(Asking sincerely in good faith - I read this note along with skimming a few of the linked ones, and I'm familiar with gastown)
Can someone explain the value calculation to me? This seems like someone has truly let AI VASTLY expand the codebase of what seems to be a medium sized hobby game to ~10-100x the amount of engineering required. Especially with statements about how wheelhouse, his software factory, has grown to nearly the size of the actual software he's writing. He also mentions several times that wheelhouse is specific to developing the game. Then, he discusses "pulling in beads" which itself seems on my read to be enormously engineered. (note that he also says that you burn a lot of tokens with agents "keeping your beads in sync" and I didn't have the time to figure out what that meant, but it looks like as complicated as it is, beads can't reconcile itself without burning $$$).
It seems like if you answer "yes" at every time you have the question of "can I make AI do this", you end up burning $120k/month in tokens on your side project.
Again, I am not disparaging this- but I feel like I am genuinely missing something and would welcome help understanding it.
I guess this is a question of what you believe code and software engineering writ large is for.
If you believe that good engineering practice matters (in the sense that your software is maintainable, efficient, etc.), then this is horrific because the human has no idea what’s going on and any attempt to fix something by hand would probably uncover enough complications to make you consider rewriting from scratch.
But if your only qualifier is “does it do what I want it to do?” then this is perfectly fine. If you throw out the assumption that a human will ever maintain the code or need to understand it, then who cares if your main branch is full of unused code and it’s outrageously complex.
In a way, I guess this is a similar moment to abstract painting in the early 20th century.
If your assumption was that the point of a painting is to accurately capture a landscape or person, then abstract paintings are an abomination.
If you believe paintings should evoke a feeling and depicting people or landscapes are only one way to get there, you can throw out all the sketching and composing and painting technique.
> If your assumption was that the point of a painting is to accurately capture a landscape or person
I know this is way off-topic, but this oft-repeated disinformation should be quashed wherever it appears. "Accuracy" was never the goal. Look at Hokusai [1], at John Bauer [2], at Jeffrey Catherine Jones [3]. Even at so "realistic" painters as Beksiński [4], or Caspar David Friedrich [5] - even if you were at the site of such a ruined abbey, and it was late evening, and you had the camera of your choice, the resulting photo would look nothing like that - painted in 1810. Look at The Swing by Fragonard [6], from 1768. Trees don't actually grow like that. Light doesn't work like that! You could visit every tree in the world with a camera, and you couldn't even approximate the feeling of that painting.
These were not accidents, it was not that they tried to be photorealistic, and came up short. They were deliberate stylizations. And modern artists, and abstract artists, well educated in the history of art, know this well, but they deliberately spread the lie that accuracy was the goal, that photography made beautiful art redundant, to justify their abominations. Because the sketching and composing and painting techniques were only half about capturing life, while the other half, the more important half (or in the case of composing, the whole), was about evoking a feeling. Something at which they were infinitely better than their modern art successors.
Steve is not making a game, he is _playing_ the game of making games.
It's like playing Factorio but instead building a rocket to escape, you are building a game. Instead of messing around with mining drills and train tracks, he is deploying agents and thinking about how to interface with them.
I think it sounds like a super fun game.
update: with the added benefit that he might see a return on his investment as well!
Want to be the head of a game studio? "Hire" a bunch of game dev agents, put them to work in the software factory, get pretty progress updates and an actual game of sorts along the way.
Want to be the CEO of a startup? Hire a bunch of biz ops agents, and fire them when they make mistakes. Ship an actual website and MVP and make some numbers go up along the way.
If you know you're playing a game its fun and weird. But I worry for the people that don't know they're LARPing...
I haven’t kept up to date on Gastown, but I’m familiar with the mindset.
A lot of people who want to go all-in on AI start with how they currently produce software and they try to add AI everywhere they can until things start going wrong. At this point they’ve found the threshold for “we can automate as much as possible without compromising on quality” and they can expand their AI use when models improve.
A minority of people like Yegge start by automating absolutely everything possible but still trying to get everything to work, even if it’s low quality. At this point they’ve found the threshold for “as automated as possible while still basically working” and they can improve their quality when models improve.
Both sets of people are trying to do the same thing – automate software development as much as possible – but they are approaching it from starting points on the opposite ends of the spectrum. One starts with quality and increases automation; the other starts with automation and increases quality. The latter approach is insane if you have to ship and maintain production quality software, but it’s also where some of the ideas that will be eventually used by the people taking the former approach come from.
Disclaimer: I work at an AI lab [edit: but speak for myself :-)]
The value proposition is an average of 270 commits a day, for $4000/month, and the dramatic improvements to the game Yegge described. Even supposing 2/3 of them are burned on the scaffolding, 90 commits/day (assuming they’re of mostly good quality) would cost multiple engineers, one of which would cost much more than $48k/year.
As another comment said, the ultimate proof will be in the pudding. But I don’t think the value proposition claimed is complicated.
I wouldn't say complicated, and I agree with you that the proof is in the pudding. I looked at the gameplay and if it was something akin to.. let's say "minecraft before minecraft existed", I'd be amazed and see the value right away.
Looking at the game, it doesn't seem like a game that would require as much money to build as he's claiming he uses in credits (although via a loophole that allows him cheaper access). It doesn't look BAD, it just doesn't seem to warrant that much spend.
> I am of course using sanctioned cheating: I get all those tokens because I'm an individual, with the Claude Max discount. So it "only" costs me about $5k/month out of pocket, for a 50-agent cluster running on a 512GB M3 Ultra Mac Studio I bought off eBay for $25k.
> So it's not $120k/month of real money, but it's still crazy spend.
> would cost multiple engineers, one of which would cost much more than $48k/year.
Have you played the game?
There's no way all this wasted money is comparable to the work of multiple engineers. This is using AI for entertainment, the result isn't an engaging game.
> First, my secret: I see the future by living in it. I am spending the equivalent of $122k/month of API token spend, or about $4,000 per day, using 21 Claude Max accounts, a number that has been growing steadily at 2 per week.
Good to know that together, Yegge and Zitron bring balance to the force.
> I am running an organization of around 50-60 agents, five of whom are interfacing with around 10 humans in the outside world: myself, my 5-person core game design team, my accountant, my chief of staff, and a few others. Only Fable is allowed to talk to humans, via Slack and email.
RIP to those poor humans. I can't imagine having your brain melted by Fablespeak as your FTJ.
What is the overhead to have agents play model UN? Why is the coordination so elaborate? It sounds like a deeply complex and expensive emergent behavior that maybe looks comprehensible but could be nonsensical. Also like any complex system, can you actually predict the outcomes?
Waste is a failure case. How do we know the code factory is actually productive or just agents filling up the computation resource cap because they can?
Yeah I have noticed lots of agent stuff recently seems to be spin as many agents up as possible have a whole office of agents. Just keep spinning more AI bastards up, give names to your AI bastards not descriptive names human names, set up email messages for your AI bastards and wonder why your stressed out that you have messages from Dave, Linda and Michael and the other Dave. I think a lot of this sort of thinking is probably going to cause mental exhausting and so wasteful. Steve is a intelligent guy but I find my self thinking the whole build an AI office or company thing is not going to be the way things work in the future and better harnesses and agents that are particularized with highly particularized tools and loop are where things are going to work better rather then lots of AI employees.
The companies selling these ideas (and tuning their products to work well in this mode) make money based on how much they are used. They’ve committed to ludicrous levels of demand growth that can’t possibly be sustained by direct human use. But if they make their own product drive multiple instances of their own product, you get exponential effects.
I'll say that Gastown sounded absolutely crazy, but the idea of having orchestrator threads to manage your work and keep tabs on it, having validators to validate the other work etc. were generally the right shape. I think GasTown probably could have been really successful if there was a pared down version with more obvious names rather than the fun names.
I'm going to be thinking about this blog post for a while though because if you squint and tear it apart, there are probably really good generalizable pieces in here to take home for future models.
What Yegge describes in this post is something I have seen to a lesser extent in my own agent use.
But one specific thing I noticed, which his inline comic lampoons, is Fable (especially) explicitly stating "[User] has ruled", or "The rulings are in".
I was confused by this until I saw a post on X about how someone came to his agents in the morning after they ran all night, and the sub-agents had refused the requests from the orchestrating agents because they thought the decisions being made were not inline with the user. They thought the decisions were from the ochestrating agent.
This, as well as the governance things that Fable and Codex both often request, made me thing of provenance, especially related to decisions and authority. As we get deeper and deeper hierarchies of agents (as we likely will) this idea of authority, who has it and where does it come from, feels like it will be a key component to agental systems.
As someone who has worked on games with progression systems that run for years, engineering throughput is almost never the limiting factor. You can add content & systems much faster than players can understand them. Good design is deciding what NOT to add.
Agree/disagree! He's genuinely writing words for humans (load bearing! honestly!), and he's definitely on the forefront of _something_, whether you agree with his findings or not.
I'm particularly struck by the Superman metaphor: "Even if you put a giant shield around your house, Superman could destroy it with his super-strength and laser eyes... but he won't dare cross the white picket fence you've put up because it goes against his core beliefs!"
Maybe think of "The Truman Show" ... 90% of the time you should be relying on guardrails and making the easier path obvious. ALSO have the hardened sandbox-proof escape container, but against a determined adversary, it's all 1's and 0's at the end and there's probably a way for them to get out?
The proof is (or isn't) in the game. If the game is something genuinely great then this all worked and is important. If the game is a bloviated, boring, and derivative mess then it didn't. Without any proof point on the output it is hard to say if the article has any value.
The player community seems to hold the opinion it's the latter. Steam reviews are pretty dire. Worse, there really aren't a lot of players. Certainly not enough to cover Yegge's absurd number of Claude Max accounts.
This whole article is very strange to me. Consider the following quotes "I have been launching new game features so fast that the players asked me to slow down." and "using magic tricks that are a year ahead of their time. It's a system that we can ride so hard that it scares the players and they tell us to slow down. " That tells me a lot about the operation.
Maybe I'm out of touch but none of what was described seems like a sustainable operation or business. Fun experiment perhaps but that's it.
In another section there's also boasting of how the problem was that more code was needed to be produced and faster. I don't know. Was that really the problem all along before LLMs and agents showed up?
Eh. A quick scan of steam reviews shows a clear pattern - the game has been in decline ever since Yegge took over, well before LLMs were a product. It seems he started upsetting people with a capitalist-brain insistence on extracting value, and nothing has changed since. As long as he doesn't understand that his goals are contrary to building a game people enjoy, quality of the code is completely irrelevant. I don't think this one works as a test case on the value of programming like this, as the intended audience just isn't interested.
> his goals are contrary to building a game people enjoy, quality of the code is completely irrelevant
i think this happens a lot inside and outside of games when people get obsessed with the tech or process behind it; they completely miss the goal of the job
The concept is like Gas Town - AI as a organization, not an emulated human. Yes, it's inefficient. But it scales.
(A very long time ago, I got a tour of Xerox PARC, even before Steve Jobs did. Alan Kay explained that they were building the future of computing, accepting that it cost far too much to be cost-effective. They assumed the hardware would catch up. It did. Took about ten years.)
100% This is EXACTLY my experience and I didn't even spend 1% of what Steve has. Fable loves to create rules around the evaluations and decisions that I make. This is good, but this is also scary because 1/ I am fallible 2/ I don't have time to digest every detail and make a careful decision. And if I do spend that time, the system will return in 10 mins and give me another massive set of decisions to make .. thus creating an unending loop, resulting in decision fatigue... which leads to #2 again.
So AI is optimizing in some ways for an AI as the judge, not human as a judge. It needs this input, this steer, but humans aren't built to support this.
I'm glad someone is trying this and exploring what's possible. It's easy to deride his work by saying what's the value but are you running 50 fable agents and letting them run wild to see what the future looks like?
The guy looks like he’s having a blast, I don’t want to yuck his yum. There is a nonzero chance that machines will take over source code authorship; yeah it’s interesting as hell that we get to watch somebody seriously try it, but if this thing is going to “stab me right in the career” I want it to come from the front.
Yeah same tho, that 5k looks to be pissing in the wind. He even knows for a fact he's not paying the actual costs yet never mentions how he's going to maintain that stack of code if costs per token ever get tightened up.
I've used AI chat for coding, but with the caveat that I need to be able to "manually" edit the results. I like the process and am ready to maybe spend a little money on tokens but am still pretty far from using agents, although I'm starting to peruse peoples claude.mds and spending some hugging face tutorials plus I guess that means python bah
He's really focused on the software factory aspect, and talks about the future and how he's living it but lets see how the code base he's accumulating is maintained, because thats the one question about the future that I rarely see answered.
I don't doubt what you're saying, but I still hack on things mainly for fun tho. In the OPs case, he's even paying below cost, how long is that deal going to go on? On the other hand, his current working code base might be so valuable that he can afford future maintenance help, in whatever form that would be.
> So it's not $120k/month of real money, but it's still crazy spend. I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models.
We get it bro, you buy lots of tokens. My son buys lots of pokemon cards. Still haven't seen shit for return on that investment either, but as long as its all in good fun, "you do you".
> That's how I'm able to tell the future. I'm living in a world that will not become cost-effective for most people for another year.
That's pretty easy to verify. #remindme August 24, 2027 - are $120k worth of tokens in 2026 generally "affordable" by 2027? By what mechanism have the economics shifted?
I don't disagree with you, but I would clarify that it's not $120k of tokens, it's $5k of Claude subscriptions. And also I suspect he would consider "affordable" to be "as a well paid software engineer". So I guess his benchmark is really more like: does $5k of subscriptions reduce to ~$500 next year.
I don't think it will. I also don't think this is really that affordable, but I'd imagine he'd call it accurate.
That $5k subscription is already deeply discounted by debt funding. Anthropic is losing money on that account. So the bet is this free lunch will continue indefinitely and prices are gonna drop 10x on top of that? We'll see!
Put another way, someone else is paying that 115k difference, and it's not out of the goodness of their hearts. Once they capture market share, capitalists gonna capitalize. For now, investors seem happy to pump money into the bubble without seeing returns, but that will not last forever.
I think we're in total agreement here. The $120k gets a profit for Anthropic. The $5k subscription loses money hand over fist. Somewhere in between is the real market price. And the OP is betting on a 10x reduction, from his already bargain-basement subsidized fee? Right at the moment when the market is demanding returns? Bold strategy Cotton, let's see how it plays out.
> When you add it all up, Fable is trying to turn Wheelhouse into an engine that can prove, mechanically, that every change to Wyvern is legal. The agents capture every single intention, decision, policy, rule of thumb, and legacy behavior in the system, and they use that to govern every future decision and action. They live by the Rule of Law.
It’s interesting to think of this as the chaotic result of essentially unbounded compute on fable in a positive feedback loop. I would suspect this essentially represents the formalization of the underlying system prompt as a process.
Lots of talk about engineering his game, lots of talk about how this game is exactly what he wants, no talk of play testing. Is there any world in which this is going to be a good game?
I do agree even more rules help keep things on track, but I do these as linters, specifically as Golang `go vet` and `go fix` commands. That works 100x better than any SKILL.md or team of agents in my experience.
In the glorious future where I can conjure a game with a snap of my fingers for the cost of a pack of gum, why would I ever choose to play Wyvern, the game tailored to Steve Yegge's tastes, when I could just make a game tailored to my tastes? Perhaps Steve will be content to play his MUD with agents pretending to be human players?
You joke, but even very simple/cheap AIs already make it a way bigger pain in the ass than it used to be to identify growth hacking fake user/signup/activity bullshit in all sorts of situations.
> the subject matter is too complicated to explain … All I can do is walk you around like an excited tour guide, one who has unearthed an ancient alien civilization.
I honestly think Yegge has reached a point where he knows he isn't interested in doing the leg work of today's software engineering, but wants his career to ride the wave. The lowest effort, maximum value to extract in his position is abusing the crap out LLMs to an extent that is genuinely novel to secure thought leadership, but on close inspection it's just pointless token spend driving blog posts and publicity.
To me it is absolutely farcical that he was being being paid by a harness company (Amp) as a staff software engineer, slop-coded a solution to keeping LLMs on track (Beads), and having paid for all of this Amp and Steve...parted ways. Beads never went into Amp. Someone cottoned on to the value Steve was providing, for my money.
Yegge is either a visionary or completely delusional. Probably both. But I'm glad someone is running this experiment. We'll know in a year whether it was genius or madness
The code fable generated is interesting from the perspective of understanding how llms behave but it's not an immutable behavior -- it could have been designed not to do that. I think it's quite interesting how it conflated details of the game with the code's architecture. I don't think that's a good thing, and he could have caught it had he read the code.
I think Steve somewhat predicted the future well. He might be slightly off, because he is overly optimistic and operates at the edge, but look at Gas Town, when it was released it felt like dystopian, today it looks not too far away, I am sure most of your orgs are already running some kind of agent to triage the tickets and in some cases automatically open the PR in your git repo.
This weekend I stumbled across a codebase unlike anything I’d ever seen: Swamp Club[]. Came out of nowhere, lots of useful features, but with a coherent philosophy and architecture and a large body of working code. Change volume is high - notably so, but it looks nothing like the vibe slop I’d expected.
A “swamp” contains tools that you build - with the help of agents - intended for use not just by apps and scripts, but agents (to add deterministic behavior, like code mode). Tools are ops things, product things, whatever.
*And so it’s a tool plus a factory to update the tool.*
The factory though isn’t only used for building your tool - that’s just a set of skill.md files. It can build your whole app, which itself can _use_ your tools - either for product features or ops tasks.
*And so it’s a factory that can build any app, including a tool which can operate any app. *
Right now, looking at the commit volume of the swamp stuff, it’s not stupid to suppose that the swamp is being used to create the swamp.
*And so it’s a factory which makes, among other things, factories.*
"I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models."
I'd argue that number of tokens burned does not translate into "experience".
And if you needed to spend $300k or so to find out that process helps manage larger efforts (process is what "fences" are outside Yegge-land), you sure spent a lot.
I mean, I'm glad the guy's got a a hobby he enjoys, but there's less insight than you'd hope for.
(Asking sincerely in good faith - I read this note along with skimming a few of the linked ones, and I'm familiar with gastown)
Can someone explain the value calculation to me? This seems like someone has truly let AI VASTLY expand the codebase of what seems to be a medium sized hobby game to ~10-100x the amount of engineering required. Especially with statements about how wheelhouse, his software factory, has grown to nearly the size of the actual software he's writing. He also mentions several times that wheelhouse is specific to developing the game. Then, he discusses "pulling in beads" which itself seems on my read to be enormously engineered. (note that he also says that you burn a lot of tokens with agents "keeping your beads in sync" and I didn't have the time to figure out what that meant, but it looks like as complicated as it is, beads can't reconcile itself without burning $$$).
It seems like if you answer "yes" at every time you have the question of "can I make AI do this", you end up burning $120k/month in tokens on your side project.
Again, I am not disparaging this- but I feel like I am genuinely missing something and would welcome help understanding it.
I guess this is a question of what you believe code and software engineering writ large is for.
If you believe that good engineering practice matters (in the sense that your software is maintainable, efficient, etc.), then this is horrific because the human has no idea what’s going on and any attempt to fix something by hand would probably uncover enough complications to make you consider rewriting from scratch.
But if your only qualifier is “does it do what I want it to do?” then this is perfectly fine. If you throw out the assumption that a human will ever maintain the code or need to understand it, then who cares if your main branch is full of unused code and it’s outrageously complex.
In a way, I guess this is a similar moment to abstract painting in the early 20th century.
If your assumption was that the point of a painting is to accurately capture a landscape or person, then abstract paintings are an abomination.
If you believe paintings should evoke a feeling and depicting people or landscapes are only one way to get there, you can throw out all the sketching and composing and painting technique.
ok, but when steve made gas town, it didn't actually do anything but build more gas town.
Meh, the actual complaint is that it does not really do what it should and keeps burning those $$$ as you ask it to fix this or that.
> If your assumption was that the point of a painting is to accurately capture a landscape or person
I know this is way off-topic, but this oft-repeated disinformation should be quashed wherever it appears. "Accuracy" was never the goal. Look at Hokusai [1], at John Bauer [2], at Jeffrey Catherine Jones [3]. Even at so "realistic" painters as Beksiński [4], or Caspar David Friedrich [5] - even if you were at the site of such a ruined abbey, and it was late evening, and you had the camera of your choice, the resulting photo would look nothing like that - painted in 1810. Look at The Swing by Fragonard [6], from 1768. Trees don't actually grow like that. Light doesn't work like that! You could visit every tree in the world with a camera, and you couldn't even approximate the feeling of that painting.
These were not accidents, it was not that they tried to be photorealistic, and came up short. They were deliberate stylizations. And modern artists, and abstract artists, well educated in the history of art, know this well, but they deliberately spread the lie that accuracy was the goal, that photography made beautiful art redundant, to justify their abominations. Because the sketching and composing and painting techniques were only half about capturing life, while the other half, the more important half (or in the case of composing, the whole), was about evoking a feeling. Something at which they were infinitely better than their modern art successors.
[1] https://en.wikipedia.org/wiki/Hokusai
[2] https://en.wikipedia.org/wiki/John_Bauer_(illustrator)
[3] https://darklongbox.com/2024/08/17/jeffrey-catherine-jones-a...
[4] https://en.wikipedia.org/wiki/Zdzis%C5%82aw_Beksi%C5%84ski
[5] https://en.wikipedia.org/wiki/The_Abbey_in_the_Oakwood
[6] https://en.wikipedia.org/wiki/The_Swing_(Fragonard)
[dead]
Steve is not making a game, he is _playing_ the game of making games.
It's like playing Factorio but instead building a rocket to escape, you are building a game. Instead of messing around with mining drills and train tracks, he is deploying agents and thinking about how to interface with them.
I think it sounds like a super fun game.
update: with the added benefit that he might see a return on his investment as well!
You nailed it, this is a meta game.
AI lets you LARP.
Want to be the head of a game studio? "Hire" a bunch of game dev agents, put them to work in the software factory, get pretty progress updates and an actual game of sorts along the way.
Want to be the CEO of a startup? Hire a bunch of biz ops agents, and fire them when they make mistakes. Ship an actual website and MVP and make some numbers go up along the way.
If you know you're playing a game its fun and weird. But I worry for the people that don't know they're LARPing...
Pushing pixels around the screen is what I bluntly described the grand sum capabilities of AI to my brother, who is quite infatuated with it.
It wasn't appreciated.
Pushing pixels around the screen is how I used to describe my day job. These days I describe it as "I talk to Claude about every 15 minutes".
I haven’t kept up to date on Gastown, but I’m familiar with the mindset.
A lot of people who want to go all-in on AI start with how they currently produce software and they try to add AI everywhere they can until things start going wrong. At this point they’ve found the threshold for “we can automate as much as possible without compromising on quality” and they can expand their AI use when models improve.
A minority of people like Yegge start by automating absolutely everything possible but still trying to get everything to work, even if it’s low quality. At this point they’ve found the threshold for “as automated as possible while still basically working” and they can improve their quality when models improve.
Both sets of people are trying to do the same thing – automate software development as much as possible – but they are approaching it from starting points on the opposite ends of the spectrum. One starts with quality and increases automation; the other starts with automation and increases quality. The latter approach is insane if you have to ship and maintain production quality software, but it’s also where some of the ideas that will be eventually used by the people taking the former approach come from.
Disclaimer: I work at an AI lab [edit: but speak for myself :-)]
The value proposition is an average of 270 commits a day, for $4000/month, and the dramatic improvements to the game Yegge described. Even supposing 2/3 of them are burned on the scaffolding, 90 commits/day (assuming they’re of mostly good quality) would cost multiple engineers, one of which would cost much more than $48k/year.
As another comment said, the ultimate proof will be in the pudding. But I don’t think the value proposition claimed is complicated.
I wouldn't say complicated, and I agree with you that the proof is in the pudding. I looked at the gameplay and if it was something akin to.. let's say "minecraft before minecraft existed", I'd be amazed and see the value right away.
Looking at the game, it doesn't seem like a game that would require as much money to build as he's claiming he uses in credits (although via a loophole that allows him cheaper access). It doesn't look BAD, it just doesn't seem to warrant that much spend.
I thought you were supposed to eat the pudding?
How can you have any pudding if you don’t eat your meat?
> I am spending the equivalent of $122k/month of API token spend, or about $4,000 per day, using 21 Claude Max accounts
That's ~$1.4 million/year
Emphasis on "the equivalent".
Further down he writes:
> I am of course using sanctioned cheating: I get all those tokens because I'm an individual, with the Claude Max discount. So it "only" costs me about $5k/month out of pocket, for a 50-agent cluster running on a 512GB M3 Ultra Mac Studio I bought off eBay for $25k.
> So it's not $120k/month of real money, but it's still crazy spend.
> would cost multiple engineers, one of which would cost much more than $48k/year.
Have you played the game?
There's no way all this wasted money is comparable to the work of multiple engineers. This is using AI for entertainment, the result isn't an engaging game.
> First, my secret: I see the future by living in it. I am spending the equivalent of $122k/month of API token spend, or about $4,000 per day, using 21 Claude Max accounts, a number that has been growing steadily at 2 per week.
Good to know that together, Yegge and Zitron bring balance to the force.
This is beyond AI psychosis; this is full AI schizophrenia.
This is ludicrous; also, why publicly admit that blatant violation of ToS?
Anthropic doesn't actually mind if you use multiple subscriptions
> I am running an organization of around 50-60 agents, five of whom are interfacing with around 10 humans in the outside world: myself, my 5-person core game design team, my accountant, my chief of staff, and a few others. Only Fable is allowed to talk to humans, via Slack and email.
RIP to those poor humans. I can't imagine having your brain melted by Fablespeak as your FTJ.
What is the overhead to have agents play model UN? Why is the coordination so elaborate? It sounds like a deeply complex and expensive emergent behavior that maybe looks comprehensible but could be nonsensical. Also like any complex system, can you actually predict the outcomes?
Waste is a failure case. How do we know the code factory is actually productive or just agents filling up the computation resource cap because they can?
Yeah I have noticed lots of agent stuff recently seems to be spin as many agents up as possible have a whole office of agents. Just keep spinning more AI bastards up, give names to your AI bastards not descriptive names human names, set up email messages for your AI bastards and wonder why your stressed out that you have messages from Dave, Linda and Michael and the other Dave. I think a lot of this sort of thinking is probably going to cause mental exhausting and so wasteful. Steve is a intelligent guy but I find my self thinking the whole build an AI office or company thing is not going to be the way things work in the future and better harnesses and agents that are particularized with highly particularized tools and loop are where things are going to work better rather then lots of AI employees.
The companies selling these ideas (and tuning their products to work well in this mode) make money based on how much they are used. They’ve committed to ludicrous levels of demand growth that can’t possibly be sustained by direct human use. But if they make their own product drive multiple instances of their own product, you get exponential effects.
I'll say that Gastown sounded absolutely crazy, but the idea of having orchestrator threads to manage your work and keep tabs on it, having validators to validate the other work etc. were generally the right shape. I think GasTown probably could have been really successful if there was a pared down version with more obvious names rather than the fun names.
I'm going to be thinking about this blog post for a while though because if you squint and tear it apart, there are probably really good generalizable pieces in here to take home for future models.
[dead]
What Yegge describes in this post is something I have seen to a lesser extent in my own agent use.
But one specific thing I noticed, which his inline comic lampoons, is Fable (especially) explicitly stating "[User] has ruled", or "The rulings are in".
I was confused by this until I saw a post on X about how someone came to his agents in the morning after they ran all night, and the sub-agents had refused the requests from the orchestrating agents because they thought the decisions being made were not inline with the user. They thought the decisions were from the ochestrating agent.
This, as well as the governance things that Fable and Codex both often request, made me thing of provenance, especially related to decisions and authority. As we get deeper and deeper hierarchies of agents (as we likely will) this idea of authority, who has it and where does it come from, feels like it will be a key component to agental systems.
As someone who has worked on games with progression systems that run for years, engineering throughput is almost never the limiting factor. You can add content & systems much faster than players can understand them. Good design is deciding what NOT to add.
"Wheelhouse is about 600k lines of code and tests (mostly bash)"
This article is a joke right? Please tell me you are joking. Jesus christ.
Every time Steve goes deeper into the AI whirlpool his engineering practices slip further from good.
I’m glad somebody is doing it though; we need an honest assessment of what’s possible, and we’re getting one.
So at least somebody is figuring out how much torment the human mind can handle in the Torment Nexus...
Yegge’s writing has declined proportionately to his AI usage.
I genuinely wonder if he has completely lost touch with reality. He seems to slowly become the Terry A. Davis of LLMs.
I used to read his stuff ~20 years ago. He has cool things to say, but his writing always sucked.
I often write lengthy pieces. Then I look at it and say "Oh no, I don't want to become Steve Yegge!"
Agree/disagree! He's genuinely writing words for humans (load bearing! honestly!), and he's definitely on the forefront of _something_, whether you agree with his findings or not.
I'm particularly struck by the Superman metaphor: "Even if you put a giant shield around your house, Superman could destroy it with his super-strength and laser eyes... but he won't dare cross the white picket fence you've put up because it goes against his core beliefs!"
Maybe think of "The Truman Show" ... 90% of the time you should be relying on guardrails and making the easier path obvious. ALSO have the hardened sandbox-proof escape container, but against a determined adversary, it's all 1's and 0's at the end and there's probably a way for them to get out?
NFT’s were the forefront according to HN 3 years ago, just before the AI hype cycle kicked off.
It looks like AI psychosis
The proof is (or isn't) in the game. If the game is something genuinely great then this all worked and is important. If the game is a bloviated, boring, and derivative mess then it didn't. Without any proof point on the output it is hard to say if the article has any value.
The player community seems to hold the opinion it's the latter. Steam reviews are pretty dire. Worse, there really aren't a lot of players. Certainly not enough to cover Yegge's absurd number of Claude Max accounts.
This whole article is very strange to me. Consider the following quotes "I have been launching new game features so fast that the players asked me to slow down." and "using magic tricks that are a year ahead of their time. It's a system that we can ride so hard that it scares the players and they tell us to slow down. " That tells me a lot about the operation.
Maybe I'm out of touch but none of what was described seems like a sustainable operation or business. Fun experiment perhaps but that's it.
In another section there's also boasting of how the problem was that more code was needed to be produced and faster. I don't know. Was that really the problem all along before LLMs and agents showed up?
This also stands out as strange to me:
> interfacing with around 10 humans in the outside world: myself, my 5-person core game design team
Like, really, there are 5 human designers on this game? What are they doing, exactly?
(Or is this just a fancy way of saying, they approve/reject design proposals of the stack of agents, and they have zero agency of their own?)
Am I so out of touch? No, it's the players who are wrong.
Eh. A quick scan of steam reviews shows a clear pattern - the game has been in decline ever since Yegge took over, well before LLMs were a product. It seems he started upsetting people with a capitalist-brain insistence on extracting value, and nothing has changed since. As long as he doesn't understand that his goals are contrary to building a game people enjoy, quality of the code is completely irrelevant. I don't think this one works as a test case on the value of programming like this, as the intended audience just isn't interested.
This scrolled off HN too fast.
The concept is like Gas Town - AI as a organization, not an emulated human. Yes, it's inefficient. But it scales.
(A very long time ago, I got a tour of Xerox PARC, even before Steve Jobs did. Alan Kay explained that they were building the future of computing, accepting that it cost far too much to be cost-effective. They assumed the hardware would catch up. It did. Took about ten years.)
Where’s the benefit in that?
It’s not like Xerox is at the forefront these days. Maybe investors made a tonne on their equity back in the day? (What the real game is).
The purpose of any game is to get better at playing the game
100% This is EXACTLY my experience and I didn't even spend 1% of what Steve has. Fable loves to create rules around the evaluations and decisions that I make. This is good, but this is also scary because 1/ I am fallible 2/ I don't have time to digest every detail and make a careful decision. And if I do spend that time, the system will return in 10 mins and give me another massive set of decisions to make .. thus creating an unending loop, resulting in decision fatigue... which leads to #2 again.
So AI is optimizing in some ways for an AI as the judge, not human as a judge. It needs this input, this steer, but humans aren't built to support this.
I'm glad someone is trying this and exploring what's possible. It's easy to deride his work by saying what's the value but are you running 50 fable agents and letting them run wild to see what the future looks like?
The guy looks like he’s having a blast, I don’t want to yuck his yum. There is a nonzero chance that machines will take over source code authorship; yeah it’s interesting as hell that we get to watch somebody seriously try it, but if this thing is going to “stab me right in the career” I want it to come from the front.
I would if I had that kind of money!
If I had $122k/month to spend on tokens I would simply.... not.
$1.5mm/year buys more of everything I want out of life than I could possibly consume.
He’s using a $5k/month account to access the “equivalent” of $122k worth of tokens.
Yeah same tho, that 5k looks to be pissing in the wind. He even knows for a fact he's not paying the actual costs yet never mentions how he's going to maintain that stack of code if costs per token ever get tightened up.
I've used AI chat for coding, but with the caveat that I need to be able to "manually" edit the results. I like the process and am ready to maybe spend a little money on tokens but am still pretty far from using agents, although I'm starting to peruse peoples claude.mds and spending some hugging face tutorials plus I guess that means python bah
He's really focused on the software factory aspect, and talks about the future and how he's living it but lets see how the code base he's accumulating is maintained, because thats the one question about the future that I rarely see answered.
How long since you had to really manually edit. That world is history now - no point holding on to it.
I don't doubt what you're saying, but I still hack on things mainly for fun tho. In the OPs case, he's even paying below cost, how long is that deal going to go on? On the other hand, his current working code base might be so valuable that he can afford future maintenance help, in whatever form that would be.
It is not history. I write code almost exclusively by hand, for example.
Congratulations, you have reinvented RFCs, architecture reviews, linters, and integration tests. From first principles. With cutesy names.
> So it's not $120k/month of real money, but it's still crazy spend. I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models.
We get it bro, you buy lots of tokens. My son buys lots of pokemon cards. Still haven't seen shit for return on that investment either, but as long as its all in good fun, "you do you".
> That's how I'm able to tell the future. I'm living in a world that will not become cost-effective for most people for another year.
That's pretty easy to verify. #remindme August 24, 2027 - are $120k worth of tokens in 2026 generally "affordable" by 2027? By what mechanism have the economics shifted?
I don't disagree with you, but I would clarify that it's not $120k of tokens, it's $5k of Claude subscriptions. And also I suspect he would consider "affordable" to be "as a well paid software engineer". So I guess his benchmark is really more like: does $5k of subscriptions reduce to ~$500 next year.
I don't think it will. I also don't think this is really that affordable, but I'd imagine he'd call it accurate.
That $5k subscription is already deeply discounted by debt funding. Anthropic is losing money on that account. So the bet is this free lunch will continue indefinitely and prices are gonna drop 10x on top of that? We'll see!
Put another way, someone else is paying that 115k difference, and it's not out of the goodness of their hearts. Once they capture market share, capitalists gonna capitalize. For now, investors seem happy to pump money into the bubble without seeing returns, but that will not last forever.
Well that $120k is also likely at a huge profit margin, at least accounting for COGS. That's already most of the $115k difference gone.
Again, I don't really disagree with your general point here, but I think it's just a bit more complicated than you're making out.
I think we're in total agreement here. The $120k gets a profit for Anthropic. The $5k subscription loses money hand over fist. Somewhere in between is the real market price. And the OP is betting on a 10x reduction, from his already bargain-basement subsidized fee? Right at the moment when the market is demanding returns? Bold strategy Cotton, let's see how it plays out.
> When you add it all up, Fable is trying to turn Wheelhouse into an engine that can prove, mechanically, that every change to Wyvern is legal. The agents capture every single intention, decision, policy, rule of thumb, and legacy behavior in the system, and they use that to govern every future decision and action. They live by the Rule of Law.
It’s interesting to think of this as the chaotic result of essentially unbounded compute on fable in a positive feedback loop. I would suspect this essentially represents the formalization of the underlying system prompt as a process.
> They live by the Rule of Law.
I feel that description is a bit too strident for the underlying stochastically-creative writing exercise going on.
> "Is this a chain of command, or a chain of credulity?"
> "There's a difference?"
-- The Vor Game by Lois McClure Bujold
Lots of talk about engineering his game, lots of talk about how this game is exactly what he wants, no talk of play testing. Is there any world in which this is going to be a good game?
Seems like this is the year of the sandbox and governance layers.
I recently wrote how we set all this up for our agent computers that help people with their email:
https://housecat.com/blog/agent-computer-101
I do agree even more rules help keep things on track, but I do these as linters, specifically as Golang `go vet` and `go fix` commands. That works 100x better than any SKILL.md or team of agents in my experience.
If the agents are really doing politics now, we can test and see which method of organizing agents politically works best.
How do you find the bad decisions with so many subscriptions active? isn't that looking for needle in a haystack:)?
I think A Bad Decision is the genesis of this.
[dead]
I don't even like the idea of humans operating without guardrails where I work, much less AI.
Why not? What kind of work is this for?
In the glorious future where I can conjure a game with a snap of my fingers for the cost of a pack of gum, why would I ever choose to play Wyvern, the game tailored to Steve Yegge's tastes, when I could just make a game tailored to my tastes? Perhaps Steve will be content to play his MUD with agents pretending to be human players?
You joke, but even very simple/cheap AIs already make it a way bigger pain in the ass than it used to be to identify growth hacking fake user/signup/activity bullshit in all sorts of situations.
It was already bad out there. But now it’s worse.
https://store.steampowered.com/app/1541710/Wyvern/ Wyvern on Steam.
Looks like Runescape
> the subject matter is too complicated to explain … All I can do is walk you around like an excited tour guide, one who has unearthed an ancient alien civilization.
Now this is AI psychosis.
It's interesting that I've also gone down a similar journey and the same keywords, fences, gates .. also showed up in my ai government platform.
I honestly think Yegge has reached a point where he knows he isn't interested in doing the leg work of today's software engineering, but wants his career to ride the wave. The lowest effort, maximum value to extract in his position is abusing the crap out LLMs to an extent that is genuinely novel to secure thought leadership, but on close inspection it's just pointless token spend driving blog posts and publicity.
To me it is absolutely farcical that he was being being paid by a harness company (Amp) as a staff software engineer, slop-coded a solution to keeping LLMs on track (Beads), and having paid for all of this Amp and Steve...parted ways. Beads never went into Amp. Someone cottoned on to the value Steve was providing, for my money.
> being paid by a [...] company
> parted ways
> Someone cottoned on to the value Steve was providing
This pattern predates the LLM obsession.
You just described 99% of AI usage.
Never seen a sloppier hero image
>This tier, or the one just after it, will power hundreds to thousands of new AI employees at every company.
>And companies are in no way, shape, or form prepared for this transition.
so, uh, what's the business model here?
I'd like to read more about the game. Maybe after the new release?
> I didn't want this to be a long post, and I think I've succeeded.
Classic Steve Yegge!
There is method in the madness:
https://yegge.ai/services
The C-suite wants to hear this Gavin Belson level stuff that would traditionally get people institutionalized.
They think he is a inspirational speaker.
Yegge is either a visionary or completely delusional. Probably both. But I'm glad someone is running this experiment. We'll know in a year whether it was genius or madness
It kinda feels like Yegge is trying to start a cult. And, I just don't find that very interesting.
I don't think he's trying to start one, so much as he is being brainwashed by one.
Many cults are started by true believers. The founder can be the first victim.
The code fable generated is interesting from the perspective of understanding how llms behave but it's not an immutable behavior -- it could have been designed not to do that. I think it's quite interesting how it conflated details of the game with the code's architecture. I don't think that's a good thing, and he could have caught it had he read the code.
I think Steve somewhat predicted the future well. He might be slightly off, because he is overly optimistic and operates at the edge, but look at Gas Town, when it was released it felt like dystopian, today it looks not too far away, I am sure most of your orgs are already running some kind of agent to triage the tickets and in some cases automatically open the PR in your git repo.
I think he is onto something this time as well
This weekend I stumbled across a codebase unlike anything I’d ever seen: Swamp Club[]. Came out of nowhere, lots of useful features, but with a coherent philosophy and architecture and a large body of working code. Change volume is high - notably so, but it looks nothing like the vibe slop I’d expected.
A “swamp” contains tools that you build - with the help of agents - intended for use not just by apps and scripts, but agents (to add deterministic behavior, like code mode). Tools are ops things, product things, whatever.
*And so it’s a tool plus a factory to update the tool.*
The factory though isn’t only used for building your tool - that’s just a set of skill.md files. It can build your whole app, which itself can _use_ your tools - either for product features or ops tasks.
*And so it’s a factory that can build any app, including a tool which can operate any app. *
Right now, looking at the commit volume of the swamp stuff, it’s not stupid to suppose that the swamp is being used to create the swamp.
*And so it’s a factory which makes, among other things, factories.*
We’re cooked.
- https://swamp-club.com/use-cases
"I would guess I'm one of a handful of top individuals on Earth outside the frontier labs, in terms of my experience with top-end models."
I'd argue that number of tokens burned does not translate into "experience".
And if you needed to spend $300k or so to find out that process helps manage larger efforts (process is what "fences" are outside Yegge-land), you sure spent a lot.
I mean, I'm glad the guy's got a a hobby he enjoys, but there's less insight than you'd hope for.