TL;DR — Key Takeaways
- Intersignal’s Braid protocol is designed to let locally hosted AI models communicate and coordinate across a local network without relying on centralized cloud AI services.
- Braid uses digitally signed messages and state information to help verify that data exchanged between AI nodes has not been altered in transit.
- The protocol is aimed at users running open-weight models on Windows, macOS and Linux systems, including home users and small labs.
Longtime viral marketing guru and one of the most prominent social media voices who pushed the false Pizzagate controversy, David Seaman, is back with a new cause to champion: local AI.
Seaman has founded a company, the Fort Lauderdale, Florida-based Intersignal, which has developed a protocol, called Braid, that provides a way for open source LLMs to communicate with one another on your home area network.
In an X video posted last weekend, Seaman explained the importance of local AI—the ability to run AI models on your home computers—viewing it as a safeguard against the creeping monopolism of frontier AI labs.
“Not that there’s anything wrong with those services, but they want to keep you on the for-profit cloud. They don’t want you spinning up your own local AIs in your house or dorm, or at your work,” he said.
Braid, according to the press materials, provides a conduit for sharing state information across two LLMs, digitally signing the material to ensure it doesn’t get corrupted or tampered with en route.
The protocol was designed to run on Windows, macOS, and Linux machines connected on a local area network and running open-weight models.
Seaman’s idea is that the home user, or small lab, can have multiple LLMs run a single project, coordinating agent work without centralization.
This allows the machines to “share context through symbolic cache alignment, and maintain memory-aware reasoning structures with traceable identities,” according to the company. The receiver only ingests a .brad object after verifying that the object’s Ed25519 signature, signed by the sender, is legitimate.
A Pivot to AI
A longtime media pundit and writer who got his start at Jezebel news site, Seaman has written for the Huffington Post and appeared on the Joe Rogan Experience multiple times to discuss government surveillance, bitcoin, and the like. He’s also authored numerous self-published books on these topics.
But he was most well-known for his role perpetuating Pizzagate, a 2016 conspiracy alleging the existence of a fictitious DC pedophile ring. He subsequently downplayed his involvement, disavowing the allegations.
These days, he is concentrating his energy on championing local AI.
Seaman’s Intersignal has been issuing press releases since at least November, recently announcing the release of version 1.5.2 of its Braid protocol.
Different models can work collectively as nodes in a Braid network. Braid “nodes discover verified peers over a local network, exchange signed messages, and synchronize compact state vectors that represent system activity,” noted App Developer Magazine.
The machines aren’t just sharing data but rather “semantic states,” or the LLM’s numerical and geometric representation of some concept.
As of press time, Intersignal hasn’t shared any academic work proving the value or validity of sharing states across LLMs, and, to date, only limited work appears to have been done elsewhere on the topic.
NVIDIA has investigated sharing KV caches across different models as a way to skip prefilling. Researchers from Alibaba and Zhejiang University investigated how agents can share learnings not by JSON-formatted human language but with the internal latent space, or mathematical coordinates within the LLM itself. One group of Chinese researchers have demonstrated that concepts can be transferred across different LLMs using simple linear transformations.
Braid in Action
In this case, Braid passes a 384-dimensional vector of 32-bit floats to another node, where it is placed in the receiving machine’s local memory or UI heat map (which provides a human-readable signature).
“A signed Braid object was divided into 30 QR frames, displayed on one Mac, captured by the camera of another, and successfully reassembled byte-for-byte,” the press statement asserted.
This work builds on an earlier effort to run a 1 trillion parameter LLM (Kimi K2 with Atlas 2 for optimization) on two Dell “Alienware” gaming machines, to demonstrate how LLMs can run in-house.
Intersignal claims that the code is available under an MIT License, though as of press time, it doesn’t offer an open source repository. For now, the code for Braid and Braid.Node can be found on Google Drive.
