“[P]eople who have been unbelievably successful in one space, therefore thinking they have totally figured out another world that is totally different.”1
- Anonymous, reading Silicon Valley’s foray into politics
[Disclosure: I was previously a registered lobbyist on AI and related issues, including in-house for an advocacy group opposed to the proposed moratorium. Earlier I was involved in the first ever AI Executive Order as a White House employee.]
“All failure is failure to adapt, all success is successful adaptation”2
What would be the optimal strategy for a coalition of AI companies and investors seeking a national policy conducive to innovation? In the case of Little Tech/Leading the Future, the play was a single federal moratorium on non-federal laws.3 Having thus far been unsuccessful in codifying the first goal, though with intermediate success in an Executive Order to that effect, I was prepared to argue the optimal strategy would focus primarily on the state level.4
Turns out they agree. While drafting this piece, Politico published a feature outlining the explicit strategy to do just that. Nonetheless let’s explore how I and they got there.
Read the room because it gets a say
First they failed to realize their creatio ex nihilo policy may have been substantively optimal but strategically misaligned. They’d have to acknowledge that this is not a static environment but a dynamic one with multiple players. With some backward induction they’d anticipate that even “in war the loser gets a vote.” That is, the Silicon Valley mindset of “0 to 1” doesn’t account for naysayers and objectors, but politics does (to be fair, even sophisticated AI models have trouble with backward induction leading to the “turtle trap”).
Personally I discount the recent surveys and commencement speech theatrics indicating a strong resentment of AI proliferation. Like former colleague Evan’s point about distinguishing between stated and revealed preference, my guess is these data points also reflect an unresolved tension between the two. Still it’s a long leap to characterize the current popular sentiment as overwhelmingly positive. And after all, stated preferences feed into political incentives. So the strategy of coming off the top-rope with a decadelong Frankensteiner was not going to land in a receptive political arena, even with a friendly White House.
Which means arguments for such a law even if in good faith would meet significant resistance.
For instance, the core message that a “patchwork”5 of State laws risks an unworkable barrier to innovation may be true (I have doubts), but it’s incoherent alongside anecdotal evidence we see everyday. Again from Evan:
Here is what companies are doing under conditions they argue are an existential threat. AI hyperscalers are spending hundreds of billions on AI infrastructure. Venture capital into U.S. AI startups hit $211 billion in 2025, up 85% from the prior year. The total value of U.S. unicorns nearly doubled in 2025, from $2.4 trillion to $4.4 trillion, with AI startups driving the surge. OpenAI alone hit a $500 billion valuation in October, becoming the world’s most valuable private company. None of this is what an industry being strangled by state regulation looks like.
Weeks later, already out-of-date.
While drafting this sentence, another AI chip maker IPO’d to $5B+ market cap while the implied valuations of others continued to increase (SpaceX, OpenAI, and Anthropic all sprinting toward record-setting IPOs as well).
Of course in regulatory economics, one must acknowledge the invisible cost to future innovators: the garage-tinkerers (ok, WeWork vibe-coders) and college senior capstones, not just the big movers right in front of us.
[ed. This might be a good time to point out that at one time or another, the following companies officially identified as members of the Little Tech coalition contra Big Tech (Facebook, Google, et. al.): SpaceX ($1.75T), Palantir ($312B), Anduril ($61B), Scale AI ($29B after acqui-hire by Meta, unclear if they returned their decoder ring).6 Also funny, one of those companies ran ads against a Congressional candidate attacking him for…previously working for that company, which may be backfiring. Points for self-awareness?]
To the degree we observe binding constraints to AI development, it’s pushback on data center construction. That’s a knot a moratorium can’t easily untangle as it’s tied up in hyperlocal NIMBYism, zoning, energy costs, environmentalists, and so on. Not overzealous state AI laws. The other element that reads to me as significant are a growing number of lawsuits, again not attributable to new state AI laws. They certainly do not point toward lethargic deployment of technology.
“Say what you mean, mean what you say”
Let’s concede that it’s a sincerely held belief–state regs are existential–because of effort, cost, and political alienation. But it also tells us, because all policy is future-looking and there’s no apparent state-level binding constraint right now, that their optimal solution is moratorium full stop.
Bust through the wall a la Kool-Aid Man, lock in the status quo, exit like Homer backing into the hedges. Re-emerging every 10 years like a DC cicada to prompt agentic AI lobbyists to re-up the statute
(ed. Brood X would be a pretty boss tech company name).
Nods toward a national framework after establishing a moratorium were never credible to most people.7 The decade-long window was also what we in the industry call “a tell.” This technology which they often characterize as transformational, world-altering, with major geostrategic and national security implications, hence the urgent need for a hothouse policy environment, simultaneously requires a 10-year window? That implies a fairly low discount rate in an industry more frequently characterized by more…hyperbolic rhetoric.
Before the fabled birth of the great god Mythos, it was difficult to parse specific elements making up a hypothetically workable national governance framework. For example, an identified onerous requirement of one state law (the big one) is compulsory incident reporting. At the risk of talking my book, one element they might consider is a federal AI incident reporting system - no patchwork necessary. Compliance still too onerous? Well you’re in luck because it’s totally voluntary (“...now 100% mandate-free!”). Let’s just say there was no purchase among the moratorium-ites when developing a legislative proposal with Congressional offices. If your starting point is: “many transparency proposals we think are probably unlawful, they probably violate the First Amendment, and are not particularly useful for consumers,” I wouldn’t expect there to be. I’ll concede he’s referencing mandatory safety assessments; the point remains - unwillingness to engage on even de minimis governance doesn’t buy you credibility.
Later in that same conversation, having insisted any new laws should regulate uses and not development and uses can be largely governed by existing laws, they explore an analogy to another frontier technological arena:
“it’s not okay to develop a virus that could cause a global pandemic and then release it into the world to cause a global pandemic….[M]y assumption is that is unlawful and should therefore be penalized.”
What does he suppose is the appropriate penalty? While I don’t worry too much about catastrophic risks, he’s opened a door to a particular syllogism leading to a specific policy problem.8
And by the way, gain-of-function research is closer to the “development” category than the “uses” one.
Suppose the same thing happened here, unintentionally. What’s the sufficient sanction for a mistake resulting in 7,000,000 deaths? (Hey, you brought it up). Whether an individual researcher, or an end user, or indeed a frontier AI lab, there’s an unavoidable judgment-proof problem. Tort law and the like ideally motivates parties to account for risk by undertaking optimal levels of special care and risk-reducing investments. But this doesn’t work when the expected cost of liability exceeds the total ability to pay it.
So sometimes we get around this problem through regulation (including transparency). Having hand-waved the COVID example as a real problem that’s already or easily resolved, he then sidesteps by distinguishing AI development as “just science[,] it’s math.” Plenty of biological researchers would bristle at the suggestion that their work is somehow not science.9 Plus someone may want to tell this guy.
Two advocates (both friends) concede on this general point, but also delineate a uniqueness between AI and more familiar articles of commerce (e.g., “AI is Not Pencils or Pork Bellies”).10 It is in fact that uniqueness that sometimes has to be explicitly incorporated into existing law. On the premise that existing law does much of the work already - fraud is fraud, stealing is stealing, restrictions and penalties already exist whether by analog or digital means - it’s simply not true in all cases. If originalism remains the primary mode of legal interpretation, some laws will not account for certain novelties of emerging technologies like AI. For example, when photoshop and later social networks were new, definitions applicable to CSAM, revenge porn, and libel required affirmative changes to account for new modes of harm. This is true now for AI. Terminology like “digital alteration” and “likeness” are not cleanly forward-compatible to all the new uses and implications.
We saw an actual case in the TAKE IT DOWN Act and state analogs. Prior federal law was not addressable to AI-generated revenge porn. I’ll assume advocates are not talking about these kinds of new laws as the problem. So maybe we can dispense with the “1200+ state AI bills” canard (for this and a bunch of other reasons11).
A patchwork, or a rug?
So back to the states as the dominant strategy. The revealed threat is not really 50 states, but a few, and really just one: California. Patchwork rhetoric notwithstanding, the “California effect,” whereby the world’s fifth largest economy becomes de facto national policy, is the real target. I’m ignoring that laments of a “regulatory floor” ought to largely dispense the patchwork concern–that’s not how floors work–both of which they simultaneously caution against. We can add a couple other states to this state strategy, say New York and Colorado (I presume ya’ll good with Texas).
Perhaps they feared California was already lost given the breathless complaints thereon.12 But DC was never going to be an easier lift. Even adjusting for a favorable Trump, the outcome volatility is so high it’s not obvious the risk-adjusted outcome is positive. As one of their fellow travelers noted, assessing their foray into the CA governors race, “This is first grade, second grade for me, personally. We won’t be first and second graders forever.”13 Noted. In later grades you’ll learn in politics it is exceedingly difficult to replace something with nothing, and success requires reading the room (even if you think you are correct and smarter than everyone).
So after commencement, the coalition should have walked off stage diploma in hand and straight into redoubling the effort to optimize the CA law to their purpose. It would still require significant resources, but inexorably less than an optimal federal law - though it would require more than “no rules for 10 years” type language - then use that success both as proof-of-concept and momentum elsewhere.
The sticking point, of course, would require acknowledging an affirmative policy scheme they are in favor of, and not just a negative one. Maybe that’s actually the binding constraint that prevented this strategy, because as I noted above, there never really was one beyond messaging.
But solving problems by building fast and an aggressive go-to-market strategy is supposed to be what y’all are good at.
I was noodling a piece that engaged the AI state moratorium debate by steelmanning the arguments in favor, framing it though a self-administered ideological Turing test. I may still write it, but a recent essay by former colleague Evan Swartrauber prompted me to think in earnest more on the incentives for this policy preference. And in turn it led me to think about an optimal strategy thereon.
McKeown, M. (2012). Adaptability: The art of winning in an age of uncertainty. Kogan Page.
The word “preemption” often substitutes for “moratorium” in this context. I believe the latter is more informative and accurate as preemption implies (though is not strictly legally the case) replacing one thing with another. In this case the proposed federal law is explicitly a ban on any state action without a substantive replacement.
A reminder, Executive Orders are neither self-executing nor binding law.
At the risk of being overly precious, a patchwork refers to a design or method of connecting multiple parts into a unified whole, an intentional single object. To quote Msgr. Montoya: “I do not think this word means what you think it means.”
Implied valuations provided by a frontier LLM. Any inaccuracies, take it up with them.
We all know the Parable of the Underpants Gnomes: https://www.forbes.com/sites/artcarden/2011/07/14/underpants-gnomes-political-economy/. To be fair, the advocates have more recently committed to more specific elements, but concepts a plan do not make.
I learned a while ago lobbying in a different arena, even the DOE National Labs don’t have a uniform security protocol. Yes Oak Ridge, Los Alamos, etc…the places doing nuclear, bio, chemical science, and even AI.
Elsewhere - and I don’t mean to pick on a16z, they’re just the most egregious - advocates have made an argument that actually, these state laws may already be per se unconstitutional. Speaking not as a lawyer, true or not (and recent jurisprudence indicates the courts are not going to be fond of this dormant commerce clause argument) that’s not a convincing argument in favor of a moratorium. https://a16z.com/the-evidence-gap-why-courts-cant-balance-state-ai-regulation/.
Also I hate to bring this up, but… https://www.nytimes.com/2026/04/29/us/ai-chatbots-biological-weapons.html
I want to make clear I am not implying the two authors referenced are arguing in bad faith or are representative of the broader claims made overall in this piece. They are both thoughtful and straightforward advocates for their position, which is why I’m pulling in this one iteration of a particular argument that is echoed widely elsewhere.
Observing a surge of interest in a novel topic is not unusual (e.g., recent examples: AVs, crypto). Moreover many of those and the current AI ones are not substantive in nature, certainly not regulatory by any measure. Some are even favorable: e.g., DOT will study infrastructure needs for widespread AV adoption; parks department will purchase and pilot an AV for future fleet; the state agencies will develop a plan to accept crypto payments for certain official uses. Beyond the non-passage rate in each case exceeding 95%, how many were (1) never going past intro in earnest, (2) had no substantive regulatory impact (in making a patchwork, the fabric can’t be entirely threadbare); or (3) actually encouraging (either money, avenues for deployment, updating existing laws to put new tech equal footing with old)?
Please believe this example is not unique. The arguments made here are particularly tedious given that they’re entirely familiar to the arguments I would make in favor of broadly deregulatory efforts in other contexts – it’s upsetting that there seems to be little innovation on this margin, which amounts to formal sophistry…e.g., the statute’s language is inflexible and imprecise. Indeed, that’s what we call [checks notes] “a law.”
https://www.politico.com/news/2026/05/26/matt-mahan-governor-campaign-tech-00935650