Would you be willing to share these links? ROT13 or base64 encode the URLs.

Here's a counterthesis:

This is people having fun with a new technology that is far from perfect, is full of unknowns, but is ripe for exploration and discovery.

Gas Town itself is a piece of speculative fiction: throwing out a hypothesis as to what might be possible were inference to drastically drop in price. Its supervisor + isolated worker + merge factory approach is an experimental spike into how agentic coding could play out at scale.

And funnily enough, it is also the approach that Anysphere arrived at through their own experimentation.

Karpathy's alien technology metaphor is particularly apt. No one knows how to use these tools properly yet. We're having some success and a lot of fun, but really we're only going to find out by experimenting in public and sharing our results. Which means the positive and negative.

I don't understand why people see basic automation of the SDLC and think to themselves "this dude cracked the orchestration code" as if it's something profound.

I would not call this “basic automation”. I’m also not saying “this dude cracked the orchestration code” (you’re free to be mad at people who are, but I feel like it’s more interesting to engage with the people who aren’t).

I will say that Gas Town is the most maximalist approach I’ve seen to accounting for the myriad flaws of current generation agents, essentially treating them as cattle and seeing if something of worth can be gained from a sort of brute force approach. I think that’s interesting, and I’m glad that someone built a (somewhat) working system to show what happens if you do that, because no one has built something like this (in public) before.

Overall I think it’s way better to think about this as a big gift basket full of ideas. Take the ones you like, regift the almonds to your cousin if you don’t like them. If someone sees me eating Gas Town banana cream truffle and goes “ZOMG, I NEED TO BUY $GAS NOW.” then that’s their problem, as neither Steve Yegge nor I are telling them to do that.

Because they aren't saying that.

They're saying "this is a very thoughtful way to approach orchestration for using AI coding agents". This experiment is profound not because it works or because its THE end game, but rather it's a novel approach worth testing...but so is Ralph AI.

The supervisor-worker architecture is standard for distributed systems, but I'm not sure the unit economics make sense yet. Given current latency and inference costs, that specific pattern seems significantly more expensive and slower than a human developer.

They don't make sense yet. It works out very expensive at present. Far too expensive to run as anything other than an experiment.

> The supervisor-worker architecture is standard for distributed systems, but I'm not sure the unit economics make sense yet. ... [T]hat specific pattern seems significantly more expensive and slower than a human developer.

Yeggie very explicitly states that Gas Town is for people who give zero shits about how much money they're forking over to their LLM company. If I remember correctly, he said that he had to get a second entire account because of how much money he was spending.

The better comp is how much gastown would cost if it weren't on subsidized the 200$ per month claude max plans (even if he is using 2 of them)

Fair point. I suspect if you priced that workload out on per-token API costs, it would be completely unviable for a bootstrapped business. The flat-rate subscription is really the only thing making it accessible right now.

> The flat-rate subscription is really the only thing making it accessible right now.

So this paragraph from the "Welcome To Gas Town" article [0] suggests to me that real, sustained users of a Gas Town instance are paying far, far more than -say- 600USD per month:

  Gas Town is also expensive as hell. You won’t like Gas Town if you ever have to think, even for a moment, about where money comes from. I had to get my second Claude Code account, finally; they don’t let you siphon unlimited dollars from a single account, so you need multiple emails and siphons, it’s all very silly. My calculations show that now that Gas Town has finally achieved liftoff, I will need a third Claude Code account by the end of next week. It is a cash guzzler.

200 USD per month is something that I -as a working programmer- wouldn't think twice about spending for a fantastically useful tool (even if I had to spend it from my own pocket). If I had to pay 600 USD/month out-of-pocket, it would have me thinking for a bit to see if it was really worth it, but if the company was footing the bill, I'd expense it without a second thought.

Compared to USian programmer pay (especially Yeggie-level pay), 600 USD/month absolutely does not qualify as "a cash guzzler". Hell, that's less than the cost of the sort of health insurance you usually get at nice software companies.

I suppose that there's an alternative interpretation where Yeggie is concerned about the actual cost to the LLM company for the queries that Gas Town makes... but that seems unlikely to me. First, why would he care? Second, why would he say "You won’t like Gas Town if you ever have to think, even for a moment, about where money comes from."? I would give zero shits about where my LLM company's money comes from... that's not my problem.

[0] <https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16d...>

He would care because the 200/m relies on users not using the whole allotment and is likely heavily subsidized. What if the true cost is 4x? (Feel free to add api pricing numbers and correct). Is a programmer willing to spend 2400/month?

is it not also one of the most logical approaches to experiment with first? it's loosely how I've been working with agents as well.

I make this point to say, if someone were to try to claim this approach as IP we should expect it to be denied right?

I guess you could say that is perhaps why the suggestion made by the post being replied to doesn't make much sense to me.

I fail to see how anyone had fun with cursors Web browser.

Gastown is fun in the same way time cube is.

If it smells like bullshit and looks like bullshit there's little need to eat it to make sure it tastes like bullshit too.

You don't think Anysphere did?

I looked at the code base. Did you?

> I fail to see how anyone had fun with cursors Web browser

The codebase that versions funness intervals of people?

Take a look at CodeRabbit and Sourcery if you want to give that a go.

They're leveraging the (relative) ubiquity of npm amongst developers.

Is paying $20 to read this the only option?

Addendum:

Tables from the study available here: https://www.pnas.org/action/downloadSupplement?doi=10.1073%2...

I find some of it interesting. I'm very interested in understanding why others' experience of using genAI is so vastly different to my own.

(For me it's been as transformational a change as discovering I could do my high school homework on a word processor in the 90s when what I suspect was undiagnosed dyspraxia made writing large volumes of text by hand very painful).

I'm also interested in understanding if the envisaged transformation of developers into orchestrators, supervisors, tastemakers and curators is realistic, desirable or possible. And if that is even the correct mental model.

I don’t like “AI is useless” as an argument because

* it is basically invalidated by somebody saying “well I find it useful”

* it is easy to believe it’s on a path toward usefulness

OTOH it is worth keeping in mind that we haven’t seen what a profitable AI company looks like. If nothing else this technology has massive potential for enshittification…

I agree with you entirely. On the other hand, I love that nobody will ever be able to take the current open models away from us.

I tried GLM 4.7 in Opencode today. In terms of capability and autonomy, it's about on par with Sonnet 3.7. Not terrible for a 10th the price of an Anthropic plan, but not a replacement.

It's all my fault. I make sure I burn through my weekly limit every week.

I pretty much do most of my AI coding in Rust. Although I do still use Python or Typescript where appropriate.

It's the following that is problematic: "I asked each of them to fix the error, specifying that I wanted completed code only, without commentary."

GPT-5 has been trained to adhere to instructions more strictly than GPT-4. If it is given nonsense or contradictory instructions, it is a known issue that it will produce unereliable results.

A more realistic scenario would have been for him to have requested a plan or proposal as to how the model might fix the problem.