Count Stacula

Local AI Enthusiast

9x Asus Ascent GX10s Clustered with a Mikrotik 812 Switch
1x Nvidia DGX Spark
5x Mac Mini
2x RTX Pro 6000 WS
2x RTX Pro 6000 Max-Q

[Count Stacula @CountStaculaAI]

@fute_nukem @inherentlyso @gabriel1 You do realize you can run them locally, offline, exposing nothing, right?

[Count Stacula @CountStaculaAI]

@dummerspast39 I'll eat an RTX 6000 live on stream if this ends up being true. No shot, but I'd be very happy if so.

[Count Stacula @CountStaculaAI]

In the age of AI, open weight models and small/efficient models it is comical that Siri is still this bad. @Apple , you are a 5 trillion dollar company, figure it out already.

[Count Stacula @CountStaculaAI]

I will bet you any amount of money the following is all true 1 year from now:
1) My local AI hardware will have appreciated in value at least 50%, (probably 100%+)
2) The current $200 plan will have its usage allowance cut by at least 50% or more
3) Anthropic (and others) will continue training on your data/input
4) The local models I will continue using, will surpass all of the current benchmarks of Opus 5.5 (objectively the best frontier model currently), which is capable of handling the vast majority of coding and general tasks
5) I’ll be getting tokens for free still, minus upfront cost (that appreciated) and electricity, all 100% private

You rent, I buy. While your argument has some strength right now, as these huge labs stop subsidizing their subscriptions, the local AI hardware runners will be the ones laughing and there will be no return to the current norms of the rental market, they will continue to rug pull.

[Count Stacula @CountStaculaAI]

@SteveSTechMan @ciruai This. I won't even consider selling anything until the end of next year, which is when I believe resell value will top out, right before the new fabs come online and start to flood the market.

[Count Stacula @CountStaculaAI]

@loktar00 @JTBeers I do, if you send me the prompt I can run it to compare if you'd like.

[Count Stacula @CountStaculaAI]

Last weekend, after getting fed up with CodeX, I decided to finally try DeepSeek Harness. I hate change, so it took a lot for me to leave CodeX, but I'm glad I did.

DSH is just better all around. It's faster, each task (coding in my case) is handled more thoroughly, the connectivity for local AI is much more seamless, detailed inference statistics.

The only things that bother me, but are not a huge deal are:
1) No timestamp per output.
2) Minimal customization/settings.
3) No chat outside of a workspace.

But maybe less is more and that is all intentional.

Well done @deepseek_ai @DeepSeekHarness

Note: My CodeX and DeepSeek Harness experience are both based on a local hardware + open weight models, so my experience is specific to that.

[Count Stacula @CountStaculaAI]

@TrustyProtoMage @MiaAI_lab Cheapest I could find and what I bought and works perfectly is:
200g direct:
https://t.co/pl6TqnXnNh
400g out -> 2x 200g:
https://t.co/i434aM40J0

[Count Stacula @CountStaculaAI]

@net_termina This is incredible.

[Count Stacula @CountStaculaAI]

Finally got the last RTX PRO 6000 Workstation delivered today and have the 4 RTX 6000 rig setup. Time to fly.

[Count Stacula @CountStaculaAI]

@net_termina This is incredible. I would love to try this on my setup, if you have a link?

[Count Stacula @CountStaculaAI]

@Hikari_07_jp I'm running mine on that exact same mining rig, that was not fun to put together. Looks good though, congrats!

[Count Stacula @CountStaculaAI]

@loktar00 Wow this is incredible.

[Count Stacula @CountStaculaAI]

@stefanopineda What was the consensus here? I was considering doing this as well, but couldn't find any details on the improvement with this setup vs risers.

[Count Stacula @CountStaculaAI]

@SteveSTechMan Here was my total on the last 4 I purchased:
200G QSFP56 to QSFP56 Twinax Copper Passive DAC Cable 0.5~3 meter
-
Subtotal $109.48
Shipping & Handling $35.00
Import tariffs $15.89
Estimated Total $160.37

[Count Stacula @CountStaculaAI]

@jun_song This is not accurate. You can buy a Mac M3 Ultra 512GB for $16,000 on ebay right now. You can get RTX 6000s for $15,999 new from MicroCenter, even less on ebay. You can get sparks for $5,xxx on Ebay as well.

[Count Stacula @CountStaculaAI]

In the last 24 hours, the DGX Spark has sold out from many of the MicroCenter locations. My guess is there's 1 week max left until its sold out everywhere. If you're interested in buying one, I would get it while you can.

[Count Stacula @CountStaculaAI]

@LocalAiCherry @Hikari_07_jp WS's are going for a little more right now so used I'd say roughly:
$14,000 for a WS
$13,500 for a Max-Q

Last week I sold a new Max-Q I wasn't using for $14,250 but used vs new does not seem to make a huge difference in demand and price for these at the moment.

[Count Stacula @CountStaculaAI]

@LocalAiCherry @Hikari_07_jp New or used? Roughly $14,000 right now on ebay.