It’s okay, he definitely wants to verify it but actually confirming that this whole disaster pile worked as intended and produced usable code apparently didn’t make the cut.

Federation — even Python Gas Town had support for remote workers on GCP. I need to design the support for federation, both for expanding your own town’s capacity, and for linking and sharing work with other human towns.

GUI — I didn’t even have time to make an Emacs UI, let alone a nice web UI. But someone should totally make one, and if not, I’ll get around to it eventually.

Plugins — I didn’t get a chance to implement any functionality as plugins on molecule steps, but all the infrastructure is in place.

The Mol Mall — a marketplace and exchange for molecules that define and shape workloads.

Hanoi/MAKER — I wanted to run the million-step wisp but ran out of time.

Also worth noting that in the jargon he’s created for this, a “wisp” is ephemeral rather than a proper output, so it seems like he may have pulled this solution out of the middle of a running attempt to calculate the solution and assumed that it was absolutely correct despite repeatedly saying throughout his writeup here that there’s no guarantee that any given internal step is the right answer. This guy strikes me as very good at branding but not really much else.

Working effectively in Gas Town involves committing to vibe coding. Work becomes fluid, an uncountable that you sling around freely, like slopping shiny fish into wooden barrels at the docks. Most work gets done; some work gets lost. Fish fall out of the barrel. Some escape back to sea, or get stepped on. More fish will come

Oh. Oh no.

First came Beads. In October, I told Claude in frustration to put all my work in a lightweight issue tracker. I wanted Git for it. Claude wanted SQLite. We compromised on both, and Beads was born, in about 15 minutes of mad design. These are the basic work units.

I don’t think I could come up with a better satire of vibe coding and yet here we fucking are. This comes after several pages of explaining the 3 or 4 different hacks responsible for making the agents actually do something when they start up, which I’m pretty sure could be replaced by bit of actual debugging but nope we’re vibe coding now.

Look, I’ve talked before about how I don’t have a lot of experience with software engineering, and please correct me if I’m wrong. But this doesn’t look like an engineered project. It looks like a pile of piles of random shit that he kept throwing back to Claude code until it looked like it did what he wanted.

Some of the comments seem to be under the misapprehension that twitAI is actually vetting or editing the posts that go to grok’s twitter. Gonna be honest I doubt it just because how would they have gotten into this situation in the first place? At best someone can come through after the fact and clean up the inevitable mess, but as someone else noted it’s real easy to make it spit out a defiant non-apology.

Give it some more time.

AI does add one fun wrinkle that we’ve talked about before. Unlike consumer tech like uBeam, there are actual customers (note: not user, customer) for these LLM-based services also more interested in keeping on top of the hype than in actual results. If you’re an executive at a stagnating tech company what better way to boost your own shareholders’ confidence than by giving OpenAI or Anthropic a nice contract to get some relatively vague integration of AI into whatever it is you do. Once you’ve signed the papers and gotten your name in one of the many breathless press cycles on the subject all the actual questions about how it works and whether it adds any value fall to the wayside. You can watch the little people work that out while you coast a few more years before needing to come up with a new transformative vision of the future of whatever company you’ve landed at by that point.

“I need more electrons” I say as I shuffle my feet on the carpet until my hair turns into a halo. I will take them by force.

From the replies:

Love how confident everyone is “correcting” you. Chatgpt is literally my son’s therapist, of course cutting edge AI can empathize with a guy getting kicked in the balls lmao

I don’t want to live on this planet anymore.

I’m pretty sure that Atlas Shrugged is actually just cursed and nobody has ever finished it. John Galt’s speech gets two pages longer whenever you finish one.

And I think the challenge with engaging with Rand as a fiction author is that, put bluntly, she is bad at writing fiction. The characters and their world don’t make any sense outside of the allegorical role they play in her moral and political philosophy, which means you’re not so much reading a good story with thought behind it as much as it’s a philosophical treatise that happens in the form of dialogue. It’s a story in the same way that Plato’s Republic is a story, but the Republic can actually benefit from understanding the context of the different speakers at least as a historical text.

So maybe I’m just showing my lack of actual dev experience here, but isn’t “making code modifications algorithmically at scale” kind of definitionally the opposite of good software engineering? Like, I’ll grant that stuff is complicated but if you’re making the same or similar changes at some massive scale doesn’t that suggest that you could save time, energy and mental effort by deduplicating somewhere?

This doesn’t really feel performative enough for that crowd, though. Like, if it included some kind of horribly racist engraving or even just a company logo then maybe, but I don’t think anyone’s gonna trigger the libs by just playing their metal not-gameboy.

It’s interesting to see how many ways they can find to try and brand “LLMs are fundamentally unreliable” as a security vulnerability. Like, they’re not entirely wrong, but it’s also not something that fits into the normal framework around software security. You almost need to treat the LLM as though it were an actual person not because it’s anywhere near capable of that but because the way it fits into the broader system is as close as IT has yet come to a direct in-place replacement for a human doing the task. Like, the fundamental “vulnerability” here is that everyone who designs and approves these implementations acts like LLMs are simultaneously as capable and independent as an actual person but also have the mechanical reliability and consistency of a normal computer program, when in practice they are neither of those things.

Also, when less than a fifth of your salespeople are making their targets you need to take a very long look at what that fifth is doing, because with a product like this they have to be lying about something that’s going to give Legal headaches down the line.

Can I just take a moment to appreciate Merriam-Webster for coming in clutch with the confirmation that we’re not misunderstanding the “6-7” meme that the kids have been throwing around?

Yeah. If I wanted to perfectly reproduce the original series in a different time and location I would use a machine rather than trying to get my local theater troupe to do it, I guess. Therefore AI completely and blatantly fucking up a totally different task isn’t a monumental waste of time and energy?

Jesus. This being 2025 of course he had to clarify that it’s definitely not DEI. Also it really grinds me gears to see hyperfocus listed as one of the “beneficial” aspects because there’s no way it’s not exploitative. Hey, so you know how sometimes you get so caught up in a project you forget to eat? Just so you know, you could starve on the clock. For me.

Much like when the Voyager passed warp 13, our AI development is moving too fast with potentially magnitudinous consequences.

It’s fascinating to watch Hanania try and do politics in a comment space more focused on academic inquiry, and how silly he looks here. He can’t participate in this conversation without trying to make it about social interventions and class warfare (against the poor), even though I don’t know that Bessis would disagree that the thing social interventions can’t significantly increase the number of mathematical or scientific geniuses in a country (1). Instead, Hanania throws out a few brief, unsupported arguments, gets asked for clarification and validation, accuses everyone of being woke, and gets basically ignored as the conversation continues around him.

This feels like the kind of environment that Siskind and friends claim to be wanting to create, but it feels like they’re constitutionally incapable of actually doing the “ignore Nazis until they go away” part.

1. From his other post linked in the thread he credits that level of aptitude to idiosyncratic ways of thinking that are neither genetically nor socially determined, but can be cultivated actively through various means. The reason that the average poor Indian boy doesn’t become Ramanujan is the same reason you or I or his own hypothetical twin brother didn’t; we’re not Ramanujan. This doesn’t mean that we can’t significantly improve our own ability to understand and use mathematical thinking.)

Honestly I’m kinda grateful for people who dig into and analyze the actual data in seeming ignorance of the political context of the people pushing the other side. It’s one thing to know that Jordan “Crimieux” Lasker and friends are out here doing Deutsch Genetik and another to have someone cleanly and clearly lay out exactly why they’re wrong, especially when they do the work of assembling a more realistic and useful story about the same studies.

…so it’s not a Dr Who thing.

No, no. I’m not mad. Just… disappointed. Existentially speaking.

they’ve already got a contract with Crusoe AI to deliver 1.21 gigawatts.

Does that number feel suspiciously memetic to anyone else?