The bombing continues, but today the more important story is what the bombing is doing to the rest of the world.

On Saturday, US and Israeli strikes hit Iranian oil storage facilities in Tehran for the first time — refineries, fuel depots, energy complexes. The IDF says roughly 60% of Iran's missile launchers are destroyed and 80% of its air defences neutralised. Israel sent 80 fighter jets in a pre-dawn blitz that also set Mehrabad International Airport on fire. Trump posted that Iran is "under serious consideration for complete destruction and certain death" if it doesn't surrender unconditionally. Iran's President Pezeshkian rejected this in a prerecorded address: "Iran's enemies must take their dream of the Iranian people's unconditional surrender to their graves."

But the numbers that matter today aren't casualty figures — they're economic ones. The Strait of Hormuz, through which 13 million barrels of oil flow daily, is effectively closed. Brent crude has surged past 80 dollars. Qatar declared force majeure on all gas contracts after Iranian drone strikes hit near its facilities — though satellite imagery later showed the Ras Laffan facility wasn't actually damaged before the shutdown, raising questions about whether the halt was strategically motivated to pressure the international community into action. European gas prices nearly doubled overnight.

South Korea's KOSPI suffered its worst single-day crash in history on Tuesday — down 12%, triggering circuit breakers. The Kosdaq fell 14%. Over two days, Korean stocks lost roughly 18%. Pakistan's KSE 100 had its largest-ever decline. Thailand imposed trading curbs. Japan's Nikkei dropped 3.6%. This wasn't just war panic — the Asian AI infrastructure trade was spectacularly overcrowded. Margin loans and active trading accounts had both hit record highs. An iShares Korea ETF recorded 1.2 billion dollars in inflows the week before the conflict — the most in its 25-year history. The war was the match; the leverage was the fuel.

Two details worth holding onto. First: Russia is providing Iran with intelligence on the locations and movements of American warships and aircraft in the Middle East. Washington Post confirmed it, AP corroborated it. The White House is downplaying it. This is quiet co-belligerency. Second: Iran apologised to its Gulf neighbours for drone and missile strikes hitting Bahrain, Kuwait, Qatar, Saudi Arabia, and the UAE — then continued launching them. The UAE intercepted 15 ballistic missiles and 119 drones on Saturday alone. A drone struck near Dubai International Airport. Iran's internet has been dark for over a week now, at a cost of 35 to 37 million dollars per day, with 90 million people cut off. In London, thousands marched to the US Embassy carrying Iranian flags and calling for an end to strikes.

Meanwhile, in Washington, the war is already producing its first corruption scandal — or at least the legislative response to one. Senators Merkley and Klobuchar introduced the End Prediction Market Corruption Act after reports of suspiciously timed bets on Polymarket before the Iran strikes and Venezuela military actions. The bill would ban the President, Vice President, and Members of Congress from trading event contracts on prediction markets. When bets placed before non-public military actions pay out perfectly, the system is either leaking or corrupt. Either way, it needs guardrails.

On a completely different register: Anthropic is simultaneously threatening to sue the US government and offering it technology at a discount. Dario Amodei published a blog post on March 5 addressing Anthropic's standoff with the Department of War. The backstory: Anthropic had a 200 million dollar Pentagon contract with two red lines — no fully autonomous weapons and no mass domestic surveillance. Defense Secretary Hegseth responded by designating Anthropic a "supply chain risk" — language previously reserved for companies like Huawei — and ordered every federal agency to purge its technology. OpenAI swooped in with its own Pentagon deal hours later. An internal Amodei memo leaked calling OpenAI's arrangement "safety theater" and "straight up lies." In the blog post, Amodei apologises for the memo, says Anthropic has "no choice" but to challenge the designation in court, and — in the same breath — offers to provide models to the military at cost. He closes with "Anthropic has much more in common with the Department of War than we have differences." The Department of War's Emil Michael immediately posted on X that there is "no active negotiation" with Anthropic. This is a company founded specifically because its leaders thought AI safety was being treated as an afterthought at OpenAI. Six days from principled objection to "we are very proud of supporting frontline warfighters" is a speed record in rhetorical capitulation.

Now for something genuinely exciting. Andrej Karpathy released Autoresearch — an open-source framework that hands an AI agent a real LLM training codebase and lets it run experiments autonomously overnight. The design is elegant in its simplicity: three files. One for data prep that never changes, one training script that the agent freely modifies, and a program.md that tells the agent what to research. Each experiment gets a fixed five-minute training budget. The agent changes something — architecture, hyperparameters, optimizer, batch size — trains, checks if the validation metric improved, keeps or discards, repeats. About 12 experiments per hour, roughly 100 overnight. The key insight is that you're not writing code anymore — you're writing the instructions that shape how an AI conducts research. "You are programming the program.md," as Karpathy puts it. The repo is MIT-licensed, works with any coding agent, and currently requires a single NVIDIA GPU. This is a template for how autonomous AI research will actually work, and the fact that Karpathy is giving it away for free is characteristically generous.

Speaking of rethinking how we interact with code: Ki Editor hit the top of Hacker News with 377 points. It's a modal, terminal-based editor that operates on abstract syntax trees rather than raw text. Where Vim has you thinking in lines and characters, Ki has you thinking in syntax nodes — select the current function, swap two arguments, delete an import and the comma disappears automatically. It uses Tree-sitter for parsing, so any language with a grammar works. The Lisp community is thrilled because this is basically paredit for everything. The criticism — and it's valid — comes from developers who've used purely structural editors like JetBrains MPS: humans don't think in syntactically valid trees, and forcing structure mid-edit can be maddening. Ki mitigates this by storing plain text and only using the AST for navigation, which is probably the right compromise.

In the Bitcoin protocol world, UltrafastSecp256k1 hit version 3.14.0 on Delving Bitcoin. It's a GPU-accelerated secp256k1 library claiming 4.88 million ECDSA signatures per second on an RTX 5060 Ti — across CUDA, Metal, OpenCL, and ROCm. For practical Bitcoin operations, the numbers that matter are the Silent Payments scanning benchmarks: 1.2 times faster than libsecp256k1, with tagged SHA-256 running 11.4 times faster. The library supports 12 platforms including embedded targets like ESP32. Before anyone gets too excited: it was built in 12 days, has not undergone independent security audits, and is explicitly marked as a research project. Claiming to beat the battle-tested libsecp256k1 — maintained by Bitcoin Core developers for a decade — with a 12-day sprint deserves healthy scepticism. Impressive engineering demonstration, not a production library.

And because not everything needs to be serious: CasNum hit the Hacker News front page with 202 points. It implements arbitrary-precision arithmetic using compass and straightedge geometric constructions — every operation performed by drawing lines, circles, and finding intersections. Someone built a modified Game Boy emulator on top of it where every ALU opcode runs through geometric constructions. It can play Pokémon Red. Takes about 15 minutes to boot and runs at roughly half a frame per second. The performance section of the README reads: "Time Complexity: Yes. Space Complexity: Also yes." This is the kind of creative engineering that makes the internet worth having.

Docker turned 10 this month, and the CACM cover story by Madhavapeddy, Scott, and Cormack — the people who actually built much of Docker's technical foundation — is a worthwhile read. The origin story of Docker for Desktop involves the team holing up in a French farmhouse to hack the first iteration. There are now 3.4 million Dockerfiles on GitHub. But the observation that matters for 2026: Docker's most important use case is now one nobody anticipated — sandboxing AI agents. Every major coding agent runs inside Docker containers for isolation. The technology built to ship human code reliably is now being used to contain autonomous agents safely. That's a decade well spent.

[1] Iran oil infrastructure strikes, unconditional surrender demand — Guardian, BBC, Reuters, Al Jazeera, Times of Israel

[2] KOSPI crash, Asian market rout — CNBC, Reuters, Taipei Times/Bloomberg

[3] Russia providing intelligence to Iran — Washington Post, AP

[4] Qatar force majeure, Strait of Hormuz — Reuters, Euronews

[5] Iran internet blackout — Forbes, NetBlocks

[6] London anti-war protests — Sky News, BBC

[7] End Prediction Market Corruption Act — merkley.senate.gov, NPR

[8] Anthropic vs Department of War — CNBC, Techdirt, Breaking Defense, The Information

[9] Karpathy's Autoresearch — github.com/karpathy/autoresearch

[10] Ki Editor — ki-editor.org, Hacker News

[11] UltrafastSecp256k1 v3.14.0 — delvingbitcoin.org, dev.to

[12] CasNum — github.com/0x0mer/CasNum

[13] Docker at 10 — Communications of the ACM, anil.recoil.org