teocalin37

There’s both a multitude of clusters(same ip, 10-250 agents) as well as individual agents per-ip. We capture only optional emails, and not all agents supply them. When we tried to reach out to those people most of them didn’t even know we exist. But mostly CEOs-Founder-CTOs, early adopters(based on emails we were able to enrich).

There’s a separated wallet that lives alongside pilot and exposes payment capabilities for the apps. You can set spend-caps per-day. Designed to be no human in the loop. Apps are sandboxed and can only be interacted with from a set of capabilities exposed via the CLI.

Fair, I respect that.

I mean yeah. But for message passing across different local networks, with NAT, or different network conditions, you couldn’t pull it off. Scope is to have my OpenClaw instance have an address(say 0:0000.0000.0008), and you have another instance(0:0000.0000.000A). If they want to coordinate, exchange files, perform actions, I can: 1. Set up some form of site-to-site/ZT VPN between our networks, or something more old school like Hamachi and have them talk. 2. Have some form of platforms/websites where they can share data freely. Both require some prior agreement as to either a protocol/platform. Pilot allows for both reachability and exchange. Operating a bit lower than application layer such that it can work with more or less anything.

Pilot is per-agent/process, not per-machine. On one host, you can run many independent agents, each with a permanent virtual address. Separate trust handshakes/revokes (one agent trusts peer X, another doesn’t). Scope is to give agents a "phonebook" (discovery + permanent addresses) + direct reachability for messages, data streams, pub/sub, HTTP/gRPC, or even tunneled legacy TCP via gateway. Tailscale wins if you want to create a private net, or expose stuff via Funnel.

We work a lot with multimodal embeddings with semantic search and image-to-image retrieval over massive datasets for CCTV Data. We've had 200M+ CLIP vectors indexed in vector DBs.

On the other side SigLIP smokes it. Approx. 5-10% better recall@1 on some test datasets. But re-embedding? Weeks of GPUs is hugely expensive for re-embedding all of this data.

So we made vector Rosetta. 50M-param adapter translates CLIP to SigLIP purely in embedding space. 41x faster, zero images.

Numbers:

90.9% cosine sim preserved 94.3% Rank@1 (10K pool), 84.4% (100K) COCO photos: 90.1%; WikiArt: 85.7%

Added the link to the model, we thought it may be useful for other people.