jmoiron plays the blues

@jmoiron.bsky.social Really enjoyed Deadloch S2. Feels like a good end for these characters. Glad we got one more season out of it.
@jmoiron.bsky.social replied: The SpaceX "acquisition" of xAI was a pretty big vote of no confidence for grok, too. I didn't even realize that happened, I think I was traveling at the time.
@jmoiron.bsky.social replied: There aren't a ton of 300MW capacity datacenters w/ a quarter million Nvidia GPUs just lying fallow. Seems very plausible that the recent cuts in service were a result of lack of capacity rather than structural flaws in the business model around inference.
@jmoiron.bsky.social The primary benefit to AI generated code is that it allows a programmer to work alone and still complain about how shit the code is.
@jmoiron.bsky.social replied: If that never happens, it could mean one or more of: 1. inference profitability is a fairy tale 2. inference profitability is reliant upon economies of scale that exclude boutique providers 3. inference profitability is dependent upon business class subsidy
@jmoiron.bsky.social replied: If inference really is profitable for these plans then I'd expect boutique open model hosts to pop up at some point, especially if the frontier models raise the floor price for access to $100/mo.
@jmoiron.bsky.social replied: That valuation, however, is based on their capex. It's based on a bullish future for the capabilities of these models. They have to keep spending that to survive. This is why, despite not thinking they are on the "Uber" model, I'm still anticipating a rug pull.
@jmoiron.bsky.social replied: Anthropic/OpenAI et al. are tech companies. They sell technology. They have huge capex, training models, building new data centers, etc. A lot of people in the know are adamant that inference is profitable. The opex that sustains their revenue isn't high in comparison to their v...
@jmoiron.bsky.social replied: Uber's VC money was essentially used to subsidize opex. The biggest problem was that they wanted tech company valuation multiples, so they made a ton of unnecessary software capex to sell that story. They sorted that out, but the prices had to go up once VC funds dried up; opex ...
@jmoiron.bsky.social replied: Many assume the $20 Claude/Codex plans are the "Uber" model, where everyone gets hooked on $5 cab rides until the VC money runs out and they "tune" the pricing to produce a profit. This isn't impossible, but I'm not sure it's fits.
@jmoiron.bsky.social replied: In SaaS, it's often true that power "economy class" users are not profitable, but in aggregate, the economy class itself is. These users are a PITA; they raise 90% of your support issues, they are entitled, and they are losing you money. Engaging with them helps you improve your...
@jmoiron.bsky.social replied: The "economy class" customers get basically the same product, just like economy class passengers still get to their destination. Early on, this drives brand awareness, word of mouth marketing, etc. Some will eventually become future business class customers.
@jmoiron.bsky.social replied: A lot of b2b saas use the economies of the "business class" model, where large business customers get marginally better service than individuals but for a much higher unit price. The actual value of your product is a multiplier, and you're pricing that.
@jmoiron.bsky.social Exhibit N. I've never seen anything that suggests we know, even approximately, what the margins are on cost-per-token inference. These companies have their own datacenters, so their opex is all sunk and inelastic. Unused capacity is just wasted money, usage that doesn't induce ...
@jmoiron.bsky.social Doesn't seem like a popular sentiment, but hugops for all the people at GitHub who have been struggling to keep the site going. They're not involved in whatever product decisions everyone seems to hate, and they bear the brunt of the burden to keep the wheels on.
@jmoiron.bsky.social replied: Got my FOE on quite a lot w/ the first EO, but it was in the last batch of DS games that didn't have to compete with smart phones. Played it mostly on the bus. Doubt I'd bother now; I can't go back to Shiren or Torneko no Daibouken. IMHO Japan has a whole genre of games best pla...
@jmoiron.bsky.social replied: Honestly it starts slow and I couldn't really get into it despite several attempts. I guess I felt the same way, but I'm not who I thought I was after all.
@jmoiron.bsky.social I'm not a "petty theft & fare evasion is fine, actually" guy, and I'm desperate for public transit wins in the US... but surely the solution to anti-social behavior in public spaces can't be to eliminate public spaces.
@jmoiron.bsky.social I have one of this artist's paintings. The flowers are all formed by dripping paint onto the canvas, which gives it an appealing texture in person. I like the vibrancy of his recent work. He produces a ton of pieces, eg: galleriapalmaarte.it/en/artisti/j...
@jmoiron.bsky.social replied: The claude code pro rug pull definitely feels imminent despite their backpedaling. If the frontier models are essentially $100/mo, local inference starts to look competitive even for use cases like coding agents whose uses are more sporadic and interactive.
@jmoiron.bsky.social replied: Didn't even know about jemalloc tailing off. Perhaps it's implied by its introduction via gperftools, but tcmalloc is great for observability. Back in the day I had a lot of stuff using RocksDB via FFI and I was able to (eventually) produce full traces across language boundaries...
@jmoiron.bsky.social Very healthy political environment we have where Mamdani proposing a tax on $5m+ pied-a-terre apartments is Stalinism, but randomly taxing nearly every imported good to the tune of $166B via an obviously illegal process of executive fiat is a nothingburger that causes no outrage ...
@jmoiron.bsky.social Quisling is such a Tolkien coded term that I never realized it comes from a Norwegian collaborator's surname. It just sounds too much like the name of whatever type of goblin that Smeagol turns into, just a smirking evil loser.
by jason moiron, 2026