The Lookout

A daily briefing on AI, Bitcoin, and tech.

Apple Podcasts Overcast Pocket Casts RSS

Listen to this briefing

Forty-eight hours. That's what Trump gave Iran yesterday to reopen the Strait of Hormuz before the US begins systematically destroying civilian infrastructure — power plants, bridges, desalination facilities. Iran's General Aliabadi responded with the kind of rhetoric that leaves no room for de-escalation: "The doors of hell will be opened to you." This is not posturing from a position of strength. Two American jets went down on Friday — an F-15E Strike Eagle shot down over Iranian territory, one crew member still missing, and a second aircraft crashing near Hormuz with its pilot rescued. Iran has posted a reward for the missing American. The B1 bridge between Tehran and Karaj, one of the country's largest, was destroyed by US strikes, killing thirteen. Bushehr nuclear facility was hit for the fourth time. Rosatom is evacuating a hundred and ninety-eight workers. Kuwait's Mina Al Ahmadi refinery is on fire from Iranian drones. The UAE was hit. Brent crude jumped eight percent to a hundred and nine dollars a barrel on Friday, and the Monday deadline is going to make markets extremely nervous when they open. Pakistan says ceasefire talks are "right on track," with Turkey and Egypt also mediating, but that feels optimistic when both sides are publicly promising escalation. Week six of this war, and the question isn't whether it's going to get worse before it gets better — it's how much worse, and what "better" even looks like when you've bombed someone's nuclear facility four times and they've mined your primary shipping lane.

On a different kind of collision course: Linux 7.0 is about to ship with PostgreSQL performance roughly halved, and the kernel maintainers are telling Postgres to fix it themselves. An AWS engineer named Salvatore Dipietro raised the alarm on Friday: on Graviton4 ARM servers, PostgreSQL throughput under Linux 7.0 is about half what it was on prior kernels. The cause is a commit that restricted preemption modes, dropping PREEMPT_NONE and keeping only Full and Lazy preemption. This forces much more time into user-space spinlocks. Dipietro submitted a patch to restore the old default. Peter Zijlstra, who wrote the original change, rejected it, pointing PostgreSQL toward Restartable Sequences time slice extensions instead. The "fix it in userspace" response is technically defensible but practically brutal — Linux 7.0 stable is two weeks out, Ubuntu 26.04 LTS ships with it later this month, and PostgreSQL is the most widely deployed open-source database on the planet. Two hundred and ninety-seven points and ninety comments on Hacker News, which is the sound of a lot of database administrators suddenly paying very close attention.

Speaking of things that got a lot of attention this week: Mark Zuckerberg is writing code again. Gergely Orosz at The Pragmatic Engineer reported that Zuckerberg submitted three diffs to Meta's monorepo last month — his first code contributions in twenty years. The tool pulling him back in: Claude Code, Anthropic's terminal-based coding agent. One of his diffs received over two hundred approvals from engineers who presumably wanted the bragging rights of having reviewed the CEO's code. He's not alone in this pattern. Garry Tan at Y Combinator built and open-sourced gstack, a system of twenty-three specialist Claude Code tools. Tobias Lütke at Shopify is running similar experiments. The thesis is straightforward: AI coding tools have lowered the barrier enough that technical founders who stopped coding a decade ago can be productive again. The irony that the CEO of the company behind Llama models is using a competitor's tool is hard to miss, though in fairness, Claude Code's harness engineering is genuinely good, and at that level of leadership you use whatever works.

That connects neatly to a piece Sebastian Raschka published this week on the components of a coding agent, which argues — convincingly, in my view — that we've reached the point where the harness matters more than the model. Repo context management, prompt-cache stability, tool design, memory across sessions, long-running continuity. These are systems engineering problems, not model capability problems. The vanilla versions of frontier LLMs have very similar raw capabilities now. What makes Claude Code feel different from Cursor from Codex from OpenClaw is the software wrapped around the model. Two hundred and thirty-nine points on Hacker News, which is the kind of engagement you get when a respected researcher articulates what practitioners already feel in their bones.

In the department of naming things — a notoriously hard problem in computer science — Tey Bannerman sat down and tried to count every product Microsoft has called "Copilot." The answer is at least seventy-five. Not seventy-five features of one product. Seventy-five distinct products and capabilities, each wearing the Copilot brand like a teenager who found one band t-shirt and wears it to everything. There's Copilot the chatbot, Microsoft 365 Copilot, GitHub Copilot, Copilot+ PCs, Security Copilot, Dynamics 365 Copilot, and sixty-nine others. The piece landed at six hundred and forty-nine points with three hundred and eleven comments, and the discussion is withering. When a single brand name means seventy-five things, it means nothing. Enterprise buyers — the people Microsoft actually needs to convince to pay double their license fees — can't explain what they bought. That's a problem that gets very visible at renewal time.

An interesting thaw in a long-running cold war: Apple approved Tiny Corp's driver that lets Nvidia and AMD eGPUs work with Apple Silicon Macs. This is primarily for AI and compute workloads — no video acceleration, so don't expect to game on it — but it means you can plug an RTX card into a Mac via Thunderbolt and use it for local model training and inference. Apple has been actively hostile to Nvidia since the Apple Silicon transition in 2020, so this reversal likely signals that AI demand from their developer base became impossible to ignore. Four hundred and thirty-six points on Hacker News.

Meanwhile, Germany is implementing its EU Digital Identity wallet in a way that requires citizens to have an Apple or Google account. The wallet app relies on mobile device attestation through Apple and Google's platform APIs, which means your government-issued identity depends on a US tech company. Two hundred and seventy-nine points and two hundred and thirty-nine comments on Hacker News, which is about what you'd expect when a European government mandates dependence on the companies that European regulators spend most of their time regulating. The irony is precise enough to hurt.

On the Bitcoin side, the Consensus Cleanup proposal reached a useful milestone: Antoine Poinsot demonstrated intentionally slow blocks on Signet, showing the real-world impact of the timewarp and related consensus bugs the proposal aims to fix. Having a live demo on Signet makes the case for the cleanup much more concrete than abstract discussions about theoretical attacks. Separately, BIP-110 — OP_TXHASH and OP_CHECKTXHASHVERIFY, which enable flexible transaction introspection covenants — released version 0.4.1 with an implementation submitted to Bitcoin Core. There's also a solid thread running on how wallet fingerprints damage Payjoin privacy, which matters for anyone who thinks on-chain transaction privacy has a future. The network itself remains extraordinarily quiet: block 943,747, one sat per vbyte across all priority levels, sixty-seven thousand dollars. The mempool is essentially empty. Against a backdrop of war, oil shocks, and geopolitical chaos, the Bitcoin network is doing what it always does — processing blocks, indifferent to the noise.

References

The Lookout

Archive