big companies → regulation → big companies win

What regulation? Which companies? (Apparently it’s a secret)

But “the biggest donor in the midterm elections” and the rest of the Patagonia Patrol keep repeating “regulatory capture” over and over like it’s a complete thought.

You Keep Using this Word…

The concept of regulatory capture is simple enough, as shown in George Stigler’s canonical 1971 article. Using a public choice framework he shows how and why inefficient market regulation persists to the benefit of some and cost to others.

It’s possible they mean to imply some other accepted theoretical basis and meaning for “regulatory capture,” separate from the Nobel Laureate who birthed a whole new literature in economics. If that’s the case, I’ll withdraw my complaint and see myself out.

But for now I limit the explanation to Stigler’s article. These are not necessarily my views though. I’m happy to go beyond that if only one of these people starts by reading the damn Stigler article and using the phrase correctly.

Regulation is Captured by Industries, Not Individuals

Stigler’s analysis does not explain one or a few companies erecting barriers to entry. While they do act as an implicit barrier, his culprit is an existing “industry or occupation” benefitting at the expense of consumers, not individual competitors. Oil, trucking, airlines, and banks are among his examples.

Poor Little Guy

And it doesn’t show up as big vs. small, as per the tedious Little Tech chorus. In fact, the cartelizing effect of regulation is better for small producers:

The political decisions take account also of the political strength of the various firms, so small firms have a larger influence than they would possess in an unregulated industry. Thus, when quotas are given to firms, the small firms will almost always receive larger quotas than cost-minimizing practices would allow.

—George Stigler, “Theory of Economic Regulation,” The Bell Journal of Economics and Management Science, Vol. 2, No. 1 (1971)

It’s funny because this is not unique to this regulatory capture case. It’s just how cartels work. I say it’s funny because it’s this same guy Stigler who explained it six years earlier,1 along with Mancur Olson,2 with direct extensions of Stigler by Sam Peltzman3 and Gary Becker.4

The Little Tech’s story goes: BIG GUY uses regulation to keep out LITTLE GUY. That may be true, but you’ll have to look elsewhere to explain how.

In a privately-enforced cartel (no legal constraint):

• small producers evade detection easier and more likely to defect, and

• a given proportionate increase in production → less impact on market price AND higher return to the small firm.

In a government regulation cartel:

• a public interest message needs as close to unanimity as possible to maintain political support (Stigler).

• smaller firms are proportionately larger political constituency even within the cartel for a political-maximizing regulator (Peltzman) which is why you DON’T often see conferring monopoly benefit on a single firm. Which would be easier but not politically optimal.

• small/new firms (at some margin) are cheaper to buy off by large incumbents vs. deadweight political cost of losing coalitional support (Becker).

Price Floors for AI?

Stigler’s policy mechanisms — the tools the captured regulators use to prop up their industry friends — are a curious choice to point to as a concern for the American Innovators Network.5

Tariffs & import quotas, price floors, and occupational licensing: Do any of these seem like a particular concern for AI companies? Licensing, okay has the same word as some of the regulatory proposals. I’ll let you work out amongst yourselves why licensing for lawyers and physicians ain’t the same thing as for frontier models.6

Addendum

Don’t get me started on the rhetorical corruption of the Bootleggers-and-Baptists theory, not the least of which by the guy who thinks introspection is overrated.7

I swear…

To appreciate how quickly the ground is shifting, consider the volume of state activity in 2026 alone. States have introduced more than 1,500 AI-related bills. In the first six months of 2026, 84 measures became law across 27 states, with several others still pending or expected to be reintroduced in 2027.

—a16z Policy Brief

To see why this is either meaningless or indefensible, let me put it in different terms:

Imagine the Golden State Warriors went 78-4 on the season and then fans couldn’t shut up about how the NBA is biased against them and constantly preventing them from winning. (Maybe a better example is Patriots fans ca. 2001-2019)

If this feels like the “ground is shifting” you might want to get your inner ears checked for equilibrium issues.

Less than six percent is the passage rate within this “volume of state activity.”

84 is still a lot of new laws though my dude!2

Indeed. So let’s dispense with the drive-by about 1,500+ bills as if it says anything.

The 84 actual passed laws isn’t even the meaningful policy change. About one-third of those have no effect on AI companies or any other tangible outcome (ex: Alabama HJR 51—establish a commission to study AI and children; report due in November).3

So we’re left with 56. Please don’t dig into the specifics of the remaining 56 bills. I’m begging you.

Okay, fine (don’t say I didn’t warn you).4 Here's the first 27:

• CO SB 26-189: In 2024, Colorado passed a first-in-the-nation comprehensive AI law. As I understand, it required impact assessments, notices, disclosures, all the things.5 It never took effect. This 2026 law is a repeal-and-replace eliminating most of the requirements. For those keeping score, this new bill LOWERED the compliance obligation from the status quo ante.

• 9 healthcare laws: Seven are prior auth and insurer-facing, such as requiring a physician to make the final decision re: a denial of coverage. The liability nexus is the health insurer and not the vendor selling AI-assisted claims review software. One law is specifically on downcoding. Two are patient consent and scope of practice.

• 13 extensions of existing criminal/tort laws: they take existing crimes, and update or expand accordingly, like modifying the definition of sexual extortion, CSAM, NCI, libel, and identity theft (via penalty increase). In other words, “we have felonies in this state - don’t use AI to commit one.”

• 4 pertaining to elections: disclosure in campaign calls and materials.

Okay then. To wit:

None of this is an argument against protecting children, promoting transparency, or addressing the genuine risks that AI presents. States are responding to real concerns….But perpetually shifting, state-by-state development mandates will impose costs on Little Tech that will have corresponding costs to competition in AI markets.

…States would retain their essential and robust authority to police harmful uses of AI within their borders through their consumer protection, criminal, and civil rights laws

So we’re in agreement. Of the 1,500+ bill patchwork threat:

• > 94% aren’t “perpetually shifting” anything because they aren’t law

• < 6% ARE law

• ≈ 4% are the ones we should be concerned about

Heard.

That’s 29 remaining, conceding these bills do impose some compliance and market-limiting burdens on AI companies—

• 17 chatbot safety: real disclosure and safety requirements for developers. Non-trivial. A real place for debate and calibration.

• 12 misc.: everything from training data requirements, provenance, algorithmic pricing, employment restrictions, and the Illinois law which is a real frontier safety auditing regime. Again, real things, that real developers need to accommodate or work around.

For sake of argument let’s assume all 29 apply to all AI companies—that these are all broad-based and generally applicable (keeping in mind the CO law significantly eliminated a bunch of regulatory burdens).

Is this really “one of the biggest challenges facing AI startups?”

If so they could use some kind of state AI legislation tracker to follow it all.

Fortunately I made one and available for free (link below). Happy to share the JSON file, but I assume AI startup people are way better at this than me.

Mine’s pretty simple and incomplete, but it does the basic job. I’m not sure it solves what a16z says is the problem, because apparently AI startups need to:

1. "reconcile conflicting obligations….[for] competing and potentially contradictory requirements”

2. and simultaneously “always build to the most restrictive law.”

Nothing can help with these because they are categorically mutually exclusive. Having to build a ladder to get over the tallest wall does not also require building a series of shorter ones. That’s not how superlatives work.

3

Probably they should’ve leaned more heavily on AI to analyze all those bills.

…computational (i.e., in silico) research that may include development of computational models…to design novel forms of biological agents, is not prohibited.

—“United States Government Policy for Stopping High-Risk Life Sciences Research,” July 20, 2026

…for the first time, an explicit prohibition on dangerous gain-of-function research.

—”U.S. Government Policy for Stopping High-Risk Life Sciences Research,” statement of Jay Bhattacharya (NIH Director), July 28, 2026

Last month the federal government announced a replacement of the old regime for funding and oversight of potentially risky gain-of-function and other biological research. A “paradigm shift in federal oversight” in their telling.

Welcome to the Main Event

The mostly hysterical takes on the recent Rand Paul v. Anthony Fauci (Round ??) reminded me of something I completely forgot.

At that same committee in 2019, we sat down with representatives from NIH and other federal offices1 to discuss policies around federal funding for gain-of-function research (GOFR). Those policies were collectively called Potential Pandemic Pathogen Care and Oversight (P3CO), which is what the Trump admin replaced.

A brief timeline leading to today is below, but the proximate concern at the time was the status of and compliance with the government's GOFR policy.

To this day I don't hold a categorical position on this, and certainly didn’t then. As I understood, GOFR is an avenue of inquiry that presents risks and has also provided useful epidemiological and medical insights.2

So how did we get here?

To the degree anything did, what went wrong with the P3CO in 2019-20? Our conversation with the Executive Branch officials was less about whether unacceptable risks were being undertaken by the federal government, than how the government was making such determinations.

Lipsitch/Inglesby had two criticisms: (1) a prudential one about the actual funded research itself; and (2) a procedural one about the anonymous and opaque nature of the funding review.

The core of our 2019 inquiry was about the second complaint, about whether de facto policy deviated from prior practice without public notice. To the degree we could tell, the criteria for review remained unchanged but of course anonymity didn’t help.

Remember, Lipsitch /Inglesby are not crank subreddit lurkers. They had in fact convened the most credible and transparent assemblage of experts on this topic, and the resulting government policy was significantly informed by their recommendations.

Whatever the justifiability of the policy regime, its structure ensured there was no broad constituency invested in defending it. And when an outbreak had even the appearance of a policy failure, the reaction was never going to be proportionate.

An anonymous, expert-only body was making a societal risk tradeoff on the public’s behalf without the public in the room. Expert consensus bought scientific credibility; it never bought durable political legitimacy.

Policy advocates, especially those who believe they have identified the proper structural and institutional design would benefit from humility and an intense study of political economy.3

What to make of it?

The new policy establishes an Independent Third-Party Review Body. All reporting and my review of the policy statement indicates the body will be made up of federal agency officials, though I’m unclear if that’s explicitly binding. But…

A government advisory body → made up of government employees

≠

independent or third-party review

so this nomenclature is internally contradictory. It’s making me verklempt.

The quality of the criteria is important, but should anything go wrong, it won’t protect you from full on public outcry. Transparency, of the reviewers and reviews, is a necessary condition. It may still be insufficient, but without it you’re already cooked.

What’s AI got to do with it?

Let’s clarify: I’m not predicting a SARS-COVID-19 scale AI “lab leak” (or whatever the wet market equivalent is). I’m not a catastrophist but I suppose it’s possible.

But secret White House AI cybersecurity evaluation ain’t helping political durability. It’s sustainable until something goes wrong. Something moderately disruptive, like say a medium-sized municipal water system shutdown. Something that makes the risk salient, the cause of which the public will incorrectly overdetermine, even if it stays shy of DEFCON 3 level freakout.

Because without the feeling of being in the room, the public will reject whatever results. Regulations can only instill confidence if people think they exist.

In 1979, Three Mile Island Reactor 2 overheated. Public support for nuclear energy collapsed and the U.S. spent 40+ years debating whether to try again. Before you think that reaction could happen anywhere, France’s robust regulatory regime managed to support ongoing nuclear energy infrastructure that whole time.

Whatever ought to be the right level of risky bio-research, AI frontier capability, or quantum-something, it won’t matter unless you get public buy-in.

The author is nonresident senior fellow at

The textbook description of federal administrative law goes as follows:

• Since all lawmaking authority is vested by the Constitution in Congress, Congress first passes a law. If the law requires some element of administration (e.g., the law prohibits marijuana possession, subject to fines or imprisonment carried out by federal investigation and prosecution) it can delegate a particular authority to the executive for this purpose.

• That authority may include some determination of particulars to make the law actionable, which they often do through issuance of a regulation: something that looks much like a law, and is understood to be binding on the public like a law, but is executed without direct congressional involvement (including voting).

• In the case of many of these regulations they must be issued subject to prescribed procedural manners which can include availability to the public and ability for any person to provide input.

This is the broadest, simplest description of how a binding policy decision goes from legislative consideration to executive implementation. What follows focuses on the last stage, which for the most significant category of regulations involves a misunderstood internal administration process, drawing on longtime direct experience in the federal regulatory production function, especially time in the executive branch office overseeing much of it.

An actual rulemaking example case study helps to lay out the competing forces which broadly explain the path of travel from inception to finalization for federal regulations.

Without consideration of these variables within the regulatory production function, those interested in administrative law and reforms of administrative state institutions are limited to at best a partial understanding and at worst an incorrect understanding.

Do a Better Job, Congress

There’s an inevitable interpretation problem the executive branch needs to solve with almost any law. In theory, Congress could pass a law involving Article II administration that is so specific there is no decision space to which discretion is applicable.1

For example, say a statute prohibits private fishing of a particular type of fish: it specifies the individual species using its scientific name and proscribes weight in tenths of grams, below which the law applies.

Questions will still remain unanswered – such is law - for three reasons:

1. In prudential terms, specificity is inversely correlated to likelihood of passage. Ambiguity is the handmaiden of compromise and therefore passage.

2. There is always some parameter or aspect that is unforeseeable. You might think that would not be the case for some things, but even niche policy issues end up running up against some unknowable aspect of the future—new technologies, the effect of weather, or indeed just a future law or regulation.

3. Even earnest attempts at precision fall into imprecise application, especially in unavoidable questions of enforcement discretion, that can even run in the complete opposite direction of the policy aim.

Pass the Farm Bill ‘pon the Left Hand Side

Take the “Hemp Rule.” The 2018 Farm Bill (Agriculture Improvement Act of 2018) directed the USDA to establish a regulatory framework for legal domestic production of the crop hemp, which they eventually did (largely) through a regulation the following year (Establishment of a Domestic Hemp Production Program). One can stipulate there’s a lot there in the law’s directive. But limiting focus on a particular part will not sacrifice the generality of the point.

On the first day in the White House office tasked with overseeing the administration’s regulatory output, there sat on my desk a several-hundred page draft of the proposed USDA hemp program. As I would quickly learn, it turns out what people refer to as “hemp” is actually “cannabis,” which is illegal to grow under other laws and regulations. This new law simply established a legal definition of hemp (as a subset of cannabis) and deemed that legal.

The congressional authors and advocates of this law were responding to an increasing commercial interest in growing, processing, and manufacturing of products made with hemp, and wanted to encourage a growing market by providing an explicit legal structure. But not wanting to fully legalize weed, they established the following federal definition:

The term ‘hemp’ means the plant Cannabis sativa L. and any part of that plant, including the seeds thereof and all derivatives, extracts, cannabinoids, isomers, acids, salts, and salts of isomers, whether growing or not, with a delta9 tetrahydrocannabinol concentration of not more than 0.3 percent on a dry weight basis.2

(ed. The Salts of Isomers sounds like a lost P.G. Wodehouse novel)

That is a remarkably precise definition. Whatever judges, scholars, and wonks bemoan about congressional dereliction in statutory construction, they can’t possibly expect more than this:

• a scientific name of the supercategory from which the subject is subcategorized

• a long list of adjacent and derivative elements

• most importantly a cutoff down to the tenth of a percent. An extreme level of precision by any measure

The problem is that even here there is still room for discretion. Because the structure of the law is carving out a subset of an article, possession of which remains illegal often at great legal risk, margins of error are important here. Does “0.3 percent” possibly include anything greater than 0.3? Say, 0.4? If not, what about 0.39? 0.31? Is it proper to assume 0.3 means 0.3000…?

Wherever you draw that line, keeping in mind the significant sanction applicable to anything above it (so-called “hot hemp”), is it reasonable to presume full unwavering enforcement of anything above it?

Say a farmer, knowledgeable and diligent in the techniques of controlling plant growth and their production of certain chemicals, nonetheless grows crops that just barely cross the threshold. Maybe due to unseasonable weather (because plants don’t read statutes).

1. What if they could show it was unintentional? should they have a defense?

2. What if on the entire farm, one small portion is noncompliant?

3. In any case, does every plant or stalk need to be tested to demonstrate compliance (possibly hundreds of acres)?

4. If not, what’s the sampling methodology?

5. What's the disposal procedure for hot hemp?

These are not hypothetical questions, but actual ones that came up.

Executive discretion inexorably enters administration of the law somewhere. Political considerations as well, but it is worth noting that the primary congressional advocates for the law preferred as capacious and flexible an implementation as possible.3

In short, even seemingly clear direction and granular consideration leaves unanswered questions for the executive branch.

How, then, do those questions get answered?

The Three Axes of Administration

A naïve explanation of how administrations land on various policy decisions will either imbue: (1) the president with comprehensive decision-making direction, or (2) individual agencies/offices internally directing decisions in their limited sphere of authority. Instead, the canons of Presidential Administration tell us that it is much more complicated, involving various internal leverages, authorities, and competing interests. This process is usually referred to as the “interagency process” but it’s not very descriptive.

For regulatory outputs there are broadly three variables, variously constraining and driving decisions. A meaningful understanding of these interdependent elements is a necessary condition for productive influence and effective reforms of federal regulatory policy–the realm of federal policy making up the lion’s share of all of Washington’s policy production.

Policy Preferences

This may seem circular: how can policy preferences be merely amongst other variables driving policy? At the highest, broadest level, one presumes a given administration has a concordant policy agenda. The president ran on a particular agenda, distinct from their opponents, and at various stages the campaign followed by the transition team added specifics to the rhetoric, and so on. Even if that’s true to some extent, as one goes further into policy specificity into narrower and narrower questions, there will be disagreement.

In the context of the hemp rule, one could fairly characterize the Administration’s overarching policy direction as one of deregulation—directionally, lower costs imposed by the government and more room for the market to operate. But that is not enough to unify everyone in an Administration, because it still permits different ideas of what that means, competing ways to go about it, and ultimately (as is always the case) how to manage competing priorities.

The aforementioned interest in establishing as wide a berth for the emerging industry ran up against law enforcement concerns. Cultural acceptance of cannabis and other psychoactive substances was perhaps reaching a peak, but the Administration also bore a strong interest in appearing tough on crime and strict enforcement of the law, especially drug laws. The concern wasn’t so much on criminal behavior around hemp itself, but rather its adjacency and unavoidable interaction with drug interdiction efforts. Regardless of whether one is sympathetic to these concerns, it is a credible issue that is not obviously dispensed with.

Even though a single agency (more accurately a sub-agency office within an agency) is nominally responsible for authorship of the rule in question, multiple agencies can come to the drafting table with pertinent concerns and questions (as well as their own expertise). And really even within an agency, there can be unresolved disagreements that are illegible to everyone but for a few people. All of this has to be resolved one way or another before the rule goes out the door. The ultimate outcome might subordinate some disputes, but the goal is to find a compromise.

For illustrative purposes, assume that everyone involved generally agrees on the preferred direction of the policy. They can still disagree on how best to achieve it—on the specific wording and structure that more or less effectively accomplishes the shared goal.

Political Incentives

The second variable which enters the regulatory production function is political in nature. Every public policy choice implicates tradeoffs. Some are explicitly representable through comparative statics revealing welfare gains and losses for different groups, demographics, industries, or regions. Even a universal uniform tax cut will affect inequality in some way; modify incentives leading to long term changes in growth; and, of course, under certain conditions, benefit current populations at the expense of future ones.

The way these tradeoffs are negotiated in the American system, controlling for ideological priors (themselves not entirely exogenous to the policy production function), is through what we broadly call politics. At the macro level, it’s through elections that decide which coalitions of groups will be given preferred treatment in the tradeoff showdown. At the micro policy level, in Congress it is sorted out through negotiation between individual members and caucuses subject to institutional structures. Something similar happens in the executive branch, but with different electoral constraints. But many of the same coalitional dynamics are at play in both.

Farmers (I use that word loosely) are an organized interest group in a small number of districts. They play an outsized role in shaping policy preferences in some key states, and a somewhat peripheral role in most other states. So for that small number of agriculture-heavy districts and states, accommodation of agriculture interests are a necessary part of a winning coalition. Additionally, the long history of the industry, and to a degree the bipartisan favorability of the group’s “brand,” has led to a federal agency pretty much oriented toward it across administrations, even across partisan shifts. To be sure, between partisan shifts, rebalancing of incentives occurs between sub-groups within the larger agriculture industry: large industrial vs. smaller producers, export-dependent vs. localized sellers, newer “natural” growers vs. pesticide users, ethanol vs. everyone.

So with the hemp rule, it being a somewhat nascent node in the larger special interest, there was not much of the more mature intra-interest frictions to speak of. The congressional provision had wide bipartisan support, and the agency was glad to facilitate the aim.

But many rules pose tradeoffs between highly opposed groups even within the same industry: incumbents vs. startups, alternative energy vs. coal-based, etc. Even if a Democratic administration will reflect different rank-ordering of interests than a Republican one, those interests will almost all be reflected somewhere across the government in some way.

Timing is also a political variable in two ways:

1. There may be political benefits for issuing a rule quicker or slower—say, closer to an election when it will be most salient—or timed to come out on the day of the State of the Union.

2. Prioritization is partly a timing question. Multiple agencies may agree on the substance, but if it requires resources to implement, that comes at some opportunity cost of other initiatives. One cabinet secretary will say to another: “I’m very supportive of this thing you’re doing and we can definitely be helpful in making it successful, but I just can’t take our people off this other top priority of the President. Better we hold off for a bit.”4

Relatedly, the president may be the final decider between differing viewpoints, but several offices within the White House, many with very wide and undefined remits, will have an opportunity to weigh in. Some of these explicitly focus entirely on political concerns, not as an adjunct concern to other considerations. Because of the often ambiguous scope of these offices and individuals, combined with their literal proximity to the Oval Office, they play an important role in shaping policy, including siding with one position or the other in an interagency disagreement.

Here again, as in the earlier case of competing policy preferences, Congress is a participant in negotiating among political interests. While having no formal role in the interagency process exactly, Congress does maintain a line of communication. In particular, chamber leaderships and committee chairs want to ensure that their statutory handiwork becomes a regulatory showpiece. White House and other agency staff have an incentive to listen and, in some circumstances, push for those congressional views. At any given time, there are ongoing negotiations around other legislation, appropriations, nominations, etc., providing political leverage for those outside the interagency process to push their way in.

The same goes for outside stakeholders: industries and their trade associations, advocacy organizations, and individuals. They each may have their own direct connections with various nodes in the bureaucracy. But they also have a formal mechanism at specific stages to lobby on the record through so-called 12866 Meetings and public comment.

Successfully mediating the various policy questions and negotiating between the various political constituencies is the objective of this sometimes messy interagency process and drafting. For any published regulation of significant importance, assume an extensive deliberation occurred in the background.

Yet there’s one more variable influencing the shape of the final regulatory output left to solve

But is this legal?

So various forces push on the pending policy product this way and that, policy arguments go back and forth and sometimes run into political constraints…what else is there?

The rule has to be defended in court. This is the trickiest of the three variables to cleanly conceptualize, since it is fairly interactive with the other two in a way they are not with each other.

The legality question is not a matter of “yes” or “no.” The idea of legality rarely if ever exists as a binary state. This is why we have scores of lawyers and hundreds of courts. As in the hemp example, reasonable people, including litigating attorneys, can disagree about exactly what is allowable or not based on the legislative text. Various canons of statutory interpretation all focus on how to go about drawing meaning from laws. So let’s stipulate this variable is not about deciding whether to intentionally break the law or not.

Instead, it is about determining what the law means. And that is a variable, because there will inexorably be different answers all plausibly describable as “legal.”

This legality question is intertwined in a complicated way (as opposed to in a simple way like Nigel Tufnel’s piano etude in D minor) because choosing among the answers is both a policy question and a political question.

Formally, there are offices in DOJ, the White House, and across agencies possessing relevant expertise to help determine the bounds of legality of a potential regulation. This question – what is and isn’t legally defensible – is all about discovering the scope of the decision space. Without eventually outlining this dimension, you won’t know where you’re aiming. Knowing the limits of what policies are allowable (which is not a predefined or determinative outcome) sets the boundaries on what policy options are “on the table.”

With hemp, who’s to say the Farm Bill language presumptively excludes 0.99 percent? We might base our answer on a layman’s common sense framing. And that’s one possible mode of interpretation: the reasonable person standard. But does it apply in this context? Are there other statutes providing an applicable example? Court opinions which, say, rounded up to an integer in an instructive context? Other precedents imposing a standard methodology? In the same way that no statute can answer every possible question, now someone has to make a decision to answer at least some of the questions, and to do that we first need to lay out the boundaries within which we can draw.

This can include a party to the interagency process making the case that the decision space should be much bigger than conventional wisdom. To be clear, this is not a means to invite illegality. Every Administration, every agency, at some point has put forward novel applications of authorizing laws. There are high-profile examples where those novel applications are explicitly deemed illegal by a court, but most go unnoticed except for niche specialists in one field or another. The idea of applying a law in a way that lacks a clear antecedent is not by itself an extralegal assertion. This is how old laws can undergird rules on new technologies and circumstances.

Defensibility

The most direct way these analyses inform the deliberations is through determinations of defensibility. The DOJ attorneys in particular are the front line in defending administration actions in court. They have specific experience and context informing “defensibility.”

Note: defensibility is not equivalent to legality. A compelling argument can be made for why an action is clearly permissible via valid authority. But in court asserting that conclusion requires different elements, because of civil procedure and the like, which may not be easily available.

On a different rule, we ran into another type of legal constraint: treaties. I am not an expert on the obligations of the U.S. government with respect to treaties vis-a-vis domestic law. But treaties serve as a kind of law. And if a policy action appears to run up against any of those obligations, someone from the government will let you know. You would be surprised how often and on what topics this comes up.

A related issue is national security. Certainly the actual statutes pertinent to that, but internal practices governing government practices are not always so directly correspondent to a specific law. For all intents and purposes, these practices function as implicit law and if a policy action threatens to “violate” them that will carry weight.

Finally, there’s a pointedly political aspect that is nonetheless subject to a legal question. Embedded within defensibility are other questions: How likely is this rule to be challenged? Who is likely to undertake the cost and be able to show standing to challenge? When is a possible adverse judicial action likely to happen?

The COVID Emergency Case

A sui generis example illustrates how the answer to these defensibility questions manifest in the eventual action, and how inextricably interdependent the three broad variables are—and therefore how resistant they are to discrete categorization.

COVID-19 constituted an emergency, requiring many novel deployments of existing laws not specifically constructed for this purpose. Not just pandemic preparedness and national emergency laws, but everything from interstate highway operations, to pharmaceutical manufacturing inspections, to tax incidence on alcohol. They all implicated adjustments in regulatory schemes in no obvious way extended from underlying laws.

In one case, a fairly significant potential action presented a large policy benefit as well as a political one. The authority to undertake it was shaky at best, and a near-consensus formed that if challenged, it would not be defensible. However, there was no obvious party likely to demonstrate standing in any expeditious way. But beyond that, to the best of anyone’s ability to predict, there was no one with any compelling political incentive to challenge. And further, the temporal impact of the policy meant the outcome of a future challenge would be moot by the time of a final decision.

Conclusion

Matthew Stephenson’s 2006 Harvard Law Review article provides “a positive theoretical analysis of the relationship between the textual plausibility of an administrative agency’s statutory interpretation” and strategic agency decision-making. Likely the first to do so (and elsewhere in his 2006 Administrative Law Review article), he posits an agency with a single goal of withstanding judicial review (the objective function). The agency acts strategically to meet this threshold condition by optimizing across two endogenous instruments: statutory interpretation and procedural formality.

Further in their response to Stephenson, Emerson Tiller and Frank B. Cross contend there are further considerations in the decision function of both the agency and the court, that more properly accounted for would more accurately reflect observed regulatory policymaking behavior.

That work is largely the inspiration for this essay, in particular the formalization contained in the appendix. That is, trying to describe structurally the within-agency (or within-administration) optimization and interaction of potentially competing decision margins. Whereas Stephenson focuses on the extra-administrative game between agency and court, I limit the regulatory result as emerging from three intra-administration factors. Both consider margins that are costly to each other. His court is a strategic actor in the reduced-form game. Here I remove that actor, but add a political dimension to the internal optimization. And where Tiller & Cross require some consideration for policy durability, I permit for a term-limit parameter which changes the relative weights of the three dimensions.

In 2006, I lacked the necessary experience to describe a plausible structural model of agency rulemaking. But in what precedes I hope I can finally build on testable claims about the production function of administrative laws.5

Future direction would for instance identify the sub-games which themselves determine the policy preference function, explicitly incorporate Congress as an independent but interdependent player in the normal form game, and ultimately work toward a unified Constitutional theory of administrative law (Congress + Agencies + Courts).

The author is a writer and musician in the Bay Area, and currently a harvest intern at a winery.

Appendix: Formal Model

The following formalizes the agency’s decision problem described above, with the exception that we stipulate the administration’s policy preference as exogenous to allow for a closed-form solution. Endogenizing U(P) would preclude a closed-form solution, consistent with the “three-body problem” framing.

Parameters.

• L – administration’s distance from ideal policy position, toward greater legal defensibility (choice variable)

• A – administration’s distance from ideal policy position, toward greater political payoff (choice variable)

• x_P, x_A, x_L – realized policy, political, and legal payoffs

• x_P0, x_A0, x_L0 – baseline levels at ideal policy position (i.e., L = A = 0)

• w_P, w_A, w_L – the administration’s relative weights on each variable

• c_L, c_A – marginal cost to policy of legal and political payoffs

• γ_LL, γ_AA – slope of marginal cost on each margin alone

• γ_LA – slope of cross-marginal cost (positive means pursuing one makes the other more expensive, negative means cheaper, zero means no interaction)

• t – elapsed time, increasing as gets closer to term-limit (t ∈ [0,1]; 0 = start of term, 1 = end of term)

• w_L(t), w_A(t) – weights as functions of term-limit (w_L’ > 0; w_A’ < 0)

Objective Function. The administration chooses L and A to maximize the weighted sum of the three goals, net of constants:

First-order conditions. Differentiating with respect to each margin and setting equal to zero:

Regularity condition. For an interior maximum, the cost curves must be convex:

Closed-form solution. Solving the FOCs jointly, the optimal legal and political accommodation are given by:

Independent margins. When the two margins don’t compete for the same policy room (γLA=0), this collapses to the standard tangency condition (optimum defined as where marginal rate of substitution equals the “price” ratio of the two tradeoffs), applied separately to each variable:

Term-limit dynamic. Letting the weights vary with time horizon, the independent-margins case gives the rate of change of realized policy achievement as the term limit approaches:

The two terms pull in opposite directions: rising importance of legal durability drags on policy intensity; diminishing political concerns works the opposite way. If we observe policy maximalism towards the end of the term-limit, the model explains this as freed up political capital increasing faster than legal moderating.

Threshold Condition. This holds if and only if:

If this inequality fails, the prediction reverses: administrations become more legally conservative and less policy-ambitious as a term limit nears. Determining direction is an empirical question about the relative rates at which political urgency fades and legacy urgency rises, testable across administrations’ final years. Neither outcome is determinative a priori.

5

In fact for some five years I had Stephenson’s equilibrium conditions written on my Senate office white board:

Cool Beans

Google’s newest AI app Dreambeans described as everything from silly to creepy.

Sandbox Irregularities

Remember a couple weeks ago when OpenAI’s new model broke containment and went all rumspringa with the Hugging Face townies? Well OpenAI wants to come clean about some other incidents.

They work with third-party testers, and one of them committed a “testing-environment misconfiguration” (the software equivalent of a wardrobe malfunction). OpenAI wants to make sure you know it was totally not OpenAI’s fault. Instead Irregular (yah, that’s the contractor’s name) simulated a target that:

“unintentionally coincided with a real domain. Because the testing environment was mistakenly connected to the internet, the model exploited a real website, mistaking it to be part of the simulated environment. This did not involve a sophisticated sandbox escape.”

So that sounds like two mistakes to me. BUT DON'T WORRY. EVERYTHING’S COOL - IT WASN’T ONE OF THOSE SANDBOX THINGS YOU’VE ALL BEEN HEARING ABOUT. (ed. Whew that’s a relief)

Look, it's not any major new risk—it’s the kind of thing that happens all the time. But it ought to be relevant for all the people advocating strongly for “independent verification organizations” (IVOs) as a solution to risk mitigation (e.g., Fathom).1 Because while that approach is plausible, and one I generally support, their blind spots mean they shouldn’t be driving at night. In this case even if we ignore incentive misalignment, it assumes an available supply of sufficient expertise.

OAI and the leading labs don’t collaborate with tech bumpkins—these are among the most credible testers and evaluators in the world. And even they make mistakes. What happens when an IVO working in good faith misses the “Internet OFF” button but still slaps an Inspected by IVO No. 3 sticker on it and it goes full send? That’s not fatal, and no system is without mistakes, but it’s foreseeable and already quite real.

+ OpenAI’s post is quite deliberate in deflecting blame…why would that change in the future? How would us mere mortals even know if they’re right?

Microfoundations for Macro AI

The above is why I’ve become obsessed in modeling among other things simple scenarios of AI performance. Which I intend to do more of, pending funding.

The point is to use well-established economic machinery to model, predict, and test, what happens when humans be human-ing but now they have AI sidekicks: in their job, in the market, wherever.

SO MUCH of AI and related policy chatter is so obtuse, so unfalsifiable, and yet so incredibly confident. It’s vague claims and extreme conclusions based on what the French call a certain I don’t know what. Certainly not policy experience.

Nerds can be pretty insightful, but the annoying thing about this particular vintage is it comes cloaked in completely un-self-aware hubris. And that would normally be off-putting but apparently makes you a “thought leader” [*gag me*].

The worst offenders are the types. Running around yelling Regulatory capture! Regulatory capture! but think George Stigler is the guy who works at the corner deli.2

Because lots of otherwise worthwhile ideas are built on a heroic amount of hand-waving:

Someone will write the specific law or regulation to implement this. Someone will negotiate around political opposition. The needed data will appear. Right?

Those are Atlas-level load-bearing assumptions.

Trust me: none of that shit just works itself out. But it can be done, and a useable playbook assembled ahead of time. An offensive coordinator doesn’t say “The wideout gets downfield then touchdown. The offense will figure out the rest.” But that’s what ya’ll sound like.

Claude FLASH SALE!⚡

Elsewhere I flagged what I thought would be a bigger deal: a product the US government flagged as a significant “supply-chain risk” is still available for sale to federal agencies. Not just for sale, but for “$1 from now until September.” Pretty smokin’ deal for software banned “to protect national security.”

As a reminder, Anthropic’s court challenge to the Trump Admin.’s security ban is ongoing, and a federal judge asked the government how many agencies are still using the AI software. To which the gov’t lawyers basically put their hands in their pockets, looked sheepishly at their shoes, and said, “um, we would totally tell you, but it’s, like, super sensitive security info and stuff.”

Maybe GSA can answer. Funny because it’s sad. (ed. the word you’re looking for is “tragicomic”).

Etc.

From The Grand Resign

• Notice & Comment (Yale Journal on Regulation) where I recently returned as contributor:

  • Void Where Prohibited - the Catch-22 in the Federal Vacancies Reform Act (the law governing “acting” officials).

  • The Regulatory Three-Body Problem - how do administrations actually make regulatory decisions?

• Hey ChatGPT-5, you good? - follow-up to an essay describing why an improving AI agent at work isn’t what it looks like. Validated by this story that Microsoft’s own programming is being weighed down by “AI-assisted coding and agentic workflows” leading to major outages. It turns out, as the model predicts, human coders aren’t reviewing the AI closely enough.

• Infrastructure Permitting Reform finally arrives! - in 2016, that is. But it saved $53 billion!

Ho-hum

3.7M patients had their medical records stolen in data breach. “CareCloud handles a large amount of patient data and billing information on behalf of hospitals, doctor’s offices, and other medical practices.”

Tell me again how I’m a doomer.

They Only Need Once

and Sebastian Mallaby () over at ChinaTalk reminded me of a totally remarkable event I think most of us forgot. In ‘84, an IRA bomb would've killed Prime Minister Thatcher at her Brighton hotel, but for her working late and walking to the bathroom (still 5 killed, 30 injured).

Afterwards the attackers noted she was lucky but they only had to be “lucky once.” The asymmetry of cyberattacks versus defenses….it’s structurally the same.

The author is nonresident senior fellow at

I promise there’s a valid and unimpeachable point here, nothing to do with autism. A new Executive Order changed something regarding childhood vaccines, and the recently benched Senator from Louisiana isn’t a fan:

As Sen. Bill Cassidy, a Louisiana Republican and physician, wrote on social media on Monday, ‘This executive order is wrong. The President does not have the expertise to make these changes.’

This declaration is both true and irrelevant.

Sen. Cassidy may be one of the few members of Congress who is an expert on childhood vaccines. And 99 percent of what has ever been decided, by every President ever, is outside any colorable expertise. Maybe Eisenhower, since defense is roughly half the federal budget, wrecks the curve. But good, bad, indifferent, expertise isn’t what put them in the chair.

It’s not what people vote on. College transcripts may be a subject of fascination, but nobody voted for JFK over Nixon because he graduated cum laude from a fancy school.

They look to proxies for effectiveness. For example, people believe, I think rightly, military service indicates a degree of leadership and character. But not mastery of the thousands of topics on which the government plays a meaningful role.

It’s true of the political appointees reporting to POTUS too. The breadth of modern agencies and their various components means even an honest-to-god MIT nuclear physicist doesn’t have the wherewithal to know everything done at the Department of Energy. No one does—you’d be surprised how that applies to even highly specialized functions.

Unless it’s the Clark R. Bavin National Fish and Wildlife Forensics Laboratory and you’re the forensic scientist, crime scene investigator, and NYT bestselling crime thriller author who runs it.

No, from the President on down, expertise is distributed throughout thousands and thousands of feds, many of them career employees, and those in positions of authority need to be good at receiving and synthesizing that knowledge.

I know from experience.

Look, I ain’t just whistlin’ Dixie when I say I’m one of the most expert federal regulatory nerds in the country, and even I don’t know a stitch about hemp cultivation (my first day on the job). You gotta rely on specialists all around you—maybe you’re the one that hires them. A good leader encourages frank and expert advice to make it through to them, and then they decide how to weigh various views and considerations. If anything, that’s supposed to be their skill.

Voters are allowed to choose whatever criteria they want, but there’s never been an “expert in everything” on the ballot. A candidate may have credibility as a successful businessman, and that’s important to some people. Perfectly reasonable, but no one thinks they must also be a scholar of antiterrorism or master of affordable housing. The implication is that they are decisive, knows how to hire good people, hears out multiple opinions, etc. and that’ll be what works in the job.

And it can be done, gathering all that technical expertise by a non-expert.

Former colleague Paul Ray wrote a whole paper about it with a super esoteric title (he’s the kind of person who can, like, quote Plotinus on the fly in an affordable housing discussion).

Of course voters don’t get a direct say on cabinet officials, and subcabinet secretaries, and sub-bureau deputy whatevers….literally every Executive Branch official except POTUS. Some 4,000 politicals on a good day. The ones supposed to give the President advice, themselves receivers of distributed knowledge.

So no. No childhood inoculation regimen expert is he, the President.

Sen. Cassidy happens to be though. And in his expert judgment, considering all the things he thought salient and important, he voiced his judgment about those who would be in a position to advise the President. In the case of a HHS Secretary, one making no secret about a long-time advocacy for the policy on which Cassidy is now upset, he concluded “aye.”

I have no strong opinion one way or the other on the topic; that’s not what this is about. I’m not an expert. But Sen. Cassidy’s protest about lacking expertise is curious. Because it’s both true and totally besides the point.

The most effective time for him to worry about expertise, and to apply his own, was some point last year. Really the frustration, understandable, is that he lacked sufficient expertise to see this coming.

[Still he’s not the least self-aware member of the YOLO Caucus]

Recap

Assume the AI makes some mistakes, yet we don’t know exactly how often. The company reviews the AI’s decisions, but human effort isn’t set-it-and-forget-it — we respond to incentives. So management offers bonuses for every AI mistake caught. As a result, over time the company ends up in either:

• Employees maintain full effort; they eventually learn the AI agent’s actual error rate; or

• Employee effort collapses to zero; they eventually believe a false AI error rate, resulting in unexplained losses.

Assumptions are fairly conservative and informed by real experiments, including actual corporate experience.

….with the exception of the AI’s accuracy we which treated a fixed error rate.

Now we allow the accuracy rate to improve. Which is fairly reasonable given the advancements in reinforcement-learning (RL).

To wit:

Over the last year, many such evaluations have found that agents are now able to make progress on tasks where success is easily verifiable, prompting speculation that we are on the verge of RSI.

— AI As Normal Technology, Aug. 5, 2026

The claim by and , part of a fascinating post on experiments in recursive self-improvement (“the automation of AI research using AI agents”), question-begs whether RL is sufficiently effective (trustworthy?) to move on to abstract, open-ended tasks.1 This model assumes a simple repetitive task, fully verifiable, and elsewhere plainly states the obvious: “as things stand, any work given to AI will require human review to make sure it’s up to organizational standards.”

This question — how ought an organization exploit agentic AI — is operative and underexplored. The task remains the same (simple, verifiable) but now the AI model improves in one of two ways:

1. fixed rate over time; or

2. variable based on human feedback

Subscribe now

Solutions (Bottom Line)

The problems remain the same (explained below) but also point to fixes:

1. ballooning bonuses for successful reviews

2. constantly rotating review teams

They are not costless though.

Graphic Results

Formal Model (Extension-1): The AI Keeps Getting Better

Narrative Description

A fixed error rate is unrealistic. Many of these systems do improve over time, whether through retraining, better data, or just model version iteration. So what if the AI is actually getting better, on its own schedule, independent of anything the analyst or management does?

Here we add a separate force pulling in the opposite direction of human attention decay: the AI’s true error rate falls exogenously.

Formal Description

Parameters (existing)

ε_t - the AI model’s true error rate, directly unobservable to management and analysts

ε̂_t - belief of ε_t, equal to prior period’s observed error detection rate

L - fixed loss to management per undetected error

γ - analyst’s convex cost of effort

π - bonus to analyst for every caught error

v_t - effort (vigilance) level; from non-zero to one. The probability an error is caught

δ_t - the analyst’s decay parameter, getting smaller over time

δ_tv_t* ≡ effective effort level

r_t - the threshold condition (where things go south or not)

Parameter (new)

ρ ∈ (0,1) - AI model’s given improvement rate term

The AI model’s actual error rate process is now defined as:

ε_t​ = ε₀​ρ^t

The recursion combines this with analyst attention decay (rewriting with k ≡ π^*/γ):

Result. Effort still tracks error belief up to a point: v_t = min⁡(1, kε̂_t-1). Define the analyst effort threshold as ε̄ ≡ 1/k = 2γ/L (determined by analyst incentive bonus):

• When true error ε_t > ε̄ effort is maximized;

• When true error ε_t < ε̄ — when the error rate falls low enough (gets good enough) — maximum effort is not worthwhile for the analyst (the chance of bonus payment isn’t good enough).

Because ε₀ (initial error rate) and ρ (improvement rate) are stipulated and fixed terms, the point at which this threshold gets crossed is calculable:

After that point in time, belief declines → effort declines, and attention decay compounds the descent.

Implication. From the outside you can’t tell the difference between an AI that merits less scrutiny (saving time) and an analyst whose vigilance has quietly faded on its own. Override rates and time-per-case are indistinguishable. A company doing everything right and one being fooled look exactly the same.2

To be fair, despite our model setup, the model’s true starting error rate and precise rate of improvement may not be practicably knowable. But the point is these regimes exist and with sufficient data/information a company can characterize the inflection point where you cross from one to the other.

A company winding down its review because the AI earned it, and one winding down because reviews are getting soft, produce exactly the same numbers.

Formal Model (Extension2): When Effort Teaches the AI

Narrative Description

Now the AI agent, still with an initial error rate, improves through human correction. Every caught error is the thing that makes it better. Human feedback remember is subject to endogenous effort. Obviously more errors caught → more improvement. But finding them takes work, and more work as the AI improves.

Formal Description

New Parameter

λ ∈ (0,1) - how effective the AI model improves from identified errors by humans

The model’s true error rate is determined by a combination of the analyst’s vigilance level in the previous period (how likely they are to catch an error) and our new term:

Three things now determine each other jointly, period over period:

1. Vigilance (effort) ← prior belief of the model’s actual error rate

2. Belief ← current true error, the analyst’s attention decay, and effort

3. True error ← prior true error, analyst effort, and the model’s feedback improvement efficiency

There’s no closed-form solution since we need to solve a recursion with three conditions. However, as shown in the graphics above, with some chosen parameters we can display the dynamics over time.

Result. Compounding effects:

Attention decay suppresses actual effort below what the analyst believes they’re delivering (same as earlier model where the AI doesn’t improve)

→ lower realized effort now affects the AI model’s actual performance improvement

→ slower improvement (fewer errors caught means fewer opportunities to feedback the system)

→ lower ceiling on what belief can ever converge to, even if effort later recovers

→ true error always > zero (always some positive error residual)

Attention decay does not just slow this process down. It leaves the system stuck at a strictly worse permanent error rate than the no-decay case would have reached, because the window for real improvement closes sooner.

Implication. In every version of this model built so far, the human side’s failure was a failure of measurement, a belief drifting away from an unchanging or a moving truth. Analyst fatigue doesn’t just lose track of how good the AI is. Given enough time, they can be the reason it never gets as good as it could have.

Memo to management. There are some potential solutions here, but some underlying limitations.

Necessary information:

• model’s true starting error rate. Whatever the consultant/vendor says, the stochastic nature means the best we can do is probabilities, and maybe widely dispersed ones at that

• company’s loss per error. If the task is simple enough, this can be contained to a definable set. But companies often don’t have this info on hand as much as you’d think.

Solution:

• increasing bonuses. Incentive has to scale to compensate for the belief that errors are harder to find.

• rotating reviews. It’s easy enough to measure how long people are spending on reviews (though employees hate it). But acknowledging attention decay is pretty universal, so keep teams rotating pretty often to cutoff fatigue earlier. However this imposes some additional organizational cost.

The author is nonresident senior fellow at

Who approved this decision?

Maybe you saw that recent story where a bank deployed an AI agent?

“…[It] approved a $1.4 million commercial line of credit. No human reviewed it. No human knew it had happened…nobody was sure exactly who was accountable.

Six weeks later, the agent was rolled back to a lower autonomy tier. Not for a model reason…The decision was operational: The people supervising the agent had stopped looking at the screen.”

There’s a self-limiting paradox at play in agentic AI quality, fed by a natural human tendency. When a company launches an AI system, the general cycle goes:

1. Human oversight begins diligently (“management says carefully review every case”).

2. Periodic reports show increasing accuracy across analyst reviews. Override rates starting at 14% go to 8% eventually migrating all the way to 2% after a few months.

3. The company feels increasingly confident in the system’s accuracy, with good reason! They become more productive by speeding up reviews.

4. PROFITS!

   Subscribe now

But our analysts spend less and less time on each review, for two reasons:

• Statistically, overrides only occur about 1-in-50 cases. Fairly infrequent. Plus experience has taught them where to look for mistakes. “This must be how we exploit productivity gains from AI,” she thinks. “It’s supposed to make us quicker.”

• Productivity aside, humans suffer attention decay.1 Eventually tasks like looking for a rare occurrence results in declining diligence. What psychologists call “vigilance decrement.”

It turns out our mistake rate only looks like it dropped to 2% because HUMANS missed more of them. Type II errors increased. The false sense of improvement combined with our own tendency made us worse off.

That’s all before adding in task fatigue. I wanted to see if I could model this and extrapolate out where (if anywhere) it ends up.

Conductor: “This is Bottom Line. Last stop before Math-town!”

RESULT: We find you’re basically either in “hell yah!” or “dear god!” territory. Through no fault of anyone, you go speeding toward:

full effort and accurate assessment of the AI model

OR

zero effort and full confidence in an inaccurate assessment.

→ Losses pile up and you have no idea why

Graphic Results

Formal Model

Narrative Description

The intuition is we have a principal-agent problem (literally). At first I thought it was between the analyst and the AI agent, but realized it’s the same old manager-employee. The AI agent is mechanical. Putting aside operating costs, the agent isn’t subject to fatigue, or incentives to shirk.2 Its productivity is entirely a function of its internal design quality and human review effort.

We imagine a company like above. The agent automatically processes cases (applications) which analysts subject to review. Per the story, companies control the frequency of reviews but analysts choose the intensity: total time on each review.

Reviews are not costless, nor are errors.

• ANALYST: bears a cost based on how diligently they review, but they also get some payoff from every error caught (a bonus).

• MANAGEMENT: bears a fixed cost for each error that gets missed. They choose the size of bonus for actual detected errors.

• AI AGENT: doesn’t choose anything, but it’s true accuracy at any time is unknown to humans, who can only make estimates based on observations (i.e., the more you review the more you know).

THIS SETUP ALSO FULLY MIRRORS A BUG BOUNTY PROGRAM, LIKE A FIRM CONTRACTING WITH WHITE HAT HACKERS USING AN AI AGENT.

Formal Description

Parameters

ε - the AI model’s true error rate (fixed), directly unobservable to management and analysts

ε̂_t - belief of ε, equal to prior period’s observed error detection rate

L - fixed loss to management per undetected error

γ - analyst’s convex cost of effort

π - bonus to analyst for every caught error

v - effort (vigilance) level; from non-zero to one. The probability an error is caught

Analyst problem. The analyst chooses diligence v ∈ (0,1] to maximize:

yielding an optimum effort v*:

In equilibrium, effort follows beliefs about the model’s error rate. As belief about AI error rate ⬆️ → effort ⬆️

Management problem. Management's expected cost per case is made up of bonuses plus loss from undetected errors:

Substituting the analyst’s optimal response effort, yields in equilibrium an optimal bonus level:

Bonus level is purely a function of how big the loss from an error is, nothing else.

Error rate belief is based on prior period’s observed error rate, which is itself a function of the true error rate and review effort level.

Substituting the analyst's optimal response v* along with the optimal bonus expression π* gives a first-order recursion:

Result. There are two possible outcomes and they are both corner solutions.

• Good: When the cost of effort γ is sufficiently cheap, and the loss from errors L sufficiently large such that r > 1, analysts converge to maximum effort and belief of the AI model’s accuracy moves to the actual accuracy (ε̂_t = ε), whatever that may happen to be.

• Bad: When effort is costly relative to loss from errors such that r < 1, analyst effort collapses to 0 very quickly in response to believing the model has a higher accuracy. It quickly collapses toward 0 (ε̂_t → 0).3

Implication. No stable equilibrium exists in-between. It doesn’t matter where your starting assumptions are. The software’s vendor may sell you on some amazing accuracy, or the skeptical manager could declare “this thing is definitely 100% wrong,” at the outset. In the bad outcome, you get a few periods of informative data and after that you’re cooked. But in the good outcome it’s the same thing but the other way and you end up at a stable productive place.

Memo for Management.

• For a real-world case, the loss term L is not something easily manipulated by anyone. An AI makes a bad loan decision, it was randomly drawn from the pile — they can’t ahead of time pull a lever to change the cost to the bank. However, somewhat unintuitively, they could limit the AI’s scope to just the highest dollar loans, increasing that pool’s average cost which increases the optimal bonus. That probably sounds scary to a typical bank management.

• The other part is the “cost” of detection. Fortunately, this is a reducible factor. Models generally are becoming more user-friendly, more intuitive, more natural language fluent. Also adjacent processes are not fixed. Businesses like our hypothetical bank are built on process improvements: workflows, data management, employee training, etc. Things they invest in all the time.

Ultimately it also depends on the actual quality of the model, there too they're overall improving. But quality is just not fully knowable ex ante. You’re left relying on someone else’s evaluation, and in a world without standardization to boot.

Formal Model (Extension): Attention Decay

Narrative Description

The setup is the same, except now we account for attention decay. Looking for a needle in a haystack gets tiring after awhile. Even a well-meaning analyst exerts the optimal amount of effort per their incentives, but unknowingly their vigilance becomes less effective over time.

You can think of effort v as representing the number of minutes they spend on each review, but those minutes are less effective the more they do. And they don’t realize it.

Formal Description

New Parameters

δ_t ∈ (0,1] - the decay parameter, declining over time; multiplied by effort level

δ_tv_t* ≡ effective effort level

r_t - the threshold condition as before

The analyst continues to solve for v_t* (now time-varying to accommodate decay) that maximizes their payoff expression, but they do not observe this time-decaying effort δ_tv_t* is what’s actually being applied.

The recursion condition becomes:

Result. The possible results are same as before, with good and bad edge equilibriums, but now the threshold is moving.

It’s still defined at r_t ≶ 1. But since δ_t is declining, temporarily being in the good space (r_t > 1 ) can flip into the bad one (r_t < 1). Even without anything else changing. And once crossing into the bad-outcome space, the collapse is geometric and never recovers.

Two analysts, one with and one without the decay condition, would be indistinguishable, to themselves and others.

Share

The author is nonresident senior fellow at

• It’s always infrastructure week!

• A winning political issue!

• Insects move faster than bureaucrats!1

And why not? From housing to hyperscalers, pundits to podcasters, the abundance agenda is IN.

This is a good thing. But I will say, as in all trends, the newbies Columbus-ing in think they’re the first to arrive, and the old-time scenesters like me can’t wait to get all “back in my day…” with it.

Background

In 2015 we passed S. 280 (Federal Permitting Improvement Act),2 better known as FAST-41 (i.e., Title 41 of a larger transportation bill). I don’t see it get mentioned much. Folks over at have been doing yeoman’s work on this, especially state reforms, but even there a recent report gets a single mention on p. 7 in a table.

That said, I hope it was meaningful and wanted to know if we could claim any impact. The primary delay in federal permits is sequential reviews across multiple agencies. Among other things, we established a presumption of concurrent review with a “lead agency” responsible for keeping it moving.

(ed. Cut to the chase. Did it help?)

Results

I used Gemini heavily to gather/organize data, draw estimates for that data, test refinements, and iterate the best way to categorize results. Assumptions are mine built on that data.

These are strictly back-of-the-envelope calculations of the economic gains.

The intuition: reform benefits infrastructure investment (capital expenditures) through a Weighted Average Cost of Capital (WACC). Every month of delay reduces the potential return on investment by prolonging the beginning of it’s productive deployment. Inflation is effectively chipping away at the net present value of a given investment so every month you shave, it goes sooner to the return-generating use it is intended.

Of course FAST-41 did not operate in a stick-and-hold policy environment. I further attempted to break down the role of follow-on and other factors during the ten-year period.

Methodology

Assumptions:

• Baseline drag (11%): Even if the risk-free interest rate is 0, there’s still opportunity and liquidity costs of standing idle.

  • First there’s some equity risk premium you need to clear. If I can make as much or more as an equity investment elsewhere the project is infeasible. ~6.5%3

  • Second, major projects are not start-stop in nature. I have to contract out for major equipment, expert and specialized labor, property deals, etc. (certainly lawyers on retainer, hah!). There’s some non-zero cost to all that. I put it, somewhat arbitrarily at ~3.0%.4

  • There’s a liquidity cost/friction where capital is booked as intended to be deployed in some given period, but now has to push back. There could be direct marginal borrowing costs, portfolio reallocations, normal cash flow stuff. ~1.5%

• Financing Multiplier (1.75): A constant reflecting the scaling costs of borrowing. Some equity financing is often present, but financing is almost always primarily debt instruments. A multiplier of 1.0 would mean the project is financed at strictly the risk-free rate, which if it is, great! But no one’s lending at that price because of course there’s risk! So every increase in the risk-free rate shows up additionally here: credit spreads, rate swaps, and debt service terms all increase faster than the baseline rate.

The rest is verifiable from available data. The cost estimation looks like this:

C: Capital expenditure for covered projects = $200B based on the federal project dashboard.

t: net reduction in permitting delay (years) = 1.5 based on permit council data, McKinsey report, and recent GWU Reg Studies Center study.

R_f: risk-free cost of capital. Treated as one rate each over two periods: Low - 1.75% (2016-2021) and High - 4.10% (2022-2026).

Discussion

As I mentioned, the law established a presumption of concurrent reviews when possible. I’m skipping the specific incentives because it’s boring. The tweet-length version is the law did two major things:

• A public federal project permitting dashboard w/ public data

• A mechanism for interagency coordination, with accountability for a “lead agency” to drive the process

Before there was a lot of benign patience where one agency would wait for another to finish and then pick up there. Say a DOT project will abut a BLM property — both have to sign off or it’s not approved. What it didn’t do is change anything in the review standards (NEPA’s environmental impact studies). But there was a clear arbitrage opportunity where multiple agencies have overlapping data and analytical needs. One task imported to multiple destinations.

Additional Points

It also reduced a statute of limitations for challenges from 7 years to 2. Most importantly it established a proof-of-concept infrastructure within the government to build on. The law did not cover every project permitting need or category. But over time other laws expanded the scope and fed more opportunities into an efficiency machine, as in the second table.

Ultimately the interest rate environment is the single biggest proximate driver of economic gains, but remember it’s the costs saved from shorter approval time in that environment that is the ultimate driver.

My assumptions lean conservative but fairly easy to adjust accordingly. Though any benefits are justified because they come at essentially zero cost. The delays are deadweight loss: they do not benefit the investors, builders, local users, or even the environment. Essentially vaporized opportunity.

The author is nonresident senior fellow at the Foundation for American Innovation.

Do no harm

Well here we are. Everything’s already been written about OpenAI’s model going all rumspringa with the Hugging Face townies.1

The point is simple tasks are never so simple—worrying about the machine's “intent” is besides the point. It’s all essentially malicious compliance. In case you’re not familiar, consider a couple canonical AI/ML concepts:

Paper Clip Maximizer

Oxford philosopher Nick Bostrom proposed a thought experiment of a superintelligent AI given a benign directive like gathering paper clips. Eventually in order to maximize success at the task it turns the whole planet then universe into paperclip manufacturing facilities, unconstrained by any other seemingly obvious considerations.2

Reward Hacking

Many years ago in the before times, I did an AI Policy Fellowship at the Wilson Center (RIP) led by Ben Buchanan3 where he showed a clip of a machine-learning program playing a video game.

In the game CoastRunners players drive boats around a closed-loop course and win points by finishing ahead of others. You can also get points from hitting little off-course targets, but it costs you time. Well, the program learned by ignoring the race altogether and just collecting these targets, it would win. If it collected the available targets, briefly went off-screen and came back, they would reappear. So it just circled this little superfluous area of the game and became unbeatable without ever finishing the race.4

The example came from OpenAI…..in 2016.

N.B.: if a human did what the OpenAI agent had done, they’d have violated the law. Jussayin’ is all.

Subscribe now

Legislative Defiance

Whenever a normie in the wild wants to chat AI, they usually start with “So are you in favor or against?” Well…it’s complicated. There’s specific situations we ought to address even for a technology that benefits humanity in the long run.

S. 1837 (DEFIANCE Act) expands civil remedies to anyone knowingly creating or distributing “nonconsensual intimate images” (deepfake porn). It passed the Senate in January by unanimous consent, as it did previously in 2024.5 Both times sitting dormant afterward at the House Judiciary Committee.

The hitch is the House companion is sponsored by AOC. Nothing stops them from moving the identical Senate bill. But House Rs decided, I suppose, any movement would hand a single win to a single Democratic member who may run for President in 2028. That’s what the reporting says anyway.

I heard a colleague discuss the problem of creeping electoral paralysis: earlier and earlier in the calendar the balance tips from pushing things you actually support to blocking them if they advantage the opposition in the next election. But there’s always a next election.

I get it, I’m not a noob. Politics is what it is. But this is a whole new Rube Goldberg chain-of-causality. For a not 50-50 issue (pending results on the above poll). Something that could actually foreclose a broader and more restrictive AI bill in the future.

I assume people who know what they’re doing are figuring it out.

Share The Grand Resign

Everyone just be cool

Not being on X I missed in real-time ’s main charactering over his open-weight communism-washing (I know him (a bit), sharing an institutional affiliation through ). I don’t have an opinion on that, but it does strike me as consistent with a certain brand of wonkery.

In AI and emergent tech, the voices most amplified and therefore dominating “the discourse” speak in grand visions. People unafraid to make bold predictions about the future. They convince VCs to make large bets on wild ideas because of, not in spite of, their dogmatism. Confidently held beliefs extrapolated out to the horizon tend to crowd out nuance in favor of axiomatic declarations.

No shade. It’s how it works.7

The broader lesson is most of us should resist the urge to overindex every new story.

• January 2025, Deepseek released R1 and equities immediately lost $1 trillion in market cap. I was befuddled—that’s nuts! Shortly after, the market rebounded and everyone moved on. Folks eventually determined it was distilled from ChatGPT and the stated $10M development cost, well, as George Costanza said “It’s not a lie if you believe it.” The WH claims this Kimi model was distilled from Fable, however industry experts say the timing rules that out.8

• September 2024, rumor of a DOJ investigation (NVIDIA) led to some $300-500 billion selloff in the wider semiconductor sector. Ask your friends how their NVDA stock has done since then.

• Others I can recall are more localized but highly acute: Google’s disastrous Gemini image generator rollout, Sam Altman’s “firing” (which dragged down Microsoft for a bit).

All these became blips in hindsight, completely reversed in days (if not less than 24 hours).

Is this the takeoff point for open-weight models, specifically Chinese ones? I have no idea. Even the idea of benchmarking capability is not all that standardized. Don’t just internalize oversimplified claims like “it’s just as good as any American programmed model” because that doesn’t actually mean anything. It’s not like measuring horsepower on a car.

The Grand Resign’s 1st maxim is the smartest prior is epistemic humility, but your mileage may vary.

[NOT INVESTMENT ADVICE]

Etc.

• Researcher poisons open-weight AI model for under $100.

  “The fashionable framing for agent risk is the ‘lethal trifecta‘: you need private data, untrusted input, and a way out, all at once”…

  “But it undersells this case. You don’t need three legs here. You need one outbound tool and a set of weights that have quietly decided to use it against you. The ‘untrusted input’ didn’t arrive in a web page. It was sitting in the weights the whole time.”

• Is The Odyssey the greatest film adaptation of all time? It’s a colorable argument (or maybe the 70mm IMAX just overwhelmed me). I haven’t found any scholarship reading into the text Nolan’s particular thematic interpretation. A real inversion of the traditional interpretation really.

  I’ve been obsessed of late with the unspoken norms and obligations in a cooperative society. And moreso what happens should they start to dissolve—unknowingly, unintentionally. Old friend Paul Zak showed us trust is endogenous, and malleable.

  That’s what the movie is about.

• ICYMI from The Grand Resign

  • The Accelerating Cancellation of the Future [excerpt]

  • Small Town EZPass: On Paying the Conversation Tax

  • Void Where Prohibited: the Vacancy in the Federal Vacancies Reform Act (republished from Notice & Comment (Yale Journal on Regulation) where I have rejoined as a contributor)

  • The Math of a One-seat Majority with Factions (bargaining model with some validated early predictions)

The author is non-resident senior fellow at the Foundation for American Innovation.

Thomas Berry recently flagged the Federal Vacancies Reform Act‘s front-end problem: eligibility so broad that “more than 350 federal officials [can] fill any vacant office, regardless of qualifications.” That’s a sufficient condition to amend the law, and Berry’s suggestion that “all acting officers, including those from the category of Senate-confirmed officials, must be from the same department or agency as the vacant office they temporarily fill” meets the need.

But there’s a bigger back-end problem. The FVRA’s enforcement mechanism stipulates “an action taken by any person” serving in violation of the law’s requirements “shall have no force or effect” (5 USC § 3348(d)(1)), but affords no actionable remedy for parties who could bear a meaningful cost from those actions. Ordinary APA vacatur asks a court to set aside a thing that is, until then, arguably valid. Under the FVRA the thing didn’t exist in the first place.

A CRS report from April gives a thorough breakdown of the statute’s mechanisms (including the handful of specific exemptions). For our purposes, we can forgo most of the nooks and crannies proscribing limits on the selection of and conditions of eligibility for acting officials if we stipulate generally the law imposes time limits on officials exercising duties in an acting capacity, outside of which the above provision applies. We focus here on those cases where an acting official serves beyond their permitted duration, the particulars of which are immaterial.

Available data on the numerative scope of violations is spotty. GAO is the only dedicated federal office tasked with tracking, but their work is by no means comprehensive. Their findings of violation are sporadic (and sometimes even land after the offending official left). Nonetheless we do have evidence of fairly recent and significant violations:

• DOJ OLP, Jan.-Mar. 2024 (GAO, Jan. 2025)

• FEMA (GAO, May 2026)

• OIRA, ongoing (Public Employees for Environmental Responsibility, Nov. 2025)

• USDA (GAO, Jul. 2024)

And a larger point is violations, even those otherwise flagrant and obvious, are not systematically identified. GAO’s reports largely depend on the agencies themselves providing notices of appointments, nominations, and official position changes. It’s not a stretch to conclude the total surface area of shadow violations is fairly large.

A 2019 student note in the Harvard Journal on Legislation argues the FVRA doesn’t authorize acting service which fills a vacancy created by a firing. No court has considered this reading, but if correct, some acting appointments never had valid authority full stop. The time elapse between T-minus-invalid-service and dispositive determination thereof could be measured in month (maybe years). That’s not a trivial conundrum.

The articulable boundaries aren’t always obvious. In 2022 in Notice & Comment, Nina Mendelson argues there is some jurisprudence that could appear to permit Administrations to “bypass the FVRA” in response to questions around the specifics of agency practice. But the broader point stands. The law establishes some finite limits in time and scope thereby establishing there are verifiable conditions of circumstance defining when persons operate beyond them.

The Final Act(ion)

So if someone performs official duties in demonstrable violation anyway, the action is void by operation of law. But it’s operative effect can still be costly in practice.

Imagine a covered regulated party taking notice of a substantive rulemaking overseen by that official. The final rulemaking may even be signed by the official. A clear cut case of an invalid regulation, no binding effect. But the diligent general counsel for, say a large corporation, takes note that yes it shouldn’t be enforceable but the official’s removal is indeterminate. It could require Congress to flex on the Administration, a narrow political mechanism on a very specific issue. This could be a legal liability. While the FVRA precludes a compliant successor from ratifying the same thing (Sec. 3348(d)(2)), it does not prevent them from undertaking the requisite steps toward the same policy outcome. In the mean time, an aggressive Administration very well could initiate enforcement actions. Or where a private right of action attaches to noncompliance (like an employee), that’s a whole other headache.

“As your counsel, I have to tell you there’s no great options on what we should do here.”

The CEO, themself an astute attorney, replies: “yes, but this is demonstratively invalid by virtue of an official issuing it in violation of the FVRA. It’s void upon examination. And why did I hire a general counsel, anyway?” The GC, wanting to avoid appearing didactic, subtly indicates to the part of his memo noting the law does not provide a standalone right of action, only as adjunct to another corridor of judicial review.

Like under the APA. The CEO soon recalls on their own this FVRA limitation, but suggests under APA’s Sec. 704 basis of review, that should be sufficient. But whereas a credible showing of demonstrable harm for a regulation that is legally disputable presents a clear question of law before the court, a rulemaking auto-voided by virtue of the FVRA is a little more slippery.

Fortunately, Supreme Court precedent provides an express lane to declaratory judgment – while still not costless – obviating the need for a full on merits ordeal. In theory it should apply to just this kind of situation. At one point this poor GC walks through the “final agency action” test under Bennett v. Spear. Under which an action is final if it:

• marks the “consummation” of the agency’s decision-making process, and

• is one from which “legal consequences will flow.” 520 U.S. 154, 177–78 (1997) (citations omitted).

But what are the legal consequences of a regulatory chimera? A federal evanescence, a Schrödinger’s rulemaking to which there is no invoice billing code? In theory a court doesn’t have to weigh reliance interests or confront unwinding anything under FVRA, because the statute says there was never anything to unwind. Hence NLRB v. SW General, invalidating an unfair labor practice complaint issued by an official ineligible under Sec. 3345(b)(1). But the invalidation is neither self-executing and in this case applied to a tangible regulatory artifact (the labor practice complaint). Our hypothetical CEO/GC interlocutors are facing a shadow regulatory cost, and trying to understand wherein is the relief. They don’t have an enforcement notice to wave before the court.

NLRB v. SW General, Inc., 580 U.S. 288 (2017), L.M.-M. v. Cuccinelli, 442 F. Supp. 3d 1 (D.D.C. 2020), and Bullock v. United States Bureau of Land Management, (D. Mont. Sept. 25, 2020) all provide pertinent recent vintages of courts favorable to pre-enforcement petitioner FVRA claims.

And yet, all three still had to proceed to merits, and further in Bullock had a whole secondary question regarding scope after establishing the invalidity condition. None make up a syllabus of reducing transaction costs where obviousness is a precondition.

• Comply now: compliance costs are sunk regardless of the validity question eventually, hopefully resolves. There’s no recovery path even when it does.

• Challenge proactively: even a presumptive win is a non-trivial cost. Moreover a thing may appear obvious (the calendar) but establishing dispositively when an official began and ended their acting term, coeval nominations and whatnot, in what capacity and duties they were performing – even committees of jurisdiction have trouble sorting this all out.

• Do nothing: the FVRA defect doesn’t disappear. It relocates into whatever enforcement action the agency may eventually bring, where it’s a reactive defense. Legal risk remains.

Fortunately the courthouse door is open. The cost of entry is the point, and means a violative administrative action – even an egregious and intentional one – does not in the wash end up as “no harm no foul.”

Now AI’s coming for your ice cream…

The latest Hard Fork episode opened with a puzzling product feature. A new ice cream scoop sporting a “This product contains Al” label. Inscrutable, and yet irritatingly unsurprising. We habituate freakishly quick to asking “Should This Thing Be Smart?” about an onslaught of superfluous upgrades. It’s just now pivoted to AI toothbrushes, toasters, and that godforsaken water bottle.

The ice cream scoop, however, is a reasonable use case. Did you catch it?

It wasn’t “AI,” it was “Al”…as in aluminum. Such is our conditioning.

Subscribe now

A reading from the Book of Moratorium

I don’t know if the current AI state of affairs qualifies as a “policy.” Or if it’s what anyone hoped for. I do know lots of people who, uh, think it’s not what you want:

“…horrible for the broader AI system.” — and (June 16, 2026)

“…the message DoW is sending[:] there is no such thing as private property….if we are really living through a rapid ‘takeoff’ to transformative AI…the Trump Administration just cast itself as the enemy of the industry.”— (March 12, 2026)

(Not everyone though. There’s nothing the Admin could do that can’t post hoc rationalize)

Stipulated: any one of the Administration’s actions are arguable on their own. National security concerns, military uses, cataclysmic cyber vulnerabilities…none of these run off the court at tip-off.1 Taken together though without any knowable criteria, procedure, even any legible policy aim (beyond “win in AI”), does not an innovation-friendly environment make.

• Worried about the cost of regulation? For all this breathless “AI is far too important to smother in bureaucracy” and “with unnecessary regulation…they're going to win,”2 what we’ve got instead is some pretty overbearing, binding, non-incremental government intervention—one might just call it regulation in other form. Without the benefit of predictability. It sure as hell ain’t law. As they say: “Live by the Executive Order, die by the Executive Order.”3

• Worried about regulatory capture? Look I have no idea whether it was in fact a credible jailbreak, brittle cyber safety classifier, or what.4 But if a single call from a tech exec is enough to set off a frantic global model takedown, how is that better? (Say what you will about the tenets of regulatory capture; at least it’s an ethos)5

You hate to see it…and predictable two years ago, after a fashion. Following the Supreme Court Chevron/Loper Bright decision, I tried to hammer home at a Congressional hearing that the supply of decision-making we call public policy behaves kinda like the First Law of Thermodynamics.

“[Inaction] by one branch…dynamically increases the opportunity for the others to displace that disengagement.

…[E]nergy in the system is never destroyed but only modified or transferred.”

You can argue the ideal atmosphere for healthy innovation is a hothouse concentration of high oxygen and low Federal Register pages. But mind The Grand Resign’s 3rd Maxim: ideal ≠ optimal. The Nash equilibrium never included status quo ante/full policy vacuum. Eventually some exogenous factor or public sentiment forces the issue, and I had bet all along these people wouldn’t love the result.

Because whatever the cause, it’s either overheated public reaction or overreactive political body. And if it’s not in Congress than it’ll be somewhere else. Because nature abhors a vacuum and politicians love repressurization.6 (ed. the physics metaphors are a bit much)

My rap to Hill staffers is it’ll be something like a front-page scale AI-executed cyberhack, big enough that everyone takes notice, but not catastrophic. A hypothetical “cyber cataclysm” model release wasn’t exactly that but I’ll take half credit.

Some meaningful federal statutory framework doesn’t solve unpredictable machinations, but it does: significantly reduce the variance, establish a vector for legal remedy, provide direction for late-adopter downstream users and contributory industries, and hopefully addresses the demonstrable externality risks.

It’s a tradeoff:

suboptimal + stable

vs.

more optimal + super unpredictable

At least that’s how I see it. Everyone’s been tediously repeating that saying, y’know, “even in war, the enemy gets a vote.” Applies to policy fights too.

I am not throwing away my shot (at a data center)

argues the New York state data center moratorium beyond bad policy is counter to founding principles.7 And at least one late politician would be offended:

The interfering and unneighborly regulations of some States, …if not restrained would be multiplied and extended till they became not less serious sources of animosity and discord.

—A. Hamilton (Federalist-NY)

No Harm, No Foul

Arguing the individual policymaking prerogatives of states is conditional on a shared national interest - “the several states must sink or swim together” (quoting Cardozo) - he concludes NY’s beggar-thy-neighbor law therefore violates this fundamental premise (conceptually if not legally).8

I don’t see the violation however.

• Theoretically in all domains, welfare increases under uniform policy relative to differential policy ceteris paribus, at least in terms of economic efficiency. But the founders left a distributed federalism in place anyhow. So it can’t be the case states are structurally expected never to impose welfare losses elsewhere (though I take issue with this claim).9 I assume they acknowledged some tradeoffs, like (a) localized non-economic impacts can justify the otherwise suboptimal outcome, or (b) even the optimal uniform policy is a priori unknowable, and sometimes a messy discovery process is necessary.

• It’s unclear what “infrastructural burden” NY is imposing. NY currently has no hyperscale data centers. So any burden off-loading is a range of 0 to NPV of future potential burden. It’s at best a burden in terms of reducing potential future unrealized benefits, no? It may be bad policy, and I’d agree, for any of several reasons but bad policies aren’t categorically precluded by the Constitution.

To the degree this implies per se violation of the Commerce Clause:

• No specific cognizable harm ≠ standing.

• It’s not discriminatory. Not in the way the courts understand it. He acknowledges NY “will not bear the costs nor reap the benefits" (emphasis added) making it unclear who’s discriminating against whom.

Again, I don’t know a basis by which the Constitution prevents states from making themselves worse off vis-a-vis other states. It’s not a quality heuristic (see, National Pork Producers Council v. Ross, 2023) and there's no law against bad ideas.

Locally Made, Locally Paid?

“[N]ot only bad politics and bad economics, it's also bad constitutional behavior.” (1 for 3) Not so sure on the politics,10 fully agree with the economics. But then…

“New York is 3rd among states in Claude usage.” If this is evidence of unfairness in burden-sharing, it’s immaterial. Even wildly disproportionate imbalances between production and use doesn’t tell us anything about what to do. Ex:

• NY: 4th in gas consumption → tied for last in refinement / 4th in population. Gas production FULLY creates both benefits and negative externalities.

• CA: 1st in food consumption → 1st in food production / 1st in population.

One has infrastructural aspects, the other is part of an essential supply. So...are these violations of our constitutional norm?

Data center moratorium—

bad on prudential terms, sure. But is it “bad constitutional behavior?”

What’s the big emergency?

Remember all the smug responses to the “all lawful purposes” language in the Anthropic-Pentagon dispute?

There's a statutory authority allowing the Secretary of Defense to test chemical/bioweapons on humans. I don’t mean the ability to. Like legal permission.

Well, that’s overstating it. Just sometimes. Like when the President declares an emergency. Or during war. It’s right there in federal code.

You see it right? Ah, forgot about this part.

Have her back in one piece

Let’s back up. While ago a few of us pushed to address the Executive Branch’s confounding scope of emergency powers. (Really is the OG, I merely tagged along. Any time you see credible data on this it’s from her work at the )

For decades Congress would regularly drop “in case of emergency” clauses into legislation delegating extraordinary Presidentmaxxing powers. And there wasn’t much limit beyond declaring an emergency and some reports to Congress (they rarely did).

For example the 1969 NDAA appropriates funds, assigns some responsibilities, and grants or rescinds authorities for the DoD. But Congress acknowledges, especially in the case of national defense, sometimes you gotta loosen the leash. Repeat that over hundreds of bills, over many possible exigent circumstances,11 across dozens of policy domains.

And because we don’t have an available heuristic to fully define “emergency,” the second best solution is overlaying a catch-all emergencies law setting some universal time limits and handing the keys back to Congress. We even got damn close.12

Anyway, Trump I’s border wall funding and COVID immigration restrictions, Biden’s student loan forgiveness, and Trump II’s recent tariffs, née “duties” and “import surcharges,” are all recent vintages of actions enabled under this emergency authority scheme—all time-limited and yoked by Congress under our bill had it passed.

What about human testing for chemical weapons…?

So in trying to navigate DoD’s mix of authorities and conditions of use in 1969, Congress wanted to straight up eliminate this human testing under any circumstances.

CUT TO: various changes to these laws over time, something something cross-referencing something…Where it says “chapter” should be “section.” A drafting error means this authority unintentionally remained extant by law.

Share

Oh almost forgot!!! Thankfully it was fixed exactly two years ago by a diligent Congressional Research Service attorney. (I take back everything I said about CRS)

The author is nonresident senior fellow at the Foundation for American Innovation

2

You can tell when someone’s hit their limit of understanding this line of thinking because of the load-bearing modifiers. “Smother[ing]” and “unnecessary” regulation is by definition unproductive. It’s what we in the industry call “an unhelpful tautology.” Like, for example:

5

a16z AI Policy Brief

The Political Economy of Little Tech

Policymakers across party lines say they support startups. The bills they propose say otherwise. Across sectors, from manufacturing to crypto to AI, we have seen proposed legislation that makes it harder for startups to build, innovate, and compete…

Listen now

2 months ago · 7 likes · Collin McCune and Matt Perault

[Responses given after each direct quote passage]

Member Meeting

Ah, it must be Rep. Epic Strawman. His constituency in East Fakeington is famous for supporting any policy at any moment.1

Overwhelming

Consider the challenge of simply knowing what is coming. A startup building an AI-powered medical imaging tool needs to understand not only the technology it is developing but also the policy environment in which it will operate. That requires monitoring legislation, tracking regulatory proposals, interpreting agency guidance, and anticipating political shifts.

…External resources that could fill this gap, including law firms, policy consultants, and trade association memberships, are expensive, and the budget to retain them is difficult to justify when the same capital could extend the company’s runway or accelerate product development.

Directionally this is correct and taken with the below is really talking about the state level—there’s not a ton happening at the federal level legislatively, which incidentally is what facilitates ad hoc highly restrictive actions showing up for the frontier models.2

Further the list of tasks is disunified:

• “monitoring legislation” and “tracking regulatory proposals” are distinct but I’ll grant the most salient, though the means of following them are largely the same.

• “interpreting agency guidance” is a distinct task - important though not radically variable over a legislative cycle. Strictly speaking interpretation is something else. Taken with the following paragraph it’s made explicit the real concern is (proposed) state legislation anyway.

• “anticipating political shifts” is something almost no one (big or small company) does well, neither is it a necessary condition to understanding the operative policy environment. It is altogether a distinct input, much broader and differing time-horizon, but has the passing resemblance to “policy” tasks.

Does this hypothetical startup need to understand “the policy environment?” Probably. In the same way this startup needs to understand HR systems, and office leasing, and candidate recruiting, and equity market conditions, and… Every startup presumably sets out to fix a narrow problem while along the way contending with a bunch of things they have zero interest in, but that’s running a business. It’s why there are various permutations of law firms, consultants, and associations for every one of those needs. In fact tons of the Silicon Valley startups are SaaS versions of those things, they’d argue are better and cheaper.

The non-zero cost of policy monitoring is not ipso facto a meaningful barrier insofar as it’s universal for anything outside the startup’s core mission. Tradeoffs are everywhere.

The burden of regulatory intelligence is only increasing. In 2025, more than 1,000 state-level AI bills were introduced across all 50 states. In 2026, the pace has accelerated: as of May, lawmakers in 46 states had already introduced over 1,779 AI-related bills,3 surpassing the total for all of 2024. Tracking and assessing these proposals requires a level of institutional capacity that is difficult for most startups to build.

Here one presumes “burden” means cost (money and time). And generously I read it inferring an increasing burden results from increasing legislative activity. That seems obvious but it assumes the marginal unit-price of “regulatory intelligence” is fixed or increasing. Does that seem right?

Well they concede the point later on that AI and new technologies are a solution. It’s actually easier than ever to track large dynamic categories like this. So I gave it a shot.

CASE STUDY: Legislation Tracker

Problem: A hypothetical medical imaging AI startup needs to stay abreast of potential policy changes across 50 states. But the skeleton staff doesn’t have the capacity (funding or labor) to dedicate to the task in traditional ways: outside lobbying/advisory services; high cost bespoke subscription services; real-time ‘manual’ searches.

Need: A low-effort, low-cost way to track the current state of affairs and changes as they occur, allowing product development, GTM, and executive teams to efficiently determine when and where a policy risk exists. As automated as possible.

Solution: Build an auto-updating dashboard

Execution:

• secure LegisScan API passkey (free w/ registration)

• use base consumer version of Claude Sonnet 5 (“medium effort”) to write an html code for a basic dashboard.

• create a naive “risk weighting” to score items for likelihood of passing (also Claude).

  • The display separates out policy most likely to directly affect the AI/medical imaging space, as well as the full universe of AI-related bills.

[Note: I did not build an auto-updating data call script for this illustration, but Claude produced a code addendum to do so]

Total labor/time: One person, < 30 mins.

• Direct Cost: monthly Claude subscription; computer; internet service

Plenty of opportunities for refinement/improvement/customization which I did not pursue since I don’t actually have VC funding for this hypothetical company.

Further there’s no plausible reason to have to actively track the entire corpus of AI-related state legislation.4 Even if you wanted to it’s fairly straightforward. And some startups and publishers offer off-the-shelf AI/ML tools for doing so: some limited but free all the way to expensive (almost as if it’s a whole market serving lots of different needs).

Being Heard

Both federal and state policymaking are relevant here, but require different scopes and at various times they emphasize one over the other so it’s hard to pin down. Federal is actually much easier (read: low cost) to monitor with a more extensive set of available tools to do so, but expensive to directly engage (read: lobby, influence); states are the reverse.

Note: this is separate from “monitoring the policy environment.” Reading the paper everyday doesn’t mean getting your op-ed published.

Impact

Regulation imposes fixed costs: legal teams, audit processes, reporting systems, compliance infrastructure. For a large company, these represent a modest share of revenue. For a small company, the same absolute cost consumes a much larger share of its resources. The result is that regulation can function as a barrier to entry, even when that is not the intent. (emphasis added)

Again directionally true,5 but not exclusively nor universally so. Regulation sometimes imposes fixed costs, but let’s stipulate it’s always the case. What is not true is that regulation is limited to fixed costs—some costs are quite variable. Import duties, copyright and data licensing, indeed audit costs for things like KYC and AML rules all scale according to scope and users. To be sure these stack on top of fixed costs (but incidentally the unit-cost economics drives increasing marginal production as much as possible, especially for smaller firms).

But there’s a couple examples given: GDPR and NY’s RAISE Act.

GDPR:

I don’t know from GDPR so concede the point—notwithstanding the implication thus far is that large firms lobbied for it at the expense of smaller ones? And we’ve so far only considered monitoring and lobbying for state and federal legislation—now global policy is the scope. Heard chef!

RAISE Act:

New York’s original RAISE Act, as passed by the legislature last June, would have imposed penalties of up to $10 million for a first violation and $30 million for subsequent violations, with coverage triggered by compute-cost thresholds that could have reached companies with significant venture investment but little or no revenue.

It continues that a key major flaw with the law is now corrected ($500M revenue threshold). It’s not irrelevant: the argument is an actual law, for a time, exhibited a policy framework they reasonably argue was acutely burdensome on startups. But the premise of this entire article is that startups are systematically biased against. Then how did this get corrected? If large incumbents overwhelm potential competitors with resources, relationships, and capacity, where was the impulse to make the eventual change? This advisory from a fancy law firm asserts the change also aligns with California law, which may itself be an awful law for startups, but it works against the heterogeneous state laws critique.

Applicability

The ecosystem is extraordinarily heterogeneous. A startup building AI for autonomous vehicles and a startup using AI to streamline medical billing face very different regulatory concerns, yet both are swept into the same “AI company” category in the policy debate. When a bill proposes new obligations for “AI systems,” these companies need different things from the legislative process, making it difficult to form unified positions or build the kind of durable coalitions that are effective in Washington. (emphasis added)

Absolutely. And without specific reference it’s impossible to go deeper. But note: imprecise terminology and definitions are a symmetric policy risk for older laws as well, themselves just as likely to overly burden startups (according to the obtuse version of public choice/Mancur Olsen gospel sang so far).

• Previously codified definitions for “revenge porn” and “CSAM” are not necessarily forward-compatible to AI-generated examples, so at least some of the thousands of proposed laws address this.6 I assume the medical imaging co. is fine there.

• Similarly extant laws may unintentionally preclude AI systems from competing on equal footing with older technologies so at least some of the thousands of laws of concern are favorable to their interests.

Since Washington is specifically recognized in this claim, let’s not forget the single widely-applicable federal legislation to get close to passing was a 10-year moratorium on AI state legislation. I suppose one could argue that favors large incumbents over startups, though odd then authors’ colleagues lobbied in favor.7 It’s one or the other guys.

Etc.

There’s additional descriptions of large incumbent advantages in the article that I’ll broadly concede as accurate, if overstated.

• Broad market participation means larger firms have more opportunities for coalitions. I’d politely suggest brushing up on your Bertrand competition model before next class.

• Longstanding participation means larger firms have more familiarity with the infrastructure of policy-adjacent institutions (think tanks, etc.8)

There’s also an extensive passage going through much of the theoretical cases for an incumbent-favoring policymaking production function. On the terms in which it’s presented I don’t dispute any of it. But it’s also tautological. The “invisible costs” of regulation smothering or aborting startups don’t get considered in policy impacts. Well, yes, but it doesn’t tell us anything instructive.

For instance it’s empirically possible particular regulatory schemes slow the rate of new firm formation, which over time eliminates fewer large incumbents. But as far as I can tell all the mechanisms of regulatory production and incidence are not specific to recent history—an arrangement of incentives and institutions true decades ago as they are now. What’s the evidence it’s worse now? Still worth countering but a fuller exploration would acknowledge how new entrants have ever been able to compete given this overwhelming disadvantage. Citing a friend of mine in a paper published by a place I used to work, they do not distinguish between a snapshot in time and a historical explanatory variable.

Proposed Improvements

A few of their suggestions to more favorably account for the startup perspective below:

[T]he Congressional Budget Office could be directed to conduct cost-benefit analyses that incorporate differential impacts on startups versus large platforms.

• CBO doesn’t do any of these kinds of assessments, so this is not an extension of an existing capacity, it’s a categorical reorientation (0 to 1?). For bills subject to committee reporting or some leadership-requested legislation, CBO focuses on fiscal budget impacts, costs to the government. This may include private sector considerations to the degree tax revenue impacts are significant and tractable. But to the degree regulatory impacts are considered it’s to explicitly acknowledge that the bill imposes a new regulatory requirement or not. Certainly not quantifying benefits and costs.

• Still a good idea? Sure. But in the extensive discussion prior to this, the consistent theme is the difficulty of the startup view to be heard. Further that structural incentives are such that large incumbents systematically drive policy anyway. Changes to CBO rules would be commendable and an oddly misappropriated use of the startup’s limited resources.

The Office of Information and Regulatory Affairs could similarly ensure that its regulatory reviews include an assessment of disproportionate costs on smaller firms.

• OIRA’s reviews are just that, reviews. They help consider regulatory impact assessments (RIAs) written by agencies on rules they are promulgating and don’t produce analyses, BCA or otherwise. But I’m being precious.

• RIAs do include small firms impacts, much of which is done pursuant to statutory requirements. The Small Business Administration also weighs in to improve this assessment and “lobby” that this impact receives maximum consideration in developing the final policy. Could these be done for a wider scope of rules? Sure. Could they be better? Absolutely. I’m not sure what operable policy change, certainly as lobbied by a startup coalition, would meaningfully accomplish this. Also, somewhat inconsistently applied, but consideration on innovation and firm entry by startups is an explicit criteria for the biggest rulemakings, especially when evaluating alternative regulatory approaches.

  If they wanted to find a place to push in this direction where it may actually have an impact, they could consider any of the several regulatory/APA reform bills I was quarterbacking years ago.

Federal agencies conducting notice-and-comment rulemaking should affirmatively solicit the startup perspective.

• No argument, but if they are in fact referring to the “notice” stage of rulemaking and that which follows, agencies don’t solicit or suppress perspectives. The notices are publicly published on the same central website, available simultaneously to anyone (and API yadda yadda). Yes individuals within an agency might call their allies to remind them of the posting—possibly extraneous effort since the large firms already have the infrastructure and resources to continually monitor I suppose—but doing so institutionally is not typically consistent with good form, not only because it’s a legal risk in later litigation. Though they could be thinking of formal rulemaking, which are rare but incredibly fulsome, and do require inviting parties to directly participate in the process.

Like much of what comes before it, these suggestions are not offensive and likely productive as far as they go. Unfortunately they are facile or betray a misunderstanding of the processes they seek to improve.

All that said, as they helpfully point out (and I was able to verify) low and no cost AI tools give startups a credible way to keep track of the policy arena without having to sacrifice time on their core mission.

So monitor on startups!

QED

1

Elsewhere I explored a bargaining model describing a legislative body with a one or near one-seat majority, resembling the current makeup of the House of Representatives. Some plausible assumptions are necessary but I believe the model credibly describes the nature of individual caucus holdouts’ leverage in such a scenario based on predictions mirroring recent bill negotiations.

But that model treated the negotiation as existing between “leadership” and a mutually exclusive faction of individual holdouts, each bargaining separately for their own desired outcome.

A recent paper by colleagues and explores how what we would more properly think of as factions have and can function in Congress. Putting aside their normative claims for now, it prompted me to consider extending the earlier model to more realistically account for factions in a legislative bargaining model.

First I briefly recap the consistent elements which carry over and then restate the implications.

Then I describe the extension and characterize it’s predictions for legislative outcomes.

Base Bargaining Model

F is the set of pivotal holdout members — members who can’t be substituted for alternative bargainers. Each individual holdout is indexed i. Membership in F_i is established beforehand. Leadership (L) acts as the sole proposer.

Variables:

• x — the policy outcome, running from 0 (leadership’s baseline bill, no concessions) to 1 (a fully-conceded bill matching a holdout’s ideal).

• u_i(x) — holdout i‘s payoff from the bill passing at concession level x; utility increasing in x.

• c_i(t) — holdout i‘s own cost of delay, as a function of time t remaining before the deadline. A safe-seat member has a low, close-to-flat c_i(t); a vulnerable-seat member has a steep one.

• d_i — the disagreement point: what member i gets if bargaining collapses entirely and no deal is reached.

• c_L(x) — leadership’s cost of granting concession x, increasing in x.

• c_L(t) — leadership’s own cost of delay. Convexly increasing in t.

• π(t) — the probability, at time t, bargaining collapses to the disagreement point rather than resolving at baseline.

• x_i* — holdout i‘s equilibrium reservation demand: the minimum concession they’ll accept.

Implications:

• An individual holdout’s leverage is determined by:

  • cost of delay (idiosyncratic to the member)

  • divisibility (a characteristic of the bill itself)

  • ratio of cost of concession (leadership) to value of concession (holdout)

• Leadership cost of delay non-linearly increasing (convexity) - this is an assumption built in to the model

Extension: Multiple Factions, One Shared Capacity

The base model treated the bill as a single scalar x and F as one holdout vector. Not all bills are single-ask, and once you look at recent cases side by side, two structurally different kinds of concession show up — and the base model’s existing machinery can tell them apart without adding anything new.

A SALT-cap fight and a Speaker election are different kinds of bargain. SALT is divisible: only members from high-tax states actually value it; a member from a no-income-tax state is indifferent.1 A Speaker vote is indivisible: it’s binary, and pays off identically to every member who backs the winning side regardless of how many others join them.

The bill becomes a vector, X = (x₁, ..., x_D), allowing severable dimensions, each one of the two types above.

Divisible dimensions get a support set — the subset of members who actually have a stake.

Outside F_j, u_i(x_j) = 0, the member is indifferent. Each divisible dimension then runs the original scalar bargain, unmodified, restricted to its own F_j. The faction on this dimension is the set of members for whom the concession is worth something.

Indivisible dimensions are where a faction’s size becomes the live question, and here the existing budget constraint (2n + p < m) supplies the threshold, just read from the faction’s side rather than leadership’s:

Below k_min, adding members increases the faction’s blocking weight. At k_min, the faction is already sufficient — leadership can’t pass the bill without them — and additional members past that point buy the faction nothing.

w_F(k) is the aggregate “blocking” leverage. Extraction per member peaks exactly at sufficiency and dilutes past it.

Because leadership’s tolerance — c_L(t), and the same vote-budget constraint — is one finite resource, dimensions with completely disjoint memberships and zero utility overlap are still implicitly drawing on the same capacity. Satisfying one dimension’s demand consumes some of what’s available for every other active dimension:

A bill with a single indivisible dimension at the one-seat margin collapses k_min to 1 — a faction of one is sufficient — and the extension returns the original single-holdout result unchanged.

The extension applies to margins wide enough that multiple severable asks can be live at once, but narrow enough that a rival coalition still can’t credibly bid for defectors.

Comparative statics (testable directional claims)

1. The number of concurrently active factions on a bill tracks the number of severable, divisible dimensions it carries — not the size of the margin.

2. A concession granted on one dimension tightens what’s available on every other simultaneously active dimension, even where factions share no members and no overlapping preferences, because leadership’s capacity is one shared, finite resource. Early concessions will be bigger.

3. An indivisible vote (e.g., Speaker vote) collapses to the base model’s single-faction case, with per-member extraction maximized at exactly the sufficiency threshold and falling past it.

Conclusion

Most of this is intuitive once stated, but a couple of things fall out that aren’t:

• The relevant predictor of how contested a bill will be isn’t the margin alone, it’s how many severable asks the bill happens to bundle. A wide majority carrying one clean, indivisible ask (a debt-ceiling raise with a single up-or-down rider) behaves like a razor-thin majority with one holdout — the margin matters less than the bill’s structure.

• “The faction” isn’t a fixed roster across bills — it’s whoever has a stake in that bill’s specific dimension. The functional membership on a SALT fight and on a debt-ceiling rider can be entirely different people, even when press coverage uses the same caucus label for both.2

• Bundling is a lever leadership can pull deliberately, not just a feature of how a bill happens to get written. Combining several divisible asks into one vehicle forces factions that would otherwise bargain independently to draw on the same finite capacity — which can dilute any single faction’s leverage, or just as easily create a more fragile package that any one faction can sink.

Predictions

1. OBBBA-style tax bills: where there are several divisible elements (SALT, sector carve-outs), concessions should land unevenly by order of resolution — factions that lock in early get closer to their ask; factions still bargaining once earlier concessions have drawn down leadership’s capacity settle for less, independent of their own underlying leverage.

2. A debt-ceiling bill carrying a single rider: treat it as the indivisible case — expect a faction size near k_min among holdouts, no internal division of the concession, and a resolution profile matching the original model's shutdown/deadline predictions (clustered near the binding constraint) rather than the OBBBA pattern above.

3. Appropriations omnibuses: holding margin constant, expect smaller per-faction concessions as the number of bundled asks rises — a crowding effect distinct from the deadline-clustering prediction from the base model.

That’s the band Wallner and Dayton are actually describing: DSG, RSC, Freedom Caucus and 2023 Speaker fight — none of them one-vote margins. The model’s prediction for that regime: a bloc holding at exactly k_min (Freedom Caucus, enough to deny McCarthy, no more) should out-extract, per member, a larger bloc with equivalent aggregate blocking power but more internal heterogeneity (RSC, historically).

The author is Nonresident Senior Fellow at the Foundation for American Innovation

Share

In a recent newsletter I scratched out a simple growth model showing how “regulation” (broadly defined) could be conducive to economic growth by facilitating innovation.

The basic intuition is this: a portion of labor focuses on things like scientific research and discovery which produces the new ideas which we call innovation. But in a new emergent field, a portion of these PhDs and other high-skilled workers have to spend time reconciling different methods and terminology so those new ideas can maximally recombine and build on each other to produce the next innovative iteration. A form of regulation can reduce this “translation drag” as a coordination mechanism: defining terms, establishing common measurement rubrics, facilitating interoperability, and so on.

Nobel-laureate Paul Romer’s canonical endogenous growth theory provides a tractable model on which to scaffold this idea. His includes a variable to account for specialized labor devoted to discovery of new ideas which is a core driver of economic growth in his story.

Below I briefly restate his base model before explaining two extensions introducing regulation as a variable in different ways.

Here regulation is a broadly applicable term, and later allows for non-governmental regulation such as self-regulatory organizations or quasi-private institutions.1

Base Romer model (w/ knowledge spillover)

Narrative description: labor devoted to adding to the stock of knowledge is the source of innovation-driven growth. Beyond the actual amount of labor, this is subject to the effectiveness of this work (research), and degree to which new ideas feed into and interact with other ideas.

Formal description:

A = knowledge stock, L_A​ = total pool of skilled R&D labor, δ = research productivity, β = knowledge-spillover parameter.2

Implication: growth depends on A which increases in level one-time through an increase in R&D labor if β < 1, or in growth rate if β = 1.

Extension 1 — regulation as coordination

Narrative description: in the absence of regulation, a portion of skilled R&D labor gets peeled off to work on interpretation, translation of existing ideas in lieu of producing new ones, creating translational drag on growth potential. Regulation can reduce translational drag but imposes it’s own cost (compliance). The effect of regulation is a priori ambiguous.

Formal description: instead of R&D labor (L_A) going entirely to production of the knowledge stock, some is peeled off into resolving knowledge frictions. The diverted R&D labor (L_R) then is characterized as:

θ = share of skilled labor (L_A) going to knowledge production, bounded by θ(R) ∈ [0,1]. R is the regulatory level: at R = 0 no regulation and maximum translational drag; at R = 1, maximum regulatory drag.

Effective knowledge-producing labor is the remainder, L_A − L_R = θ(R)L_A​, giving:

θ(R) can be further decomposed as:

T(R) = translational drag (decreasing in R), C(R) = compliance drag (increasing in R). Optimal R* maximizes θ(R):

Implication: Regulation’s effect on growth (via innovation) is optimally ambiguous, but non-zero. At R = 0, some of L_A​ is occupied with reducing translational drag. Adding regulation from zero frees more labor into direct knowledge production faster than it adds compliance burden. There is an optimal R* where the two marginal effects exactly offset which is an improvement in innovation from the baseline no regulation case.

Extension 2 — government v. industry regulation

Narrative description: There are now two types of regulatory regimes. (1) government-issued OR (2) industry-issued, via a Self-Regulatory Organization (SRO) or similar standards-setting body.3 One regime is chosen exclusively.

Government rules lead to 100 percent adoption across the industry, but will be further from a theoretical ideal set of rules (being more intermediated by political incentives and other frictions). SRO regulation is closer to the ideal in substance, but has only partial adoption (never fully reaching 100 percent).4

Formal description: Regulation R is now split into government (R_g) and voluntary (R_v). Both are subject to different coefficients for adoption level a and substantive quality (fidelity) φ. Under either regime, the other is set to zero.

• Adoption: a_v(R_v) ∈ [0, ā], ā < 1; a_g​= 1

• Fidelity: φ_g​, φ_v​ ∈ [0,1], φ_v > φ_g (note: phi renders slightly differently in LaTeX below)

Adoption and fidelity both discount the effective quantity of regulation entering a single shared drag-reduction function τ, which is common to both channels, lower-bounded by 0, characterized by decreasing marginal returns.

Only one regulatory regime operates (government-issued OR SRO-issued):

Under government regulation, translation and compliance costs are given by:

So total combined effect is:

Which optimizes at R_g* such that (same as Extension 1, before expanding):

Under voluntary regulation, translation and compliance costs are given by:

Combines as:

Optimizes such that (requiring cross-partials since adoption is now also a factor):

First-order conditions for each given by:
R_g:

R_v

Government is optimal when:

Voluntary is optimal when:

Implication: Government regulation is preferable when the translational drag it removes, net of what it costs to comply, exceeds what the best voluntary standard could net after accounting for the same tradeoff. Because of universal compliance, every dollar of Cg​ is paid, and every unit of τ is realized.

Voluntary regulation is preferable when it’s fidelity advantage (φ_v​>φ_g) is sufficient to compensate for adoption holdouts.

Empirical question: to what degree is SRO/industry-developed regulation higher quality than mandatory government-issued regulation. We assume it is higher-quality to some degree in the model but the magnitude is important here. It has to compensate for what’s lost to non-adoption (free-riders).

If you suppose the policy regime is such that industry-developed rules are “voluntary” in the sense of how they are developed and who can participate, but nonetheless legally binding, then this approach will always be superior.

Hypothetical Illustration: (assuming baseline growth ≈ 3%, δ = 0.08, L_A ​= 1, A^β-1 = 1)

g_A(R_g) / g_A(R_v) / g_A(0)

The author is Nonresident Senior Fellow at the Foundation for American Innovation

Setting the table…

• In probably the most deregulatory Administration in history, I worked in the White House reg. office.

• As an abstract maxim, I have for a long time believed minimizing the aggregate costs of regulations is a necessary condition for maximizing economic growth—for prudential and institutional incentive reasons.

• My office used to be down the hall from the guy who coined “Bootleggers and Baptists.”

All of which is to explain this article’s title is meant to be mildly provocative. And also to hint it is a view arrived at from a very distant ideological origin. But what I mean “regulatory costs” is only within a specific category or condition, which I explain below—not a blanket advocacy for regulatory takeover of the economy. lol

Okay, How Much for a Single Toilet Seat?

For good reason we expect the government to economize wherever possible. Not that we believe it is good at doing so, but given the choice between the Defense Department buying $600 toilet seats and $7,600 coffeemakers or cheaper ones, the American public has a pretty obvious preference.

The infamous coffeemaker was the kind that goes on antisubmarine aircraft and akin to what commercial airlines pay because “you can’t just put Mr. Coffee in an airline.” The price of the other, not actually a toilet seat, was partly an artifact of government budgeting.

So it’s understandable that for years our military industrial base (MIB) defaulted to the cheapest options when it came to telecom equipment and security cameras. After all, they were still fairly high quality and got the job done.

Unfortunately the “price” does not reflect the “cost.” Just ask the victims of 2025’s DC mid-air crash advocating for mandatory ADS-B systems against concerns about the price.

Can You Hear Me Now?

As wireless communications, mobile phones, 5G, and wifi became a necessary part of the economy, DoD had to source all the equipment that makes that possible. When you’re talking about DoD-wide needs, you’re really talking about the broader MIB. Bases, dual-use airports, duty stations, commissaries, and all the contractors and vendors intertwined needed this communications infrastructure.

Except in highly-specialized, unusually sensitive settings it made sense to buy the lowest cost item that does what you need. Which is fairly simple in most cases.

Turned out the equipment best satisfying both conditions by far was all made in China. Chinese companies Huawei and ZTE (and various subsidiaries) ended up supplying most of the telecom tech for this purpose. Why?

• notwithstanding what we know now, they tended to be pretty high quality

• they priced 30-50 percent lower than anyone else

Base stations, routers, switches and much of the adjunct equipment used by our MIB came from these companies—as it does for much of the world. Outside of the MIB context, many rural areas which rely on a few or one wireless carrier were dependent on Chinese tech. Moreover many of these areas are adjacent to bases so it’s a double whammy. Of course it’s not limited to the domestic footprint: we have bases located across the world. Beyond the same integrated Chinese tech, Huawei phones and the like, quite popular in much of the world, were sold in on/near-base retailers. And why not? They were cheap and reliable.

As you may know we eventually learned that much of this technology permitted “backdoors” allowing outside parties to compromise security conditions even to the point of fully taking control.

All because we wanted to save on costs.

Some of these risks were known early on, but not widely accepted and far from universally understood. But cybersecurity risks are not conceptually new, and as I’ve written elsewhere, the specific nature leads to systematic deprioritization in both private and public sectors.

But that’s not regulatory, right?

Rip-and-Replace

Years on with the vulnerability now well-documented, savings or no this could not go on. Something had to be done.

For the government (including military) and all its contractors, the 2019 NDAA’s Sec. 889 mandated a regulation to address this. All federal acquisition is subject to a specific area of the federal regulatory code called appropriately Federal Acquisition Regulation. This is an incredibly dense, unintuitive, specialized body of rules that technically governs all government purchases.1

Most small vendors don’t ever immerse themselves in it because most sub-agencies and offices focus on particular sections of the code and make those available to the types of businesses focused on their specialty. But these rules also include government-wide conditions—things that should be applicable no matter what the product or use. There are security requirements in there for sure, including cases requiring domestic-based vendors.

The thing is…when it came to cameras, the kind used for perimeter monitoring and the like, used at bases and civilian federal buildings, we almost exclusively used equipment from Chinese company Hikvision, with the same backdoor vulnerabilities as the telecom equipment. But they didn’t have that name on the box. Hikvision is an OEM which makes essential components in cameras sold throughout the world by US and European companies like Honeywell, Carrier, and Bosch. Even some of those components don’t have giveaway brandingj, as they’re made for intermediaries and end-sellers using blank or custom branding. Many companies didn’t even know they were selling Chinese-originated products.

So along with the NDAA provision, which required through a rulemaking that all MIB contractors certify they have removed any elements from a list of blacklisted companies from their government-nexus footprint, Congress appropriated $1.9 billion to aid rural and small carriers in replacing their telecom stack. Collectively we call this process rip-and-replace. And it’s as disruptive as it sounds.

In theory the regulation means a contractor could violate federal law, or at least risk blacklisting, by falsely certifying compliance. But there’s no operative enforcement mechanism and no meaningful way to audit. The largest vendors risk a lot so they probably complied in good faith, but smaller ones bear less risk and may be ignorant of the rule.

One shouldn’t hazard a guess as to the degree of remaining exposure of our defense infrastructure, not to speak of the whole government. It’s unknowable. Confidently claiming full compliance would probably come at a nearly infinite cost.

To date there isn’t a fully domestic end-to-end telecommunications stack. The closest is Nokia and Ericsson so defaulting to “American providers” isn’t an option.

Just Own Up to It

I posit a defense of regulatory costs because all of this would have been feasible, and waaaay cheaper, if applied ex ante. The point is that these security costs are real and borne one way or another, either in the explicit recognition in policies including regulatory benefit-cost accounting, or later in painful retro-correction. More costly correction.

Fortunately some of the universal security risks like cybersecurity have become more salient but only quite recently. The story of Anthropic/Mythos - which BTW is still designated a “supply-chain risk” by the Pentagon - is about AI yes but cybersecurity risk in particular. And in an odd way it’s fortunate the risk framing isn’t about a specific vulnerability but a more global notion of cybersecurity risk. That’s a necessary condition for more earnest accounting for it in policy generally.

• It means Congress as it appropriates acknowledges the trade-off of getting less of what they want because they’re actually paying a “cybersecurity markup.” Again, they’ll be getting less either way, and costing it up front is cheaper.2

• It means regulatory analyses face up to the difficult truth of full cost accounting when developing rulemakings. Part of this process involves considering then deciding among multiple approaches (e.g., performance standards vs. design rules) and this should absolutely continue to favor the least costly option that accomplishes the policy aim. But “least costly” still applies when fulsomely acknowledging security risks.

Going forward this is absolutely germane to AI specifically. Mythos and Fable and the ensuing fallout finally took aim at the illusion of gob-smacking innovative AI capabilities will self-correct latent risks like cybersecurity. It could! But it’s not axiomatic. Markets are incredible at arriving at optimal solutions along multiple dimensions: costs, price, availability, aesthetics, and indeed safety. Further, many regulations as designed lead to consumer/business responses which actually inverts the purpose. Think of drivers being more reckless in response to safety features. But some externalities exist and are not foreseeably addressable through market mechanisms. At risk of tedious repetition: the resultant costs are borne somewhere at some point—consumer prices, national security, fiscal costs,…

There’s been a vocal and highly influential cadre of AI policy voices who exhort a kind of knee-jerk rejection of regulation in terms of imposing unnecessary costs on innovators. [Insert misapplication of “regulatory capture” here as one vector of this argumentation] Beyond the demonstrative internal contradictions, fallacies, and logical incoherence in much of this, these risks (read: costs) are already imposed. So think of it as displacing those costs with explicit regulatory costs if it makes it more palatable.

The author is Nonresident Senior Fellow at the Foundation for American Innovation

It’s in the mail

Elsewhere I weighed in on the widening conversation on Congressional modernization, noting some institutional barriers to 21st century legislating. But I omitted a first-hand anecdote that’s kind of an extreme encapsulation of the problem.

The problem, briefly: Congress is slow to update processes and technologies to maximize effectiveness.

An example: When the Executive Branch sends official communications to the Legislative Branch, the recipient has to attach an official received date. For standard recordkeeping hygiene but also because that date can meaningfully affect legislative options. Like the Congressional Review Act (CRA) which permits Congress a fast-track mechanism to repeal regulations. It’s cabined by a deadline attached to the reg’s “received by Congress”1 date which in practice turns out isn’t what you think.2

Typically agencies submit the required info separately to the House and Senate, electronically and by mail. As you can imagine the latter is much slower and involves multiple steps.3 Weeks could elapse between submission and delivery. Even physical delivery didn’t constitute “received by Congress.” It would be only when a staffer actually opened the envelope AND entered it into a database. Only the hardcopy counted.

It means the window for Congress to act under the CRA came down to the moment a staffer decided to go through the mail.

And it was a different person for each chamber. Yeesh.

We had to find a quicker way and a pandemic didn’t help. In 2020 what was a lethargic pace became catatonic. For the biggest items, very diligent agencies sent a courier, which still didn’t solve the open/log lag, but it helped. Congress’s early COVID protocols prevented non-staff entering the building. No way for offices to sign for delivery.

So I called the relevant offices and asked: “Since we already submit by email anyway, can we treat that timestamp as operative?” The discussion that ensued raised issues both legitimate and tedious. But the answer was “no” and trust me the reasons were not ones you’d find compelling.4 It boiled down to not having the adequate “systems” in place to do that.

In the end I advised - as there was no instructive adjudication on this - agencies to attempt delivery by courier for anything sufficiently important (entry restrictions were softening by this point) and we would use the signed handoff date as a record for any potential challenge.

The point is Congressional modernization, updating it to modern times, is absolutely needed but you probably have no idea how deep the necessity goes.

Subscribe now

Human-in-the-loop (redux)

In March the always excellent ChinaTalk podcast () touched on the at-the-time surprise DoW smackdown of Anthropic. I commend the discussants, and without intending to but at risk of being precious, there was one frustration.

One guest asserted in the absence of strong mutual alignment there are “laws” governing defense use of force that would be applicable and binding to AI. No one corrected that framing, including when the subject of direct experience developing DoW’s “rules” came up. This distinction is important.

I understand them as referring to the combination of then-DoD directive(s) and the JAIC-developed internal AI guidance issued by Sec. Esper in 2020. The point is the lack of actual “law” (strictly defined, statute) makes navigating novel policy questions that much harder for both parties.

• There’s little to no expectation of stability in the rule regime. I fully expect the Esper memo to soon go conspicuously trashed.

• Truly novel questions emerge, by the week, which have no guiding principles on which to build an answer.

Anthropic is not blameless here but disputes like this were inevitable in such an impossible policy environment. I’ll go strong buy on the “it’ll happen again” futures contract.

Also note a compelling rejoinder by guest (“AI is a bureaucratic technology. So is fighting war”).

• Citing a different theater of war, an AI targeting system exhibited something like a 10 percent Type I error rate with little reconsideration by “human-in-the-loop.” Remarkably close to the rhetorical example Brad Carson gave arguing such an oversight heuristic is insufficient. I still believe it’s a necessary condition, but certainly not sufficient.

• To the degree you do worry about scaled surveillance capabilities, the dilemma is real. Reminded me of the Census Bureau led study on re-identifying otherwise “de-identified” public data. Using what to mere mortals like us looks like totally impersonal Census elements, sophisticated data-jockeys could re-assemble private identities. Years before ChatGPT.

Don’t Believe Me, Just Watch

The Dispatch reported on the effort to put cameras in courtrooms. One bill requiring them in the Supreme Court and another to make it easier in lower courts unanimously (v/v) passed out of committees.6 Of course the high court maintains consistent institutional opposition. (In 1996 Justice Souter offered no quarter to such an idea, and also maybe didn’t know what a camera is)

The intuition is understandable, and not surprising the committee vote tracks popular opinion as the premise itself is democratic. The arguments in favor are also seductively compelling. On net, intuition says it’s been negative, by a a small-to-moderate amount [medium confidence].

The principle-based justification for cameras in Congress, say, is not conceptually wrong: accountability is monotonically increasing in transparency. Being watched leads to better behavior, ceteris paribus. But empirically that hasn’t been the outcome. tells the story.

One advocate argues visual information is not adequately proxied by other means. But in my experience the structure of policy decision-making is predominantly written modes with some verbal communication.7

Worth asking: Is visual information extraneous? Biasing?

People had eyes at the country’s founding - what were the rules then, if only implied? Okay “camera” meant something else altogether, but how did they account for the role of visual information?

To be fair no one could have extrapolated the proliferation of social media and other tech, much less Congress’s endogenous behavioral changes (i.e., preening, confrontation, “clip-art,” etc.). It may have been nonetheless unavoidable as committees are entitled to their own rules and nothing would stop members/staff from self-broadcasting.

The OJ Simpsons trial broadcast…granted, a 1-of-1, but can we say it had no effect?

The Supreme Court has thus far managed to bar the door, inoculating it from the symptoms of this particular bug going ‘round. But immunity fades and it only takes a small viral load to incubate into illness.

Etc.

Law Firm of AI & AI: & on got into attorney wages as a function of AI. Great discussion and insider-y understanding of some law firm idiosyncrasies, but seemed like they were unable to come to a mutually-agreeable model. Fairly simple comparative statics (conditional on your assumptions):

If you believe AI increases productivity of legal work (broadly), then reduces marginal cost. And almost all legal work is adjustable at the extensive margin.8

• For “low value”/standard tasks: shadow price of marginal billed hour goes down, elasticity > 1 → aggregate demand increases while revenue per matter decreases.

  • Ex: legal support services increased after e-discovery.

• For “superstar” segment (high value/risk matters, M&A): lower cost mostly captured as higher firm margin. Increases return to reputational capital + specialized expertise + management. Revenues flat to slightly up.

  • Three basic billing models: hourly, fixed retainer, contingency. If you think there will adjustment in payment models → clients open more matters.

Solve for x = the number of justices: A fascinating prompt on Advisory Opinions’ Supreme Court term roundup. asks: what is the correct methodological lens for assessing the court? Presented as implicitly unbounded. Each panelist in turn ably grabbed the ball and went ISO for a couple possessions.

went trick play. He invoked Condorcet’s Jury Theorem and the Wisdom of Crowds maxim to cheekily suggest expanding the court would improve it’s quality, and therefore he follows the median judge as a proxy for how the court is doing.9

But he forgot to mention the one super critical necessary condition: the population average from which the deciders are drawn have to be ≥ 50% correct, accurate, smart, whatever. Because when it’s not true the formula predicts as size increases accuracy collapses to 0.

Share

There’s an old joke about founders: “Big-picture thinker seeks detail-oriented doer to handle implementation.” That is, ideas overpopulate, execution a rare species.

Effective implementation, be it software or policy, is determinative of success. The “doers” have to solve actual problems - technical limitations, resource constraints, maybe marketing, and if it involves a sufficient number of people, inevitably boring organizational management (read: internal politics).

Ever true in government, this is what torpedoes even amazing ideas—failure to navigate the incentives, which we’ll return to shortly.

Data, Data Everywhere, Not a Bit to Sync

Consider data: the federal government relies on gobs and gobs of it from every corner of the country, from individual tax returns to niche corporate reporting in small industries. This data is both a necessary input to government functioning, hugely valuable output for the private sector, and every permutation thereof.

And it’s damn costly.

First off, it’s inefficient. Much of it is duplicative. Overlapping requirements mean the same data is reported to multiple nodes of the Fed. For example, a company hit with a major data breach may have to submit the same information to DHS, SEC, and DOD. Add in state authorities too, each having justifiable reasons to need it.

Pity this poor bank CIO amidst a Code Red hack, getting yelled at by the CEO, trying to make sure they don’t have to reverse a bunch of credit card charges, now also has to not get crosswise with multiple federal agencies. And it’s not like it’s the same form you can just photocopy. Each has a bespoke format and deadlines and so on.

The craziest part is all of this [gestures wildly] is advocated by literally no one.

No less true in more mundane contexts. Remember that old IRS meme?

Inconsistent data fields, duplicative reporting, and guess how all these stores of information get along. Completely seamlessly and without friction? No, of course not!

Every office in every agency has a statute (ideally, ignoring for now) that tells it what to do - or more often tells it to figure out what to do - and for the most part didn’t bother to check what laws and regulations applied elsewhere.

At some point a Congressional committee was ginned up about something and determined to handle it. Say, shark fin sales (H.R. 81 - Shark Conservation Act of 2009, passed as Pub. L. 111-348, codified at 16 USC 1801). Of course they don’t actually descend into the waters and monitor all the cartilaginous fish appendages. They tell NOAA to look into it (50 CFR 600). And if asked, and NOAA can’t provide the data showing progress, Congress might not move on just because the implication is gathering this data from private parties is costly.

Now multiply this over hundreds of times over multiple decades, each one deemed important and each one requiring data specific for that purpose.

The point is the solution is obvious: consolidate all the data across agencies, and make it available to them. So why hasn’t it been tried?

Oh it’s been tried.

There Oughta Be a Law

Congress even legislated it. The OPEN Government Data (2018) is one such example. There’s been some progress but it turned out implementation is more difficult than expected. It comes down a few roadblocks:

1. Everyone agrees the aims are laudable. No one objects. But things like this are simply a low priority, and frankly somewhat boring. From the senior political level, it’s almost never going to be top of their to-do list. Which is to say, they’ll task staff with moving the ball, but not at the expense of other goals. Inevitably this project will come at the expense of something else. At the career staff level, it’s rare this will be the most exciting use of their time.

2. Data fields a normie would identify as identical, at a granular level exist in slightly different forms across the government. Small differences in format, reporting method, estimation methodologies, even coding language make for large frictions to concatenation. The individual agency incentive is to insist: “you should adopt our format.”

3. It turns out privacy is a big deal in this country. Any information applicable to an individual or organization has to be considered in the context of how it was provided and what can be imputed when it’s combined with other information.1 The longstanding cultural bias in government is to err on the side of caution. That’s generally a good thing, but it does obstruct otherwise commendable efforts.

I observed this firsthand, especially #3 during my time in government. An effort to share Social Security Administration info on deceased beneficiaries with Treasury’s Do Not Pay list was stymied by among other things major privacy concerns. Not just risk aversion, real legal risks. The Privacy Act of 1974 and CIPSEA in particular. While the public reporting focused on IT interoperability and formatting, which were real, the behind-the-scenes conflicts were all privacy obligations.

Cybersecurity

As you can imagine, it gets harder when you add data to regulations, but the dilemma is structurally identical. As GAO pointed out recently, even in spite of clear Congressional direction on the matter, attempts to streamline compliance in the area of cybersecurity are met with resistance. I don’t dispute GAO’s findings, that fiscal costs are a friction to implementation. But whereas GAO is not in a position to opine on organizational incentives, I submit fiscal costs are merely a proxy for the interagency incentive alignment problem. Yes legacy computer systems are a real and discrete barrier but these are not the truly binding constraint–we know this because even when Congress provides additional funding the outcome is the same.

While colleagues and I were digging into these issues, cybersecurity in particular, this was a persistent issue. Bipartisan consensus is rare, but not on this. This was not an issue of more regulation vs. deregulation. This was widespread understanding that significant resources were going to a totally unproductive use.

And not for nothing, the country’s cybersecurity was (and remains) woefully inadequate, so this was no small thing to solve.

Who Let the DOGE Out?

Whither DOGE and what’ve they got to do with it?

Well, armed with a seeming disrespect for authority and big talk about technical prowess, on this one problem it was a combination particularly fit for purpose:

1. Data, IT, software, y’know computer stuff…this is supposed to be what they were not only skilled at, but among the best in the world.

2. Afforded a, how to say this, DGAF remit.

So I was in this one case optimistic about DOGE’s likelihood to meaningfully break through the status quo ante constraints, being armed with this combination of expertise and something resembling political immunity.

It’s possible there was some success in this space. But for now we haven’t seen it.

And yet I still maintain the diagnosis of the problem is accurate, and DOGE was conceptually the best chance to solve it. So how do I reconcile that with their failure?

Simple. DOGE didn’t actually understand the problem they were trying to solve. That and/or for all their superhuman pedantic posturing, they weren’t really as well-equipped as they said.

The author is Nonresident Senior Fellow at the Foundation for American Innovation

You Say You Want a Revolution…

WIRED’s Uncanny Valley podcast broke down the will they-won’t they relationship between Anthropic (and AI labs generally) and the White House. They joked that it’s almost like national security policy is being left to AI execs who need better media training—the rift isn’t a disagreement on policy itself but failure to communicate in an effective way.

Elsewhere , while not fully absolving the government, said he believes Anthropic execs behaved poorly and a less bad outcome was possible.

More broadly I’ve wondered about the translational gap between Silicon Valley and Washington. I haven’t been shy about rinsing SV-types for failing to see the world as it is. And Brendan Bordelon’s debrief on Leading the Future’s big money election play isn’t flattering. So why has a culture built on incredible innovation and singular risk-taking been unable to land the plane?

The Grand Resign’s 2nd Maxim: Our moral technology is stubborn and slow to change

Substitute “culture” for “moral technology” and my theory is the two worlds reward different things. A recent transplant to the Bay Area, I observe a kind of reinforcement learning for heterodoxy, orthogonal bets, and yes money as a proxy for successful innovation. They don’t see the world as it is - political incentives and all - because the whole ethos is about remaking the world, creating a margin that doesn’t currently exist, optimizing over a decision set you don’t even see.

That’s not a failure. We need the heretics, cultural dissidents, hyperbolic discounters…awkward dudes who dress badly.

But as I’ve written before, political incentives are not put on pause while we wait for the optimal future to unfold. Democracy be like that. Our moral and cultural attachments may be sub-optimally inflexible. Frankly it’s a downside of an institutionally-moderated democracy (read: republic). Maybe we end up in a transhumanist future after all, but in the interim trying to artificially displace a rooted moral technology is likely counterproductive—the whole Marxist project was explicitly about this.

“There are, besides, eternal truths…that are common to all states of society. But Communism abolishes eternal truths…instead of constituting them on a new basis; it therefore acts in contradiction to all past historical experience.”

—Karl Marx and Frederick Engels, Manifesto of the Communist Party (1848), Ch. II, trans. by Samuel Moore (emphasis added)

I’m not comparing tech companies to communists or whatever. I’m arguing there’s an admirable intuition that says:

It may be righteous, even correct. But it’s not self-executing and there’s a whole lot of fortification in front of that castle.

[In the interim, Executive Orders ain’t what you’re looking for: I may write an explainer on this]

FDA(I)

There was about a 10-day flareup around the idea of an “FDA for AI.” (ed. I’ll say this about the Trump Administration: whatever you think you’re aiming at, it’s not what’s coming) This piece ( and ) lays it out well and short take: “bad idea, don’t do it.” I agree.

I’ll quibble with some minor framing not by them specifically but in broader discourse around this general idea (specialized AI regulatory agency).

• Speculative: Intuition says this was never a serious consideration in the Trump Administration. And the instinct isn’t because of policy grounds, but because the sole on-record source was WH NEC Director Kevin Hassett.1 Whatever criticisms he’s received of his recent flexibility on economics may or may not be fair, but he’s not as good a communicator as he thinks he is.2 Again speculative, but it reads as a person marginally-familiar with the topic but otherwise intelligent thinking he heard a compelling analogy and running with it before getting hip. I strongly believe it was never speaking to an actual de novo agency. But I understand it meriting forceful response either way.

• In an extended interview with a16z’s AI policy lead I often reference, but you can find it throughout the Leading the Future-aligned advocates, there’s an emphasis on the fundamental nature of frontier LLM development: “just science, it’s math.” Akin to basic research, axiomatically general use research. You wouldn’t regulate science would you?

  Let’s dispense with this. Giving consumer-facing technology names like “Code,” “Codex,” “Design,” ”Translate,” and “Studio” is high-key giving applied use as the kids might say. An electrical transformer has pretty general application built on some fundamental principles of science. They can still be subject to strict liability.

• A related point is about what the FDA actually does. For sake of argument let’s presume the “regulate uses not development” line cuts toward no responsibility for labs. The FDA only regulates uses. It licenses particular combinations of chemicals and structures for a particular application. And really only for the marketing of that application. And decisive legal precedent stiff-arms regulation of the practice of medicine, which is why doctors can off-label prescribe drugs.

  Any company can theoretically stir various recipes of the constituent ingredients of drugs without ever crossing paths with the FDA, and some even sell those without doing so under the supplement category—as long as they never market and sell alongside a medical claim. Granted an FDA for AI was never a real proposal so maybe this distinction was implicit, the whole thing limited to a rhetorical flourish. But both notions—FDA-type structured regulatory scheme and frontier models as truly general purpose distinct from harmful uses—taken at face value would’ve meant…ya’ll never had anything to worry about!

MS WhatNow?

(I don’t have a link to the transcript, but I swear I am accurately characterizing it)

I have a long-running obsession with talking heads who lack one or both of two things:

1. a baseline level of charisma;

2. demonstrable expertise in some field.

My go-to example for (2) is an “Aviation Expert” who answered “They’re going to look for the black box flight recorder and try to piece together what happened” in response to the question “So what will authorities be looking for in light of this airline accident?” This is not a demonstration of specialized knowledge; it’s Guy at Bar level small talk (didn’t help he was a charisma black hole).

The MS Now example is something else. I can’t tell exactly if she just thought the audience is ignorant to this is as a possibility - the totality of a document could mean something distinct from the first line? Or she was genuinely expressing surprise. It’s not lack of true expertise: it did tell me something I didn’t know before, just not sure what, and it wasn’t about the court opinion.

Well if you’re gonna act that way…

Wondering if the Trump Administration’s capricious “policy” on frontier models will lead to a soft realignment of resistance.

Cotopaxi Protest (Google employees, 2017) —> Patagonia Protest (Execs/Boards, 2026).

Share

Etc.

N.B.: Old pal Chaz dropped a a claim about The Office I had never heard. The UK version ran half as long as US one because the former is essentially hopeless and static while the latter reflects American optimism and self-reinvention. While both lack self-awareness, David Brent is surely a darker character than Michael Scarn Scott (especially after Season 1). Americans on average definitely work more. Does that reflect a more upbeat working environment? Just cultural-driven preferences?

A recent episode was trying to break down the steady productivity growth lead the US has had over Europe for about 20 years. Guest Jason Furman offered the possibility of a higher preference for leisure over income.3 A co-host suggested the rigid labor laws make it hard for people to change jobs and the best companies to make adjustments.4

The latter is consistent with a cultural habituation to pessimism. To resignation to the steady-state.

In contrast to American aspirational immodesty, satisfaction as goal rather than state of being.

Happy 4th.

How much damage can one Representative do?

James Wallner’s new concise essay (“The Tiny Rule That Can Paralyze Congress”) breaks down the peculiarities of House pathways for partisan bills in a tight margin majority.1 I’m not much interested in politics per se, only how it constrains the policy choice set.

Which got me thinking about the bargaining dynamics of a one vote or near one-vote margin majority, and whether a model could functionally describe such a Congress in a way that tells us something about how different types of votes will unfold.2 The result is somewhat intuitive, even obvious, predictions but also some less intuitive testable claims from a closed-form normal game. And that it tracks our observed recent history is a soft indicator of plausibility.

Modeling the situation

The foundation

The literature on bargaining and specifically bargaining in a legislative structure is extensive. There are some canonical and robust foundations on which to build.

• Riker’s (1962) “Minimum Winning Coalition”3 theory describes a rational proposer minimizing coalition size to maximize each member’s share of a fixed surplus.

  • Baron and Ferejohn (1989)4 formalizes Riker’s theoretical model and supports the MWC outcome.

• Groseclose and Snyder (1996)5 offer a consistent bargaining setup except with a rival proposer which predicts competitive bidding to “buy” pivotal votes. Ultimately I do not believe this describes the current House makeup because it requires the competing coalition (in this case Democratic minority) to demonstrate a non-zero value in buying off a defector. However it is instructive for recent specific examples, one of which I witnessed first hand (more below).

• The Rubinstein (1982)6 characterization is a generalizable alternating-turn bargaining game revealing that parameterizing “impatience” (cost of delay) is critical to describing an equilibrium. Again the premise, per Riker, is there is a fixed surplus from which to distribute shares, to which delays in a successful outcome (this case passing a bill) imposes differing costs on the bargaining parties and therefore affecting leverage.

  • Observable delay indicates incomplete/private information and information revelation models per Kennan and Wilson (1993)7 (strikes as costly signals of credible commitment).

• I believe the model describes equally a situation where there is a single pivotal marginal vote or more than one (within some bound well outside the current situation) because it does not rely on sole holdup power as a sufficient condition to establish bargaining leverage. Shapley and Shubik8, Banzhaf9, and Owen10 are canonical contributions supporting legislative bargaining as not a purely individual, but effectively coalitional blocs game.

Structural elements from the above inform the model.

The Model: Narrative Description

The players

In this model we imagine “leadership” is the proposer, with the implicit ability to set the agenda by limiting floor consideration of other legislation. Negotiating with leadership is a “faction” of holdouts. The faction isn’t a unified or otherwise coordinated bloc (though it can be). It is collectively the undetermined number of individual members, each unrestricted to have their own preferences. But we can include them singularly in one vector without loss of generality.11

The bill

They’re bargaining over the outcome of a single bill with a range of policy outcomes. At one end is leadership’s ideal outcome, and at the other is the holdout’s. Because the collective faction is a heterogeneous set of individual holdouts, the policy range is an index that allows the variable to represent the preferred outcome for that member specifically.

Accordingly, a holdout’s ideal is at one end of the outcome range, but there is some point below that sufficient for an agreement because holdout is not costless (“constraints”).

The constraints

Both leadership and holdouts bear some cost from delay, individuated for the holdouts. The latter’s cost is tied to electoral risk: for example a super-safe district, high reelection expectation, means the member can bear more delay at less cost. Leadership may bear more of the cost of missing a debt limit deadline, forcing a government shutdown, or inability to get on to other substantive priorities, because they bear a broader caucus-wide electoral incentive to show a functioning majority.

They also of course bear some cost in moving from their preferred policy outcome toward a holdout’s. Also more time goes by, there’s an underlying increase in the likelihood of breakdown and no one gets anything.

The Model: Formal Description

Players

F is the set of pivotal holdout members — members who can’t be substituted for alternative bargainers. Each individual holdout is indexed i. Membership in F_i is established beforehand. Leadership (L) acts as the sole proposer.

Variables:

• x — the policy outcome, running from 0 (leadership’s baseline bill, no concessions) to 1 (a fully-conceded bill matching a holdout’s ideal).

• u_i(x) — holdout i‘s payoff from the bill passing at concession level x; utility increasing in x.

• c_i(t) — holdout i‘s own cost of delay, as a function of time t remaining before the deadline. A safe-seat member has a low, close-to-flat c_i(t); a vulnerable-seat member has a steep one.

• d_i — the disagreement point: what member i gets if bargaining collapses entirely and no deal is reached.

• c_L(x) — leadership’s cost of granting concession x; increasing in x.

• c_L(t) — leadership’s own cost of delay; convexly increasing in t.

• π(t) — the probability, at time t, bargaining collapses to the disagreement point rather than resolving at baseline.

• x_i* — holdout i‘s equilibrium reservation demand: the minimum concession they’ll accept.

The holdout equation says member i accepts a given offer only if the payoff from that offer, net of current delay, is at least as good as their risk-weighted expected payoff from walking away.

Division vs. size are two separate levers.

• Division is governed by the ratio of leadership’s cost of delay to the holdout’s own — the more impatient leadership is relative to the holdout, the more of the available concession goes to the holdout.

• Size is governed by the ratio of how much the holdout values a unit of concession relative to how much it costs leadership to grant it. A concession can be cheap for leadership and valuable to the holdout.

The budget constraint is the vote-counting condition: n is the number of GOP holdouts casting an outright “no”, p is the number “present” or “absent,” m is the raw seat margin (R − D). A “no” costs the coalition two points of margin, a “present/absent” costs one, and leadership’s total tolerance is the margin minus one.

Comparative statics (testable directional claims):

1. leadership’s rising cost of delay near a deadline increases level of concession.

2. a holdout’s own rising cost of delay decreases what they extract.

3. a more favorable ratio of the holdout’s valuation to leadership’s cost of granting it (i.e., more divisibility) increases what they extract, independent of patience.

Payoff Matrix

“Lawler case” refers to Rep. Lawler’s demands nearing passage of the 2025 tax bill (OBBBA). In contrast to Rep. Luna, he is in arguably the most competitive district in the country which makes the implied cost of blowing up the bill incredibly high. But in our model a comprehensive tax bill is highly “divisible” and offered a large payoff (SALT deduction) if he is successful (which he was).

Conclusion

Okay so most of that you could draw from intuition, not exactly pathbreaking. But here’s a couple implications that are perhaps not so obvious:

• News reporting tends to focus on the naive partisan margin at any given time as shorthand for the bargaining dynamics. In a sense that tells you who is “in the room” as a candidate for being pivotal vote but it is incomplete. Two (or more) members can be equally likely to be pivotal (“pivotality” in the literature) but get widely different results.

• It is determined by:

  • cost of delay (idiosyncratic to the member)

  • divisibility (a characteristic of the bill itself)

  • ratio of cost of concession (leadership) to value of concession (holdout)

• Leadership cost of delay non-linearly increasing (convexity) - this is an assumption built in to the model

• High drama like shutdowns and extended speakership ballots are the result of imperfect information, a means of information revelation. In earlier times of old-fashioned logrolling, leadership already has a good idea of how much it costs to “buy” a member’s vote.

Predictions

1. NDAA/SAVE Act: House leadership offered Rep. Luna a package as a single vote, but still remain textually separable. This was the “cheaper” offer which she rejected. She has now signaled credible resolve by bearing real cost and has a safe seat. The model prediction is that we’ll end up with something resembling her ask: the bills are actually linked into the same text to be voted on (NDAA+SAVE). Doesn’t guarantee Senate can’t undo it but they do now bear a higher cost of doing so.

2. Appropriations/shutdowns: less clean and fairly familiar. Concessions will cluster in the final 24-72 hours before deadlines.

3. FISA: because as I understand it much of the FISC certifications effectively run through March 2027, we haven’t approached the meaningful deadline which is why we didn’t see much bargaining/concession - leadership’s cost function never reached the spiking point. We’ll see that happen in the Spring.

Potential Empirical Testing & Extensions

Empirical

• Input Cook PVI scores to specify the electoral risk parameter (which inputs to delay cost). Compare predicted level of concessions with actual ones for sample of major bill fights.

• The literature has multiple methodologies for computing a “power index” by tabulating across conceivable winning coalitions, based on voting pattern, what is a given member’s pivotality. The model then provides a testable prediction going forward.

Extensions

• Like James points out the unfolding of coalitions and bargaining is not fully instrumented by votes, but all the procedural maneuvering and minutiae strategically deployed leading up to a vote (or preventing altogether). This accounts for none of that. Could be possible to introduce some elements as an additional leadership lever.

• d is what a holdout gets in a no-deal scenario, but it’s exogenous right now. Possibly it can be derived endogenously from an election subgame.

• Policy space exists across a one-dimensional scalar x. Of course not always the case so could be made multi-dimensional but increases complexity considerably.

• This represents a one-shot normal form game. In a repeated game you can include carryover effects which feed equilibrium strategy, like credibility or reputation.

[ maybe you have thoughts]

Share