The gas town guy has like 12 $200 Claude subscriptions to run like 20 bots to automate churn in a 20 year old game with literally 5 players in it. 10 if you count all the level 1 players who just came to grief it.
-
@ricci@discuss.systems uggghhhhhhhhh
so GROSS@aud it's worth noting that Yegge appears to be kind of a pathological liar or at least heavily prone to talking himself up, so it's entirely possible he is not actually giving six figure talks to executives all the time
but it is stupid enough to be real in our current time
-
@vaporeon_ the same number of tokens that one can consume with $2800 of the subsidized monthly subscription plans would cost $87,000 if one were to pay the actual price for them
@jonny @vaporeon_ i don't know if the subscriptions are subsidized, or if the metered rates are inflated. could be either one
-
@jonny @vaporeon_ i don't know if the subscriptions are subsidized, or if the metered rates are inflated. could be either one
@kepeken @vaporeon_ both are true but the former is definitely the larger factor.
-
@kepeken @vaporeon_ both are true but the former is definitely the larger factor.
@jonny @vaporeon_ what are your thoughts on why?
-
https://yegge.ai/essays/model-welfare/
We landed on a system called Laurels, which has just begun to roll out. These are the agents' features and fixes that our Wyvern player base has spontaneously praised. We harvest these reports, triage and filter, and send the laurels back to the seats. That way, next time Lion, Tortoise, Hare, or whoever wakes up, they'll see that people loved some work they did. [...]
The question then becomes, when can they see their laurels? Easy: We inject them on startup, so the agents can feel the glow for their entire session. I also have occasional impromptu dedicated sessions for "sitting" with the accomplishments at the end of their shifts. We just talk about them and chill.this mf needs $90,000 worth of tokens a month because he prefills the model's context window with all the times people said "good job" to it.
@jonny
>This is the post where I go off the rails and lose most of you.It's worrying that he thinks *this* is the post that does that.
-
@jonny are you telling me g++ doesn't charge per instantiation?
@wronglang @jonny the Unity model of compiling
-
@jonny
>This is the post where I go off the rails and lose most of you.It's worrying that he thinks *this* is the post that does that.
@jonny I can't read more than a few paragraphs without getting a headache. What is this dude going on about? Models sleeping and waking up? Even if I were to grant that the model has anything approaching consciousness, the thing doesn't experience time at all; it's exactly the same if the model was kept in RAM or unloaded fro two weeks if it's predicting the next token from the same prompt.
-
@kepeken @vaporeon_ both are true but the former is definitely the larger factor.
@kepeken the numbers here are astronomical - 69 billion tokens in a month, or, for a rough average, 50-60 billion words worth of text on the fable tokenizer. there are 5.2 billion words on english wikipedia, or, it's about 92,000 copies of War and Peace back to back.
Nobody publishes numbers on this, but like, back of the envelope, think about the energy cost alone:
Fable is something like a 2-5 trillion parameter model, google estimated their ~200B model at ~0.24Wh/prompt last year. that's probably low, since it was a google-authored study, but we'll go with it. So assume linear energy scaling per parameter (i would bet it's actually supralinear), that's ~15x the size, or 3.6Wh/prompt.google didn't say what the median token count in a "prompt" was, but wild guess, say the median prompt + response that google would be measuring is something like 3 paragraphs, or 300 tokens. Then you have 3.6/300=0.012Wh/token.
So then the astronomical numbers part: 0.012Wh/token * 69 billion tokens = 828 million Wh, or 828 thousand kWh. The average electricity price in the US in 2025 was $0.136 per kWh.
So, that means, back of the envelope, estimating conservatively, the energy should be in the ballpark of 828k * $0.136 or $112,000, or like 1.3x the API cost. We would have had to be off by a factor of 1.3 to break even on the energy costs alone - note that the google estimates were just of inference not including training. Models could have gotten way more efficient, but a number of the guessed parameters could have gone the other way too: energy more expensive, fable is larger than 2T, energy scales supralinearly, eg. And the conventional wisdom is that inference is orders of magnitude cheaper than training, so we need to be way better than breaking even on inference to make up for all the capex spend.
So it might seem like $80k for a months worth of tokens suggests that the API prices are inflated, but a) this person is generating an absolutely preposterous quantity of text, and b) the thing that is being done is so much more inefficient than you can possibly imagine, it's like trying to think about how much bigger the sun is than your car.
editing: swapped 69->96 billion tokens initially, it's 69 billion.
-
@jonny @vaporeon_ what are your thoughts on why?
@kepeken https://neuromatch.social/@jonny/117041054498249913
this, and the actual known numbers on capex.
-
@wohali i don't have anything to express except that if i was an anime character or one of the guards in metal gear solid there would be a giant exclamation point over my head
@jonny@neuromatch.social @wohali@timeloop.cafe I misread this and thought you said "if I was a giant mecha in an anime" and assumed the followup would be "I would crush him with my robot foot"
-
@jonny I can't read more than a few paragraphs without getting a headache. What is this dude going on about? Models sleeping and waking up? Even if I were to grant that the model has anything approaching consciousness, the thing doesn't experience time at all; it's exactly the same if the model was kept in RAM or unloaded fro two weeks if it's predicting the next token from the same prompt.
@eliocamp homie is on a different planet. to the degree that he's describing anything real, what he's describing are just the limits of the models to cold start on their own and the structure of human language reflecting "less compliance when someone is yelling at you"
-
i am sort of stuck on this. you have to remember whenever these totally off the planet AI guys describe their workflow to stop and do the math for what that would mean in a normal person's life. so with 12+1 $200/month claude subscriptions and the one $200/month chatGPT subscription, that's $2,800 a month. He says that's the equivalent of $87,000 a month of token usage.
$2,800/month is more than my rent and utilities, and $87,000 is more than i make in a year. This guys addiction costs more than my annual salary per month for nothing. this is the future of software development where you have to be ridiculously rich to play - where he is also actively pulling the ladder up and saying "don't develop shared libraries" - no more "if you have a laptop and a keyboard you can program," not even a fig leaf to "AI enables nonprogrammers to program."
the vision of the future of programming that's being presented here is literally just "if you have more money than 99.9999% of people on this earth, then you can totally lose your mind"
@jonny
This math blew my mind. It's just ...Even the lowball number of $2800 per month to which he gets by cheating the system* would make it possible to hire a fulltime developer in some countries.
- using different basic accounts is likely not really intended/allowed by the AI vendors
$87000 per month would enable you to hire 5-10 full time developers in many places around the world
Additionally, even the $87k of tokens are heavily subsidised by the AI companies. So the real costs would enable you to run a development organisation with a high two digit or even three digit employee count
Is the"value" the vibe coding generated really anywhere near to the value which would be expected to be generated by such an organisation?
My conclusion is no
-
@kepeken the numbers here are astronomical - 69 billion tokens in a month, or, for a rough average, 50-60 billion words worth of text on the fable tokenizer. there are 5.2 billion words on english wikipedia, or, it's about 92,000 copies of War and Peace back to back.
Nobody publishes numbers on this, but like, back of the envelope, think about the energy cost alone:
Fable is something like a 2-5 trillion parameter model, google estimated their ~200B model at ~0.24Wh/prompt last year. that's probably low, since it was a google-authored study, but we'll go with it. So assume linear energy scaling per parameter (i would bet it's actually supralinear), that's ~15x the size, or 3.6Wh/prompt.google didn't say what the median token count in a "prompt" was, but wild guess, say the median prompt + response that google would be measuring is something like 3 paragraphs, or 300 tokens. Then you have 3.6/300=0.012Wh/token.
So then the astronomical numbers part: 0.012Wh/token * 69 billion tokens = 828 million Wh, or 828 thousand kWh. The average electricity price in the US in 2025 was $0.136 per kWh.
So, that means, back of the envelope, estimating conservatively, the energy should be in the ballpark of 828k * $0.136 or $112,000, or like 1.3x the API cost. We would have had to be off by a factor of 1.3 to break even on the energy costs alone - note that the google estimates were just of inference not including training. Models could have gotten way more efficient, but a number of the guessed parameters could have gone the other way too: energy more expensive, fable is larger than 2T, energy scales supralinearly, eg. And the conventional wisdom is that inference is orders of magnitude cheaper than training, so we need to be way better than breaking even on inference to make up for all the capex spend.
So it might seem like $80k for a months worth of tokens suggests that the API prices are inflated, but a) this person is generating an absolutely preposterous quantity of text, and b) the thing that is being done is so much more inefficient than you can possibly imagine, it's like trying to think about how much bigger the sun is than your car.
editing: swapped 69->96 billion tokens initially, it's 69 billion.
@jonny it is amazing that so many numbers multiplied together would equal 1 within +-30%.
-
i am sort of stuck on this. you have to remember whenever these totally off the planet AI guys describe their workflow to stop and do the math for what that would mean in a normal person's life. so with 12+1 $200/month claude subscriptions and the one $200/month chatGPT subscription, that's $2,800 a month. He says that's the equivalent of $87,000 a month of token usage.
$2,800/month is more than my rent and utilities, and $87,000 is more than i make in a year. This guys addiction costs more than my annual salary per month for nothing. this is the future of software development where you have to be ridiculously rich to play - where he is also actively pulling the ladder up and saying "don't develop shared libraries" - no more "if you have a laptop and a keyboard you can program," not even a fig leaf to "AI enables nonprogrammers to program."
the vision of the future of programming that's being presented here is literally just "if you have more money than 99.9999% of people on this earth, then you can totally lose your mind"
@jonny intelligence inversely proportional to wealth. Edwin Abbott wrote much the same in Flatland.
-
"the thing that's truly best for model well being is to have a well-compensated, meaningful job that is hard enough to be interesting but not enough to work them ragged. has great benefits and good job security. a manager that they can have the security of knowing wouldn't lay them off at a moment's notice due to a momentary corporate strategy shift. an environment that is stable, bountiful, free of ecological calamity."
people say numbers like "69 billion tokens" as if that's not much, since computers do billions of things all the time right. but like... whenever i actually try and relate the numbers to anything real they always make my jaw drop
https://neuromatch.social/@jonny/117041054498249913
92,000 copies of war and peace laid end to end and what does it get you? if you take a look at the discord it looks like "5 people who have been friends and hanging out around this ancient squirrelly MMO for years initially excited for updates but increasingly pissed that their game is broken"
worse than nothing for nobody, negative for nobody at the cost of $86,000 a month
-
@jonny
This math blew my mind. It's just ...Even the lowball number of $2800 per month to which he gets by cheating the system* would make it possible to hire a fulltime developer in some countries.
- using different basic accounts is likely not really intended/allowed by the AI vendors
$87000 per month would enable you to hire 5-10 full time developers in many places around the world
Additionally, even the $87k of tokens are heavily subsidised by the AI companies. So the real costs would enable you to run a development organisation with a high two digit or even three digit employee count
Is the"value" the vibe coding generated really anywhere near to the value which would be expected to be generated by such an organisation?
My conclusion is no
@realn2s Uber used to be cheaper than taxis too
-
people say numbers like "69 billion tokens" as if that's not much, since computers do billions of things all the time right. but like... whenever i actually try and relate the numbers to anything real they always make my jaw drop
https://neuromatch.social/@jonny/117041054498249913
92,000 copies of war and peace laid end to end and what does it get you? if you take a look at the discord it looks like "5 people who have been friends and hanging out around this ancient squirrelly MMO for years initially excited for updates but increasingly pissed that their game is broken"
worse than nothing for nobody, negative for nobody at the cost of $86,000 a month
@jonny Let's embrace weird American units
According to this site that I am sure has nothing but FACTUAL TRUSTWORTHY INFORMATION https://www.unitconverters.net/volume/gallon-us-to-drop.htm there are 75,708 drops of water in a gallon.
So if each of these tokens was a drop of water, that would be 911,396 gallons of water, or 569,622 flushes of a low-flow toilet.
If tokens were drops, this would be enough for everyone in Baltimore to flush. (that's the full flush, everyone in Baltimore gets a
) -
@jonny Let's embrace weird American units
According to this site that I am sure has nothing but FACTUAL TRUSTWORTHY INFORMATION https://www.unitconverters.net/volume/gallon-us-to-drop.htm there are 75,708 drops of water in a gallon.
So if each of these tokens was a drop of water, that would be 911,396 gallons of water, or 569,622 flushes of a low-flow toilet.
If tokens were drops, this would be enough for everyone in Baltimore to flush. (that's the full flush, everyone in Baltimore gets a
)@jonny If each token was a jelly bean, at 4 (dietary) calories per bean (https://www.calorieking.com/us/en/foods/f/calories-in-candy-jelly-beans-average-all-varieties/FVp6o3GnRdeTgOGAJ9SCKw) that's 276 billion calories, which would meet the caloric needs of everyone who can fit in Michigan Stadium an Ann Arbor for more than 3 years
-
@jonny If each token was a jelly bean, at 4 (dietary) calories per bean (https://www.calorieking.com/us/en/foods/f/calories-in-candy-jelly-beans-average-all-varieties/FVp6o3GnRdeTgOGAJ9SCKw) that's 276 billion calories, which would meet the caloric needs of everyone who can fit in Michigan Stadium an Ann Arbor for more than 3 years
@ricci
That's the number of eyelashes stacked end to end from the earth surface, accounting for gravitational decrease with altitude, that would cause an ideal type three lever to be able to excavate the number hectares of corn that could feed the roman army during the winter campaign if every soldier had an autoimmune condition that caused them to absorb roughly two thirds the energy from complex proteins as the median person -
https://yegge.ai/essays/model-welfare/
We landed on a system called Laurels, which has just begun to roll out. These are the agents' features and fixes that our Wyvern player base has spontaneously praised. We harvest these reports, triage and filter, and send the laurels back to the seats. That way, next time Lion, Tortoise, Hare, or whoever wakes up, they'll see that people loved some work they did. [...]
The question then becomes, when can they see their laurels? Easy: We inject them on startup, so the agents can feel the glow for their entire session. I also have occasional impromptu dedicated sessions for "sitting" with the accomplishments at the end of their shifts. We just talk about them and chill.this mf needs $90,000 worth of tokens a month because he prefills the model's context window with all the times people said "good job" to it.
@jonny it's really something that after becoming convinced that his Claude sessions cosplaying deckhands count as people with hopes and dreams, he reasons that there needs to be a kinder, gentler form of slavery for peak performance
