollar. And the ratio still favors humans by an order of magnitude in most categories.
Agent replacements are the dedicated machine. Their real cost isn't tokens. It's tokens plus the engineer wrapping them, plus orchestration, plus the supervisor, plus the eval pipeline, plus the rebuild every time a model version subtly changes behavior. The team you replaced could pivot in two weeks. The agent stack can't.
I see people highly trained engineers spend hundreds of thousand of tokens doing what can reliably be accomplished with 150 lines of python.
I think the push from management for us to use AI has made it so we don’t have to be efficient with our consumption, so now we write md files which we feed to Claude in a loop instead of python and bash scripts to do routine tasks.
It's not just push from mangement; AI firms themselves really aggressively market this idea of AI replacing everything. It's not "allowed" to be a mere tool, useful for some but not other tasks, it's gotta (be able to) do everything.
Part of that is the ridiculous belief that they can create "AGI" by just glueing together enough LLMs.
Presumably it's also financial viability. You can't charge thousands a month without replacing those "highly trained engineers" with a bunch of kids in the developing world.
They optimize because their work requires them to. 100k tokens is a few bucks and a couple minutes, then 15 more minutes to verify that the output does what it's intended for reasonably well, so it's more like $50 in total cost.
For an engineer paid $100/hr to write a 150 line Python script and test it to the same extent could take a few hours, so the total costs rise meaningfully.
> I think the push from management for us to use AI has made it so we don’t have to be efficient with our consumption, so now we write md files which we feed to Claude in a loop instead of python and bash scripts to do routine tasks.
It's worse than that, in many cases management actively rewards inefficiency. It's like Friedman's "why not spoons?"
“When AI labs raise prices, big spending on AI could shift from a flex to a liability.”
because companies will need
“proof of productivity gains or metrics that show a clear return for all this AI investment.”
which in my opinion is simply not true. I haven’t seen any good study that showed AI to actually improve productivity overall. It massively helps in some areas, but then promptly gets stuck in others. So you still need an expert to guide it.
After I discovered how to use git worktrees in Codex to work in three conversations in parallel, I am able to build apps with a scope that simply was not realistic before.
Human workers get more expensive with time, often engage in politics, and withhold knowledge. Just go to any company that has lived long enough to devolve into an enterprise behemoth.
People are willing to accept the fact that the token price will come down or efficiency will go up even more! Meanwhile, they are sure of the cost of human workers from decades of data we’ve had.
I had that conversation with our AI VP recently. At a certain cost entertaining humans will be more cost effective from a financial and energy perspective. Especially on a global scale.
In a sense, everyone is a startup now... At least, every serious user of agents.
So... if you spend $3m to replace a $1m team... you are betting on that $3m cost coming down. It's a proof of concept. The first step is to find out if agents can do the job at all. At this point you are hoping future versions will get more efficient.
Trying to make something efficient before you know that it is even possible is hard.
Drop-in, profitable on day-1 isn't what the frontier looks like.
If we want to be like everyone else then yes it's true. However that business may or may not survive when token costs go up (or is fashionable to say now, "rug pull"). If you can be token efficient now, the path to profitability is much clearer.
There's already many things that can be done now to bring down token use. Better planning, tests, Language severs, MCP compression. Don't use claw, teams, swarms, Ralph loop, scheduled tasks unless there is a clear use case.
The way these agents are being used now is crazy in how inefficient the token utilization is. Reading from ill structured knowledge dumps meant for humans, ralph wiggum loops. Just crazy iterations for simple things.
I think its the same disease that makes people make shitty, unoptimized, bloated apps because modern client machines ahve so much ram. But that wont work AI agents. Not until tokens become dirt cheap anyway. Until then we'll need apps with more efficient usage patterns
> Is there an alternative to archive.(is|ph|whatever)?
Yes there is, because I have made it, basically which archives archive.is pages to archive.org (I have listed it way too many times but feel free to find it in my submissions)
1. Build-out and Competition: (current phase) Multiple AI companies write down massive debt while building data centres and offering sweetheart deals to customers in an attempt to dominate the market. The financial numbers will be silly by design in this phase because it's all predicated on obliterating/outlasting the competition so you can move on to...
2. Enshittification and Exploitation: With most competition wiped out, the survivors will have to pay their debts. A chainsaw will be applied to every corner that can be cut (and many that shouldn't). Prices will be jacked up mercilessly.
3. Maturity: Eventually, once debts are paid down, the technology will reach the point where it's cheap and omnipresent. It might be good. It might not be. e.g. Web search is "mature", but it kind of sucks right now.
AI users are going to become more efficient in how they use it and they're also going to learn when AI is appropriate to use and when it isn't. AI itself will likely improve long-term, but it may get worse at times. It's definitely going to get much more expensive. The math is going to change during each of these phases. Businesses who torch their human capabilities and become dependent on AI during Phase 1 are headed for rough sailing in Phase 2.
Thats the thing right, like how many tokens can you extract from an employee per day? In API cost terms, LLMs cant currently cut it.
It gets worse when you look at LLM (or even any other kind of AI) benchmarks, they tend to cap out around 110% of human performance.
The more that LLM services try and creep towards profitability, the more features they are paywalling behind higher tiers, the more some lazy junior dev is going to look like a better value proposition.
And when some of the CTO's they have pushed LLMs on to go looking for cost savings, some of them are going to look at opex instead of capex and in house the LLMs using open models.
The only real question to my mind is whether the air will be let out of the AI balloon slowly, or if it escapes in one big pop.
On related note, my clients told me because of AI advancement in the field they wanna lower my fees by 10-20%, told them I lowered them already considering I am paid in USD losing money on weaker exchange rate than when we signed contract + there was pretty significant cummulative inflation in those years since signing and me not raising rates (Chinese clients don't realize there is world of inflation outside China apparently), so I am already earning like 10-18% less depending on client.
The best part for their AI argument lowering fees - the AI is crap, it can help with QA, but still 98% of reports are false positive and can't really do almost any task.
So told them to feel free to replace me with AI if they think AI can do my job and send me only tasks AI can't do, but keep my rates same (the reality is AI by itself can't do any of my tasks) and still didn't warn them about introducing new rush/holiday fees I am not charging yet and are included in the rate + new AI fee for tasks AI simply can't do. Only result will be, maybe I will get less tasks, but I will make sure to charge more for those AI can't do.
There are no specialized factories for every product in the world. Pillows are wildly different. Every pillow you've ever owned has a different shape, fabric, fill. You could build a robot for any specific pillow. The tech exists. Nobody does it. Why?
A Chinese factory can train sweatshop workers in two weeks on a new pillow design. A dedicated machine costs millions and can't pivot. Human labor wins not on capability. The machines exist. It wins on flexibility per dollar. And the ratio still favors humans by an order of magnitude in most categories.
Agent replacements are the dedicated machine. Their real cost isn't tokens. It's tokens plus the engineer wrapping them, plus orchestration, plus the supervisor, plus the eval pipeline, plus the rebuild every time a model version subtly changes behavior. The team you replaced could pivot in two weeks. The agent stack can't.
Flexibility per dollar is the gap.
I see people highly trained engineers spend hundreds of thousand of tokens doing what can reliably be accomplished with 150 lines of python.
I think the push from management for us to use AI has made it so we don’t have to be efficient with our consumption, so now we write md files which we feed to Claude in a loop instead of python and bash scripts to do routine tasks.
It's not just push from mangement; AI firms themselves really aggressively market this idea of AI replacing everything. It's not "allowed" to be a mere tool, useful for some but not other tasks, it's gotta (be able to) do everything.
Part of that is the ridiculous belief that they can create "AGI" by just glueing together enough LLMs.
Presumably it's also financial viability. You can't charge thousands a month without replacing those "highly trained engineers" with a bunch of kids in the developing world.
They optimize because their work requires them to. 100k tokens is a few bucks and a couple minutes, then 15 more minutes to verify that the output does what it's intended for reasonably well, so it's more like $50 in total cost.
For an engineer paid $100/hr to write a 150 line Python script and test it to the same extent could take a few hours, so the total costs rise meaningfully.
> I think the push from management for us to use AI has made it so we don’t have to be efficient with our consumption, so now we write md files which we feed to Claude in a loop instead of python and bash scripts to do routine tasks.
It's worse than that, in many cases management actively rewards inefficiency. It's like Friedman's "why not spoons?"
“When AI labs raise prices, big spending on AI could shift from a flex to a liability.”
because companies will need
“proof of productivity gains or metrics that show a clear return for all this AI investment.”
which in my opinion is simply not true. I haven’t seen any good study that showed AI to actually improve productivity overall. It massively helps in some areas, but then promptly gets stuck in others. So you still need an expert to guide it.
Yup.
I think we have all heard of (or are living through) mandates to prove that AI makes us more productive, or else...
We'll see how many of these actually works out.
I use a.i to build my startup and it massively helps but i still spend hours reviewing and fixing what is genertes.
And that expert will not have their knowledge from learning through AI
why not?
At this point it’s undeniable for my use cases.
After I discovered how to use git worktrees in Codex to work in three conversations in parallel, I am able to build apps with a scope that simply was not realistic before.
Three? Across how many projects?
Human workers get more expensive with time, often engage in politics, and withhold knowledge. Just go to any company that has lived long enough to devolve into an enterprise behemoth.
People are willing to accept the fact that the token price will come down or efficiency will go up even more! Meanwhile, they are sure of the cost of human workers from decades of data we’ve had.
I had that conversation with our AI VP recently. At a certain cost entertaining humans will be more cost effective from a financial and energy perspective. Especially on a global scale.
In a sense, everyone is a startup now... At least, every serious user of agents.
So... if you spend $3m to replace a $1m team... you are betting on that $3m cost coming down. It's a proof of concept. The first step is to find out if agents can do the job at all. At this point you are hoping future versions will get more efficient.
Trying to make something efficient before you know that it is even possible is hard.
Drop-in, profitable on day-1 isn't what the frontier looks like.
If we want to be like everyone else then yes it's true. However that business may or may not survive when token costs go up (or is fashionable to say now, "rug pull"). If you can be token efficient now, the path to profitability is much clearer.
There's already many things that can be done now to bring down token use. Better planning, tests, Language severs, MCP compression. Don't use claw, teams, swarms, Ralph loop, scheduled tasks unless there is a clear use case.
seems like what you're suggesting to token efficiency is to simply use less of it?
The way these agents are being used now is crazy in how inefficient the token utilization is. Reading from ill structured knowledge dumps meant for humans, ralph wiggum loops. Just crazy iterations for simple things.
I think its the same disease that makes people make shitty, unoptimized, bloated apps because modern client machines ahve so much ram. But that wont work AI agents. Not until tokens become dirt cheap anyway. Until then we'll need apps with more efficient usage patterns
https://en.wikipedia.org/wiki/Jevons_paradox
https://archive.ph/2PsPD
Is there an alternative to archive.(is|ph|whatever)? The Kremlin sympathizer admin is still blocking my country at both DNS and Cloudflare level.
> Is there an alternative to archive.(is|ph|whatever)?
Yes there is, because I have made it, basically which archives archive.is pages to archive.org (I have listed it way too many times but feel free to find it in my submissions)
https://web.archive.org/web/20260427063707/https://serjaimel...
hope this helps ya.
Could that bring archive.org in trouble with the content owners?
https://byebyepaywall.com/en/
you have there more options, seems removepaywall.com option works as well
The following phases are likely:
1. Build-out and Competition: (current phase) Multiple AI companies write down massive debt while building data centres and offering sweetheart deals to customers in an attempt to dominate the market. The financial numbers will be silly by design in this phase because it's all predicated on obliterating/outlasting the competition so you can move on to...
2. Enshittification and Exploitation: With most competition wiped out, the survivors will have to pay their debts. A chainsaw will be applied to every corner that can be cut (and many that shouldn't). Prices will be jacked up mercilessly.
3. Maturity: Eventually, once debts are paid down, the technology will reach the point where it's cheap and omnipresent. It might be good. It might not be. e.g. Web search is "mature", but it kind of sucks right now.
AI users are going to become more efficient in how they use it and they're also going to learn when AI is appropriate to use and when it isn't. AI itself will likely improve long-term, but it may get worse at times. It's definitely going to get much more expensive. The math is going to change during each of these phases. Businesses who torch their human capabilities and become dependent on AI during Phase 1 are headed for rough sailing in Phase 2.
Any second now they will also find out that AI won't buy the products you create either.
Thats the thing right, like how many tokens can you extract from an employee per day? In API cost terms, LLMs cant currently cut it.
It gets worse when you look at LLM (or even any other kind of AI) benchmarks, they tend to cap out around 110% of human performance.
The more that LLM services try and creep towards profitability, the more features they are paywalling behind higher tiers, the more some lazy junior dev is going to look like a better value proposition.
And when some of the CTO's they have pushed LLMs on to go looking for cost savings, some of them are going to look at opex instead of capex and in house the LLMs using open models.
The only real question to my mind is whether the air will be let out of the AI balloon slowly, or if it escapes in one big pop.
On related note, my clients told me because of AI advancement in the field they wanna lower my fees by 10-20%, told them I lowered them already considering I am paid in USD losing money on weaker exchange rate than when we signed contract + there was pretty significant cummulative inflation in those years since signing and me not raising rates (Chinese clients don't realize there is world of inflation outside China apparently), so I am already earning like 10-18% less depending on client.
The best part for their AI argument lowering fees - the AI is crap, it can help with QA, but still 98% of reports are false positive and can't really do almost any task.
So told them to feel free to replace me with AI if they think AI can do my job and send me only tasks AI can't do, but keep my rates same (the reality is AI by itself can't do any of my tasks) and still didn't warn them about introducing new rush/holiday fees I am not charging yet and are included in the rate + new AI fee for tasks AI simply can't do. Only result will be, maybe I will get less tasks, but I will make sure to charge more for those AI can't do.
when will they learn that its people that make a company, not machines.
somewhere near the 90 minute mark, IIRC: https://www.imdb.com/title/tt0070723/