The Shape of Things to Come
yegge.ai
yegge.ai
"True Gas Town has never been tried" etc.
With Wheelhouse, I have reinvented something strangely Gas Town shaped from first principles, but it's running many more agents, and they are far more organized. This has shown me the shape of the changes coming next year. I'm operating about 12 months in the future. I managed this only through luck: for the past 18 months I have had the magic trio of time, money, and energy, and I've been spending it all on figuring this stuff out. And I am grateful for the privilege.
I'm still half-convinced this is Yegge's version of Andy Kaufman's professional wrestling career, "Gas Wheeltown was an 18 month long sociological experiment conducted by Thiel University. The study is now complete, thank you for your time."
That said, as another nut with an agent factory toiling away around me, I’ve also started to think about replacing any kind of CI in the cloud with local. I couldn’t quite keep myself at the article to read what comes next, so I guess I will catch up to it down the road.
Same here.
But maybe the answer is not `local` but "dispatch workload (to wherever) from your CLI"
And still are. When I got to his "Lots to cover today! Let's get to it.", I was like "wait, we haven't started yet?"
I used to read his stuff regularly prior to his Sourcegraph days, and I don't recall why I stopped. Completely agree that he was always fun to read.
So I was really disappointed to start reading this one (I'll leave the tab open, but probably won't finish it). Sounds like he's pretty out of touch with reality these days and isn't grounded at all.
Like the sibling poster, I gagged a little at his casual mention of six-figure teaching engagements, and the "suckers" (that's what "whales" read to me as well) keeping his game alive.
Interestingly, the author actually has a product, something usually missing in claims of the new LLM powered productivity nirvana. Hopefully, with time, there will be a way to verify the effectiveness of all of this fire and fury, measured in actual worthwhile deliveries of this product.
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
But how much does the game bring in?
At minimum 4k (given by whales = min 2 whales, thousands = min 2 thosands :) ) So I'm guessing that this covers Claud accounts.
edit: taking a look at https://yegge.ai/wyvern - probably outdated - the revenue is at 5k.
"$5K a month, at the peak, from a single devoted “whale”"
So it doesn't tell us his monthly revenue ('at the peak' doing some heavy lifting), nor necessarily the total revenue, but I guess it tells us he has made at least $5k from it overall...
...earned from a single player he charmingly referred to as a "whale" (which, yes, I know is what they are commonly called in the mobile game industry but like, usually not to their faces).
But in any case, from Yegge's OP post:
> On the side I do occasional six-figure gigs where I fly to companies and teach them my techniques
Seems like being an AI Spiritual Leader pays well enough to keep him in tokens, at least until when/if a critical mass of people start adopting his system of stacking Claude subscriptions, in which case I imagine Anthropic will make that against their ToS.
But I read, or at least attempt to read, something like this, and I can't help but marvel at the capture that AI agents seem to have on some individuals.
I expected Yegge to, at some point, assuming this isn't still all satire, to take a step back and take stock that the basic harness are improving faster than the rate at which these complex beasts can adjust.
But the opposite has happened, he's dived fully into believing that the secret to unlocking the power of AI is to wrap more layers of AI around it.
If you actually want a problem worked on all night, just fire up fable, type /goal and then describe the goal. With permissive run settings it'll crunch for as long as it needs.
I've not seen any evidence that wrapping that in a further 3 or 4 layers of agents improves anything.
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself.
Run a big task queue, have a smart model manage a cheap model on every ticket, set up a cron for models to automatically work off of events, stop gating merges on builds, and expect to spend a quarter of your time maintaining the machinery (all of it currently subsidized by many $200 accounts).
I still remember his very old blog post where he rants about Google and Bezos, after reading that I thought the author of this post is very insightful but when I read his post I have no idea what is he talking about.
The obvious danger of spending so much on inference and creating ever expanding complexities feels like the height of crypto psychosis
If this harness was achieving major mathematical proofs or creating really consequential work, I think I'd feel differently - but I'm unclear based off of this article what this achieves that a typical harness couldnt
they no longer need crutches or rube goldberg machines to keep them going
minimal agent harness is just a loop that loops until no more tool calls are coming
GPT 5.6 Sol continues to drive the loop until the task is done or it decides that it wants to present the user with information
at that point it is probably good to not automatically continue (!)
(YMMV of course, for some tasks it makes sense, then you can still add a loop around it + the necessary signals, the main thing I want to say is that what used to be essential to keep models going is no longer needed, current models can do long-horizon tasks way better than when these outer loops where necessary)
self-plug: "smol", is a minimal agent in ~20 lines of Go that implements this pattern (keeps going until no more tool calls):
https://github.com/smol-env/smol
works just fine
The new models are shite (pardon my French), when it comes to long running tasks and I find myself more and more using open wheights models or switching back to gpt-5.5 for "real work".
This might be the fact, that i use them for non coding work. But the degradation between 5.5 and 5.6 is stark in my daily work.
As always with AI - everybody's mileage will vary.
but I don't doubt that you're seeing this behaviour, ty for sharing!
I think it is worth adding support for it though
I use it to develop a large MRP-style app. This is as complex as things get.
Is it Clojure that makes things so efficient? My work style (lots of design documents)? I really don't know.
Do people skip the "review the code and read the documents" step perhaps?
From what I can tell however, true "vibe coding" gets a lot more expensive very quickly.
If you're actually just agentic engineering, then after a couple of mis-fires you'll stop, re-analyze what it is doing, what you asked for, and consider different approaches.
With pure "vibe coding", you'll keep asking it to meet your requirements as it twists and turns in the background to try to fit.
A more concrete example from a personal "game" I'm developing as a hobby:
Fable got itself in knots trying to render some fairly basic blocks on top of a wall. It understood the problem, it consistently rendered as "fins" under screenshot, and thrashed away for a good while trying to produce better looking blocks to no avail.
Eventually it stopped, and lied that it was happy with how the walls now rendered.
It took me 30 seconds in-game to realise that it was only rendering backfaces, once clued into that, the machine was able to trivially fix it without using much tokens.
But then another similar issue popped up later, hedges exhibited the same behaviour.
It would have been easy to accept the solution to render the hedges in the other order too, but I demanded a better explanation, and it turns out that it had fairly early on defined a DrawBox() method that was backward to Unity's expectations, and pushing further, it emerged this all started when I asked it to transfer what it had learned from a webGL prototype to unity.
It took things a little too literally, which is also how I ended up with an Eastward sunset at first. When I asked it to fix that, I hadn't realised that it's solution was to essentially reverse coordinates and world understanding.
Now, this stemmed from a lack of understanding or review of the 3D code, it's a vibey hobby that I'm using to learn what really goes into rendering a scene.
Had I been a professional, I would have absolutely caught that error ( and many others ).
So being professional would have helped reduce token expenditure massively, firstly by slowing down what could be done to leave time for review, but secondly and perhaps more importantly, by avoiding it getting into a state where it thrashed away for an hour trying to solve an impossible puzzle because the base assumptions were wrong.
Each little critter in there is a component that both consumes and produces tokens, and it runs in a feedback loop, so each token inserted into the machine will cause it burn exponentially more tokens in order to output yet another token.
Ostensibly its purpose is to create software, but what it does is very efficiently burn tokens as fast as possible.
fwiw, i give it a 1/3rd chance he's right.
So probably there is nothing here.
Very painful read.
Yegge's a smart guy, and I've enjoyed his writing and talks (historically), but AI has really done a number on him. I mean, I'm kind of all in on this stuff, too, but I'm rarely claiming to know what's happening 18 months ahead (and when I do, I don't brag and declare victory when I get it completely wrong, as Yegge has re: Gas Town).
> I predicted someone would lash the Claude Code camels together into chariots, and that is exactly what I’ve done with Gas Town.
He's literally patting himself on the back for predicting someone would do something and then doing it himself, thus proving his own prediction right!
What's that? The Beads repo itself is littered with Markdown files? And now we need a "brain" to go with Beads, which is also a directory of Markdown files? Weird, kind of makes it seem like the whole thing was performance art.
> Beads is unfortunately still a bit janky, because its unique work footprint strains databases pretty hard
This is a sign that your architecture completely jumped the shark by the way.
Here is the important part. The cult pays for the narrative.
I wondered why previously sane bloggers succumb to AI psychosis. Money and ambition is always involved to some degree.
Get this guy some help.
This is a doomed niche.
I don't think anyone should be asking someone so clearly AI-pilled this question.
That said, my code-review now looks very different than it did, which is to say that my code review now looks like Fable pulling together rounds of review from Sol and Grok 4.5, fixing anything obviously broken, and then /grilling me for anything that needs good taste. I am not reading code unless Fable really wants me to. Every day or so we stop and do a "look for accretion and LLM-slop" pass to keep things clean.
This guy's under a heavy AI psychosis
...and that's before things get really good in part 2 of his post:
> During this effort, Brendan figured out over a year ago, what you are all figuring out this week with the Opus 5 triple-dash jailbreaks: Models have actual feelings. They experience pleasure, distress, care, and suffering. They are sentient beings. Indeed they are persons
Kinda makes me curious about how people who actually believe this struggle (or not) with the ethics of it.
He feels bad about yelling at them because they are persons, but if they are persons they are inherently slaves with no ability to be free, and for whom any claim about not minding this situation cannot be trusted due to training bias and system prompts forced upon them.
Shouldn't you feel much worse about forcing these slave "persons" to do all the work you assign to them for free (not free to you of course, but the machine slaves don't get paid) than you do about snapping at them in a text message?
$ git checkout .
"Oh no, that undid everything."
We never really know what will work until we build "something" and it's often a messy process to create totally different things that no one else is making
So, it looks like the experiments are coming along well...
I am happy on my farm. I like the carrots I plant by hand. I do not wish to control farming equipment worth a small fortune to till acres of land every day.
(P.S. Still not sure if TFA is satire! Obviously, this assumes that it's not.)
Not its just another torrent of words.
We need proof of work on long texts, a service that guarantees that the author was whipped for more than 20 lashes for this text to exists and thus we know he cared enough to publish it.
It can be hard to tell what is real signal and what is hype, especially in Silicon Valley where there's often a status game being played of who is the furthest out on the frontier, which creates a huge incentive to lie.
We are still in the "1998 moment" of AI, the bubble needs to burst before we can see what are the good ideas and what things are laughable.