AI-Wiki

Department of War AI Adoption, a Superintelligence Compact Against Regulatory Capture, and Procurement Reform for Startups (Emil Michael with Jason Miller)

12 min read

Department of War AI Adoption, a Superintelligence Compact Against Regulatory Capture, and Procurement Reform for Startups (Emil Michael with Jason Miller)

▶ Watch on YouTube · YC Root Access · 26:34

The defense industrial base of the future will be built by small, fast-moving teams working on some of the hardest problems in national security.

Live from Washington, D.C., Y Combinator brings together founders, senior government officials, and military leaders for a day of conversations about rebuilding America’s defense industrial base, working with the government, and closing critical capability gaps across autonomy, munitions, space, secure communications, manufacturing, and more.

— Channel description, YC Root Access (event description; this page covers one chapter)

The day’s first conversation with a government official at YC Root Access’s Washington livestream (see the event page). Emil Michael is introduced by YC’s head of public policy, Luther Lowe, as “the under secretary of war for research and engineering at the department and the department’s chief technology officer, leading its research, development and prototyping efforts”, formerly Uber’s chief business officer. Jason Miller, who asks the questions, is introduced as “a longtime senior adviser to President Donald Trump and a veteran political strategist and communications adviser” with senior roles in Trump’s campaigns and transitions. An audience member later addresses him as “special adviser Miller”, and the on-screen caption reads “Special Advisor to the P…” (cut off). He speaks as the interviewer only; all claims below are Michael’s unless marked.

Most of the half hour is defense procurement: drones, hypersonics, directed energy, supply chains. It earns its own page for three AI passages. Michael gives a head count of AI users inside the department (under 50,000 of 3.5 million when the current leadership arrived, about 1.7 million now) and says it had been “single-threaded on one AI provider”. He gives a first-hand account of a White House compact on superintelligence, signed by AI-industry leaders the day before, which he says should keep AI open to competitors and to open source. And he describes the procurement barriers that keep new entrants out, and what the department has changed. That last part is adoption seen from the buyer’s procurement rules, which the wiki rarely covers.

TL;DR

  • A compact on superintelligence, signed the day before. Michael describes a White House meeting of “Jensen to Elon to OpenAI’s leadership, Anthropic’s leadership, Google” that agreed “on a framework for how we develop these technologies for the benefit of the American people and do it safely without sort of an overbearing regulatory scheme” (27:20). He expects people to “look back on this moment where they signed this White House compact for the safe development of superintelligence as a watershed moment” (27:50); the industry agreed to “auditing and independent evaluations and those things” (28:01).
  • Against regulatory capture, for open source. What he hopes the compact does for founders: “ensure that we’re not headed for some notion of regulatory capture by one or two companies in the AI, in the SI world. We are headed for a world where we’re going to have a lot of competitors. We’re going to promote open source” (28:26). Labs that talk about risks should “internalize those externalities” and spend enough on “the safety part of their model development process and model deployment process” (28:43). “What you never want is one or two companies sort of driving regulation that keeps the small companies out” (28:57).
  • From under 50,000 to about 1.7 million AI users. On arrival he found “an allergy to business, an allergy to AI and technologies” (33:50): “we had less than 50,000 people out of a department of three and a half million people using AI when the rest of the world was using ChatGPT on your phone” (33:54). “Today we have about 1.7 million out of our three and a half million. So just bringing the basic technologies to make people more efficient was totally ignored” (34:07).
  • One AI provider. In the same list of inherited failings: “We were single-threaded on one AI provider” (34:18). He does not name the provider or say how many the department uses now.
  • Procurement written for incumbents, rewritten as problem statements. The old route was a “300-page RFP, and that RFP was designed for a lawyer or a lobbyist and a prime contractor” (39:17). Under “requirements reform” the department now writes “a one or two page description of the problem I’m trying to solve. You come back to me with solutions” (39:44), and tries to shorten a cycle that could run “two years” “into 90 days” (40:18).
  • Competitions with objective metrics. In the Drone Dominance programme the department takes “the top 20. We put you in a flyoff situation. We give you objective metrics. We measure you against other companies. And at the end, the top three get some money” (37:30), in successive rounds he calls “gauntlets”, with “half a billion dollars of orders” at the end (38:01). He wants more of this “in every category”.

Why this matters to the wiki

1. A head count of AI users inside a government department. The corpus’s adoption figures are mostly organisation-level, such as the AI Index’s 88% of organisations, or staged, such as MIT CISR’s four maturity stages. This is a user count inside one very large public organisation, the first such figure for a government department in the wiki: from under 50,000 to about 1.7 million of 3.5 million people, or roughly 1.4% to 49% (the percentages are this page’s arithmetic, not Michael’s). The only use he names is “bringing the basic technologies to make people more efficient”. Frey’s question, “adoption for what?”, recorded on enterprise-ai-adoption, is the question the talk leaves open: it does not say what the 1.7 million use AI for, or how “using AI” was counted.

2. Dependence on one provider, named as a failing. “Single-threaded on one AI provider” is the public-sector version of a dependency the corpus discusses for firms. Ng advises firms to “preserve optionality … I don’t let myself be locked into any one of them”, and he and Michael both see a few leading companies shaping AI rules as a risk, with open models as the answer. ai-sovereignty treats this kind of dependency layer by layer, mostly for states buying from foreign suppliers; here it is a state department dependent on one domestic supplier.

3. The compact, from someone who was in the room. Michael’s account adds a government voice to two concept pages. On open-source-ai, the “concentration of power” argument has so far come from Delangue, who cites an AI company’s leverage over the Department of War, and from Ng’s complaint about lobbying. Here a sitting official in that department says the administration wants “a lot of competitors”, will “promote open source”, and does not want “one or two companies” driving regulation. On responsible-ai, the page holds two ways of testing models: the state tests, or the lab tests itself. The compact as Michael describes it is a third arrangement: the government convenes, the leading firms commit to “auditing and independent evaluations” and to safety spending, and he stresses the absence of “an overbearing regulatory scheme”. Reuters’ China correspondents describe the US as “very focused on trying to reach AGI” while China pushes diffusion; the conversation here frames the contest with China the same way, as “the superintelligence race” that “we decided we’re going to win”.

4. The buyer’s side of procurement. Procurement rules rarely appear in the wiki’s adoption sources. Here a very large buyer describes its own barriers: “rules, qualification standards … written 20 years ago”, 300-page RFPs suited to “a lawyer or a lobbyist and a prime contractor”, a $50 billion programme whose contractor “got one meeting every four years”. His remedies are problem statements instead of specifications, 90-day feedback, competitions scored on objective metrics with orders at the end, and one office (his) for the department’s innovation funding routes: SBIR/STTR small grants, DIU, APFIT and the Office of Strategic Capital, previously “all over the department”. Collison, recorded on enterprise-ai-adoption, reports enterprises now buying from unvalidated startups because “the risk of the status quo is actually extremely high”; this talk describes a public buyer changing its rules to make that possible. Michael’s examples are drones and hypersonics, not AI.

Dynamic capabilities. contextual/internal-barriers: the department he describes inheriting has “an allergy to AI and technologies”, 20-year-old qualification rules, 300-page RFPs and four-yearly programme reviews, which is rigid planning and change resistance in W&W’s terms. digital-transforming/improving-digital-maturity: the move from under 50,000 to about 1.7 million AI users is a workforce-wide rollout of basic AI tools, the “leveraging digital knowledge inside the firm” activity at the scale of 3.5 million people. digital-transforming/navigating-innovation-ecosystems: the talk is about how a very large buyer deals with many outside partners: naming single-provider dependence as a failing, backing a compact meant to keep AI open to competitors, and replacing incumbent-shaped procurement with open competitions, problem statements and one front door for SBIR/STTR, DIU, APFIT and the Office of Strategic Capital.

Linked entities and concepts

Debates and supersession

  • The adoption figures are self-reported from the stage. The under-50,000 and about-1.7-million figures come from the official responsible for the change, with no source, date or definition of “using AI” (an account, a licence, monthly activity). The start point is “when we got in there”, the current administration. The wiki has not checked them against any department publication.
  • A user count, not a use. Read against Frey (“if people adopt it for email, it’s not going to drive growth in any meaningful way, but it may look good in the sense that the adoption rate is high”) and MIT CISR (the financial inflection comes between pilots at Stage 2 and scaled AI ways of working at Stage 3), the figure measures breadth only. The talk offers nothing on depth or results.
  • The compact is known here only through Michael. The list of signatories, the commitments (audits, independent evaluations, safety spending) and the claim that it avoids “an overbearing regulatory scheme” are his paraphrase on the day after. The text was not fetched. Open question: Ng and Delangue locate the capture risk in the influence of a few leading firms, and by Michael’s account those firms signed the compact. Whether it constrains them or is shaped by them cannot be judged from one description.
  • Two pictures of US AI governance in one quarter. The corpus records the US government restricting access to an Anthropic model by nationality, a state instrument the vendor did not choose (see Anthropic, the Mythos sequence, via Sundararajan and Frey). Michael’s account of the compact stresses light regulation and open source. They concern different instruments, and both rest on single accounts.
  • Political framing. Michael is a sitting official describing his own administration’s record, interviewed by a political adviser to the President. “Sheer neglect” and “an allergy to AI” describe the previous administration and are not evidenced in the talk.
  • “Five times” is not the AI Index ratio. Michael says US “capital markets are still five times as robust as theirs” (32:34). The AI Index puts US private AI investment at about 23 times China’s in 2025. The measures differ (capital markets and venture capital in general, against private AI investment), so the two figures do not contradict each other, but they should not be cited interchangeably.
  • China’s inventiveness. Michael’s claim that China copies until “they’ve run out of stuff to copy, they start inventing” (30:07) comes from his Uber years (2014–16) and is about manufacturing and technology generally, not AI models. The corpus’s AI-specific dispute over distillation is on open-source-ai.

What was actually ingested

The chapter 0:26:19–0:53:37 of the livestream: the full conversation and two audience questions (722 ASR segments; speech ends at about 52:53, followed by about 45 seconds without captions). Speaker roles come from Luther Lowe’s introduction at 25:14–26:04, which falls at the end of the previous chapter, and from an audience question at 51:31; the cut-off on-screen caption for Miller was read from a single frame at about 32:57 supplied by the coordinator. The transcript is raw auto-generated captions. Quotes were checked against it at the cited timestamps and corrected only for obvious ASR errors: “independent valuations” → evaluations, “cyber sitter” → SBIR/STTR, “appetit” → APFIT, “Castellian” → Castelion, “Androl” → Anduril, “super intelligence” → superintelligence, and doubled words. The interviewer’s and Michael’s use of “SI” for superintelligence is kept. The two audience questions (brushless motors and a drone-components deal by the Office of Strategic Capital with a company the ASR renders as “PDX”; Ukrainian companies in Drone Dominance) and the non-AI passages (directed energy, Golden Dome, hypersonics, critical minerals, five new primes, clearances for foreign-born founders) are summarised only where they bear on the AI content. Not checked: the compact’s text and signatories, the department’s adoption figures, the Drone Dominance programme’s terms and order volume, and the named companies’ contracts. No stills were extracted (a fireside with no slides).

Relationships

Part of

Supports

  • China, Open Source & AI Competitiveness I Andrew Ng — Both treat a few leading AI companies shaping the rules as a risk and back open models against it. Ng says lobbying by 'a handful of businesses' against open models has hampered American AI and advises firms not to be 'locked into any one' model provider; in this talk a US official says the White House superintelligence compact should prevent 'regulatory capture by one or two companies' and 'promote open source', and that his department had been 'single-threaded on one AI provider'.
  • Hugging Face's CEO on why companies are done renting their AI | Equity Podcast — Both name concentration of AI in one or two companies as the risk and open source as the counterweight. Delangue, CEO of the largest open-model platform, cites an AI company's leverage over the Department of War as an example of that power; in this talk the department's research chief says 'what you never want is one or two companies sort of driving regulation that keeps the small companies out'.
  • Inside China's AI race | On Assignment — Both describe the US side of the US-China AI competition as a race for the frontier. Reuters' China correspondents say the US is 'very focused on trying to reach AGI' while Beijing pushes diffusion, and that US safety testing sits inside the labs; in this talk a US official calls it 'the superintelligence race' and describes a White House compact in which leading labs commit to 'auditing and independent evaluations' without 'an overbearing regulatory scheme'.

Graph View

Backlinks