Two Good Things: 2026-06-05
Bathhouses and AI IPOs
Ever since I decided to blog well, I’ve been trying to figure out the right way to write a good “links” article. All the cool kids on Substack do it, and I read the Internet. I told myself I ought to be able to do it too. Two statistics blogs and one links blog a week sounded plausible.
Like most things, though, links blogs looked so easy to get right until I tried to do it. So, we’re experimenting with formats. I think I like this one. Two Good Things. Two stories or clips that I think our worth consuming this week. We’ll get there eventually.
This one has bathhouses and AI skeptics.
Bathhouses
Joe Hovde has a great blog on this company called Bathhouse. It’s a public sauna trying to become a third place—presumably, for the kinds of people who feel comfortable less-than-clothed around strangers: hot people.
A little trend. I’ve noticed that proposed Third Places end up getting accused of being cults—which is natural. They’re trying to replace the Church or the Bar. A Third Place has to have aspects of religion or addiction to satisfy the human need for connection via transcendence or dependency.
I don’t see it crossing generations, but fingers-crossed for them. That’s the key to a successful Third Place. There’s always the old guy at the Bar and the kids with fake ID’s. Church has Sunday school and funerals. Even coffee shops—when those were a bigger deal—cross generations. At least the marketing makes it look like a young single thing?
The website has an “AI Search” feature.
In any case, the article’s a great dive into the business and a category I didn’t know existed before. Check it out.
[I dig this blog in general, by the way. It’s always got some interesting dive into an industry I either don’t know much about or don’t know enough about. Good place to learn stuff.]
“[LLM Companies] Shouldn’t Be Allowed to IPO”
Mr. Zitron doesn’t like AI or AI companies, but he has much stronger arguments for the Company’s IPO prospects than the Product and he kind of mixes the two willy-nilly.
The part he’s probably right about is that no one has a great business model yet for how to make money by supplying LLMs (or maybe the market structure just isn’t right yet to do it: you can’t get enough of a margin), and the set of tasks LLMs are good at is much smaller than the AI-positive discourse admits.
Folks are allowed/encouraged/threatened to use AI, and, for many tasks, the price of the tokens is still way more than the value delivered by the LLM. That’s how you get places that had never used AI until a couple years ago, burning millions of dollars to let folks spin up questionable (but expensive) dashboards without knowing SQL or save some Engineer a few (very expensive) hours.
The part he’s (super?) wrong about is this whole idea he often puts out there that the LLM product itself is bad-to-useless. There are nuanced takes on how LLMs can reduce productivity (by yours truly) because workers will substitute towards the LLMs and away from effort even for tasks where effort is more productive than LLMs, but there are other tasks where LLMs absolutely deliver strong productivity gains.
I do not want to go back to debugging without LLMs.
I spent the last weekend debugging this side project (a Rails app), and it would have taken much longer and been so much worse if it had just been me searching for bugs, writing tests, etc, instead of using Codex. I think I spent around 20 hours debugging the app with Codex (i.e. not just prompting “fix all bugs, make no mistakes” like you hear from some portions of the internet…). It easily would have taken me a few weeks without it, and it probably wouldn’t have been done to the same quality because I’m human, and I miss stuff (edge-cases, various long strings of scenarios that cause odd state X, etc).
The LLMs are a little like chess computers when it comes to debugging. Humans miss possible moves in chess all the time—even grandmasters miss mates-in-one in short time control games. Finding all possible one-shot moves on an 8x8 grid is a trivial problem for a computer, though. Similarly, LLMs can come up with ways in which someone could get something funky by hitting a certain string of endpoints much better than my feeble, meatsack mind.
That being said, it is definitely true, and this is Ed’s main point: the value of debugging code, formatting docs, etcetera does not justify the stratospheric level of capital and debt being loaded into these AI Labs. They kind of have to create God or at least replace Raytheon.
I might like (very much) being able to type “replace all my poorly-formatted math in this doc with well-formatted math,” but I’m not willing to pay much for it. I doubt it drives significant company revenue. Or, for example, with the above experience, debugging a web app:
I did it all in Codex (Free).
Yeah, so it was free for me. It wasn’t close to free for OpenAI, I’m sure. How much would I have paid? Not much. It’s just a little side project. If it was a major potential revenue stream? I guess I’d pay a thou or so for access to debugging for a month or something? But I doubt I’d have to. Competition is still too intense. I’d just substitute to Gemini or Claude. So prices stay low…
Until the shakeout stage of the industry (start here for industry lifecycle literature).
I think two things are true:
LLMs are super valuable.
LLMs are over-valued.
But they won’t both be true when the market for Anthropic’s future dividends becomes liquid and well-informed…
I hope the S-1 becomes public in time for beach reading.
Yes, we do have beaches. Lake Michigan counts.
Thanks for reading!
Zach
Connect at: https://linkedin.com/in/zlflynn