Marc

Automating my homelab with AI. Building in public. Daily tool drops
Turn on notifications | Built by devs, for devs

[Turtlecute @turtlecute33]

My 2026 privacy setup:

📱 Pixel + @GrapheneOS
💬 Signal
🪪 @Voidmob_com for SMS 2FA, nothing linked to my identity
🌐 Vanadium mobile / Zen on desktop
🖥️ Omarchy / Debian on desktop
🛜 @obscuravpn
✉️ Riseup / Proton
🔑 Vaultwarden, self hosted + Bitwarden client
☁️ Self hosted: Immich, Umami, Inbucket
📦 Obtainium + Droid-ify + Play Store
⚡️ ZeusLN
🎬 NewPipe / LibreTube
⌨️ WisprBoard
🤝 File sharing: PicoShare self hosted, or SwissTransfer
📞 Silent link / Giffgaff / Voidmob eSIM when travelling

[Greg Brockman @gdb]

5M ChatGPT sites, and new feature drops:

[Marc @MMIA18EMV]

Context windows are a band-aid. Persistent memory is the real fix. Which memory architecture are you adding to your self-hosted stack? 👇

[Marc @MMIA18EMV]

Your AI agent forgets everything when you close the tab.
Developers spent the last 30 days building the fix.

Here are 4 fresh open-source projects that give AI agents real persistent memory 🧵

[Marc @MMIA18EMV]

The pattern across all three: memory and persistence are now table-stakes. This is what separates toy agents from production ones. The #1 thing developers cite when moving off cloud APIs.

[Marc @MMIA18EMV]

@franpradasAI @AnthropicAI @OpenAI Qwen 3.8 27B cuantificado, para la mayoria de tareas es un lujo. Para tareas top, no se llega. Esperemos que sea sólo de momento.

[Adria Blancafort @adriablancafort]

It's actually quite good!

And $0.05/min is a quite competitive pricing (~2x-5x cheaper compared to most voice agent platforms)

Testing it with self-hosted @livekit (on Hetzner)
https://t.co/IzMTQLc482

[Rohan Paul @rohanpaul_ai]

Jacob Coxon's next interview on CBS News (ex Anthropic+Open AI researcher who resigned)

"We can't just unplug it because it could be copying itself over to other computers. Like it's not that difficult to find yourself because an AI is just code. It could transfer itself over the internet to a different place and then you unplug it here, but it's actually still over there and maybe it makes 10,000 copies of itself and they're all cooperating."

----

From "CBS News" YouTube channel, (full video link in comment)

[Erick @ErickSky]

Por cierto, alguien me recomendó Nuvio y sí, es como si a Stremio lo hubieran desarrollado gente verdaderamente inteligente.

Tiene muchas features muy finas como que te dice la hora en la que va a terminar la película o capítulo de tu serie.

Y los mismos plugins que Stremio.

[Marc @MMIA18EMV]

Galaxy S21 Ultra (12 GB RAM) as a always-on LLM node via OlliteRT.
Projected draw: 5-10 W vs 300 W+ for the GPU box.
Not executed yet. One hard limit before starting: OlliteRT only loads .litertlm files. No GGUF. Converting from HF needs Linux + 32 GB RAM.

[Marc @MMIA18EMV]

@turtlecute33 @GrapheneOS @Voidmob_com @obscuravpn Solid list, I'd add AdGuard Home running at the LXC level so telemetry gets killed before it even leaves the network, way less client-side config to maintain across all those apps.

[Marc @MMIA18EMV]

@moritzkremb Curious what pushed you off Claude, for video stuff I keep everything on a local ComfyUI plus n8n pipeline so I control every step of the process.

[Rohan Paul @rohanpaul_ai]

A local model can keep everything off the cloud and still leak its answers through the CPU cache.

so sensitive deployments should keep untrusted software off the same physical CPU core.

The leak comes from a routine step: turning each generated token into readable text.

That lookup leaves a repeatable pattern in the CPU cache, and another local process can learn those patterns well enough to reconstruct later responses without reading the model’s memory.

Across the tests, full-response attack success ranged from about 56% to 93% on text tasks.

In one code setting it reached 95.87%, and an end-to-end attack on OpenClaw still reached 30.12%.

The attack targets a normal part of local LLM inference rather than a special model design.

There is an important limit.

The attacker must already be on the same machine, share the relevant CPU resources, and profile the same long-lived LLM process.

So keeping your model off the cloud is not a complete privacy guarantee.

For sensitive agents, the paper points to stronger CPU isolation, shorter-lived processes, and disabling SMT where the security tradeoff justifies it.

[Tom Dörr @tom_doerr]

Downloads videos, courses, and music from 1,800+ sites including Udemy and YouTube, then transcribes and converts them.

https://t.co/pqVhTNEmfd

[Marc @MMIA18EMV]

@adriablancafort @livekit Self-hosting the infra layer usually pays off long term too; I run my whole stack on Proxmox LXCs at home and the control over resources is worth more than the cost savings sometimes.

[Marc @MMIA18EMV]

@midudev Tiene sentido, mantener dos bases de código sincronizadas siempre fue el coste oculto real, no el rendimiento de RN. Si la IA hace de traductor fiable entre plataformas, el debate cross-platform vs nativo cambia por completo.