“All I want for CISA, is you….”

Happy Critical Infrastructure Security and Resilience Month for all those who celebrate! This year I wished for an “acting executive assistant director for infrastructure security.” And look at that!

(ed. I mean, a confirmed head of the CYBERSECURITY AND INFRASTRUCTURE SECURITY AGENCY would be nice…but I don’t want to appear ungrateful)

Subscribe now

Send in the Takes

I don’t know from political trends and “national mood” and all that. Per GR’s First Maxim:

The smartest prior is epistemic humility

Meaning anyone drawing high-confidence, low-nuance conclusions from a major national election is giving you a strong signal to discount their claims. But by all means, try to force the results of 470 federal offices, almost 300 statewide offices, and some 45,000 “elections across 55 states and territories” into a take on what the election means.

AI @ Work

Specialist occupations wrestling with the implications of changing technology is a fascinating sub-genre of Substack essays. The best ones are instructive, even if its got nothing to do with what you do.

Unfortunately so much content, especially on AI, focuses on sweeping, unfalsifiable claims about how “businesses” need to adapt to a “changing landscape” and “reimagine” the way they do things (see, e.g., LinkedIn). Consultant-speak dressed up as insight.

The always great saw it first hand last decade:

I spent most of the 2010s writing about enterprise technology and something called “future-proofing your business.” It was one of those ubiquitous, jargon-y phrases, usually accompanied by advice about moving to the cloud, modernizing infrastructure, going through “digital transformation” and becoming more “data-driven.”

An essay this week is the opposite. is a working architect and in an absolute gem of an essay, he systematically subjects multiple AI models to specific job tasks and reports on the results:

Architects and related professionals are poised to experience the shift that software engineers experienced during the past year, in which direct, manual production of technical drawings is replaced by supervision of AI models that grow increasingly capable of complex technical drawing tasks.

…Given their rate of improvement and given what we know about how they work, we should assume that they will get much better at drafting in the relatively near future.

But his actual conclusions aren’t what make it so great. In a few thousand words and handful of illustrations he manages to:

• explain in plain language the specific tasks making up the job and give background

• not betray any obvious preconceived notion about AI (good or bad)

• lay out the methodology and data

• draw takeaways that follow from the results, without dramatic overstatement

But what absolutely tickled me was at the end he manages to essentially work out a partial equilibrium analysis for the building industry (both goods and labor markets) and nod to implications “not limited to the architectural profession.”

I’d honestly pay good money for a series of these across occupations and would bet they’d provide more value than a 12-month consulting contract.

Dig UP!

I continue to believe tech’s pariah place in populist politics was avoidable. For instance data center buildout could’ve emphasized the possibility for a win-win like in Loudoun County, VA: higher tax revenues + lower tax rates → increased spending on schools and public safety.

My best guess is decades of benign neglect for all those NPCs (read: voters) atrophied any political muscles.

OpenAI may be the most adorable of the newest generation of tech political players. The same company that:

• after pivoting to a state-by-state policy strategy, proudly coined this process “reverse federalism”….or what the rest of us call “federalism,” and

• had to conspicuously disown the biggest midterm election PAC spender (pre-Musk)…

is now, anticipating a shift in Washington, trying not to “chase the moment”…by spending big to get “in sync with the new Washington.” Guys, announcing a splashy new strategy the week before a major change is exactly what chasing looks like.

Maybe AI execs are just bad at selling their product.1

Etc.

• “…the risks to U.S. national security of moving too slowly outweigh the impacts of imperfect alignment." That’s one view.

• Captain Obvious sighting: “Americans are increasingly against building data centers near their homes and favor candidates who oppose them.”2

  • Speaking of which (icymi): my brief writeup of the issue at Reason.

• Any readers in East London? I’ll be working from Shoreditch next week. Where do locals hang?

2

On Bloomberg Monday I hoped to dig into the nuances of this issue, but kept getting sidelined by election eve horse race stuff.

“The internet is held together by glue and dreams”

Thanks to TechCrunch Equity for the handy aphorism in re: the PHP vulnerability story.

Won’t be the last story like this.

Subscribe now

Send Me the Bill

NYT’s DealBook flagged “Big Law’s hunt to cash in on A.I”:

As A.I. threatens to upend the legal world’s business model of billable hours, firms have been racing to take advantage of new tools. Kirkland & Ellis, for instance, has committed to spending $500 million to build its own A.I. platform.

Others have partnered directly with frontier labs. (In April, Freshfields entered an agreement with Anthropic to build an in-house tool that could later be sold to other firms.)

Elsewhere TGR explored what proliferating AI would do to law firm billing practices and incentive to adopt it (results below). But it hadn’t considered the above phenomenon where firms build, buy, sell, or spinoff standalone legal AI tools. Fortunately the model can predict where and how that happens and it’s consistent with what we’re seeing.

But as DealBook notes, law firm ownership rules don’t allow access to outside equity, meaning corporate structure limits opportunity to exploit potential gains to legal tech. So workarounds emerge…

That is, law firms could be forgoing on the order of 40x returns on a hot legal tech market if they don’t figure out a way to effectively access public markets. If a particular spin out mechanism proves workable, expect Harvey/Legora’s impressive market share to flatline.

Independent Verification Organ(ization)

In another NYT piece, a relatively moving first-hand account of the structural failure of America’s organ transplant system. Not news to those who’ve worked on this issue before, but bears repeating as many times as possible.

Short version: organ donations are gathered, allocated, and delivered by a group of independent non-profit organizations (“organ procurement organizations” or “OPOs”), each of which is given a regional monopoly in which to operate. OPOs are subject to certification based on quality metrics meant to incentivize effectiveness.

Unfortunately prior to 2020 or so, OPOs lost certification basically never, in spite of well-known failures including falsified data and other deliberate behaviors which led to actual deaths. This was all well known to people who followed this stuff.

Why bring this up? Because advocates going all in on governing AI through so-called “Independent Verification Organizations (IVOs)” should slow their roll and reckon with the organ case study. Which they could, but I get the feeling they’re too far down the “We’ve solved it!” messaging to stress-test their handiwork.

OPOs were subject to a federal office with the authority (read: obligation) to review and decertify as needed, after the authority was previously delegated to a non-federal specialized authority. Both structures failed to model the unintended consequences over time. Both structures were so enamoring with the cleverness of their policy design to conceive of failure modes.

Sounds familiar. IVOs are not a magical mechanism even if they are likely one of the more promising avenues of AI governance that doesn’t suffocate the frontier. But the marketing reeks of [*sigh*] yet another group with little to no direct federal policy design, negotiation, or implementation experience.1

That Float Won’t Vote (or, a Share in Every Pot)

What would happen if every child in America got a SpaceX share? Ignore the thorny political economy questions to focus on corporate control.

Premise: Passive shareholders have the least propensity to vote.

• Most common shareholder questions like director elections and executive comp.* are decided by share of actual votes cast → not much changes (institutional and primary controller (meaning Musk) still determine outcomes)

• Changes to corporate charter and M&A decisions require a majority of all outstanding shares, so non-votes = nay votes

*as in the Tesla mega-payment vote, Musk was able to mobilize a population of otherwise passive superfan shareholders to adopt the proposal.

This all presumes voting shares are in play, but the net effect is this massive distribution scheme would marginally reduce power of institutional/13D accumulator activists, increasing Musk’s control in equal degree. But mostly not affect corporate governance of SpaceX.

Etc.

• Ho hum. “Unvetted AI agents are accessing U.S. bank accounts.”

• Appeared on NewsNation again in part to chat about this silly tweet. Silly not because of the actual content, but because of an adult human person describing as “buried” things that are (1) in the Executive Summary or (2) make up the 94 page risk factors section of Anthropic’s S-1 filing. Buried isn’t the word I’d use, and you just told on yourself.

• “[T]he risks to U.S. national security of moving too slowly outweigh the impacts of imperfect alignment." Surprised it took this long.

• The new hotness in data center design is idle GPUs. And also another reason data center moratoriums won’t have the effect you think they will.

The author is Nonresident Senior Fellow at

The Reasonable Burning Man Standard

First, some context:

On Sept. 5, Mary Herre was on her way to a math-themed camp at Burning Man for a slice of pie when she noticed two skydivers landing in an area with lots of foot and bike traffic.

…Within moments, [she] was suddenly struck by a third skydiver on the left side of her back — the same side as her sole kidney.

“I was just screaming, ‘That’s my only kidney, that’s my only kidney,’” Herre said. “I was terrified, and then I was in excruciating pain.”

Notwithstanding the story’s Mad Libs-like oddness, this is a fairly classic tort situation. As the article asks: “A skydiver landed on her at Burning Man. Who will pay the medical bills?” Who indeed?

Parties involved include the skydiver, the event’s host (Burning Man Project, a 501(c)(3) non-profit), and the woman herself in case attendance involves some waiver and/or due care.

I haven’t reviewed it, but a relatively more complicated scenario than the above is why Rep. Trahan’s (D-MA) bill on AI liability is welcome. Because that stuff is way more complicated, and it’s not limited to once a year in the Nevada desert.

The World’s Smallest AI Quiz

Back in the pre-The Social Network days when we still enthusiastically posted pictures on Facebook we would almost certainly come to regret, “The World’s Smallest Political Quiz” became one of those informative and ridiculous online curios that were nonetheless irresistible.1

I was reminded of that after coming across ’s digestible taxonomy of The Six Factions Shaping AI. Myself, I’m still sorting out the various camps and secret handshakes of Bay Area tech cliques—I’m pretty good at code-switching though it does require some language immersion. But the six categories roughly track my own mapping.

Imagine the relief when the description most closely matching my own disposition is #6, “Pragmatists.” (ed. I feel so seen)2

The point is all six categories read as reasonable on their own terms, and good faith engagement shouldn’t dismiss any of them out of hand. In fact the group I find the most frustrating (#1, “Accelerationists / Tech Right”) is the one with which I have a great deal of agreement. I just wish they were better at [*gestures wildly*] their whole deal.

Speaking of the Patagonia Patrol…

Look, the internet is not real life so put this all in proportion: the war among the various AI camps is a sophisticated influence campaign.

In “Who is behind the AI Safety Backlash?” () Tyler Johnston does an influence/social network analysis of the full court press response to the frontier lab loud quitting trend.

No surprise, it’s essentially a 1-4 ISO ball kinda thing.

No hate. You use the tools you got, and there’s no rules in this game. I’m sure some of the LessWrong folks would ring the same bell. Not being on TwiXter, I don’t know how it works.

But see above, it’s why I wish the folks I ultimately agree with were just less off-putting. Great parties tho!

Hand Covers Bruise

did the hard work and lays it all out. In what I assume is his last paper since leaving CSIS he establishes the groundwork for why a working insurance market is necessary for AI adoption (“The Insurance Industry’s Retreat from AI Threatens to Slow Innovation and Adoption”).

He overlaps much of what I was considering for a longer essay. The short version is if we don’t have a credible database of AI incidents, no one can underwrite the risk, and eventually that’ll constrain adoption/deployment of AI, especially agentic AI systems.

Knowledge of the risks will get out there, and stock prices will reflect that in a noisy, imprecise way. But without some systematic reporting system of incidents — like in cybersecurity proper — companies won’t be able to mitigate risks through underwriting and insurance. Because who’ll price a risk they can’t measure?

Lots more to say, including dunking on industry opposition to even a voluntary reporting system, and the bizarre phenomenon of cybersecurity insurance premiums declining despite acknowledged increases in risks.

For now I’m just hopeful the willful misuse of “regulatory capture” has peaked and we can get back to talking like adults.

What, Me Worry?

did the country a service by actually talking to people about their AI anxieties vis-a-vis the data center backlash. From her now high-profile reporting:

…what made data center activism such a nationwide phenom?

The only solution seemed to be hitting the road, going to the towns where data center fights were happening, and talking to as many people as I could: activists, yes, but also workers, planners, politicians, and residents. I wanted to learn: Do most people actually care about data centers, or just a passionate few?…How did local opposition become a national issue?

Beyond that piece, a stray remark on the The AI Policy Podcast gave me pause. She noted, in so many words: in SF everyone is running armies of AI sub-agents, and in DC everyone retreats to But China’s open-weights! Yet most people in America only know AI through free services.

That’s not a small thing.

Consider PNC Bank reported as of May (among their customers) “households paying for a Gen AI subscription reached [*drumroll*] 2.2%.” Meaning most people’s assessment of AI’s capabilities or threats is extrapolated from some weak sauce—the free services are frankly pretty crappy. Of course a greater number have access to more advanced models and applications through school or work. However beyond being hit over the head by the enshittification of Google search through “AI Overview,” limited context windows, lack of multi-step workflows, rampant hallucinating, and bounded automation means the “public” impression of AI is not the same as the rest of us.

The failure to recognize that disconnect is Exhibit #143 of why tech is having a conversation no one else is really interested in.

Which is not to say the Golden State Worriers are any better, what with their p(doom) and backchaining and all.

Etc.

• Lawfare: The Many Open Questions About the Trump Administration’s New De Facto AI Model Licensing System

• I’m scheduled for LiveNOW from FOX tonight to discuss, what else, cruelty to Anthropic models!

The author is Non-Resident Senior Fellow at

Tech Congestion Pricing

A new California law will fine robotaxi/AV operators if their disabled vehicle “blocks police or firefighters for more than 30 minutes.”

Believe it or not, here in San Francisco this is one of the weird things that’s actually an issue. Especially during rush hour.

I wonder if there’s a policy design lesson for frontier AI…

Read more

You’re Kidding Me

Eleven years ago, Kevin Nealon made a joke about the Kennedy Center being renamed the Trump Center.

He was speaking at the Mark Twain Prize ceremony for Eddie Murphy, which of course takes place at that very same venue.

Which calls to mind The Simpsons Prediction Theory which boils down to:

Every idea's been done, Butters, even before th…

Read more

Cool Beans

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

Read more

Read more

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 …

Read more

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 effectivenes…

Read more

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 commu…

Read more

Read more