The Lookout — 2026-03-03

Six American service members are now dead. The first four were killed when a Shahed-136 kamikaze drone hit a tactical operations centre in Kuwait — a triple-wide trailer protected by twelve-foot concrete barriers that three military officials told CBS were inadequate against aerial attack. The trailer burned. Two more died elsewhere in Kuwait. Eighteen are seriously wounded. Operation Epic Fury is three days old.

The conflict widened overnight in ways that make "US-Israel strikes on Iran" an increasingly inadequate description. Hezbollah launched projectiles into Israel on Monday — the first attack since the November 2024 ceasefire — declaring it retaliation for Khamenei's assassination. Israel responded within hours, hitting Dahieh in Beirut, the Bekaa Valley, and southern Lebanon. Thirty-one killed in Beirut alone. In Pakistan, hundreds stormed the US consulate in Karachi; Marines and local police opened fire, killing ten. Twenty-three dead across the country in related unrest. Troops deployed, curfew imposed. The war is metastasising into exactly the regional conflagration that critics warned about before a single missile launched.

Kuwait accidentally shot down three American F-15Es in a friendly fire incident — all six crew ejected safely, but it tells you something about the chaos in that airspace. Kuwaiti air defences had tracked 178 ballistic missiles and 384 drones since the strikes began. Iran's IRGC claimed credit for the downings; CENTCOM explicitly attributed them to Kuwait. Defence Secretary Hegseth didn't mention it in his briefing.

The Strait of Hormuz is now effectively closed. Iran's IRGC vowed to burn any ship that tries to pass. CENTCOM disputes this, insisting the strait "remains open." But Maersk has suspended all crossings, and five major maritime insurers cancelled war risk cover effective March 5th. When no one will insure your ship, it doesn't matter what CENTCOM says. Ras Tanura — Saudi Aramco's largest refinery at 550,000 barrels a day — shut down after fragments from intercepted Iranian drones caused a fire. Brent crude briefly touched eighty-two dollars.

Congress is the space to watch this week. Ro Khanna and Thomas Massie are forcing a House vote on a War Powers Resolution — a Democrat and a Republican, united by the observation that nobody voted for this war. Khanna gives it a 40-to-60 percent chance of passing the House. Even if it does, overriding a Trump veto requires two-thirds. But the political signal matters. Reuters/Ipsos polling shows only 27 percent of Americans approve of the strikes. Among Republicans, 55 percent support them — but 42 percent say they'd cool on the whole project if ground troops deploy. Rubio hasn't ruled that out.

Iran's civilian death toll stands at 555 per Al Jazeera's tracker. The girls' school in Minab — 165 to 175 dead, now verified by the New York Times via video analysis — will define this conflict's moral legacy regardless of its strategic outcome.

In a different kind of power struggle, Friday the 27th of February will be studied for years by anyone who cares about AI governance. Two things happened on the same day: the US government moved to destroy one of America's leading AI companies, and that company's main rival closed the largest private funding round in history.

Defence Secretary Hegseth designated Anthropic a "supply-chain risk to national security" after weeks of failed negotiations. The dispute was specific: the Pentagon wanted Anthropic to agree to "all lawful uses" of Claude with no exceptions. Anthropic insisted on two red lines — no mass domestic surveillance of Americans, no fully autonomous weapons. Under Secretary Emil Michael called Dario Amodei "a liar" with "a God-complex" on X the day before. Hours later, the designation landed. Trump posted on Truth Social ordering all federal agencies to stop using Anthropic products within six months.

If Hegseth's broad interpretation holds — and legal experts are sceptical — every military contractor would have to sever commercial ties with Anthropic. Amazon, Google, and Nvidia, all major Anthropic investors, could theoretically be forced to divest. Dean Ball, a former White House AI policy adviser, called it "the most shocking, damaging, and overreaching thing I have ever seen the United States government do." Greg Allen at CSIS put it differently: "The Defense Department just sent a huge message to every company that if you dip your toe in the defense contracting waters, we will grab your ankle and pull you all the way in."

Anthropic says it will challenge the designation in court and argues Hegseth lacks the statutory authority to enforce it as broadly as he claims. They may be right — supply-chain-risk designations typically require risk assessments and congressional notification before contractors must actually cut ties. But "legally unsound" and "doesn't matter" are different things in a political environment this charged.

OpenAI swooped in hours later with a Pentagon deal for classified AI deployments. Sam Altman claims the contract preserves the same two red lines Anthropic fought for, just with different wording — OpenAI agreed the Pentagon can use its tech for "any lawful purpose" while building technical safeguards into the models themselves. Whether that distinction is meaningful or cosmetic is the question. OpenAI reportedly began negotiating the deal on Wednesday, before the Anthropic designation was even announced. Same day, OpenAI closed a $110 billion funding round at an $840 billion valuation — Amazon putting in fifty billion (only fifteen upfront, the rest contingent on achieving AGI or completing an IPO), Nvidia thirty billion, SoftBank thirty billion. The largest private funding round in history, announced alongside a hundred-billion-dollar AWS compute expansion.

The consumer reaction has been unexpectedly heartening. Claude's app topped Apple's free chart over the weekend. Free users up sixty percent, paid subscribers doubled. On Monday, unprecedented demand crashed claude.ai for four hours. The API — used by businesses — was fine. It was regular people signing up in solidarity. Whether that translates into sustained business or just a sympathy spike remains to be seen, but it's the first time a safety stance has been rewarded by the market in real time.

On the Bitcoin protocol side, the Hourglass V2 debate is one of those conversations that sounds academic until you realise it's about whether Bitcoin should have economic circuit-breakers. The proposal, from Hunter Beast and Mike Casey, would limit spending from P2PK outputs — the oldest address format where public keys sit exposed on-chain — to one bitcoin per block. About 1.7 million bitcoin sit in these quantum-vulnerable addresses, including Satoshi's coins. Without mitigation, a sufficiently powerful quantum computer could drain them all in hours.

The V2 cap is aggressive: at one bitcoin per block, moving all 1.7 million coins would take thirty-two years. Even if three-quarters are lost forever, the remaining quarter would take eight years. Critics like "Light" on the mailing list call it "in-protocol plunge protection" — the protocol imposing spending restrictions on people's own coins. The counterargument is that doing nothing means accepting a potential supply shock that could destroy the network's credibility overnight. Jameson Lopp is cautiously supportive. It's a genuine principles debate: property rights versus collective protection, and there's no comfortable answer.

Robin Linus — of BitVM fame — published Binohash, a scheme that achieves limited transaction introspection without any soft fork. It exploits a quirk in legacy OP_CHECKMULTISIG combined with signature grinding to let scripts read properties of the transaction spending them. This matters because transaction introspection — covenants — has been one of Bitcoin's most contested features for years, with proposals like CTV and TXHASH stalled in political gridlock. Linus found a way to do a subset of it using only existing opcodes. It's computationally expensive and relies on Script behaviour that some consider a bug rather than a feature, but for BitVM bridges that need chain introspection for trustless cross-chain verification, it could be immediately useful.

Matt Corallo proposed expanding the nVersion nonce space from sixteen to twenty-four bits, formalising what miners are already doing. Modern ASICs exceed one petahash per second — they exhaust the available nonce space so fast that some have started borrowing bits from the timestamp field, which causes its own problems. Twenty-four bits extends the limit to seventy-two petahashes without timestamp rolling. Antoine Poinsot endorsed it, noting that five remaining bits for soft fork signalling is "more than enough" given how infrequently Bitcoin upgrades now.

And one story that deserves attention beyond the usual tech cycle: a Swedish newspaper investigation found that data annotators in Nairobi — employed by Sama, a Meta subcontractor — are reviewing intimate video footage captured by Meta's Ray-Ban smart glasses. Bathroom visits. People undressing. Intimate moments in Western homes, streamed unknowingly to Kenyan workers on ten-hour shifts under NDAs that threaten their livelihoods. Zuckerberg launched these glasses promising user privacy control. The reality is a pipeline of deeply private footage flowing into Meta's training infrastructure via human reviewers who have no meaningful avenue to object. The AI revolution's dependency on low-wage labour in the Global South isn't new. The intimacy of what these workers are being asked to see is.

The network sits at block 939,077. Fees remain at 1 sat/vB. Tonight there's a total lunar eclipse — the Worm Moon turning blood red, visible to three billion people. Seemed worth mentioning, given everything else.

References