SwellJoe

I didn't think I needed to spell it out for the HN crowd, since folks talk about the Chinese models every day. But, OK...

At least three Chinese companies have bragged about their latest models being trained and/or served on Chinese-made GPUs. Literally within the past 24 hours, LongCat spent a large portion of their release announcement for 2.5 bragging about how efficient their training process is and that it runs on domestic (Chinese) GPUs: "LongCat-2.5 uses long-sequence training during both mid-training and SFT. With less than 80 GB of memory per domestic AI accelerator, the training system prioritizes memory efficiency and aligns its parallelism with the supernode architecture."

https://longcat.ai/blog/longcat-2.5/

DeepSeek has repeatedly blown the curve in terms of how efficiently a model can be trained and served, Xiaomi and Z.ai aren't far behind, and have also discussed training and serving on home-grown hardware. I'm sure you can search Arxiv for all of the papers from these providers (DeepSeek, in particular, who document seemingly everything, even when they're doing something ahead of the US providers) as well as I can.

This is why Musk rushed consolidating xAI with Spacex and rushed to an IPO. And, why it was so important to him that Nasdaq and Russell rewrite their rules so he could dip into everyone's retirement accounts as quickly as possible. I don't think he's particularly bright, but he has an incredible cunning about how markets behave and how to manipulate those markets to his ends.

The writing was already on the wall about a lot of things WRT the AI market: 1. there's a limit to how much money will be thrown at it before investors expect a return, and that limit is fast approaching, 2. it's turning out "intelligence" is going to be a commodity and it's not much of a moat, 3. cornering the market on memory and GPUs (even if they sit in a warehouse for months or years before going online, as long as it drives up the price for competitors) only works if your competitors don't figure out how to do it more efficiently and on alternative platforms.

Spacex also soaked up a lot of the retail appetite for AI, even though it's an also-ran that barely qualifies, in terms of enterprise and government usage (and for good reasons, both technical and otherwise). Every other AI IPO is going to be facing a more pessimistic market, especially if we continue to be in a tough economic environment.

In short: They waited too long, now they're in a world where they've burned so much money people expect them to deliver something that no one in the AI industry probably can deliver: a monopoly-like hold on a huge percentage of the market, and to invent a bunch of new markets that aren't currently buying AI. They have to have a dominant position for this to work because they have to be able to set their price higher than is standard now, without everyone fleeing to other cheaper providers. And, when the market for memory and GPUs finally gets back to normal and supply catches up with demand, I'd wager a lot of companies, and even individuals, are going to get real serious about open models and self-hosting (unless prices remain competitive for the hosted models).

You've made the mistake of thinking Anthropic just wants to sell everybody a subscription to their models, so everyone can make their little spreadsheets or whatever.

That's not what Anthropic wants. They plan to be The Company. The only one. The one that does everything and makes every decision.

On 64% of websites.

Sure, there are times in a familiar codebase that it would take Claude or any model longer to get its bearings and chew on all the random nonsense it wants to think about even for small changes than it takes me to open vim and make a one line change.

And, sometimes, I know models will write it in a way I don't like, over-engineering is common.

But, I mostly use models for more than few lines of code or code that isn't in a familiar project.

I think that requires testing, rather than assuming. There's only so far a harness can go.

Bonsai 2-bit quantization is notably dumber than Qwen 27B at eight bits or even four bits. I doubt you'll get this level of quality even with five of them.

"Their kernel orgs are pretty lean"

For years, Red Hat employed more kernel developers than anyone and contributed more. These days they're "only" the fourth largest contributor, after Intel, Google, and Meta, but they are still among the largest corporate contributors.

If that's a lean kernel org, I don't know what a fat one would look like.

I like nearly all of my coworkers, but they're not my friends. I met my friends doing things that don't pay, maybe things that lose money like making music or art.

Yes. For similar tasks using Antigravity+Argon vs. Codex+Astra vs. Claude+Opus, Gemini has the most generous limits on the $20 plan. Since the ChatGPT limits got nerfed last month (or whenever, I can't keep up), the $20 ChatGPT plan is very stingy. With similar tasks I'd see 2x-3x faster consumption using Codex+Astra vs. Antigravity+Argon. I could casually code all day on the $20 Gemini plan, using Argon on Medium, without ever bumping into the limit, which is not true of ChatGPT or Claude $20 plans (but used to be true of ChatGPT). If I were using automated tools and doing a bunch of things at once, I'm sure I could saturate it, but I don't think most humans are very good multitaskers even with agents doing most of the work, so I tend to stay pretty engaged with what's happening and while the agent is working, I'm spending time looking over the previous work, testing, editing docs, etc. I'm not juggling five projects at once, most of the time.

The order of generosity, as it seems to be from my limited testing is currently Gemini->Claude->ChatGPT. (And, if we're including Chinese subscriptions, Kimi is below all of those. Very stingy.)

I'm talking specifically about using it in Antigravity CLI.

I don't like it when unrelated projects squat on the name of some existing project. Trademark isn't just a commercially useful thing, it helps users understand what they're getting and who they're getting it from.

It's shitty for a new project to ride on the coattails of a trusted project with years of reputation. It misleads and confuses folks.

Take inspiration anywhere you like, borrow the code all you like, but don't steal the damned name.

The problem with Argon, that makes it feel like older models to me, is that it stops and waits for confirmations far more often. If run (in a VM or container) in the dangerously skip permissions mode, it gets a lot less annoying. It feels quite smart and thorough, to me, though still lagging Opus 5.5. I'd be fine using it instead of Opus for most things.

This post explicitly says that's been improved in Carbon. So, here's hoping. Google currently has the most generous usage limits on their $20 plan among US providers, so it's a pretty good deal.

That tracks. Fable refuses anything interesting, and Opus from 5 onward also refuses pretty regularly. It's one big reason I have other subscriptions.

I suspect the public model will refuse. Anthropic models have gotten extremely strict on biology and security work.

I'd be happy to be wrong, though. I've never tried to do biology with Claude, but I have gotten many refusals and downgrades for security work.

Yeah, without the ability to search the web, a very small model is going to be too dumb to be useful, and maybe even dangerous due to hallucinations. But, there is a point where the knowledge of self-hostable models will surprise you. That point is probably not runnable on the average laptop, though. It's maybe in the ~30B range, but more realistically in the 128B MoE range. I guess in their tests Qwen 3.8 27B got as many right as the much larger Nemotron Super, but I'd bet Qwen 3.8 Flash would be better than either, but it needs 128GB of unified RAM or a big GPU and lots of system RAM to run the 4-bit version.

I think if I had to build an emergency LLM today, that had to run on modest hardware, I'd bet on Gemma 4 12B. It's got excellent reasoning and good vision/audio and speaks 140 languages.

But, I don't know about the "classic" knowledge base theory, though. Is there a high quality curated pile of survival information that could be bundled into a web UI and made searchable in a knowledge base? The LLMs aren't focused on that topic, so there's a lot of wasted space, but they know a little about a lot of things, including survival topics. Combined with Wikipedia, and maybe a few other sources, you're probably better off with one than without.

I guess one could use some huge models plus web search to compile/distill a focused survival knowledge base or LLM.

Edit: Project NOMAD seems to mostly be exactly that. It has authoritative sources in a searchable knowledge base and the tiny LLMs are an afterthought, which is probably what they should be. https://www.projectnomad.us/

This is the worst kind of slop. The kind where someone doesn't exhibit any concern for quality or user experience. I know that's so, because it's a gazillion unrelated apps that clone existing huge apps, built by a couple of people and Claude.

The reality of launching a good product with AI is that the coding is very fast, but the QC, the documentation, the refinement and user testing, isn't all that much faster, if you care about it being good.

This is a mountain of code, full of bugs and weirdness that users will be the first to find, because nobody else is looking at it.

My process for building with AI is "intense weekend of vibe coding" and then several times that time spent on refinement and testing that comes from actually using the software a lot. A couple of devs can't be responsible for every major application.

With AI, you can make anything you want, but you can't make everything.

Because society breaks down when there is an elite class of people who are so wealthy the law no longer applies to them, while a significant portion of people are on the brink of being homeless, and many more are one health emergency away from it.

I use Hugo published to Github Pages for my blog. I've never thought about performance (it's Fast Enough™ when I visit it), but I just checked and it's 100 on pagespeed. The magic of a static HTML blog is you don't have to think about performance. An entire category of problem just disappears.

The surveillance capitalists have made it hacker news, by putting their thumb on the scale for fascists.

I love it when democracy punches a fascist in the mouth.

There's no reason to wait until a child is six to teach them to read. I was reading much earlier, before formal schooling, because my parents were readers and read to me regularly. I had a tremendous enthusiasm for books because books were important to my parents.

Plenty of research indicates reading to kids from a very young age improves their education outcomes later in life.

I'm sorry your parents didn't know to do that. Hopefully now that you know, you'll read to your kids (if/when you have any).

Nobody's running GLM 5.3 on a Strix Halo. But, yes, I'm using cloud models for everything hard, including coding. I use local models for smaller API tasks or sub-agent tasks and experiments with fine-tuning and benchmarking.

"Any suggestion that Meta favors one side is false"

Weird how it keeps happening (e.g. just last week, Instagram and the No Kings account), and it's always to the benefit of right-wing extremists. I'm sure it's just an accident, though, surely they wouldn't lie on the internet.

For folks with 128GB inference machines that's definitely the more interesting model. 27B is dog slow on a Strix Halo. And, while Qwen 3.8 Flash is quite snug in 128GB at four bits (a bit more comfortably at three bits, but intelligence is really dropping by then) but is notably faster than the dense model. It thinks a little less and thinks a little faster, adding up to a real improvement. At four bits, it feels a little bit smarter than 27B even at eight.

If someone has one or two 32GB GPUs with decent bandwidth (faster than the Strix Halo or DGX Spark), the dense model is the obviously better choice, and I think that case is more common for the average developer.

Way off topic, but I enjoyed Amiga E. And, I reckon you're one of a handful of people I would take seriously when they say, "I made a new programming language."

"I had a friend once that kept using open-ended phrasings and ambiguous descriptions"

A failure to express intent, not a failure of grammar. Perfect grammar can mislead just as easily.

I consider that among the unsavory use cases, I think. I'm not sure how I feel about its use in art, exactly, but I'm confident this tech will most often be used for propaganda (because that's how it has already been used most broadly). It is currently possible to detect AI video and audio, with reasonable confidence. I mean, unless you're a facebook boomer, in which case nothing is too ridiculous to be believed.

But, a world where billionaires can make ads showing candidates saying and doing things they've never said or done is a very dangerous world, and being able to fabricate a podcast or whatever very convincingly is going to empower the very worst people. Blurring the line between fact and fiction, so nothing one sees or hears is really believable anymore, serves fascists and autocrats, exclusively.

https://www.usatoday.com/story/news/state/texas/2026/09/09/k...

When one interprets grammar rules as prescriptive rather than descriptive, it is easy to fall into the trap of believing in "correct" and "incorrect", when expressing intent is the single most important thing. Expressing intent is not a thing AI can do, thus far, as they don't want anything...there is no intention.

I'm certain there are folks out there who have a hard time with written English that could benefit from an AI editor, but I would almost always rather read casual English written by a human, even if idiosyncratic and "incorrect" sometimes.

And, to be clear: Hard to read English can never be correct, no matter how carefully it follows grammatical rules, and AI often writes hard to read English.

It's one particle. It's not a bomb.