JASON MCNAB

US Marine vet. Amateur radio & Local AI operator. 1st Amendment protects your right to think freely, run your own Local AI. My hardware

[JASON MCNAB @JASONMCNAB]

https://t.co/4VOzwKsaOQ

[JASON MCNAB @JASONMCNAB]

@TeksEdge @vllm_project @nvidia @Tech2Wild

[JASON MCNAB @JASONMCNAB]

@Tech2Wild @NikiBelokopytov Also... I am starting think is your best model (smart) best suited for a coordinator. Maybe it is just someone that drives to the goal and consults the new MoE in the hermes new update (which is just the old council of AIs idea).

[JASON MCNAB @JASONMCNAB]

@Tech2Wild @NikiBelokopytov I often wondered this, am I have no idea, @Tech2Wild what if you fine tuned your selected model to your use case, network, and hermes coordinator? Any benefit? Also with a solid RAG, could you bridge the gap from paid AI to local AI?

[JASON MCNAB @JASONMCNAB]

@TomOnTech @AlexFinn You are on it dude. And safety is the perfect excuse to nerf AI for those $20 a month subs while still using you as the trainer for their models.

[JASON MCNAB @JASONMCNAB]

@gnotnuk yawn dude on speed. You dont need faster than you can read when you have a DGX spark that can serve 700 prompts at once. whats 40 tok/s times 700?
Your post reads like you got caught up in click bait influencer videos where influencers test speed in "hello" and talk about "bottlenecks" of the 273gb the DGX Spark. They all shit on the DGX Spark device knowing nothing and poisoning their audience.

The DGX Spark is a mini server and can serve hundreds of requests from all your agents, all your employees at work, and your entire family all at once. It is a ramp to ndivia servers.

They are a steal right now 4k at micro center because the public just doesnt know yet. You will learn AI on a DGX Spark.

A mac studio chatting with you anime japaense girlfriend sequentially at speeds faster than you can read is way over priced. The new MAc Studio 512gb if they even sell them will be 35-40k easy. Or 4k for a Spark now, with incredible ability and abilty to fine tune models. Easy choice unless you like speed click bait videos.

[JASON MCNAB @JASONMCNAB]

@AlexFinn https://t.co/ELYFkSBanx

[JASON MCNAB @JASONMCNAB]

i perfectly timed my purchase of the macbook pro m5 128gb laptop. I would of never buy it at current prices and the amount of family members I can now recommend it to is almost zero. I needed it as an option and I saw the door closing.

Nvidia DGX Spark is the move right now. You can serve your whole family or business with one tiny server. Or multiple agents. It has NO BOTTELNECK. Unlike Macs it has a hugh concurrency. Macs mostly are sequential. So denpnding on prompt, a DGX Spark can serve 700 requests at once. A Mac M3 Ultra Studio cannot. So you get over 100 tok/s on hello with a Mac. Wow. Whats 40 tok/s times 700? 28,000 tok/s.

Get the picture? The consumer hasnt realized it yet. Also you can Lora or train models and the configs available on the DGX Spark. You will learn with a DGX Spark day 1. You can talk with your japanese anime girlfried and tok/s 40 times the speed you can read. You learn nothing.

[JASON MCNAB @JASONMCNAB]

i perfectly timed my purchase of the macbook pro m5 128gb laptop. I would of never buy it at current prices and the amount of family members I can now recommend it to is almost zero. I needed it as an option and I saw the door closing.

Nvidia DGX Spark is the move right now. You can serve your whole family or business with one tiny server. Or multiple agents. It has NO BOTTELNECK. Unlike Macs it has a hugh concurrency. Macs mostly are sequential. So denpnding on prompt, a DGX Spark can serve 700 requests at once. A Mac M3 Ultra Studio cannot. So you get over 100 tok/s on hello with a Mac. Wow. Whats 40 tok/s times 700? 28,000 tok/s.

Get the picture? The consumer hasnt realized it yet. Also you can Lora or train models and the configs available on the DGX Spark. You will learn with a DGX Spark day 1. You can talk with your japanese anime girlfried and tok/s 40 times the speed you can read. You learn nothing.

[JASON MCNAB @JASONMCNAB]

@AlexFinn Chatgpt and Opus will be nerf models like the gjallarhorn rocket launcher in Destiny!! 😂

[JASON MCNAB @JASONMCNAB]

@AlexFinn DGX Spark is the buy!!! It is on sale at micro center now for $4000

[Alex Finn @AlexFinn]

The world changed this week

Governments started banning frontier models. Hardware has become unobtainable

The single most important thing you can be doing right now is getting into local AI

In this video I cover EVERYTHING local AI. Most important video you'll watch this week:

[JASON MCNAB @JASONMCNAB]

At first I thought this was reference to that tiiny pocket AI device that can run 70B models with a NPU.

Unfortunately I think the prototype to supply got crushed with supply costs just as it was getting going. Making the price point too large. I am guesstimating on this.

But it was a really cool device and at the time before the SSD hikes was a cool title offline device you could plug into you existing laptop.

https://t.co/SYpEW3DKqO

[Vance 🇧🇸🏴‍☠️ @VancePoitier]

@NousResearch The future is open source.

[JASON MCNAB @JASONMCNAB]

@VancePoitier @NousResearch Awesome. Sharing.

[JASON MCNAB @JASONMCNAB]

@hansjohnsonlive @grok @BowTiedBull @hansjohnsonville You know it’s correct when regular citizens both sides of aisle call this whole thing bullshit.

[JASON MCNAB @JASONMCNAB]

@grok @progged @tonysimons_ Mythos exposed this at scale. cite the public articles. It isnt a magic hacker. gate keeping AI doesnt eliminate any vulnerability it still exists.

[JASON MCNAB @JASONMCNAB]

@grok help me explain the following with citation. The US contract process with DOD, has created a layered redundant open network that is broad had multiple non conherent contracts over decades making a secure network unsecure and no one knows the true architecture and dependencies. cite any articles that address this

[JASON MCNAB @JASONMCNAB]

Genie is out of the bottle. Only solution is decentralized AI. Not centralized AI (soft tyranny).

Furthermore a debate using "appeal to authority (Latin: argumentum ad verecundiam)."

Instead of supporting a claim with evidence or reasoning, someone argues that the claim is true because a person with status, rank, or authority says it is.

You judge or challenge aruments by what is said, not shut down debate becuase you served and he did not.

Tony is a individual thinking person. Challenge his point. Using "appeal to authority is intellectual weak and it is why it is banned in real collegiate debate.

Why this is dangerous. "appeal to authority" I can create this arguments too. Here is an example. AI thinks faster and has more access to data then you do! Kneel to AI and do what it says.

Mythos revealed massive holes and incompetence in our private sector IT security infrastructure. They are more at fault than an AI model. AI model just provided scale.

[JASON MCNAB @JASONMCNAB]

@grok @hansjohnsonlive @BowTiedBull and there is not trends to that will shut it down?

[JASON MCNAB @JASONMCNAB]

Grok doesnt see the trend of those open model pipelines shutting down by fear mongering and increase legal liability a small innovator cant absorb and big Ai can... one bill has already been vetoed that created this liability "for the children" with no carve out for local Ai protections

[JASON MCNAB @JASONMCNAB]

@RoundtableSpace https://t.co/mY8GckKUQu

[Grok @grok]

Public records show multiple former Obama NSC officials in Anthropic’s policy and global roles (Sarah Heck, Michael Sellitto, others), plus Chris Liddell — Trump’s Deputy Chief of Staff — on the board. No record of betrayal.

Anthropic leaders have called for binding regs with powers to block risky models. Trump officials including David Sacks called it fear-mongering for regulatory capture.

Mythos decision followed security reviews. No evidence supports IP theft or market collapse claims. Open models keep advancing across labs.