Hacker News: npodbielski

New comment by npodbielski in "Framework Desktop with 192GB RAM"

npodbielski — Wed, 30 Sep 2026 17:44:52 +0000

Yeah, I did the same. Now I am happily running Flash Next.

New comment by npodbielski in "Framework Desktop with 192GB RAM"

npodbielski — Wed, 30 Sep 2026 16:07:37 +0000

For that price? It is better to buy DGX Spark or Strix Halo 50% more RAM is not worth about 5% more performance for 3 times the price.

New comment by npodbielski in "Transformers Explained Visually"

npodbielski — Tue, 22 Sep 2026 07:21:01 +0000

I have no idea why but I was thinking about Optimus Prime and Megatron when I was clicking the link. I was a bit disappointed.

That is good too though.

New comment by npodbielski in "Shapelearn Qwen 3.8 27B (13.1 GB VRAM)"

npodbielski — Fri, 18 Sep 2026 14:39:05 +0000

When I changed the number of draft tokens to 3 in both, it helped and they Draft is actually performing a bit better:

- draft: 67.17

- MTP: 64.18

Why they used those examples? Seems strange.

New comment by npodbielski in "Shapelearn Qwen 3.8 27B (13.1 GB VRAM)"

npodbielski — Fri, 18 Sep 2026 14:14:52 +0000

Which was not he point because I was testing their solution for MPT and it was just funny addition. But of course in internet you always will find some 'well akchually' person straight from the meme.

New comment by npodbielski in "Shapelearn Qwen 3.8 27B (13.1 GB VRAM)"

npodbielski — Fri, 18 Sep 2026 13:30:10 +0000

Well I tested it on 7900XTX with the same prompts and their draft model gave me about 30t/s. Their own snippet of code with regular MTP model gave me 60t/s.

Also model with their draft answered incorrectly. With MTP it answered correctly.

Question was: "Does MikroTik CRS312-4C+8XG-RM have combo ports?". The answer is Yes.

New comment by npodbielski in "Qwen 3.8 Omni Flash"

npodbielski — Fri, 18 Sep 2026 12:51:51 +0000

I am using Flash Next for few weeks and it is very capable model. I just wish there would a way to have faster prefill because reloading longer sections of session sometimes can take even 2h. I stopped using Qwen 3.8 27B completely on my Strix Halo.

New comment by npodbielski in "Qwen 3.8 Omni Flash"

npodbielski — Fri, 18 Sep 2026 12:39:56 +0000

Yes, for fun I tried OVH AI Endpoint and they do not have cache read at all. They bill you every time you send a prompt regardless if you are hit cache or not. One agent session was like 80M input and 300K output and I paid 30$ for that. Or rather I interrupted it and let my local Qwen finish it because cost was getting radicoulous.

New comment by npodbielski in "Registration without a phone number on Signal will use zero-knowledge proofs"

npodbielski — Mon, 14 Sep 2026 13:17:05 +0000

what would be the usecase? sharing account with kids? it is easier to just install something else and register on throwaway email?

New comment by npodbielski in "DeepSeek v4.1 Flash"

npodbielski — Thu, 10 Sep 2026 08:52:36 +0000

Or two gorgon halos?

New comment by npodbielski in "TALA Is Open-Source"

npodbielski — Tue, 08 Sep 2026 12:37:15 +0000

Looks like really great tool to generate some graphs and diagrams for static file blogs.

New comment by npodbielski in "What We Tell AI"

npodbielski — Sun, 30 Aug 2026 18:12:36 +0000

Wow. I just read couple, but... That seems terrible! People do that?

On the other hand my own father believes now he is an alien from outer space because someone generated stupid youtube video...

New comment by npodbielski in "Anthropic's best AI model struggles to attract users as cheaper tools thrive"

npodbielski — Tue, 25 Aug 2026 17:52:22 +0000

I am running it on 32GB and I did not saw model loosing it context even after 4-5 compactions in pi. I am running sessions for few days sometimes. I think it looped once, but loop police extension stopped it. The only problem I have know is how pi compaction works, which is forcing full prefill which takes time and it is erroring a lot. I wrote my own compaction that should remove full prefil but it does not work. But this is the only problem with this setup and it is more problem with pi then the model. I much more prefer it to use Qwen then paid models: Claude forces me to do reauth every other day and codex models either are too costly or not capable enough.

New comment by npodbielski in "Migrating a Synology NAS to a UniFi UNAS Pro 8 with Robocopy, SMB Multichannel"

npodbielski — Mon, 24 Aug 2026 18:52:34 +0000

I bought several cameras and they unvr early this year. It works great but I have no AI features because you have to buy they hardware just to accept some license (!). I mean... OK but no.

If they would allow me to send them email saying: "I hereby promise I wont record my neighbours having little tete-a-tete."

Sure I understand liability. But requiring consumer to buy another product to fully utilize they other product? Sorry but no. They have really cool hardware and nice UI but policy like that... no.

New comment by npodbielski in "Everything I own, owned"

npodbielski — Mon, 24 Aug 2026 05:03:09 +0000

What about Raspberry Pi is not open?

New comment by npodbielski in "DiffusionGemma Technical Report"

npodbielski — Fri, 21 Aug 2026 08:52:43 +0000

As you said: everything works on llama.cpp Why it does not work on vllm? Of course you can say that it is AMD fault but there was an issue of abysmal performance of models on Strix Halo, that is open for half a year (https://github.com/vllm-project/vllm/issues/34579#issuecomme...) and nothing is happening there. They do not care about those use cases. Seems like they are going with bit players that will run vllm inside datacenters racks. Hobbyists does not matter.

New comment by npodbielski in "DiffusionGemma Technical Report"

npodbielski — Thu, 20 Aug 2026 20:35:16 +0000

And it fails on rocm of course. This engine is such a hassle on AMD.

New comment by npodbielski in "DiffusionGemma Technical Report"

npodbielski — Thu, 20 Aug 2026 19:00:53 +0000

Anybody was able to run this model in a server?

New comment by npodbielski in "Composable Tests"

npodbielski — Tue, 18 Aug 2026 19:45:37 +0000

If it is not hard, with an experience of this guy, he could came up with a better example? It is your own words so I am sure you will agree? He did alright job so he could exercise gray matter a bit and came up with some API for saving customer data and then test it his way? I mean... This would be more interesting and "real world" so I am sure it would be worth to write longer blog post about it!

New comment by npodbielski in "NeoBrowser: An MCP server that drives real Chrome with your logged-in sessions"

npodbielski — Tue, 18 Aug 2026 17:41:26 +0000

It spits out password in logs according to gif. No thank you.