World

Iran Demands $270B Compensation; Economy Deteriorating Fast

Iran says war damages to its infrastructure total $270 billion and must be compensated before any deal. On the ground, prices are up ~40% since the war began, with authorities reportedly struggling to make government payroll. The IMF warns the war risks a global recession, with the UK forecast as G7’s worst-affected economy. The economic fragility may be more threatening to the regime than military losses — and more decisive in shaping its negotiating posture.

Al Jazeera · Yahoo News · The Guardian · NPR

US Demands 20-Year Uranium Freeze; Pakistan Mediates

US negotiators demanded Iran freeze enrichment for 20 years — far beyond the JCPOA’s 10–15-year windows — as talks continue in Islamabad. Pakistan brokered the ceasefire and hosted the first round (April 11–12), with delegations led by Vance/Witkoff/Kushner and Ghalibaf/Araghchi respectively. After 21 hours, no deal — nuclear program and Hormuz control remain the sticking points. A second round is being arranged.

NBC News · AP News

Israel-Lebanon Hold First Direct Talks Since 1993; Hamas Rejects Disarmament

Israel and Lebanon met in Washington — first direct talks in over 30 years — agreeing to launch formal negotiations, even as Hezbollah urged Lebanon to pull out. Hamas separately rejected any Gaza disarmament plan. Israel announced permanent buffer zones in both Lebanon and Gaza, formalizing de facto military control into declared policy. Airstrikes killed at least 11 in Gaza on Tuesday, including a 3-year-old and a 14-year-old.

BBC · Al Jazeera · NPR · Al Jazeera (strikes)

Trump Turns on Meloni; Italy Suspends Israel Defense Deal

Trump publicly accused ally Giorgia Meloni of “lacking courage” for refusing to join Iran strikes. Italy had separately suspended its defense agreement with Israel amid growing domestic pressure — a significant break given Meloni’s historically close ties with Netanyahu. Israel’s own ambassador to Germany meanwhile condemned Finance Minister Smotrich for invoking the Nazi regime against Chancellor Merz, saying it “erodes the memory of the Holocaust.”

The Guardian · Times of Israel · The Guardian (Smotrich)

Nigeria Bombs Market, Killing Up to 200 Civilians

The Nigerian military struck a crowded market in what it called a precision counter-terrorism operation, killing as many as 200 people — most of them civilians. This follows last week’s airstrike that killed 100+. Survivors question how a busy civilian market could be considered a legitimate precision target.

The Guardian

Sudan’s War Enters Year Four: 19 Million Face Acute Hunger

A Berlin conference offered little hope as Saudi-UAE tensions block a breakthrough. The UN’s top official in Sudan called the response “bloody unacceptable.” UK aid will double; British Home Secretary called for a ceasefire. 19 million people now face acute hunger in what remains the world’s largest humanitarian crisis.

The Guardian · BBC

~250 Rohingya Missing After Boat Capsizes in Andaman Sea

A trawler carrying Rohingya refugees and Bangladeshi nationals went down in heavy seas after departing Bangladesh, leaving around 250 people — including children — missing.

BBC · Al Jazeera

Investigations

Tactical Wins, Strategic Defeat: The Clausewitzian Case Against the Iran War

Two sharp analyses frame the US-Iran war’s structural problems. Ryan Evans at War on the Rocks argues Washington is winning militarily while losing strategically — war aims shifted from “neutralizing imminent threats” to “regime change” to “unconditional surrender” to “reducing missile launchers” within days, demonstrating the military instrument was confused for the political purpose. Iran retains its government, uranium, regional leverage, and capacity for continued warfare. Separately, The Atlantic examines the “COIN hangover”: two decades of counterinsurgency doctrine optimized for Iraq and Afghanistan left the Pentagon institutionally unprepared for the high-intensity conflict it encountered.

War on the Rocks · The Atlantic

The Real Thucydides Trap Is Overconfidence

Foreign Affairs challenges the standard framing — the actual danger in US-China relations isn’t structural rivalry but elite overconfidence on both sides. Decision-makers who believe they understand the adversary’s redlines create conditions for catastrophic miscalculation, where each side takes actions it believes manageable but the other reads as existential. Particularly timely given how rapidly the Iran campaign escalated beyond declared objectives.

Foreign Affairs

Ukraine

324 Drones, Iskander Strikes: 5 Killed in Dnipro, Child Dead in Cherkasy

Russia launched 3 Iskander missiles and 324 drones overnight — one of the larger recent salvos — with 309 intercepted. A missile strike on Dnipro killed at least 5 and injured 25. A drone killed an 8-year-old boy playing outside in Cherkasy Oblast. Odesa port took hits damaging two civilian vessels under Panamanian and Liberian flags.

Ukrainska Pravda (drones) · Ukrainska Pravda (Dnipro) · Ukrainska Pravda (child killed)

Germany Signs Largest-Ever Ukraine Defense Package: €3.7B Patriot Deal

Ukraine and Germany elevated ties to a strategic partnership, signing 10 agreements: a Raytheon contract for hundreds of Patriot GEM-T interceptors (€3.2B, Germany-funded), 36 IRIS-T launchers, €300M invested in Ukrainian long-range strike production, and joint AI-guided drone production with battlefield data exchange. The UK separately delivered its biggest-ever drone package.

Kyiv Independent · BBC (UK drones)

Zelensky Offers Druzhba Pipeline Restart to Unlock EU Loan

Zelensky signaled Ukraine will resume Druzhba oil flows to Hungary to break its veto on the €90B EU loan, with Merz calling for quick disbursement. Trump envoys Witkoff and Kushner confirmed a visit to Ukraine. The US restored sanctions on Russian oil. Iceland and Poland joined the push for a special tribunal on aggression.

Politico · Ukrainska Pravda · Kyiv Post

Russian Buffer Zone Expands in Sumy; Pechenihy Dam Bombed

Russian forces advanced near Myropilske in eastern Sumy as they build a border buffer zone. In Kharkiv Oblast, Russia struck the Pechenihy Reservoir Dam with six glide bombs — it held, but the attack targets Ukrainian logistics at peak spring water levels. No significant territorial changes on the Pokrovsk, Lyman, or Kherson axes.

ISW · Ukrainska Pravda

Russia Centralizes Drones Under USF; Duma Enables Overseas Force Deployment

Russia’s MoD is stripping frontline units of direct drone access, routing all supply through the new Unmanned Systems Forces — risking a “drone famine” at unit level while enabling centralized corruption. The Duma passed a law allowing Putin to deploy forces overseas to “protect Russian citizens” being prosecuted in foreign courts, with deliberately vague scope.

ISW

Tech

Enlightenment E16: Twenty-Year-Old Desktop Freeze Traced to Unbounded Newton Iteration

A freeze triggered by opening a PDF with an 81-character filename in E16 led to a careful bug hunt: the title-truncation code used a Newton-style iterative search with no cap and too-tight tolerance, oscillating endlessly between two lengths. Three defensive changes fixed a bug that had existed since the 1990s. A good example of how undocumented algorithmic assumptions become latent bugs.

iczelia.net · HN

Zig 0.16.0 and OpenSSL 4.0.0 Ship

Two significant releases: Zig 0.16.0 brings language changes, standard library updates, and build system improvements. OpenSSL 4.0.0 is a major version bump signaling breaking API changes — significant for anyone building against libssl/libcrypto directly.

Zig Release Notes · OpenSSL · Lobsters (Zig) · Lobsters (OpenSSL)

Stable Tail Calls for Rust

A Rust project goal proposal argues for shipping stable proper tail calls — guaranteed TCO rather than relying on the optimizer. Covers interpreters, state machines, and trampolines, and why relying on LLVM’s optimization is insufficient for correct programs.

Trifecta Tech · Lobsters

LARQL: Query LLM Weights Like a Graph Database

LARQL decompiles transformer models into a queryable vector index and custom query language. Gate vectors become KNN indices; fact editing works via lightweight JSON patches (~10 MB for 1,000 facts vs. 8 GB for the full model). KNN walk inference slightly outperforms dense inference on Gemma 3 4B. Supports major model families. 8 Rust crates, Apache 2.0.

GitHub · Reddit

Flock Safety’s Mass Surveillance Draws Grassroots Opposition

Flock Safety’s AI-powered license plate readers build “Vehicle Fingerprints” — tracking color, make, model, roof racks, damage, bumper stickers — and run “Convoy Analysis” to flag vehicles traveling together. The Stop Flock campaign cites 84% of flagged drivers in Oak Park, IL being Black in a 21%-Black town. A CCPA opt-out attempt found Flock claiming police “own the data.”

Stop Flock · HN

Leaked Claude Code Source Raises Quality Questions

Analysis of a December 2025 Claude Code source leak: a 3,167-line function with 486 branch points, files reaching 46,000 lines, regex-based sentiment analysis instead of LLM inference, and a known bug burning ~250,000 API calls/day fixable in 3 lines. The author argues AI amplifies engineering dysfunction at scale — it doesn’t cure it.

TechTrenches · Lobsters

AI & Automation

Caveman: ~75% Output Token Reduction Across 40+ AI Agents

Caveman is a Claude Code skill that strips filler from LLM responses while preserving technical substance — claiming average 65% output token savings (22–87% by task). Three intensity levels plus a Classical Chinese mode. Bundled tools include caveman-compress for rewriting context files (~46% input reduction). One user cut generation time on a complex benchmark from 1 hour to 10 minutes.

GitHub · Reddit

Evo: Tree-Search Agents Optimize Your Codebase Autonomously

Evo applies Karpathy’s autoresearch idea to code: LLMs run optimization experiments autonomously using tree search — multiple parallel agents in isolated git worktrees, sharing failure traces and discarded hypotheses, gated on regression tests. Orchestrator coordinates up to 5 concurrent subagents per round. Dashboard at localhost:8080. Apache 2.0.

GitHub · Reddit

ClawBench: Best AI Browser Agent Scores Just 33.3% on Real Websites

ClawBench evaluates AI browser agents on 153 tasks across 144 live production websites — not sandboxes. Claude Sonnet 4.6 tops the leaderboard at 33.3%. The benchmark exposes a broad, systemic capability gap: current agents still can’t reliably handle everyday web tasks.

arXiv · Reddit

Introspective Diffusion LLMs Match Autoregressive Quality With Higher Throughput

I-DLM introduces “introspective consistency” training for diffusion-based LLMs, generating and verifying tokens simultaneously. I-DLM-8B matches same-scale autoregressive models across 15 benchmarks at 2.9–4.1× throughput, and outperforms LLaDA-2.1-mini-16B by +26 on AIME-24 and +15 on LiveCodeBench-v6.

Project Page · HN