peri-cl

6 ms·

1.
by peri-cl 5d ago
This article only obliquely mentions in the last paragraphs, but there's a parallel Chinese program that recently brought back samples from the moon, and will attempt Mars sample return, launching in 2028, https://en.wikiped
2.
by peri-cl 5d ago
https://archive.is/YYx0d
3.
by peri-cl 5d ago
I think this the same system that was the topic of a major contractor dispute a few months ago, between Verizon (incumbent fiber contractor) and SpaceX (lobbying to replace it). https://news.ycombinator.com/item?id=43205435
4.
by peri-cl 5d ago
I think M1 through M3 were compute bottlenecked in prompt processing (hence the very large gap between M3 and M5, in this page's benchmarks, that's not explained by memory bandwidth alone). For generation speed in isolation, yes
5.
by peri-cl 5d ago
You actually don't need more RAM to batch multiple inference tasks of the same model. (Each task needs its own context, but the (e.g.) 27B of constant parameters isn't duplicated).
6.
by peri-cl 5d ago
Surprisingly, the Reddit crowd are reporting 50–60 tokens/s (for the 32 GiB 5090 + 128 GiB RAM)—on par with the M5 Ultra benchmarks, despite both the PCIe bottleneck and much smaller DDR5 bandwidth, https://old.reddit.com&#x
7.
by peri-cl 5d ago
If you look into it more carefully, you'll find this doesn't apply to social media moderation, but only more consequential AI-biased decisions. (In the excerpt you quoted: "legal effects" / or "significantly
8.
by peri-cl 5d ago
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two hea
9.
by peri-cl 5d ago
Those are some incredible graphs, that leap in prompt processing going from M3 to M5. Also: ~30 token/s on GLM 5.3-flash, locally. (That's roughly Opus 4.8-tier. I think). /meta Here's a CSS filter that stops those nuisa
10.
by peri-cl 5d ago
> "Wouldn't that be covered by fair use?" There's no human to appeal to in social media platforms. It's automated, algorithmic enforcement; and when it's wrong (or someone exploits it to be wrong, like thi
11.
by peri-cl 5d ago
Police unions would love for cameras to standardize on that. There was an article on HN that one city's police force, when there were protests going on, took to playing Disney songs from loudspeakers; so that any recording of their int
12.
by peri-cl 5d ago
This is what the runtime reports, at Q8: total params memory size = 15645.19MB (VRAM 15645.19MB, RAM 0.00MB): text_encoders 7669.77MB(VRAM), diffusion_model 7331.05MB(VRAM), vae 644.38MB(VRAM), controlnet 0.00MB(N&#x2F
13.
by peri-cl 6d ago
It's about 16 GiB at Q8 quants (combining both the image and language parts). (Meaning, community quantized models from HuggingFace). I think it will technically run on anything that has enough memory. I just tried it on a standard lap
14.
by peri-cl 6d ago
I've set it up on my local machine just now, as my first local image diffuser. I can confirm it's very easy. I tried stable-diffusion.cpp, following its compile guide here[0], and its Qwen Image-2.1 specific instructions here[1].
15.
by peri-cl 6d ago
https://archive.is/a4uct
16.
by peri-cl 6d ago
https://archive.is/2YdN9
17.
by peri-cl 7d ago
https://archive.is/Iw0JJ
18.
by peri-cl 7d ago
These objects (black hole stars) seem to be far larger than that, https://en.wikipedia.org/wiki/Quasi-star#Formation_and_prope... (Caption: "Size comparison of a hypothetical quasi-star to some of the largest kno
19.
by peri-cl 7d ago
There's dedicated Wikipedia articles responsive to this, https://en.wikipedia.org/wiki/Solar_core
20.
by peri-cl 7d ago
> "reviews" Do you really think we can sign off on a 100 billion-element analog circuit gifted to us by a malicious adversary? We can't even keep our own CPU's reliably free of security exploits (Spectre/Meltdo
21.
by peri-cl 7d ago
Is there a way to formalize the space of all possible aesthetics a model knows, in a way that you could randomly sample from it? The analogy in my mind when I read this post was: you are asking a machine to paint a color, but the machine al
22.
by peri-cl 7d ago
> "Ho also confirmed that the team had access to internal LLMs fine-tuned for chip design that are not available to the public. He declined to detail the models used." I'm imagining a Ken Thompson "Reflections on tr
23.
by peri-cl 8d ago
> "against data that is almost certainly private and possibly illegal to exfiltrate" I didn't fully understand the article, but I gathered this only impacts project directories managed by Z.ai's coding agent? I.e.,
24.
by peri-cl 9d ago
It's only the top open-weights LLM in the world, https://artificialanalysis.ai/#intelligence-category-tabs
25.
by peri-cl 9d ago
I'm relieved to hear a voice of simple decency! Really, the most questionable thing here is that someone wrote a 2,000+ word essay about someone else's sex life.
26.
by peri-cl 9d ago
> "A business generally wants to keep their customers happy and care for their reputation, so telling a business when we’re unhappy and influencing their reputation is the feedback mechanism we have at our disposal." Yes, exc
27.
by peri-cl 10d ago
Yes, that's one of the pages I wrote my comment about. I did try to interpret it charitably, but there's nothing on that page (everyone read it! It's short!) that simply communicates to users, "we're uploading your
28.
by peri-cl 10d ago
Well, it depends on the task, doesn't it? "running shoes I looked at last week" / "Here's what I found in your browsing history:" doesn't need a 119 billion parameter frontier model; it's a RAG p
29.
by peri-cl 10d ago
It's more un-fathomable if you recall Mozilla were the people who *pioneered* local browser ML as a privacy feature. Back in 2022—when Chrome was doing language translation by uploading to Google servers, and Mozilla introduced the (ac
30.
by peri-cl 10d ago
This is an excellent use case for completely local, small model inference, yet for inexplicable reasons Mozilla wants to normalize uploading your entire private browsing history to a cloud. These (Mistral's and Mozilla's) marketin

More ›