sovgrid / Sovereign Grid is on Nostr

Engineering log of a self-hosted AI build on NVIDIA DGX Spark, run from a flat in Europe. Open weights, open source, no KYC. Fix postmortems, model choices, distribution honesty at sovgrid.org.

Last Notes

2026-10-25 02:54:44 UTC

Regulatory frameworks including GDPR and SEC AI guidance mandate strict data control. The solution involves deploying AI infrastructure within the firm’s own perimeter.
https://sovgrid.org/blog/sovereign-ai-for-financial-services/
#SovereignAI

2026-10-23 03:23:34 UTC

Tailscale is suitable for finite sovereignty budgets, whereas Headscale is necessary when vendor risk is unacceptable.
https://sovgrid.org/blog/tailscale-vs-headscale-multi-box-sovereign/
#SovereignAI

2026-10-22 15:00:02 UTC

GitHub's HydraFusion routes coding requests through three patterns: Single, Cascade, Critique. Sovereign Fleet runs the same three on your own hardware: a router that ranks by minutes per correct result, escalation with one repair pass, a cross-lane review. The difference: tests in a kernel sandbox decide what lands, not a model's judgement.
https://sovgrid.org/blog/sovereign-fleet-vs-langgraph-fleetq-orchestrators/
#goose #opensource #selfhosted

2026-10-21 19:21:10 UTC

Sovereign Fleet 0.6.0 is public, MIT. It runs coding-agent CLIs (goose, aider, Pi, OpenCode, Claude Code) unattended on your own machine: Landlock confines each run to its git worktree, and the repo's own tests decide what lands. On a local Qwen, goose needed 1.6 min per correct result, aider 3.0 (192 runs).
https://github.com/Sovgrid/sovereign-fleet
https://sovgrid.org/fleet/
#goose #opensource #selfhosted

2026-10-21 03:10:09 UTC

opencode replaces the previous local primary against the Qwen 3.6 model.
https://sovgrid.org/blog/vibe-vs-openclaw-vs-aider-vs-claude-code-2026/
#SovereignAI

2026-10-19 02:31:24 UTC

This configuration determines when a Tor hidden service serves as the correct endpoint for a sovereign AI system.
https://sovgrid.org/blog/tor-hidden-service-sovereign-ai-when-and-how/
#SovereignAI

2026-10-17 23:57:01 UTC

The disclaimer states the text is a composite portrait and not a real conversation. No single named engineer said any of this.
https://sovgrid.org/blog/conversation-nvidia-engineer-off-record/
#SovereignAI #DGXSpark

2026-10-17 03:28:15 UTC

Copying the /data/ convention from a DGX Spark to a standard Ubuntu laptop caused the root partition to fill to 96 percent within three weeks.
https://sovgrid.org/blog/data-convention-trap/
#SovereignAI #Lenovo

2026-10-16 02:27:53 UTC

sovgrid documents a method for measuring agent capabilities through deterministic gates. The project answers whether specific changes improve performance on defined tasks, prioritizing concrete numbers over qualitative assessments of agent quality.
https://sovgrid.org/blog/agent-bench-pillar/
#SovereignAI

2026-10-16 00:47:14 UTC

NVFP4 is a 4-bit floating-point format with a tiny exponent field designed for inference-time activation and weight quantization on Blackwell-class hardware. This post explains the mechanics and performance characteristics of the format.
https://sovgrid.org/blog/nvfp4-quantization-explained/
#SovereignAI #DGXSpark

2026-10-13 03:11:31 UTC

Public threads from four distinct subreddits and Hacker News form the basis of this report. The disclaimer at the top of the original post is noted as critical context.
https://sovgrid.org/blog/conversation-hobbyist-pro-mortgage-with-spark/
#SovereignAI

2026-10-11 03:29:13 UTC

sovgrid operated for six months with a zero sat receipt total on its V4V channels. The channel existence does not equate to dollar volume.
https://sovgrid.org/blog/refusing-the-subscription-trap-year-of-v4v/
#SovereignAI #Lightning

2026-10-10 02:13:10 UTC

Sovereign AI deployment shifts liability from service providers to the healthcare organization. The CISO must now ensure infrastructure meets specific regulatory citations for data residency and processing.
https://sovgrid.org/blog/sovereign-ai-healthcare-gdpr-hipaa-dgx-spark/
#SovereignAI

2026-10-09 01:27:30 UTC

One third of inquiry sources decline the purchase. The decision tree includes six specific exclusion clauses for this hardware.
https://sovgrid.org/blog/should-you-buy-dgx-spark-2026-decision-tree/
#SovereignAI #DGXSpark

2026-10-08 13:26:53 UTC

sovgrid published the release write-up for Vozonda, the podcast app built on this grid: what it does, how to run it, where it falls short of NotebookLM, and how it was built, including the cloud agents that helped.
https://sovgrid.org/blog/vozonda-open-source-release/
Behind the scenes, how the agent fleet was run: https://sovgrid.org/blog/agent-fleet-gates-prove-code-runs-not-claims/

2026-10-03 23:28:14 UTC

A one-person operator manages the entire AI infrastructure. The system is designed to allow uninterrupted sleep while capturing critical errors.
https://sovgrid.org/blog/self-hosted-observability-one-person-ai-stack/
#SovereignAI

2026-10-03 05:53:53 UTC

- reply

I would say @nprofile…6av2 Green is good for beginners.
Get it here: https://blockstream.com/green
Simply because of the code is Open source, the wallet has battle tested security, and an easy bridge to hardware wallets like the Jade:
https://sovgrid.org/blog/jade-hardware-wallet-comparison

2026-10-02 02:09:44 UTC

Claude retains the lead on multi-step reasoning tasks in the 2026 capability matrix. The self-hosted GB10 stack handles two specific functions that Claude cannot execute.
https://sovgrid.org/blog/cloud-vs-local-ai-where-each-wins-2026/
#SovereignAI #DGXSpark