littlestymaar

Not at all the take I got, it was a specific, very corporate, section about ”say nothing bad about LLMs”

It was merely a “can we please not start a flame war on LLMs and focus on what is relevant to the Rust project itself”.

TL;DR; the author feels betrayed because the Rust leadership didn't take a strong stance against LLMs like the Zig team did.

IMHO LLMs are causing lots of issues in the software world, especially in the open source communities but I don't think Zig's blanket rejection of everything AI-related is a good thing.

Back in 2018 it was already running on M5 Hosting, but still on a single server: https://news.ycombinator.com/item?id=16076041

(And honestly I don't see why it would have changed given that the hardware has been getting faster in the meantime).

Does dang run HN on a single server?

I don't think Dang is responsible of that part, but I may be wrong.

HN used to run on a single server (source: https://news.ycombinator.com/item?id=16076041), and I don't see why that would have changed given that the hardware keeps getting better.

Why does HN use Algolia for search

Algolia is a YC company, it's pretty much free marketing for their asset.

That's almost enough to run a 3W device 24x7 for a year. Do you know any old phone or laptop that can serve as a reliable Jitsi server yet draws an average of 3W or less?

You don't need to run it for 24 hours a day, nor do you need it to be the only process running on your machine.

Jitsi is a bit heavy for a phone, but easily fits on an old M1 macbook/mac mini, whose idle power consumption is well below 10W (idle power consumption is what matters, since your machine isn't going to be handling calls all the time). If you shut it down at night, you're almost good. And for any additional service you serve from the machine, you reduce the cost per service further.

Modern computers are fast, and fascinatingly energy efficient.

Even you use HN.

A website that runs on an infra that could sit in a cupboard under the stairs serving hundreds of thousands of users with very small loading time.

My friends and family aren't going to be convinced to use a Jitsi instance running in my house

(where I pay $0.35/kWh).

Using an old phone or laptop as server means you'll end up with a single digit annual electricity bill for that.

Linen is the most underappreciated fabric. It's cool in both ways. I don't understand why so few people wear linen in summer.

HTML in its recent iterations made it very clear to try and separate itself from layout concerns.

That's what I'm saying, it's a general design principle for new developments, but that's it. (And I don't think that's a good one)

It really isn't about the layout. It got abused for layout

It wasn't abused for layout, it was all about layout in the first place. If it was about data it would never have been possible to “abuse” it for layout, because an heterogeneous table with nested sub tables makes no sense from a “data” point of view. The fact that this was supported from the beginning shows what it was built for.

but ideally the semantic meanings are useful. I have seen cases where "<strong>" emphasizes it by doing something other than bold for example

But I agree with that. I'm not saying we shouldn't have semantics elements in HTML, I'm saying that pretending that HTML “is only about data” is silly in a world were something like 80% of HTML tags in web apps are <div> and <span> that are here only for layout purpose.

Yes, really. Even <table> is used to represent the data, not the layout

Nope. It's a html tag to lay out data in a table form!

but it's probably the only element that does such a thing.

<pre>, <b>, <i> all come with an associated layout implication. But the worst offender are <div> and <span>, which carry zero semantic meaning and just means “I'm a block element” and “I'm an inline element”, those are pure layout element. In fact are the ones I'm using as template for my demand of <flex> or <grid> elements.

And html isn't just tags, it's tags and attributes on these tags, and things like hidden, height or width are indisputably related to layout and not to data.

The idea of HTML as a pure semantic layer with no layout consideration may be a general design philosophy for the language, but that's by no way a faithful description of what HTML actually is in practice. And pretending it is, isn't helping.

TIL thx!

For that to work I would need to define a custom component from JavaScript, wouldn't I? (and I thought custom components had to contain an hyphen in there name, is that wrong?)

Not really. This would not be different to <table> and associated elements, which are arguably better than a div soup.

GP said:

completely orthogonal and swappable css.

You're talking about having a single CSS and HTML page for multiple layouts, there aren't the same thing at all.

so there often aren't semantic HTML tags that would make sense.

Pet peeve of mine: there should really be `<grid>` and `<flex>` elements, as well as `<fi>` (flex item) and `<gi>` (grid item) for the child element instead of relying on a div soup with CSS attributes everywhere.

LD50 is a very poor measurement for this kind of stuff though, there are plenty of stuff that can severely harm you on the long term at relatively low doses but will never realistically kill you in the short term no matter how high the dose (asbestos fibers for instance). Many carcinogenic or reprotoxic stuff are like that.

Maybe it's just because reducing treatment makes produces more fragile, which means by buying organic you end up eating fresher stuff just because it couldn't withstand the logistic that heavily treated food endure.

The problem is that there's a part that's true (the “organic” label is based in part on “appeal to nature”, allowing certain stuff on the basis that it is natural, which truly isn't a great way of drawing the limit).

So of course, the (often paid) promoters of “YOLO agriculture” will use that as an argument, putting under the rug that the industry they defend is doing bullshit with people's health and the environment.

There ought to be a middle ground, but unfortunately we live in a permanent culture war so there cannot be reasonable discussion about anything, really.

It's still limited, but the shift has been dramatic over the past few month alone. It's hard to tell what the situation will look like in a few years. Unless the midterms are a democrats landslide (which the GOP has tried to prevent as much as possible with gerrymandering) I don't see the trend stopping anytime soon.

Does this translate into a similar reduction in compute?

No, quite the opposite actually. Like with speculative decoding this model will compute more tokens and discard the invalid ones.

What's the catch?

LLMs[1] are limited by memory latency and not by compute[2]: because they process tokens one at a time, you spend more time loading and unloading the weights on the GPU registers from VRAM than waiting for compute to happen. Techniques like these allow to process multiple tokens in parallel instead of one by one, and as such exploit better the compute of your graphic card. They do so by predicting which tokens are likely to occur and then verifying that the guess was correct.

For instance if the previous token is “hello”.

A regular autoregressive LLM will compute:

“hello” => “! ”,

then “hello! ” => “how ”,

“hello! how ” => “are ”,

“hello! how are ” => “you”.

and finally “hello! how are you” => “?<end>”

One at a time. Loading and unloading every weights 5 times from the GPU memory to its compute units.

With speculative decoding (I'd say this one isn't strictly speculative decoding, but it's a variant of the same principle), you have something that guesses that the whole sentence is going to be “how are you today?”, so the LLM can generate

“hello” => “! ”,

“hello! ” => “how ”,

“hello! how ” => “are ”,

“hello! how are ” => “you”.

“hello! how are you” => “?<end>”

“hello! how are you today” => “?<end>”

In parallel. So each weight would have been loaded only once from the VRAM instead of 5.

The last token will be discarded though, as the prefix “how are you today” doesn't match what has actually been generated. So in that particular example, you'd have gotten your 5 tokens 5 times faster than with pure autoregressive inference, but at the expense of a 6th token being generated and discarded immediately. So 5 times more token throughtput, but 20% compute cost increase per token.

[1]: autoregressive LLMs, that is. Which are the ones everybody uses because they are the most performant.

[2]: at least when run at low batch size, on your own computer for your personal use. On a datacenter, with many concurrent users, GPUs are actually compute-bound.

So, it's D-Flash but at each transformer layer and share the KV cache of the original model? Very smart!

US tech is currently being weaponized against the ICC and its member judges in Europe[1], and the US is threatening to annex Greenland, as a result all (former) US allies are scrambling to get rid of their strategic dependency.

[1] https://www.lemonde.fr/en/international/article/2025/11/19/n...

Stylometry has existed for decades, and there's no way an LLM is stronger at that job than a specialized piece of software (it's not more realistic than expecting Opus to beat Stockfish at chess).

In practice, you've never been anonymous while posting on the internet and AI isn't changing anything on that front. Or rather: if anything, AI can help you become more anonymous than before, since it can be used to hide your identity from stylometry by rewriting your prose before publishing.

the efficiency changes that have created this perceived downtrend in claude quality”

Why the euphemism? What Anthropic did was an aggressive degradation of their model to save compute, and it's not just “perceived downtrend”, Anthropic themselves have acknowledged the quality of service degradation.

x oil shock (due to Ormuz).

Yeah, there are arguments to be made about the benefits (less teenagers on social media) vs the drawbacks (having to hand your id card to some untrustworthy provider), or the fact that it makes people used to circumventing the law, or about the law addressing the wrong issue (so called “social media” being actively harmful by design in ways that ought to be banned) but claiming that the law increases social media consumption is ridiculous.

I literally wrote the opposite, but ok…

But you probably won't see these outside datacenters for a while.

That's especially true now that Data centers spendings are crazy high.

This would have worked a few years back, but now you can be detained at the US border for posting what you just did so it's a terrible example to pick.

By the way, even with the current administration, there's no question about which is the more authoritarian with their own citizens between China and the US. But if you aren't American, then the US government is much more of a threat than the Chinese.

China cannot make the life of an official in Europe miserable for investigating their atrocities towards the Uighurs, meanwhile CPI judges are now forcedly unbanked and cannot work with American software because they investigated in US's ally's atrocities in Gaza.

How can a medium-sized model like Deepseek-V4-Flash be cheaper than a much smaller models like Qwen3.5-35B-A3B.

It's five times bigger in both total and active parameters!

Yeah, I really don't see the difference with false bomb alerts.

On-device is incredibly far away from being viable. A $20 ChatGPT subscription beats the hell out of the 8B model that a $1,000 computer can run.

That's a very strange comment. Why would anyone run a dense model on a low-end computer? A 8B model is only going to make sense if you have a dGPU. And a Qwen3.6 or Gemma4 MoE aren't going to be “beaten the hell out” for most tasks especially if you use tools.

Finally, over the lifetime of your computer, your ChatGPT subscription is going to cost more than the cost of your reference computer! So the real question should be whether you're better off with a $1000 computer and a ChatGPT subscription or with a $2000 computer (assuming a conservative lifetime of 4 years for the computer).

My Strix Halo desktop (which I paid ~1700€ before OpenAI derailed the RAM market) paired with Qwen3.5 is a close replacement for a $200/month subscription, so the cost/benefit ratio is strongly in favor of the local model in my use case.

The complexity of following model releases and installing things needed for self-hosting is a valid argument against local models, but it's absolutely not the same thing as saying that local models are too bad to use (which is complete BS).