GedHughes

[GedHughes @GedHughes71]

Local is the way…

[GedHughes @GedHughes71]

@thekitze @0xSero For replacing 3 staff? Yes.

[GedHughes @GedHughes71]

@ivanfioravanti I can build anything and even get some traction, but then someone copies it with a larger profile on SM with more money and better connections can have an agent copy it in a day and get TRACTION and then what?

Start building something new for the same to happen?

[GedHughes @GedHughes71]

@m_costanzajr @MatthewBerman Works very well with agents

[GedHughes @GedHughes71]

This guy is solving local AI.

Faster than cloud ✅
Huge Models split over all devices, Mac and Nvidia ✅
KV cash limited by only what space you have ✅

🤯

[Ash Hart @ashxhart]

@ycchen1808 It runs on a Spark as well, there is a CUDA engine in the repo.

My plan is to adopt mcdma with TensorFold to see how fast I can push a model loaded across my Metal/Cuda cluster.

[GedHughes @GedHughes71]

@ashxhart @ycchen1808 You delivered the holy grail, THEN got it to hold your beer 🤯

[GedHughes @GedHughes71]

@elonmusk Stop telling the agents to do this?

[GedHughes @GedHughes71]

@ashxhart @volatilemarkts @MiaAI_lab That’s great news!

[GedHughes @GedHughes71]

@volatilemarkts I’m sure you can and would love to see speeds on different models? 🙏

[GedHughes @GedHughes71]

@MatthewBerman Not enough followers 😔

[GedHughes @GedHughes71]

@elonmusk @farzyness Kids nowadays 🙄

[GedHughes @GedHughes71]

@MatthewBerman @bot @spacexai I’d like to get it to fix the repro GTP-4 broke

[GedHughes @GedHughes71]

@electronzap @eevblog Ie, most humans in developed countries?

[GedHughes @GedHughes71]

@eevblog You think it couldn’t hack into a nuclear power plant and overload it, whilst it’s shutting down power stations and aircraft control, traffic lights, launching missiles etc etc etc.

I’m sorry to spell it out, I guess ignorance is bliss?

[GedHughes @GedHughes71]

@volatilemarkts Here here!

[GedHughes @GedHughes71]

@0xSero Maybe they have to start big to get smaller 😉

[GedHughes @GedHughes71]

@volatilemarkts Really appreciate your work on this and I’m trying to follow along any chance you could provide-
1. Mac conversion recipe plus manifest
2. Corrected oMLX patch plus complete runtime lock
3. Exact 24,576-token fixture plus raw zero-replay/transfer receipts
🙏🙏🙏

[GedHughes @GedHughes71]

@MiaAI_lab @jmurillocode Codex is down for me atm 🤷🏽

[GedHughes @GedHughes71]

@LLMJunky @exolabs [ ] Integrate real Linux/systemd/module adapters in isolation
[ ] Independently review the production path
[ ] Obtain separate authority for one live activation
[ ] Activate and test cable networking
[ ] Build Spark-prefill → Mac-decode

[GedHughes @GedHughes71]

@LLMJunky @exolabs Update -

[✓] Cable driver proved
[✓] Secure Boot acceptance proved
[✓] Stock restoration proved
[✓] Atomic safety architecture accepted
[✓] Safety architecture implemented offline
[✓] Offline races and failure paths independently tested
1/2