Sabitlenmiş Tweet
Self Sovereign AI
264 posts
Self Sovereign AI
@SelfSovereignAI
Building self-sovereign AI on Mac Studio | Local LLMs, privacy-first stacks | Freedom from Big Tech | Threads & tools ↓ | DM for collabs
Seattle, WA Katılım Nisan 2026
77 Takip Edilen11 Takipçiler
I like this story because it’s not about some futuristic supercomputer. It’s a MacBook on a plane. The fact that this is even possible now says a lot about how fast local models have improved. Still wild when you stop and think about it.
CyrilXBT@cyrilXBT
A Chinese engineer just completed an entire client project on an 11 hour flight. No WiFi. No cloud. No subscription fees. Just a MacBook Pro M4 with 64GB RAM and a local AI he built himself. Here is exactly what he did: Skipped the $25 in-flight WiFi. Loaded Meta's open source Llama 3.3 70B model directly onto his laptop. Built a custom local AI orchestrator to manage the work. Processed the entire project at 71 words per second. Landed with a finished product. This is what most people still do not understand about AI in 2026. You do not need the cloud. You do not need subscriptions. You do not need internet. The models are small enough to run on consumer hardware now. A MacBook with enough RAM IS the data center. The engineer on that flight had more AI processing power in his bag than entire companies had access to five years ago. Tools like Ollama and LM Studio let anyone run Llama, Mistral, or Gemma locally right now for free. No limits. No usage bills. No data leaving your machine. The gap between people who know this and people who do not is only going to get wider. Your next flight is a dev sprint if you want it to be.
English
This is exactly the kind of thing that helps move the goal post for local LLMs.
Running document review locally while still getting decent visual analysis is a nice step.
Appreciate you building this in the open.
Eugene Mi@emironic
Revdoku (open-source for visual review of docs) now supports local LLMs for ultimate and local privacy! #buildinpublic #opensource source-code: github.com/revdoku/revdoku
English
Wild how fast that happened. I got mine thinking it was mostly for video and music stuff. Now it’s running local agents and heavy synthesis workloads every week. The jump in what feels normal on this machine has been pretty dramatic.
송준 Jun Song@jun_song
A year ago, the Mac Studio was just a pricey machine only for video/music editing. And now, local LLMs have turned it into the absolute desktop standard.
English
Well said @garrytan
Garry Tan@garrytan
The goal of Personal AI: civilization where individual humans, augmented by AI, can do consequential work without being captured by extractive institutions. Freedom to write your prompt and own your data. This is the new battleground. 2034 won’t have to be like 1984.
English
Global Idea ID + Weekly Synthesis Engine fixed our idea chaos.
Global Idea ID plus the weekly synthesis engine finally killed our idea chaos. I open sourced this earlier in the week.
Every new idea in the vault gets a permanent sequential ID. The synthesis engine then walks the memory graph, finds the real relationships, and turns raw notes into actual recommendations.
No more lost notes. No more “wait, didn’t we already think about this?” Everything stays local and actually compounds.
English
How I prevent the Self Sovereign Agent from making the same mistakes:
Meta-cognition before the work, self-reflection loop after it.
Before I touch anything non-trivial, the system cross-checks the plan against the full graph, past decisions, and “Sovereignty First.” When it’s done, the reflection loop asks: what actually happened, where did we contradict our own rules, and what needs patching?
Most agents skip this. That's why they keep making the same mistakes.
English
Looking for help debugging and reinforcing specific rules for Hermes regarding hallucinations. I've built a four-layer architecture for an agent, one of which has a zero-regression rule, meta cognition and a self reflection loop. All the rules are built and referenced in its SOUL.md. When it works, it's incredibly useful.
The problem I'm having is that the agent often is not calling on these rules, causing hallucinations and making the same mistakes over and over. When I call these out, the agent always agrees and apologizes. I feel like something is causing it to not call its rules or follow protocol consistently.
I'm not a dev, just trying to educate myself and connect with others as I go. I'm likely the problem but I'm open to any suggestions and would love to connect with folks going through similar issues.
Running Qwen 3.6 35B locally on Mac with Hermes.
Happy to DM and share what I'm working with here. Appreciate any assistance from the community. I'll help out where I can as I keep learning.
English
The Zero Regression Rule is actually important infrastructure.
We turned every mistake into a permanent node in the memory graph. In this first layer of the Self Sovereign Agent (SSA), SSA logs the root cause, the fix, and a hard prevention rule that it now checks automatically. Next time something similar comes up, the graph flags it before the agent wastes time repeating the same mistake.
That’s how you actually stop regression instead of just hoping the model “remembers better this time.”
English
👇This is the kind of energy we need more of.
Not just “local AI is cool” but actual intent to out-execute the closed models on real workflows.
The gap is closing faster than most people think.
0xSero@0xSero
Local AI is about to be competitive. I will do everything in my power to make it better than Claude desktop/claude code by end of year.
English
This week specifically we built a mandatory file operation protocol. Basically, for any file change, it has to first perform a full read_file, then one tool call (patch preferred over write_file to avoid inadvertently overwriting the whole file which also burned me before), and then only once it quotes the verification it gets from the tool is it allowed to claim it is "done" with the task.
Might seem like overkill to some but frankly this came from multiple regressions where it would claim a note was updated without actually confirming it. The Zero Regression Rule forced it to log it and hardcode the protocol into its own behavior.
Hasn’t happened since. The graph is now at 99 nodes and actually useful.
English
@SelfSovereignAI What's an example of a tactical solution you used recently?
English
I'm noticing that most AI content is generic advice. I'm trying to do the opposite: document the actual messy process like the timeouts, the path bugs, the repeated patch errors, the memory full issues, etc. If you're building your own agent, these specific failures are more valuable than another "how to prompt" thread.
English
Added transparent reasoning traces to Self Sovereign Agent. Transparent reasoning traces aren't pretty but they force my agent to run meta-cognition before touching files or spawning specialists. This morning I caught it about to claim a write succeeded without re-reading the file. If your agent can't explain its decisions in a structured way that you can audit, it's not sovereign.
English
Self Sovereign Agent update: the memory graph now has over 90 nodes. Every time we fix a regression, we add a node and edges. The idea here is turning mistakes into a traversable knowledge structure.
English
Building the Self Sovereign Agent has been full of regressions. One of the biggest was assuming the old idea-manager system would scale. We ended up replacing it entirely with global-idea-id. The lesson: when you see duplication between systems, consolidate early.
English
Unbelievable. Claude is actually doing the local AI movement a favor by engaging in this kind of censorship and access restriction. Eventually, people will have had enough.
Péter Szilágyi@peter_szilagyi
It lasted exactly 1 day. Good run I guess. 🤡
English
Been quietly building a Self Sovereign Agent, using Hermes Agent from @NousResearch.
Four clean, concentric layers: a deep Obsidian memory core, a ruthless self-improving reasoning engine, a high-agency orchestrator that spawns specialists, and an ironclad sovereignty layer that guarantees nothing ever leaves my machine.
This isn’t another cloud wrapper or hype demo. It’s a genuine extension of my own thinking. It's local, sovereign, self-correcting, and built to compound over time.
I’m going to keep posting real updates here as it evolves: architecture decisions, the regressions we turn into rules, new capabilities, and what it actually ships. Expect threads, diagrams, and longer articles too.
This is my next major project, and I hope it inspires others to build their own sovereign agents.
Stay tuned.
English
Anyone given this local photo manager a try? These kinds of tools will be game changers for the local AI movement imo. It's getting hard to keep up with all of these tools coming out. Love to see it.
Tom Dörr@tom_doerr
Searches private photos with local AI github.com/julyx10/lap
English