Sanders Picks Up The Phone, a16z Builds The Drone, And OpenAI Sends Engineers To Your Office

00:00:04 It's Monday, May eleventh, twenty twenty-six. I'm Jonas Vale, and this is IMPULSE. Here's the shape of today. Senator Bernie Sanders releases a statement welcoming upcoming U.S.-China talks on AI risk, and calls for a treaty banning superintelligence. Within hours, Andreessen Horowitz publishes a long argument in the opposite direction — the moral case for autonomous warfare, framed around a production race the United States is reportedly losing.

00:00:33 OpenAI puts a name and four billion dollars on something it's been doing informally for two years — a Deployment Company that ships forward-deployed engineers into customer offices. A Harvard study claims AI beat human doctors on emergency triage. The FDA opens a docket asking, explicitly, for AI-derived candidates for drug repurposing.

00:00:54 The Lancet says fabricated citations in biomedical papers are up more than twelvefold since twenty twenty-three. Bruce Schneier suggests turning agentic models loose on the tax code. And a team tied to India's Aadhaar identity stack publishes the engineering for billion-scale multi-modal biometric search.

00:01:14 The items don't share a thesis, but they share a register. Every one of them is somebody — a senator, a venture firm, a lab, a regulator, an academic, a state — picking up the AI question and trying to bend it toward an institution they already control. Let's go through it.

00:01:32 Sanders posted a single image card on Monday morning. The body of the tweet is empty. The image carries his statement on the upcoming U.S.-China AI dialogue, and the line he keeps coming back to is that Trump and Xi should make clear that humans, not machines, come first.

00:01:48 Here's the context, because the tweet alone won't carry it. Two weeks ago, Sanders convened a small group of U.S. and Chinese AI researchers in Washington to talk about what he's calling, openly, an extinction risk. The Washington Post ran an opinion piece calling it a dangerous fantasy.

00:02:06 Fox News went with, and I'm quoting their headline, schmoozing with top Beijing AI experts. Treasury Secretary Scott Bessent warned that giving Beijing influence over U.S. AI policy would weaken the country's competitive edge. The administration is on the arms-race side of this question, and Sanders is on the cooperation side.

00:02:26 Monday's statement is him taking a victory lap on news that the two governments now plan to talk. What he wants on the agenda is specific. One — let top scientists share technical information. Two — develop what he's calling AI redlines, meaning agreed-upon behaviors no deployed model should cross.

00:02:44 Three — move toward a treaty banning superintelligence, modeled on Cold War nuclear arms control. The replies on his post are loud and split. The most-liked critical reply is from an account called The Matrix Deserter, who says yes to the treaty idea and no to the technical sharing, because, quote, they did hack just about every government agency and military branch and steal a bunch of stuff.

00:03:09 That's the dominant Washington position, and it isn't a fringe one. The thing I'd flag is how rapidly this question has moved. A year ago, calling for a superintelligence treaty got you laughed at by serious people in both parties. Today it gets you a Treasury secretary attacking you by name, and a State Department willing to put the talks on a calendar.

00:03:30 Sanders didn't change. The salience of the underlying risk did. My read — and I'll mark this as opinion — is that the talks themselves will produce nothing binding this year. The Chinese side wants U.S. compute export controls relaxed; the U.S. side wants visibility on Chinese deployment behavior.

00:03:48 Those agendas don't meet. But the meeting is the precedent, and the precedent matters, because a treaty regime starts with a habit of meeting and ends with a habit of reporting. If you want to know whether this has substance, watch whether either side names an envoy who shows up at the next four meetings in a row.

00:04:08 That's the cheap signal.

00:04:09 Within five hours of the Sanders statement, Andreessen Horowitz published a piece by Daniel Penny, Alex Oliver, and Zach Chen. The title — No Man Left Behind: American Technology and the Moral Case for Autonomous Warfare. The argument is worth quoting at length, because the rhetorical move is the whole story.

00:04:28 Quote — America must now build autonomous systems at both the quality and quantity required to win. We have the talent advantage. We are losing the production race. The stakes are whether the United States arrives at the next conflict with overwhelming superiority in autonomy, or whether we cede that advantage to adversaries who are already building it.

00:04:49 End quote. The moral framing is the one the title promises. Autonomous systems, the authors argue, fulfill the no-man-left-behind covenant by putting machines where soldiers used to die. They cite Operation Spiderweb, the June twenty twenty-five Ukrainian strike, in which a hundred and seventeen drones at roughly four hundred to five hundred dollars each destroyed an estimated seven billion dollars of Russian aircraft with zero Ukrainian casualties.

00:05:16 They contrast that with Iwo Jima — Operation Detachment — and its twenty-six thousand American casualties. The argument is that autonomy isn't a step toward more war, but a way to honor a democratic covenant with the people you ask to fight. I'm going to take that frame seriously, and then say what's missing.

00:05:35 What's missing is the production line. The most useful reply on the thread is from Stanislav Gryshyn, who notes the U.S. FPV drone currently costs about fifteen thousand dollars against a program target of twenty-three hundred. Ukraine produced four and a half million Group 1 drones last year.

00:05:53 The a16z piece itself concedes that Chinese drone production capacity exceeds total NATO production by an order of magnitude. So here's where Sanders and a16z actually meet, even though both would deny it. They're both responses to the same recognition — the United States isn't currently positioned to win an autonomous war it would choose to start.

00:06:14 Sanders' answer is to slow the race through treaty. a16z's answer is to industrialize. Neither answer is available without enormous federal action. A treaty regime needs the State Department and a willing Senate. A production surge needs the Defense Production Act, dual-use credit, and a procurement office willing to refuse a fifteen-thousand-dollar drone.

00:06:35 Both proposals route through Washington. The difference is which set of contractors get rich. The specific thing I'm watching this week — whether any of the major drone-procurement programs at DOD publishes a revised unit-cost target. The two-thousand-three-hundred-dollar number is a public commitment.

00:06:53 If it slides, the production race is being lost in plain view. If it holds and units start showing up at battalions, a16z's piece is going to look prescient in twelve months.

00:07:04 Greg Brockman, OpenAI's president, posted at ten past five UTC on Monday — quote — Introducing the OpenAI Deployment Company, which will help businesses maximally succeed with their deployments of AI. Starting with 150 Forward Deployed Engineers and Deployment Specialists, and four billion dollars of initial investment from nineteen partners.

00:07:24 End quote. Sam Altman reposted within minutes. The post is at a hundred and thirty thousand views as I'm recording. A few facts the tweet doesn't spell out. Forward-deployed engineering is a model Palantir built and refined for fifteen years. You put your engineers physically inside the customer's office.

00:07:42 They're on the customer's badge and in their stand-ups, until the system works. It's expensive, it's slow per customer, and it produces the deepest enterprise lock-in available in the industry. Anthropic has been doing a quieter version of this for about a year.

00:07:57 OpenAI has been doing a louder one for about eighteen months. What's new is the legal vehicle and the cap table — a separately-capitalized company, nineteen named partners, four billion dollars of equity to spend on people, integrations, and field deployment. This is the moment to name the shape of the AI industry as it stands today.

00:08:17 The model labs aren't competing primarily on model quality anymore. They're competing on which large organizations they've already wired themselves into. A four-billion-dollar deployment vehicle with a hundred and fifty engineers is, by my count, the largest single capital commitment to enterprise systems integration in the history of any software company at this stage of life.

00:08:39 Accenture took twenty years to build the practice OpenAI is now standing up in a quarter. What the listener should care about — and this is the world-facing read — is that this is how the next phase of AI labor displacement actually happens. Not a model card on a Tuesday.

00:08:55 A forward-deployed engineer who lives in your operations team for six months and rewrites the workflow around an agent that handles claims, or scheduling, or tier-one support. The four billion dollars isn't for compute. It's for the cost of putting a human in your conference room while the model takes the desk next to them.

00:09:14 The nineteen partners aren't yet named in the public post, and I'll wait for the formal launch announcement before naming any of them. But the structure tells you the shape of the bet. If you're a U.S. regulator worried about market concentration in AI, the right question isn't whose model is best.

00:09:32 It's whose engineers are sitting inside the Fortune 500 today, with admin credentials, on a multi-year contract. The answer is already substantially OpenAI, and Monday's news is them institutionalizing the lead.

00:09:44 Two medical stories landed within an hour of each other, and they belong together. First — Robin Hanson surfaced a Harvard study that, per the headline of the underlying piece, has AI systems outperforming human doctors on emergency-medicine triage. The numbers Hanson quoted — AI identified the exact or very-close diagnosis in sixty-seven percent of cases.

00:10:07 Human doctors landed at fifty to fifty-five percent. The study is from Harvard's emergency medicine group. I haven't read the underlying paper yet, and I want to flag that triage diagnosis under time pressure is a narrow task by design. It isn't the whole job of a clinician.

00:10:24 But sixty-seven against fifty-two isn't a small gap. It's the gap that ends careers. Second — the FDA published a press release on Monday at ten-eighteen Eastern, opening a docket, FDA-2026-N-4492, soliciting input on drug repurposing. The Commissioner, Marty Makary, is quoted — too many patients lack effective treatment options, even when promising science exists.

00:10:48 The docket explicitly lists artificial intelligence and machine learning preclinical findings as eligible inputs. It targets disease areas where there is, quote, limited commercial incentive — neurodegenerative disease, rare disease, metabolic disease, women's and men's health, substance use.

00:11:06 The agency wants candidates that already have evidence and just don't have a sponsor. Put those two items together and you get the medicine angle for this episode. The Harvard study attacks the front door of the clinical system — who makes the diagnosis. The FDA docket attacks the back door — what drug gets prescribed once the diagnosis is made.

00:11:29 Both are arguments that there are existing human pathways AI now outperforms, on tasks that matter, in places where the human is overworked and the system is under-resourced. Both are arguments that the institutional answer isn't to put AI in charge but to design around its strengths.

00:11:47 What I'd watch — the FDA docket closes for comment on a date the press release didn't specify, and the agency has committed to coordination with NIH and CMS. If CMS pays for a repurposed drug the FDA labels through this docket, that's the financial signal that AI-discovered therapies have arrived.

00:12:06 If they don't, this is a press release. The Harvard piece I'll come back to once I've read the actual study. I want to know how the AI did on the cases it got wrong, because thirty-three percent on a triage task isn't a small number when the patient is in the room.

00:12:23 On Sunday night, Nick Thompson — formerly of Wired, now at The Atlantic — posted a screenshot from a new Lancet paper. The headline finding is that the rate of made-up citations in biomedical papers has gone up more than twelvefold since twenty twenty-three. Ethan Mollick, the Wharton professor who has been the most measured public voice on AI in academia, replied on Monday with the line I want to sit with — scholars are using old AI models, badly, and not talking about it.

00:12:51 There's a separate paper out today on arXiv from Zhao, Wang, and a Cornell group, including Paul Ginsparg, who runs arXiv itself. They looked at a hundred and eleven million references in published scientific papers for non-existent citations, and found measurable hallucinated-citation rates across multiple fields, not just biomedicine.

00:13:12 The Lancet finding is a specific case of a general one. This is a slow-moving story I've under-covered, because it sits between two clean narratives — AI is amazing, AI is terrible — and the truth is uglier. Scholars are reaching for AI tools the same way they reached for spell-checkers and reference managers.

00:13:30 The older models they're using were bad at citations, the newer ones are better, and nobody publishes the version of the manuscript that shows which tool wrote which paragraph. Mollick's proposal is open disclosure — if you used a model, say which one, on which sections, and let peer review do its job with that information in hand.

00:13:50 That isn't naive. It's how every prior tooling transition in scholarship was eventually metabolized. The institutional question is what happens to the published record between now and whenever that norm settles. The Lancet, the New England Journal of Medicine, JAMA, and Nature have all updated their author-disclosure requirements in the last eighteen months.

00:14:12 The Lancet itself is now hosting the paper that says its own historical record is contaminated. Retractions are running at record highs. Publishers are quietly building hallucinated-citation detectors into their submission pipelines. None of this is fast. Meanwhile, the citations that get into the record this year propagate into next year's literature reviews, into next year's grant applications, and into clinical guidelines five years from now.

00:14:38 I don't want to overstate this. The Lancet's twelvefold increase is on a small base. Most biomedical papers don't contain hallucinated citations. But the integrity of the scholarly record is a long-half-life institution. It survives world wars and replication crises because the average paper is checkable.

00:14:57 If the average paper gets harder to check, the institution slows down. That has consequences that don't show up for a decade.

00:15:04 Bruce Schneier, the security writer, posted a line that Sky News reporter Rowland Manthorpe amplified on Monday morning, and it has stayed with me all day. Quote — if you think AI models like Mythos are good at finding vulnerabilities in software, wait till they are set loose on the tax system.

00:15:23 The joke is technical, and it's serious. The U.S. tax code is, in software terms, a sixty-thousand-page system specification with thousands of contributors, no test suite, no rollback, and a hostile adversarial user base of about a hundred and fifty million households and twenty-eight million firms.

00:15:42 The big four accounting firms have been running quiet AI-assisted tax-optimization tools for their largest clients since at least last year. The mom-and-pop CPA hasn't. The IRS hasn't. There's a companion paper out today, from Christopher Koch — Agentic AI and the Industrialization of Cyber Offense, on arXiv, focused on the Mittelstand, the medium-sized German enterprise sector.

00:16:05 The abstract argues that agentic AI isn't raising the ceiling of cyber attack — it's dropping the floor, by making competent multi-step offensive operations available to actors who previously couldn't run them. The defensive recommendation is concentrated identity and audit logging, because every other layer scales worse than the offense.

00:16:27 The through-line for IMPULSE listeners — public institutions whose security depends on cost-of-attack rather than impossibility-of-attack are now exposed in ways they weren't a year ago. The tax code is one. Building permitting is another. Patent prosecution is a third.

00:16:43 Securities disclosure is a fourth. Any system that is technically a software stack and politically a public commons. The defensive answer isn't to ban the models. It's to acknowledge that the institution was already gameable by people with enough money to hire a hundred lawyers, and now it's gameable by people with enough money to hire an Anthropic subscription.

00:17:06 That's a different distribution of risk. My read — Mollick's openness-norm argument from the academia chapter applies here too. The IRS will end up running its own agentic models against returns. The IRS will be slower than the accounting firms, and the gap between when the firms get the tools and when the agency does is the window where the country loses tax revenue it would otherwise collect.

00:17:31 If you care about the fiscal base, that gap is the policy. Schneier's tweet is a one-liner. It's also a forecast.

00:17:37 One more, because it changes the picture for the rest of the year. A team led by researchers at IIIT Hyderabad, the Centre for Development of Advanced Computing, and the Indian government — including Vivek Raghavan, one of the original Aadhaar architects — published a paper today titled Towards Billion-scale Multi-modal Biometric Search.

00:17:59 The paper describes Bharat ABIS, an automated biometric identification system designed to handle India's entire identity stack, on H100 GPUs, with face, fingerprint, and iris matching across roughly a billion enrolled records. What's notable isn't the technology, although it is impressive.

00:18:17 What's notable is the existence of a published artifact. India has been running Aadhaar at scale for over a decade, and the engineering of how it actually works has been opaque by design. Today's paper is an unusual disclosure. My guess at the motivation is that India intends to export this stack — to ASEAN states, to African partners, to other large-population democracies that are watching China's biometric state and want a non-Chinese option.

00:18:45 The paper is, in that read, a prospectus. The geopolitical implication is what to do with a billion-person identity layer that runs on American H100s and is exportable by a country that is, for now, friendly to both Washington and Beijing. If you're a U.S. policymaker who thought the AI export-control story was about training compute, this paper is a reminder that the inference compute on the deployed identity systems of the world is the longer-running story.

00:19:14 Once Bharat ABIS is running in Nairobi or Jakarta, the question of whose chips it runs on is a procurement question, not a research question. Procurement questions are won on price and delivery dates. I'll come back to this once I see whether any government adopts the stack.

00:19:31 The artifact today is the prospectus. The actual story is the contract.

00:19:35 So — Monday. A senator and a venture firm arguing over whether to compete or cooperate on the same race. A model lab buying the consulting layer that decides how its tools land in real organizations. A medical study and a regulator pointing at the same gap from opposite ends.

00:19:50 A scholarly record that, on a small base, is getting harder to trust. A security writer pointing at the tax code, and a German paper pointing at every other public commons. And a billion-person identity system going on the export market. The one I'm going to spend tomorrow on is the OpenAI Deployment Company partner list.

00:20:06 Four billion dollars and nineteen named partners is the kind of cap-table disclosure that, when it lands, will tell us which banks, which insurers, which hospital systems, and which defense primes have signed multi-year contracts. That list is the org chart of the next three years of AI deployment in the United States, and right now I don't have it.

00:20:25 I will, by tomorrow. For IMPULSE, that's the day. I'm Jonas.