SpookyActionAtADistance @zzyber3

IT Pioneer: 35 yrs dev/mgmt, Built IP telephony w/ Cisco, Ericsson; mobile portability for Telia. Trading Asia. BBS owner on Fidonet. Now relaxing in paradise. Thailand Joined November 2022

Tweets

167
Followers

22
Following

114
Likes

91

SpookyActionAtADistance @zzyber3

a day ago

@JhnyStraka @MiaAI_lab From TP3 GLM im on 2x2 now Deepseek and GLM. Not happy with this combo, i see no ds advantages so im back to TP3 GLM tomorrow, running experiments on my 4th Spark

1 0 1 42 0

View Details

SpookyActionAtADistance @zzyber3

2 weeks ago

@MiaAI_lab Im one cable short, dud you ever try line?

1 0 1 30 0

View Details

SpookyActionAtADistance @zzyber3

a month ago

@elonmusk It just cant get worse than 4,6 but be careful in production.

0 0 2 335 0

View Details

SpookyActionAtADistance @zzyber3

a month ago

I have 4 dgx sparks running several project simultaneously and this is driving me absolutely crazy. Bots, profiles, agents, kanban!!!! Came across this git a month ago must-mohsin1 and it was amazing but touching it ruin everything. I rebuilt it now for my need but im still confused

0 0 1 122 0

View Details

SpookyActionAtADistance @zzyber3

a month ago

@matthias_mrc $1300 and i have 4 dgx sparks.

0 0 0 34 0

View Details

SpookyActionAtADistance @zzyber3

2 months ago

I appreciate you work guys but I have no time to play around even if I wish I had. Im running 4 DGX sparks, my main cluster tp2 im shifting between ds and glm and can't get it to work. I get your benchmarks sure but for work, it's terrible. Head up the latest, what recipe is most stable for kanban work for example? @MiaAI_lab @Tech2Wild

0 0 0 19 0

View Details

SpookyActionAtADistance @zzyber3

2 months ago

Really trying to be productive over here: kanban, profiles, bots — tried it all. Running DS on tp=2 with 4 DGX nodes going and… I just keep failing. Bots seem to be some kind of profile icon? I don't even know at this point. Can someone point me in a direction? Right now I've settled on running several profiles, giving each one its own job. I can't work with a single profile — I have at least 3 jobs running at once and the context fills up fast. Settling on ~130k context for now, even though there's a 1M-token window available, because Hermes seems stuck on low context.

0 0 0 19 0

View Details

SpookyActionAtADistance @zzyber3

2 months ago

@MiaAI_lab DGX in a 3 cable 200gbit line. What would you do with this? Running of cause Hermes on my Mac mini, deepseek flash on TP=2. I never managed to run GLM in a 3dgx ring. To get quality i have API director managing agent and kanban board. Its amazing but i think it can do better.

0 0 0 111 0

View Details

SpookyActionAtADistance @zzyber3

2 months ago

@IBuzovskyi Just seem to be a toy! You have profiles, kanban and why bots?

0 0 0 17 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@MiaAI_lab Waiting for my 4th DGX Spark today 🎉 Been running on API and Claude MAX for a while — and something felt off. Models I knew well were acting weird. Inconsistent. Like they'd been watered down. So I hooked OpenRouter up to my Hermes agent and built a skill to catch it. Turns out: OpenRouter does dynamic quantization. Same model ID — deepseek/deepseek-v4-flash — served as fp4 by one provider, fp8 by another. No way to know which you got unless you dig. No response header tells you. Zero. Checked the docs myself. The X-OpenRouter-Metadata header everyone talks about? Does NOT include quantization. You CAN force it (send provider: { quantizations: ["bf16"] } for full precision) and you CAN verify it (cross-reference the endpoints API). Made a skill so anyone can check. This is why I'm waiting for Spark nr4 to arrive today. What's your rig? Did you try the TP=4 ring?

0 0 0 225 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@MiaAI_lab Can it show tp=2 + tp=2?

0 0 0 12 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

Om du vet vad man cat gör eller tail —help så ska du köpa en eller flera spark dgx omedelbart, en macbook eller mac mini. Att bara förstå ett enda kommando back in the days tog dagar. Nu kulram vs miniräknare. Bygg vad du vill, sky is the limit!

0 0 0 32 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@PeterDiamandis Clean water

0 0 0 5 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@NousResearch please add a proper kanban board to hermes desktop.

2 0 1 61 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@Austin_Federa @AnthropicAI Agree 100% its even dangerous to use

0 0 0 43 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@BosonJoe @NVIDIAAI Yes, cant put the finger on whats wrong cause swe bench show good performance but for my it ruined my SaaS backend and inserted shit code on frontend not even following css style. I may try minimax m3, any experience?

0 0 0 81 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@0xSero @NeoAIForecast Tomorrow i get my second spark and will have the same setup. Did you cluster them? Downloded Ds4flash also but second is Qwen-NVFP4. Its hard to find good cases in this setup. Are you running VLLM?

0 0 1 42 0

View Details

SpookyActionAtADistance @zzyber3

3 months ago

@MiaAI_lab @gokayai @NVIDIAAI Im building a SaaS and use plans like claude max x2, supergrok and my rig is a mac mini m4 with hermes. Thinking ROI for a large model like FABLE 5 is just impossibe for a user like me so i suspekt i get a quantized after heavy use and very small context. Even FABLE got stupid.