karmakaze

3 ms·

1.
by karmakaze 5d ago
They should support Mojo which might be a great fit for edge compute.
2.
by karmakaze 5d ago
I really appreciate seeing these dense model numbers. For a large unified memory system though I expect that MoE numbers are what people are more interested in. These numbers could and should get much better. As an example I can run Qwen3.8
3.
by karmakaze 7d ago
Basically the same as always--an extended interactive Turing test. Can they speak coherently and in-depth about the things they claim to have done on their CV/resume?
4.
by karmakaze 8d ago
Cluould be viewed like a fancy evoultion of CHICKEN (Cheney on the MTA) that used stack for everything.
5.
by karmakaze 10d ago
Mentally convert to downtime: 99.8 is clearly twice as bad as 99.9 Similar for LLM measures from an ideal 1.0 mark.
6.
by karmakaze 10d ago
The whole story is a joke--Microsoft has AI? Apparently they're cooking up MAI (Microsoft AI), and I'd seen their small Phi models listed online. Calling Anthropic a competitor is hilarious.
7.
by karmakaze 10d ago
Yup. Could be that everyone was rushing to get all their products and demos ready leading up to it.
8.
by karmakaze 11d ago
> i.e. whether frontier AI models aren’t generating or can’t generate new mathematical ideas. [...] I give basically zero credence to the idea that AIs are incapable of this because of some intrinsic feature of how LLMs work. I also beli
9.
by karmakaze 11d ago
Yes. Reminds me of "Poetry is the art of giving different names to the same thing" vs "Mathematics is the art of giving the same name to different things" -- Henri Poincaré. Set of spacial facts is what's left when
10.
by karmakaze 12d ago
It's part of the functional trio: map, filter, reduce--and half of MapReduce.
11.
by karmakaze 12d ago
Difference being anyone can be a "script kiddie". I don't think anyone could direct an AI to proofs like this one.
12.
by karmakaze 12d ago
I've also had this kind of topological knowingness which I couldn't name. Another comment here says "schematic conceptual" which captures what it feels like for me. Another good explanation here is that it's all t
13.
by karmakaze 12d ago
This is a bundle where you end up paying for parts that aren't as useful: - 252 GB HBM3e VRAM - 496 GB LPDDR5X RAM I would rather have a system using 2x Instinct MI350P GPUs (288GB total) for much less.
14.
by karmakaze 14d ago
Personally I'm using Qwen3.8-27B (MXFP4 quant W4A8) locally hosted on a pair of AMD GPUs (with DeepSeek Harness). It starts at 250 tokens/sec down to 120 past 128k context. At work mostly Opus 4.8 (sometimes a GPT or Gemini 3.1 Pr
15.
by karmakaze 16d ago
Busabase[0] > Languages: TypeScript 98.4%, Other 1.6% Not for me. [0] https://github.com/busabase/busabase
16.
by karmakaze 16d ago
Seems like a random rant to me. > What is the internet now to me? In many ways it’s something that I have to tolerate to do many things that functioned fine before. I have to use an app to pay for parking. I need to create an online acco
17.
by karmakaze 17d ago
I think it's largely due to psychology. If 210m is considered the best , then as they approach it they may start to tense up and choke. When the goal and possibility is known as 500m, then there's no point being concerned near 21
18.
by karmakaze 18d ago
Similar here peak ~250 and down to ~120 as it gets close to 128k (which is where I set DSH compaction) though it can readily do 256k. I just got DeepSeek Harness (DSH) set up with 2x R9700 and it's rather mind blowing that these can do
19.
by karmakaze 19d ago
The way it does tensor splitting without all-reduce cost over PCIe bus wasn't something I thought was possible. What kind of performance are you getting with 4x R9700s--what do you do with all the VRAM (batching, concurrent requests, e
20.
by karmakaze 19d ago
Thanks! Didn't expect to see this here. Exactly what I needed to run Qwen3.8-27B-Quark-AWQ-MXFP4-native.gguf as well as other experiments on one or 2x R9700's (I hope).
21.
by karmakaze 19d ago
> Anything design-y was “too Apple” and unacceptably bourgeois. That was definitely the case of the Unity desktop that really only worked well on netbooks. That's when I lost confidence in Ubuntu for design. Loss in Canonical on the
22.
by karmakaze 21d ago
The previous 'personal' AI Station I had pictured was the a16z one[0]. Each MI350P[1] in the TR Halo Station has 144GB VRAM and with 4.6 PFLOPs peak MXFP6 performance. Four liquid cooled? Yes please. [0] https://a16z.co
23.
by karmakaze 25d ago
NasaFTs are back in style!
24.
by karmakaze 25d ago
It seems we could use a new kind of memory that streams the weight data in, like GDDR in reverse.
25.
by karmakaze 25d ago
There's clearly a line running through the lake. It should only change name on the lower part.
26.
by karmakaze 26d ago
AirBnb can only be used if you are willing to accept the risk of having an unacceptable or even no reservation upon arrival. The latter did happen to me and fortunately I was able to make alternate plans--though still not recoup full paymen
27.
by karmakaze 1mo ago
Actually no. I listed all the top stories of the day and this one made the cut. Quality is a low bar these days on HN.
28.
by karmakaze 1mo ago
Why are they tying the client with the AI model instead of using OpenAI or another popular http format? Running llama.cpp locally is about the right level of complexity for most doing local AI.
29.
by karmakaze 1mo ago
I used Q8 kv cache and using 64K context but can go a bit higher. The DFlash2 model takes a few gigs and using Q6_K (rather than a Q6_K_M/Q6_K_XL that unsloth publishes) saves some more. Also using Vulkan that has less VRAM overhead, b
30.
by karmakaze 1mo ago
I did a similar thing running Q6_K model and Q8_0 DFlash2 (draft=7) quants: DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF incoai/Qwen3.8-27B-DFlash2-GGUF using llama.cpp PR/commit https:&#x2F

More ›