krasin

8 ms·

31.
by krasin 2y ago
> Only on smaller models, their numbers are all 70b in the article. No, they are 5x-10x faster for all the model sizes (because it's all just running from SRAM and they have more of it than NVIDIA/AMD), even though they benchma
32.
by krasin 2y ago
> I'm still trying to wrap my head around how these companies think they are going to do well in this market without more memory. Cerebras and Groq provide the fastest (by an order of magnitude) inference. This is very useful for ce
33.
by krasin 2y ago
(2017). Since that time, the landscape of sub-$1 microcontrollers has significantly changed. Namely, 32-bit won, RISC-V is on the rise, more RAM, more Flash, more MHz for the same price: RP2040/RP2350, CH32V003/CH32V203, STM32G0,
34.
by krasin 2y ago
> Only the US government can allow corporations to beta test unproven technology on the public. China and Russia do it too. It's not an excuse, but definitely not just the US.
35.
by krasin 2y ago
https://old.reddit.com works, https://www.reddit.com is having troubles ("no healthy upstream" and similar messages)
36.
by krasin 2y ago
Distributed training data creating & curation is more useful and feasible. Training gets cheaper 1.5x every year, but data is just as expensive, if not more, given that the era of "free web crawls of human knowledge" is over.
37.
by krasin 2y ago
Thank you very much for posting this! The code is indeed there, that's great.
38.
by krasin 2y ago
There's no code or weights released => no way to reproduce their results.
39.
by krasin 2y ago
I tried it on a coding task; this model might shine on something else, but on coding it is not.
40.
by krasin 2y ago
Outside of the US, this is usually addressed by having a dedicated police unit residing at a station. Everything is pretty calm and civil after that. The fact that it's not the case in the US does not mean that it's inevitable to
41.
by krasin 2y ago
> Four small-scale Chinese cloud providers charge local tech groups roughly $6 an hour to use a server with eight Nvidia A100 processors in a base configuration, companies and customers told the Financial Times. Small cloud vendors in th
42.
by krasin 2y ago
Here is the feature request to support Pixtral on llama.cpp: https://github.com/ggerganov/llama.cpp/issues/9440 - once it's implemented in the upstream, ollama will be able to pick it up.
43.
by krasin 2y ago
I watched the video. It's hot garbage; Source: my family has two Framework laptops (gen 11), I've been in charge of repairing / upgrading them over the last two years. Our primary issue is Intel+Microsoft breaking the S3 deep
44.
by krasin 2y ago
> Are there genuine at scale use cases for AI outside of LLM's Assuming that "outside of LLMs" means "outside of text processing". Yes. Robotics. Imitation learning with LLMs is working surprisingly well. It wil
45.
by krasin 2y ago
It's very hard to respond to a comment like that, since there's no specifics, just a plain disagreement. On my side, I would like to point that the today HN thread ([1]) that discusses a paper GameNGen ([2]) that runs Doom with di
46.
by krasin 2y ago
> like everyone else but Apple, on Nvidia (InfiniBand) Google also does not depend on NVIDIA, thx to TPUs. Rents NVIDIA GPUs to external customers - sure, it's a nice side business, but internally TPUs are king and there's no d
47.
by krasin 2y ago
> Why would you spend any of your finite attention here? Because it is research. And it's an open research, unlike OpenAI or (for the most part) modern Google DeepMind or xAI. It's completely fair game for non-researchers to
48.
by krasin 2y ago
The problem would also be solved by building affordable housing under the new state initiatives, which prohibit the cities to block construction. New residents will very quickly dominate the voter base, problem solved.
49.
by krasin 2y ago
Good point. In addition to being available right now, the Aliexpress version has a better battery (10000 mAh vs 5000 mAh on Kickstarter).
50.
by krasin 2y ago
> Why is anybody troubled by the waste then? Greenpeace had a point back in 1970-ies, but the message got into a myth and subsequent advancements in technology have not been popularized enough to change the public perception.
51.
by krasin 2y ago
Yes, breeder reactors can feed on waste: https://en.wikipedia.org/wiki/Breeder_reactor China has the plan to achieve closed nuclear fuel cycle, and there are new breeder reactors being put online to move the country cl
52.
by krasin 2y ago
Everyone doing anything has some controversy around them. At some point, we should just stop paying attention to internet dramas and focus on what matters. Like, fast issue trackers or continuous build with a novel feature to ssh into the s
53.
by krasin 2y ago
It means that it's not slow. Case in point: Jira.
54.
by krasin 2y ago
SourceHut has an issue tracker (incomplete, but very fast) and CI. The problem is that SourceHut is not becoming any better over time. With all my respect to Drew, his attention is spread too thin over many projects.
55.
by krasin 2y ago
> we don't yet have 2 competing viable players in the private commercial space transport business We kind of do have them? SpaceX and Rocket Lab: https://en.wikipedia.org/wiki/Rocket_Lab
56.
by krasin 2y ago
Twisted Fields: https://github.com/Twisted-Fields (open-source precision farming rover)
57.
by krasin 2y ago
It uses a Dust3r checkpoint under the hood, see the installation instructions for InstantSplat that include downloading DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth: https://github.com/NVlabs/InstantSplat?tab=readme-ov-file#
58.
by krasin 2y ago
Here is the Nature paper that confirms the claim: https://www.nature.com/articles/nature.2015.16771 Fantastic!
59.
by krasin 2y ago
Brilliant Labs have 640x400 glasses for $350: https://brilliant.xyz/products/frame They even have an SDK: https://github.com/brilliantlabsAR Disclaimer: I didn't try them. No idea if they work.
60.
by krasin 2y ago
These glasses don't have a camera. At all. But your bigger point stands.

More ›