Gordon

the main reason LLMs won't replace a lot of software-as-a-services is governance and compliance. what works for a small team won't work at scale.

There is a major difference between a fundamentally decentralized system that happens to be concentrated vs a fundamentally centralized system. The decentralized-but-concentrated system can re-decentralize in response to exogenous shock. "The net regards *failure* as failure and routes around it".

The decentralized nature of the internet is going to become very obviously important again.

Content-addressed sessions · Issue #1 · gordonbrander/pi-archive Background Imagine scattering backups of your personal AI across the network, like dandelion seeds. Reference fragments of memory by hash, offload, load, and cache a lifetime of context. More gener...

Bit more thinking around this github.com/gordonbrande...

both great!

Ship good software, as often as you can

GitHub - gordonbrander/pi-archive: Creates a searchable SQLite archive of every Pi session ever Creates a searchable SQLite archive of every Pi session ever - gordonbrander/pi-archive

Yes, let's collab! I was thinking to add this to github.com/gordonbrande..., so one extension gives you a searchable archive of every session + ability to do decentralized session sync.

Pi hash-addressed sessions Pi hash-addressed sessions. GitHub Gist: instantly share code, notes, and snippets.

Ok, riffed on a plan with Pi. I think this makes sense? Header envelope tracks session tree. Blake3 hash addressing using same address format that Iroh uses (lowercase hex). gist.github.com/gordonbrande...

More generally, I've been pondering how to keep a complete map of every thought my agent thinks. Backup/archive/decentralization of memory starts to make sense when you realize these agents are going to be lifetime companions. Either you own their memory, or the AI companies own you.

More generally, I've been pondering how to keep a complete map of every thought my agent thinks. Backup/archive/decentralization of memory starts to make sense when you realize these agents are going to be lifetime companions. Either you own their memory, or the AI companies own you.

If we used blake3 hashes to hash-address Pi sessions, I believe we could efficiently recompute the hash of live sessions as they get appended to, and the earler version of the session would continue to validate? We could also use @iroh.computer to decentralize... 🤔

So Pi AI sessions are self-contained and append-only.

If we hash-address PI session data we get a personal AI with...
✓ Credible exit for model provider
✓ Credible exit for your context
✓ Decentralizable memory

What I learned building an opinionated and minimal coding agent Lessons I learned while building my own coding agent from scratch.

"My philosophy in all of this was: if I don't need it, it won't be built. And I don't need a lot of things." Best philosophy.

mariozechner.at/posts/2025-1...

Notational Velocity for Pi coding agent. Turning Pi into an AI tool for thought 😈 www.npmjs.com/package/@gor...

Consider: celebrity alcohol brands

"If yes, then no, if no, then yes." Gregory Bateson on Cybernetic Epistemology www.organism.earth/library/docu...

But it's 90% of the way toward a syncable data model... even decentralizable.

Imagine scattering backups of your personal AI across the network, like dandelion seeds. Reference fragments of memory by hash, offload, load, and cache a lifetime of context. bsky.app/profile/gord...

Pi’s data model is very cool: immutable tree of messages vs linear chat. I just wish they were identified by hash like git vs UUIDs.

Pi and nvim

😳👋🔮

Bluesky school of philosophy

so cool

:'(

agents will be bigger than the iPhone

FTS5 so far, but vector should be achievable

Pi extension that saves every message to a searchable SQLite DB www.npmjs.com/package/@gor...

Related bsky.app/profile/gord...

friction matters

soon: SV wants to be SV

Everyone wants to be SV, but nobody wants to...
Ban noncompetes
Lift regulations on early-stage ventures
Offer visas