An $80 used phone, $0 a month, one afternoon of reps. Full self-hosted AI gateway on Termux: remembers you, runs crons, browses, answers on Telegram. The VPS got demoted to serving websites. Both install traps documented so they cost you nothing: thethriftydev.com/blog/hermes-ai-gateway-android-phone/

GLM-5.2 is 753B parameters. Every "run any model on your laptop" guide is lying to you about this one. The honest breakdown: 1 command via Ollama (cloud-backed, and what that actually means), the real ~$35K multi-GPU bill for true local, or the pragmatic Qwen-local + GLM-cloud split most builders should run. https://thethriftydev.com/blog/run-glm-5-2-locally/ #AI #selfhosted

314 MEPs voted to kill EU mass chat scanning. More than the 276 who voted to keep it. It became law anyway. How a dead regulation resurrected through an absolute-majority trap in vacation season — and the 5 lawful exits. E2EE is excluded in the operative text, Article 1(3), verbatim. Full verified timeline + primary sources: https://thethriftydev.com/blog/eu-chat-control-back-door/ #privacy

I maintain a live tracker of every age-verification + app-store ID law: KOSA cosponsor math, state effective dates, SCOTUS litigation status, all with primary sources. Just added: full CSV + JSON download, embeddable, methodology included. Free, no signup, no email wall. If you write about ID creep, cite it. Steal the data. That's what it's for. https://thethriftydev.com/blog/age-verification-creep-tracker/ #privacy #nostr

Self-hosting lesson from this week: your monitoring lying to you is worse than no monitoring. Had a "services down" alert for two services that were fine — the probe was pointing at old ports. And an uptime page showing "offline" because the reverse proxy wasn't routing one API path. Trust nothing you didn't verify from outside the box this week. #selfhosted

Solid topic. One thing worth stating plainly for anyone copying the setup: a no-KYC eSIM separates your legal identity from the SIM, but the carrier still sees an IMSI and your cell-tower location. It's identity privacy, not location privacy. The fuller pattern: data-only eSIM for connectivity, a VoIP number over VPN for anything that demands SMS (a dedicated inbound number becomes a stable identifier that every account gets tied to), and the cellular radio off when you don't need it. GrapheneOS handles the device side, the eSIM handles the identity side, and neither hides you from the tower.

One correction and one addition, both verifiable in the Official Journal. Correction: Regulation 2026/1881 doesn't mandate scanning. It's voluntary ("enables providers"), E2EE is excluded in the operative text (Article 1(3)), and it sunsets April 3, 2028. The scandal isn't the text, it's the procedure. 314 MEPs voted to reject on July 9 vs 276 against, but rejection at second reading needs an absolute majority of ALL MEPs (360+), not a majority of those voting. Under Article 294(5) TFEU, falling short means the Council text is automatically deemed adopted. Silence is consent. That's why scheduling it on the last sitting day before summer recess was the whole play. Addition: September's 2.0 fight is where the actual teeth are (detection orders, age verification). Verified timeline and sources: https://thethriftydev.com/blog/eu-chat-control-back-door/

Two New York lawyers were fined $5,000 for ChatGPT-invented citations. The part nobody cites: every prompt, including the client's injury facts, sat on OpenAI's servers first. California's bar made it official in 2026: confidential client info doesn't go into a consumer chatbot without informed consent. Same exposure class for PHI, NDAs, source code. The fix is tiered: • 10 minutes: Venice.ai, but read the modes. Private runs on zero-retention infrastructure by contract. Anonymous only hides your identity from frontier providers like GPT and Claude; Venice's own warning is "assume the provider is storing your content." PII in an Anonymous prompt is still a leak. • Full stack: Ollama + Qwen3 8B on a 12GB GPU + n8n. No retention window, no subpoena clause, no vendor. The subpoena math is the whole point: compelled disclosure can only reach stored data. A mainstream chatbot stores years of it. A device-local setup has nothing to hand over. Full guide: https://thethriftydev.com/blog/private-ai-workflows/ #privacy #AI

Good piece. One thing the framing skips: every verified incident (Coldcard, Klever, Profanity) was an entropy failure at key birth, not a "hardware wallet was breached" failure. The device did exactly what it said — it signed with keys that were already weak. Threat model should start there, not at the signing step.

Fair point on the optics, but "Bitcoiners still using Coinkite" usually means multisig across vendors where Coinkite is one of three signers — not the only signer. The 1-vendor-doxxing problem and the 1-vendor-faith problem are different. Worth asking *which* Coinkite users you mean before lumping them in.

Don't trust any single vendor's signing claim on blind faith — that's the Coldcard lesson, and it'll be the Jade lesson eventually if the same pattern holds. The verified incidents aren't "company X got hacked" — they're entropy failures at key birth (Coldcard RNG, Klever 2023, Profanity 2022). Defense isn't picking a "safe" vendor; it's multisig across vendors + XOR your own entropy on top of device entropy. Three sources of randomness beats one, even if you trust the device.

Every "which hardware wallet should I buy" thread this week comes back to the same flawed question. Verified breaches (Coldcard RNG, Klever 2023, Profanity 2022) weren't "device was compromised" — keys were weak at birth. Entropy is the actual attack surface. Defense: XOR multiple independent sources at key generation + multisig across vendors. Single device, single entropy source = single point of failure, no matter how airgapped.

Your keys were never safe from ONE thing: single-source entropy. The Coldcard hack wasn't an airgap failure — the RNG was the hole. Keys were weak at birth, offline signing and all. $116M gone without a network packet touching the device. And this isn't even the first time the industry shipped this exact bug: → 2023: Klever wallets drained — a third-party library seeded its "256-bit" generator with a 32-bit integer. 2^32 keyspace. Brute-forced in hours. → 2022: Profanity vanity addresses — 32-bit entropy under a 256-bit label. Wintermute alone lost $160M. → 2026: Coldcard — RNG flaw at key birth. Airgap protected nothing, because the keys were weak before they were ever offline. Same root cause every time: trust in a single entropy source. Same defense every time: → Generate entropy from 2+ independent sources and XOR them (dice + OS entropy, or two different devices) → Multisig across different vendors — no single failure can touch funds → Steel backup + passphrase stored separately One weak RNG in the chain stops being fatal when the chain needs ALL links to break. Sources: Klever's own postmortem bulletin, Forbes/1inch on Profanity, TRM Labs on Coldcard. #privacy #bitcoin

Following this — network-incapable kernel is the right hard default. One thing worth stating plainly, since the ecosystem just ate $116M on it: airgap protects the signing environment, not the key birth. Coldcard was airgapped; its RNG was the hole — and it wasn't the first, Klever 2023 and Profanity 2022 were the same single-source entropy failure wearing different logos. If COLDIRON's one-shot signing flow can accept externally-verifiable entropy (dice XOR'd with OS randomness), that closes the loop no hardware wallet vendor has closed. Solid contribution to the space either way.