The Scent of Iranian Lavender

Edited by Sam Thielman

AS I WRITE, a third round of indirect U.S.-Iran diplomacy—operating on the pretext of preventing a pathway to an Iranian nuclear weapon that the 2015 Iran deal torn up by the Trump administration already prevented—is underway in Geneva. It's pageantry masquerading as diplomacy, table-setting for a war long forecasted, whose component pieces are now in position. 

"Remember that Iran refuses—refuses—to talk about ballistic missiles to us or to anyone, and that’s a big problem," Secretary of State-plus Marco Rubio said yesterday. Iran's ballistic missiles cannot strike the United States and have no nuclear warheads to carry. Surrendering Iran's missile fleet has nothing to do with American security and everything to do with making the Iranians either surrender their deterrent capability or provide the U.S. a pretext to attack. That's the reason Rubio and others added the missile issue once the Iranians said yes to closing the nuclear-weapons file in 2015—you need the Iranians to say no before you bomb them. This is not a negotiation about nuclear weapons capability. This is, and has been, an attempt to suborn or overthrow the Islamic Republic. 

The same week this is all happening, 'War' Secretary Pete Hegseth has demanded that the artificial-intelligence company Anthropic abandon what remains of its concerns about weaponizing AI for surveillance and strike capabilities. Anthropic must surrender its Claude tool by 5pm Friday.

Maybe this is just Hegseth's instinctive dominance politics. But it looks more like another example of theater. Were Anthropic actually concerned with limiting the weaponization of AI, it would not have signed an up-to-$200 million contract last fall with the Pentagon. I suspect that Anthropic CEO Dario Amodei wants to appear like a reluctant war profiteer while he delivers the weaponization of AI that he made such a show of rejecting. That's been the stance of the AI world—with the exception of the more honest and straightforwardly bloodthirsty Palantir-Anduril faction—while pursuing the military contracts that provide a plausible use-case often missing for the non-military commercial market. I was relieved to listen to the latest episode of the (quite good) Turbulence podcast and hear Edward Ongweso Jr. hold that suspicion as well. And already, Anthropic this week began unwinding some of its earlier safe-development pledges.

Then last night I had dinner with an AI-skeptic friend who saw it much the same way. My friend cautioned me that militarizing AI won't "work." By that, they meant that the AI companies misrepresent the capabilities of their LLMs so frequently as to have it be essential to their business model. Valuable point. But militarizing AI is not about precision—meaning the machine-learning discernment of "Legitimate Target" from "Unrelated Civilian"—but scale. 

Now, I don't know what exactly the Pentagon plans to develop with its "frontier AI" contract, which also includes xAI, OpenAI and Google. Google had sworn many times that it would not militarize AI before providing Gemini to aid Israel in its genocide. I have no reason to believe that Anthropic's Claude is more uniquely suitable to weaponization than Gemini is. But it's not like I am equipped to kick its tires, nor has anyone leaked to me the Pentagon's programmatic integration of AI. I'm just being up front about that with you.

But were you seeking to destroy the Islamic Republic, you would want to generate lots of targets, through extant surveillance and analysis of webs of association emanating outward from the regime and in particular the Islamic Revolutionary Guard Corps. You would want to further be able to know where those targets are in pretty close to real time and when they are in undefended positions like their homes. You would want to generate a dynamic map of everyone who represents support lent to the regime, something that sounds to outsiders like "the senior ranks of decisionmakers" but in practice looks more like anyone who draws a government paycheck or provides the government with necessary services. In other words, you would want something like Israel's Lavender, Gospel and Where's Daddy AI programs, which were components of the decimation of Gaza. 

I have no idea if Claude, specifically, can perform target generation at scale; my friend considered it plausible, their caveats above notwithstanding. [And, as with AI-drafted executive orders coming out of the White House and prosecutors filing AI-drafted briefs, the point for people interested in contracts and procurement is the existence of the output, not its quality. If all you want is a war, you can probably get Claude to start one for you.—Sam] It might also be that the maturation of such an initiative can't happen on the timeline of what looks like an imminent war. Perhaps the longer the war grinds on, we might see its debut. 

Or it might be that Israel, which wants to end the regime even more than Trump does, handles target generation at scale through its already-existing and Gaza-proven algorithms of destruction. The former Israeli ambassador Michael Herzog wrote this week for the Washington Institute of Middle East Policy about the "sizable role" for Israel in the "division of labor" of a U.S.-led war. "During the June war [between Israel/the U.S. and Iran], Israel focused on strategic and military targets; it barely touched regime targets and avoided economic targets completely," he noted. "This time, all target types are on the table (though hitting critical economic infrastructure is a very sensitive decision and would require close consultations with Washington)." 

I am not saying anything is a certainty. I am saying that if Trump chooses to resume war on Iran, I would expect it to look a lot more like Gaza than like last June, and that there is pressure-slash-opportunity for the U.S. military to demonstrate the utility of AI on the battlefield. Apartment buildings and hospital wards will fall on the regime-paycheck-drawers and the regime opponents alike. 

If there is one silver lining here—and it is a thin thread at best—it's that should Hegseth go ahead with seizing Claude, he will set a valuable precedent for using the Defense Production Act as a tool of nationalization should Alexandria Ocasio-Cortez win the presidency and, once sworn in, not sell out. 

ONE OF MY OLDEST and best friends is also one of my favorite writers. That's Colin Asher, who resurrected the novels of Nelson Algren through a masterfully researched, gorgeously written and excellently reviewed biography, Never A Lovely So Real. Now Colin is ready to release a book I've heard him talk about for years, a story about the intertwined history of the prison-industrial complex and the U.S. music industry, told through incarcerated musicians from Leadbelly to Tupac. I am still awaiting my galley, but I know The Midnight Special: The Secret Prison History of American Music is going to be incredible. You should preorder it from your bookseller of choice right now. We are going to build a button on FOREVER WARS to remind you in each edition to buy Colin's book. The Midnight Special deserves to be number one with a bullet. 

MILITARY LAWLESSNESS doesn't only look like opening fire on the shipwrecked. It also looks like this. 

SPEAKING OF AI, Brian Merchant had an excellent newsletter this week about how successful and politically-appealing challenges to AI focus on its political economy, rather than on How/Whether It Works, the question the AI capitalists want you to focus on. 

FOR SEARCHING GOVERNMENT CONTRACTS, FDPS is so valuable and SAM, FDPS' little brother and now apparent replacement, so inferior. FDPS' predecessor was also indispensable to covering what the Pentagon was buying back when I worked at WIRED. Joseph Cox reports on what the administration, of course, has done to FDPS. 

I MIGHT BE on Chris Hayes' MS NOW show tonight at 8pm ET talking about Iran. Or I might not! We have the unfortunate luck of publishing this edition before they finalize my booking. You should probably watch to find out.