ignoramous

Possible, but pay-as-you-go Hy3 / DeepSeek v4 Pro / MiMo v2.5 Pro (from respective vendors) are genuinely good enough as daily drivers, given the costs (especially, low prices for input cache, which usually makes up 70%+ of total input for agentic workflows). I put in $10 in DeepSeek & Xiaomi MiMo, and I've barely used $1 each, in a week of coding work.

Coding Plans by MiniMax ($20/mo for 1.7b tokens) and Z.ai (~$30/week use for $17/mo) are also tremendous value for money.

This is actually what's meant but this bikeshed has been built for yak shaving.

> Salaries of every employee: https://buffer.com/salaries

Reminds me of Frappe (ERPNext) letting employees choose their own pay: https://frappe.io/blog/culture/choosing-your-own-pay-how-doe... / https://archive.vn/o7NtE

... which ended being, well ... eventful: https://frappe.io/blog/culture/in-the-aftermath-of-exuberanc... / https://archive.vn/BRaQc

> I don't need it to be trustless. If it worked well, the tradeoff would be smaller...

You mean, on a self-hosted Ente (or Immich), a off-device/server-side, capable multimodal LLM, like Gemma4 12b would do?

Ente has a CFP submitted for a India FOSS (Sep 2026) talk that will also be recorded and livestreamed (if selected): https://fossunited.org/c/indiafoss/2026/cfp/2tsmo50knt

Their design team has written a blog post, which may be of interest to you: https://ente.com/blog/ente-design-system / https://archive.vn/ucGnC

> don't really get the business plan for open weights model companies

The Chinese "Neijuan" aside, most competing labs are going for the classic, 'your margin is my opportunity': https://tomtunguz.com/is-your-margin-my-opportunity-software... / https://archive.vn/5Vmq3

> Nowadays every CIA goat herder has their own Starlink terminal.

Can the Starlink radio be sniffed? If so, it might act like a beacon, a tell-tale sign (especially, in countries where Starlink is illegal).

> will collectively effect all of us, why not apply democracy to it

That's because most execs proposing solutions are "Technocrats" or think like one: https://en.wikipedia.org/wiki/Technocracy

Besides, I don't think as a collective we're well equipped to decide one way or the other. If the collective were given a say, billions will be spent, often in consultation with technocrats, on doom / hype marketing (if it isn't happening already).

Tokenzier aside, a report shared on reddit found that the GPT 5.6 (edit: 5.5) series are incredibly thrifty with CoTs, resulting in cheaper bills than GLM 5.2 (let alone Opus/Fable): https://www.reddit.com/r/ZaiGLM/s/rUoG5adkPh

Chattiness remains an open issue for some of the SoTA open weights & (to a lesser extent) Claude.

For France, Spain, and Italy, only the inter-country and the high speed rails (TGV, Renfe, Iryo, Frecciarossa etc) require reservation?

In Switzerland, some of the touristy intra-country routes (ex: Tirano-Chur, Interlaken-Lucerne) may require (a rather expensive) reservation.

Why would you be embarassed for output by an LLM? You might be overthinking this.

Also, it would be courteous to mention in the readme just how this code came about.

> the kind of engineering that goes into Xiaomi (and Deepseeks) inference optimizations

At Xiaomi, MiMo is now led by Luo Fuli. She is a former Alibaba & DeepSeek employee: https://newsen.pku.edu.cn/news_events/news/people/15385.html (https://archive.vn/I8Pmu) / https://e.vnexpress.net/news/tech/personalities/who-is-luo-f... (https://archive.vn/sb3B6)

Don't know if it is due to Luo, but it is striking how similar performance & pricing of the models, DeepSeek v4 Pro & MiMo v2.5 Pro, is.

Curious why TFA calls out "Tencent in China".

  tencent/Hy3. New Apache 2.0 licensed model from Tencent in China

Is there a Tencent AI lab elsewhere (MiniMax have some association with Tencent, for example)?

Lower context window notwithstanding, Hy3's coding benchmarks hold their own against DeepSeek v4 Pro & MiMo v2.5 Pro. That's quite something for a model priced like DeepSeek v4 Flash & MiMo v2.5 (for non-cached tokens), which are 3x cheaper than their respective Pro variants.

Cheaper than Qwen 3.7 Max. Second indication, after Grok 4.5 ($2 in / $6 out), that the BigLabs are feeling the GLM 5.2 heat.

> This entire article is publicly and personally attacking him for choosing a different product.

Well, TFA has a conveniently titled section "Addressing the Blog Post", that raises (setting aside speculation) some good points:

  The [Bun rewrite] blog post is ... almost like the marketing department of a trillion dollar company has a lot of money riding on this article ...
  
  There's a dichotomy being presented here where you have to either choose a "style guide" or a programming language feature in order to avoid bugs. The sleight of hand misdirects the reader away from the main way bugs are eliminated: by dedicating engineering resources to it ...
  
  [TigerBeetle] put in the time to find and eliminate the bugs, they make an effort to maintain a healthy relationship with ZSF, and Bun did not do that.

  The argument for shipping all the million lines of unreviewed code is that the test suite is good enough to catch everything. Then why are you saying you have so many annoying bugs in the Zig code? What happened to the test suite being sufficient to catch everything ...

  Performance increase is attributed to LTO, which Zig has supported for all of Bun's existence. It used to be enabled by default until we ran into too many LLVM bugs, all of which also affect Rust ...

  The post claims they were fuzzing their Zig code, while during our calls the whole Bun team told us that they were not fuzzing anything. This appears to be an outright fabrication.

  The blog post outlines a bunch of engineering work done to reduce binary size, to better make the case that "Bun is better in Rust" ... you were doing the engineering work that you should have done in the Zig codebase since the beginning ...

  I noticed that you neglected to mention compilation speed. Zig compiler project is about 600,000 lines of code - roughly the same size as Bun before the rewrite, and I'm clocking 16s to build from scratch with a clean cache, followed by 90ms for each subsequent edit with incremental compilation enabled. What are the corresponding measurements of Bun post-rewrite?

> I suspect Claude might be faster and therefore cheaper, but maybe not by a lot.

While Jarred used Mythos-class model, some open weights, if they were as capable (certainly, GLM 5.2 looks the part), would have been way, way cheaper than professionals.

Approx costs:

  DeepSeek v4 Pro & Mimo v2.5 Pro   $3,426  ($2,567 / $600 / $259)
  Tencent HY3                       $3,892  ($1,180 / $552 / $2,160)
  GLM 5.2                          $30,016  ($8,260 / $3,036 / $18,720)
  Qwen 3.7 Max                     $37,925  ($14,750 / $5,175 / $18,000)
  Claude Opus 4.8 & GPT 5.5 xhigh  $82,750  ($29,500 / $17,250 / $36,000)

  5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached input token reads.

What does it mean when prompting SoTA LLMs prone to slop to be concise and precise, with respect to context at hand, not work at all? Anyone benchmarking that?

  When Workers Cache is enabled, every cacheable request to your Worker hits Cloudflare's cache first. If there's a fresh cached response, Cloudflare returns it directly — your Worker doesn't run, and you don't pay CPU time for it. On a miss, your Worker runs, and if your response is cacheable, Cloudflare stores it for the next request. The next request from anywhere on Earth can be served straight from cache.

Incredible! This is why I shoehorn all my server side usecases on to the Workers Platform. Cloudflare, since 2020 when I first went all-in, has consistently shipped features that reduce bills significantly (except for 2023 Workers usage model changes). In one case, when they shipped free Snippets (Workers but 32kb code size & 5s CPU time) for Pro accounts ($200/yr), our bills went from £15k+ to £0.

I know about the infamous "Enterprise plan" (especially, when your bandwidth is as high as ours in 100s of TBs) and know of at least one other tech shop that was required to pay for it ... but we haven't got that sales call, yet.

> By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.

There are levels to this (hermetic/reproducible builds & attestations, as one example). In fact, TFA is harsh for the only case it wants to make an example out of, WhatsApp Web:

  Code Verify works in partnership with Cloudflare, a web infrastructure and security company, to provide independent, third-party, transparent verification of the code you're being served on WhatsApp Web. We hope this gives at-risk users peace of mind. 

https://blog.cloudflare.com/cloudflare-verifies-code-whatsap... & https://engineering.fb.com/2022/03/10/security/code-verify/

> much higher base of real intelligence

Not sure how much "real intelligence" is to be found in Mythos & Sol, but at this point, ignoring the intelligence gap, I find it totally impressive that the likes of GLM, Kimi, Qwen, MiMo hold their own at 2x to 4x less cost, and work for my use case just the same.

Interesting. If I may: What was this "clean up" pass? A code review? A code review with specialized prompt? A focused review to check for edge cases / logic errors / api misuse? Or, something else specific to the codebase?

Have you settled on what the clean up pass should look like? Or, do you keep experimenting with it?

In case one might not have been aware: Composer 2 was Kimi Base 2.5 post-trained (RL'd) by Cursor: https://news.ycombinator.com/item?id=48507474. Composer 2.5 might be something totally different.

Copilot Chat supports BYOK since Oct 2025 for the VSCode plugin: https://code.visualstudio.com/blogs/2026/06/18/byok-vscode

> Fable wrote it when it spawned the background task. How am I supposed to rephrase it?

Can the harness to auto-rephrase? I imagine, doing so will burn through tokens though.

Probably busy mirroring binary search trees? ;) https://news.ycombinator.com/item?id=9695102

> Google doesn't even have enough resources to decently deploy a model like that

Probably not for free but tbf, Google did scale "AI Mode" globally to its billion+ users, with its Gemini 3 series. Pretty much broke my habit of searching the web with pplx & Chat.

Amazing! Thanks for sharing.

> I didn't look at any of the code, but I had it build me a bunch of various visualization and tuning UIs. I was basically acting as a PM/TPM/QA engineer, and what I was able to do in a couple evenings is stuff that entire teams used to spend weeks on.

Agree. The pace of improvement (for coding tasks) from all Labs vying to catch up to the Big 3 (GDM, OAI, Ant) has been relentless.

> So I asked my agent (I'm using some free agents that are pretty decent - Nemotron Ultra from OpenRouter and Big Pickle from OpenCode Zen) to build me an OpenCV classifier to try to read the digits.

I think, Meta's Muse Spark is also free to use, for now. I'd be interesting if you'd ever find yourself using that.

If I may: Did you find Big Pickle is better than Nemotron Ultra or vice versa? Which harness (or IDE) did you drive these models with?

> Mullvad believes privacy is a universal right. You might disagree but at least it's consistent ...

Is privacy inconsistent with right to dignity, liberty, and equality?