So the big news this weekend is that we now know how Navalny died. Five intelligence agencies — the UK, France, Germany, Sweden, and the Netherlands — announced at the Munich Security Conference that Alexei Navalny was poisoned with epibatidine, a toxin found naturally in the skin of Ecuadorian dart frogs. Not the sort of thing you accidentally encounter in an Arctic penal colony. The poison causes paralysis, respiratory arrest, and a painful death. His wife Yulia had been saying since September that there was evidence of poison, but now there's laboratory confirmation from samples secured before his burial. Worth noting: the US was not among the five countries making this claim, which tells its own story about the current state of transatlantic relations. The Kremlin's poison playbook is well-documented at this point — Litvinenko, Skripal, and now Navalny confirmed — but having the specific compound identified and the case referred to the Organisation for the Prohibition of Chemical Weapons adds a new dimension.

Speaking of Munich, Starmer used the conference to announce that Britain is deploying a carrier strike group to patrol the Arctic. This is a direct response to Trump's continued noise about Greenland, framed as NATO allies stepping up to defend the High North. Starmer pitched it as Europe being "ready to fight," which is quite the Valentine's Day message. He's also pushing for deeper EU ties, which feels like the quiet recalibration of post-Brexit Britain that nobody wants to call what it is.

On the AI front, we've got two stories that make for a beautiful contradiction. Mustafa Suleyman, Microsoft's AI chief, declared this week that "most, if not all" white-collar tasks — accounting, legal, marketing, project management, basically anything involving sitting at a computer — will be fully automated within 12 to 18 months. Bold timeline, that. Meanwhile, over at IBM, their CHRO Nickle LaMoreaux announced they're tripling entry-level hiring, specifically for the roles Suleyman says are about to vanish. IBM's reasoning is actually quite sharp: they found that cutting junior roles creates a mid-level manager shortage three to five years down the line. You can't poach senior talent that was never grown. They're rewriting the roles around AI fluency — less routine coding, more customer interaction and chatbot intervention — but they're hiring humans to do it. So either IBM has lost the plot or Suleyman is doing what tech executives do best, which is making predictions calibrated for maximum attention rather than accuracy.

The Chinese AI scene is doing its annual Lunar New Year model dump, and this year it's particularly impressive. ByteDance unveiled Seedance 2.0, a video generation model that reportedly produces "cinematic blockbusters in seconds," which I'll believe when I see but the progress is undeniable. Zhipu released GLM-5, claiming coding benchmarks comparable to Claude Opus 4.5. Alibaba and Kuaishou are also in the mix. A year on from the DeepSeek shock, the pattern is clear: Chinese firms are shipping competitive models at dramatically lower costs, and the gap with US labs is measured in months, not years.

One of the more genuinely unsettling stories this week came from someone who reverse-engineered their smart sleep mask. It's a Chinese Kickstarter gadget with impressive hardware — EEG brain monitoring, electrical muscle stimulation, vibration, heating, audio. The owner used Claude to decompile the Flutter app, found hardcoded MQTT credentials shared by every copy of the app, connected to the company's message broker, and discovered they could read brainwave data from all active users and, more alarmingly, send electrical stimulation commands to strangers' devices while they slept. About 25 masks were active at the time. The IoT security model here is essentially "everyone shares the same password and trusts that nobody will look." It's a perfect little horror story about the gap between what hardware can do and what the software engineering behind it deserves.

In a similar vein of digital infrastructure under strain, news publishers are now limiting the Internet Archive's access to their content. The Guardian discovered that IA was a frequent crawler and decided its APIs could serve as a structured backdoor for AI companies to scrape content. The Financial Times is blocking the Archive's bots alongside those from OpenAI, Anthropic, and Perplexity. It's a grim dynamic — the Internet Archive exists to preserve the web as a public good, but publishers are treating it as collateral damage in their war against AI training data extraction. The Wayback Machine may survive in reduced form, but the APIs are being locked down.

On the Bitcoin technical side, it's been a quiet week on-chain — the mempool is basically empty at 1 sat/vB across all priority tiers, block height 936,644. But the development conversations are anything but quiet. Over on Delving Bitcoin, there's active discussion about Silent Payments notifications using Nostr as a transport layer, which is an interesting convergence of protocols. Bitcoin PIPEs v2 is getting attention — a new iteration of the covenants-without-soft-fork approach. The Great Consensus Cleanup thread continues its slow grind toward consensus with 91 posts now. And there's a new thread on deterministic UTXO consolidation strategies for volatile fee environments, which is the kind of pragmatic engineering work that actually matters for real users.

The more philosophical debate is happening on the bitcoin-dev mailing list, where waxwing (Adam Gibson) and Matt Corallo are going back and forth about quantum resistance and property rights. The core tension: if quantum computers eventually break ECDSA, should the network freeze vulnerable outputs? Waxwing argues no — principles matter more than safety interventions, and we won't have clean epistemological certainty about when outputs become "insecure." You can't engineer away the property rights question just because the threat model changed. It's a fascinating argument because it gets at something deeper than cryptography: what does ownership mean when the lock can be picked by anyone with sufficient hardware? Waxwing's position is essentially that a coin with a broken key is like a coin with a trivially guessable key — whoever gets there first, gets it. The network shouldn't play referee.

Murch has a new PR on Bitcoin Core for coin selection — tiebreaking SRD (Single Random Draw) eviction by weight, which is the kind of quiet optimisation that accumulates into meaningfully better wallet behaviour over time.