DiabloD3

Also, do not plug any Internet-facing devices into it, have a captive SBC that connects to a LAN server hosting (legally owned) videos.

You do not, and never did, need AI for tab complete.

Might you rather be interested in https://github.com/ahatem/IoskeleyMono ?

Generally the process isn't worth it unless they think they have a near certain chance of winning against the child.

I do agree, however, that in bird culture, sharing nudes that were sent to you is considered a dick move.

I have no reason to believe this isn't another Deplorean.

https://www.aclu.org/news/juvenile-justice/minnesota-prosecu...

This is one such case, but hardly unique.

The law must recognize the right to self-determinism and bodily autonomy, and you can't inherently be both the victim and the abuser. There is no one size fits all law, and many governments are trying to act like they have one.

Also your account is 10 hours old, hows the Russian datacenters this time of year?

CSAM laws exist to punish abusers, and seek justice for the abused. Not one or the other, but both. The law cannot function without the victims, else you're just attacking people for what amounts to thought crimes.

I find CSAM and pedophiles disgusting, and I imagine 100% of HN users agree with me on this, but this just opens the door for world governments to a) claim LLM output is real and factual (!), but b) minimizes the danger to victims and equates victimhood to being a fictional character.

I'm not Danish and I'm hoping this is just some Dane-specific weird quirk of their law, but from what I can see, this is a politically-biased ruling to "think of the children".

Are you sure it isn't an issue where a specific app just doesn't know how to use ttc?

I would love to see the opinion of a copyright lawyer that would be willing to bring an LLM output copyright case to the Supreme Court, their opinion must be absolutely wild.

This whole thing was most recently affirmed in the 9th Circuit case Naruto v. Slater, where a monkey used a photographer's camera to take a selfie, and the photographer claimed he owned the Copyright because it was his camera, and then attempted to extort the Wikimedia Foundation. Monkeys are clearly intelligent and actually alive (unlike LLMs) yet that still did not alter the opinion of the Judges based on the current interpretation of the Copyright Act (which clearly states only humans can hold Copyright) and all the decades of case law proceeding this case.

As for specifically GenAI, the case involving Théâtre D'opéra Spatial also moved in this exact direction: it is a Midjourney manufactured image that the prompt-giver attempted to Copyright, and the USCO rejected the application and the USCO Review Board also rejected it, however the prompt-giver has not elected to sue the USCO over this.

The part you'd all be interested is the table at the bottom: ideal font size to get the most out of your favorite code font, but also ordered by font size (for maximum code density on screen while still being readable).

Even though vector fonts are, well, vectors and not bitmaps, they often are fuzzy or misshapen at some sizes. The only way to objectively judge a font is to view it at it's best. Some fonts are objectively bad.

Iosevka is my daily driver, but not the most dense font on the list.

I am very much paying attention to the current state of LLM acceptance.

He is now the leader of a cabal of maintainers, and the only coding he does now is minor fixups for branch merging.

I suspect he still codes his own pet projects in his off time though, that's how we got git after all.

Nobody is really against LLMs producing, for example, exploits for novel security bugs to assist security researchers in patching those bugs. The exploit itself does not need to be Copyrightable nor Open Source. Nor is anybody against LLMs accelerating debugging, and helping find the location and root cause of the bug; again, that is not producing the end product code, thus does not need to be Copyrightable nor Open Source.

The only place it works is dumb boilerplate generation (ie, a smarter replacement for templates), but the real fix for that is just jettisoning languages that beat you to death with boilerplate; its 2026, stop using C++ and Java.

I've extensively tried every LLM model, commercial or not, and for real languages like Rust or Zig, they perform poorly, and for footgun-laden languages like C++ or Java or Python, it just hands you the footgun and helps you aim the thing.

Druid is still dead, consumer DGPUs are still, and workstation DGPUs also seems to be dead, and most of the Xe driver team for Windows was laid off, as well as most of the Xe chip design team.

virtio is not enabled by default because its buggy in certain corner cases, and will be enabled by default when it stops doing that. That said, I do have that enabled, it is certainly faster.

https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_proces...

If you know what any given model uses for BW per token and flops per token, you can roughly infer for many models until you get better benchmarks.

There is also:

CUDA: https://github.com/ggml-org/llama.cpp/discussions/15013

HIP: https://github.com/ggml-org/llama.cpp/discussions/15021

Vulkan: https://github.com/ggml-org/llama.cpp/discussions/10879

Metal: https://github.com/ggml-org/llama.cpp/discussions/4167

These three are measured against the old model of Llama 2 7B, but it also means the numbers are without any per-model quirks that reduce compute and memory BW usage, thus more or less generically applicable.

You need to get more desktop GPUs on here, for the inference at home people.

It also misses most cards being used for inference, only limiting it to relatively recent Nvidia cards. Even if this was meant purely for datacenters, it would be useless for AMD customers.

And as for different models being different, it also doesn't understand kv cache size, how many concurrent sessions you need to manage, the compute cost overhead of kv cache quantization, nor the compute cost overhead of model quantization. It also doesn't know what MTP nor DFlash is, and cannot increase your effective tps to match.

As an example: compare Unsloth's quant of Qwen 3.8 Q4_K_M (4.5 BPW), a decent smaller model, without MTP, you'd get a baseline of, say, around 3-6 TPS per 100W. With the built in MTP model, you're closer to 6-12 TPS per 100W; then, switch quants to Byteshape's IQ4-XS (3.84 BPW) and use their suggested Dflash drafter, you're now in the realm of 15-30 TPS

... yet if I scribble into the calculator "27B, 27B active, 4-bit, 1B per one day", it seems to be the low end of my non-MTP figure: 1B per day = 11574 per second, it recommends 3x B200s, each B200 is 1200W, so 11574 / (1200*3) = 3.215, yet real world results would be 2x to 10x higher.

Also, quoting from the website, "A100 and V100 results for the large models are theoretical.". The small scale inference people absolutely know how these perform and it is well-enough documented.

The whole thing seems like AI slop.

Highly quantized models, especially with highly quantized KV caches, will, effectively, attend to the wrong tokens and be unable to easily discern highly similar tokens. The bastardized way of explaining this is gradient descent techniques get stuck in localized minimum and global maximums, so what happens when you turn the slopes into hard stair steps?

We need to move to smaller models and smaller caches and better samplers, not new quant methods (although I'm willing to also take those too).