Tuesday assorted links

1. Legal basis for the Pentagon’s designation?

2. Cowen’s Third Law.

3. “But what is true is that this should not be much of a surprise considering the constant rhetoric over the past few years has been that AI is a power like no other. It’s like nukes, but times a thousand. We need regulation. And when an industry repeatedly calls out for oversight, asking for someone to make the rules on how it should be used, you cannot be surprised when the Defense department take that seriously. You cannot be surprised when they make up their own interpretations of what ought to be done, because you were insufficiently prescriptive. They will listen to your articulation of any red lines and wonder, what do you mean you want to tell me how to use the mega-nuke-crazy-power that you yourself are saying you don’t know how to control?”  Rohit.

4. There are too many types of shower controls.

5. The Anthropic valuation seems pretty stable.  Plus other matters of interest from SSC, including an idea for how to improve prediction markets by inducing the sports betting to subsidize participation in other contracts.

6. You can now bet on German train delays.

7. Rohit: “OpenAI now is also the only case I know of a defense department vendor contract being negotiated in public iteratively. With plenty of object lessons on why nobody does it.”  People, there is nothing weird going on here.  It is fine to dislike various aspects of the U.S. military, after all part of their business is to kill people.  But any blame you wish to levy goes toward “the system,” do not overly spin the narrative here.

Comments

#8: https://www.nber.org/papers/w34614

"we find that a one-standard-deviation increase in broadband internet access was associated with a 9.3–16.5-percent increase in adolescent suicide ideation. While we document increases in suicide ideation for both girls and boys, the results are most pronounced for adolescent girls. "

If I recall correctly, findings were similar in Spain, but some people "don't want you to know that".

First paragraph:

"Over one third of young adults are living with mental illness"

phhhht!!! By what ever-lowering bar can that claim be made? It's preposterous.

"While the determinants of mental health are multifaceted, clinicians, advocacy groups, and policymakers have noted that the adolescent mental health crisis has occurred alongside a dramatic increase in the amount of time that adolescents spend online"

Everything that has increased in the internet age is correlated positively with time spent on-line, isn't it? :)

Bad networks

Author Robert Akerlof, Richard Holden, DJ Thornton

Abstract
There is increasing evidence that social media is detrimental to mental health and self-esteem. A puzzle is why, in spite of this, people join these platforms. One possibility is that people feel compelled to participate: they dislike these networks—in particular, the way in which they generate rat races—but they need to be on them to socialize with peers. We refer to networks that harm users as “bad networks.” We model settings with network externalities and show that, surprisingly, bad networks are easy to establish. We also show that networks tend to be both bad and easy to establish when they create rat races—as social networks often do. Amplifying the rat race boosts network size which, while harmful to consumers, may benefit the platform.

Conclusion
There is significant evidence that social networks, despite their popularity, are harmful to users. In this paper, we ask why such networks arise in the first place, and what features make them “bad.”
We show that networks with the feature are not only harmful if they get established but also get established easily. Effectively, these are parties that people do not like to attend but feel more and more pressure to attend as others choose to do so. A few “instigators” are all it takes to get such networks started.
While networks with the feature might seem counterintuitive, we argue that they arise naturally in many settings. Rat races make networks bad—yet they also create pressure to join. We argue that rat races are a pervasive feature of social networks. Moreover, amplifying the rat-race nature of social networks boosts network size which, while harmful to consumers, may benefit the platforms.
This paper (see Proposition 3) suggests that traditional policies, such as Pigouvian taxation, can serve as helpful remedies. However, once networks are established, “marginal” policies may be insufficient to induce socially optimal outcomes.

https://www.sciencedirect.com/science/article/pii/S0047272726000101

Shower controls?? Try driving my new car. There are controls for everything imaginable, of course only after you have gone through five complicated steps. And there is no uniformity to this madness. I turn off the motor and get out of the car to remove something stuck n my wiper blade and when I get back in it has reverted to my wife's seat setting. There are myriad controls and data on my windshield, some of which interfere in my vision, and almost none of which make any sense. The automatic tight parking function worked great in the dealership parking lot but refuses to function for either of us almost 9 months later. I am only surprised it doesn't have a laugh track that is triggered whenever we fail to engage some function.

It takes me ~5 touch-screen interactions to turn on my radio. So much for that 90's era moral crisis re texting while driving...

... my conspiracy theory is that it's all caused by the backup camera regulation, which takes up space better used for buttons.

Yesterday I almost collided with a car that was being backed very slowly out of a parking spot by a driver who was laser focused on his backup camera view screen, never once looking out either of his side windows to check for cross traffic before or during the maneuver.

Yes, but backup cameras all have wide-angle lenses that allow you to see cross traffic. Of course, you have to go through a bunch of touch-screen interactions to get it to that mode.

Plus the problem that all of these features can really drain the 12V battery (not too many are using the 48V architecture). I don’t drive much so I have to constantly charge my 12V. Very annoying that the 12V is the weak link on what I consider an expensive car.

You should have done what we did - buy a ten year old car. Eleven years ago.

Why a 2005 model? Because in 2006 diesels had stacks of error-prone equipment installed at the insistence of the EU.

In California, 1996 is the newest I would go for a diesel

My '25 CR-V is Honda McBeepsALot.

The only new cars that I have first-hand experience with are ones that I rent. For years the electronic controls were horrendously bad. But in recent years, the Toyotas that I've rented were not bad. Meaning I could ignore the touch screen controls and simply start driving.

The few times that I did use the electronic controls were to use the radio or adjust the climate system. And miraculously the controls for that were easy and obvious and did not involve clicking through a half dozen screens.

I drive a 23-year old vehicle partly because I want nothing to do with the screens in modern cars. But maybe Toyota at least has figured out a better UX.

What I found really annoying with my new car was the inability to remove your seat belt in a fast food drive-thru or at an ATM machine without incessant chiming - even if I have it in park!

I've had similar issues.
I recommend the GV60. It's quite intuitive.

What kind of car: year and model?

If you are asking me, it is a BMW X3, nine months old. My wife hates it too and she is going to dump it soon. I amazed that they don't also require you to pat your head and rub you tummy at the same time. My vision of the factory is a room full of 20 something designers running loose while the CEO and Chairman hiding out in their office; like Major Major, you can only see them when they are out.

1. As long as the supreme court decides to just not rule against the administration in the shadow docket (and they basically never do at that level), it doesn't matter if the policy doesn't hold up: They can keep it in the books for 3 more years through just delays and reframing things. See the tariff ruling too: It was an obvious thing that could have been done a year ago, but it stayed for a year, and now there's just more ways to delay to add even more years. We still have tariffs.

Not really. Different policies have different nuances that affect whether and when the courts might intervene; for example, the tariffs were left in place because the administration argued (disingenuously, we can now see) that there was no need for interim relief because importers could always get a refund later.

For this policy--which targets a single, identifiable party for prompt and utter destruction--the closest analog we have is the executive orders targeting law firms, which were enjoined pretty much immediately and left untouched by higher courts because Trump's DOJ knew they were basically indefensible. Whenever DoD gets around to issuing a formal order codifying the supply chain risk determination, I expect Anthropic to have a lawsuit on file the next day and an injunction in place within a week.

"prompt and utter destruction" As far as I can tell, all the ban does is tell contractors and the DoD not to use Claude for DoD contracts. Anthropic wants to ban Claude's use for some legal purposes, and the DoD said, OK, we won't pay for it. I don't know what the freak out is all about. DoD is like 3 or 4% of the economy, sell to the rest of the of it if you don't like defense work.

Maybe, if I'm being cynical, the freak out is that Anthropic's (defenders?) actually want the right to dictate AI policy to the DoD. They don't get that right, other than refusing to sell (and the DoD can refuse to buy). The people can speak through Congress to provide that policy.

What you've described is the outer limits of DoD's statutory authority with respect to supply chain risks--to direct contractors to keep prohibited sources out of the supply chain top to bottom. But what Hegseth is actually threatening goes way further:

Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.

Maybe DoD walks this back in the official determination to better conform with the boundaries of the law, but given this Administration's track record on following the law I wouldn't extend them any benefit of the doubt on that score. And even if they do walk back the scope of the ban to apply just within the DoD supply chain, it's still spitefully overbroad insofar as there are plenty of potential applications for Claude that don't implicate the disputed guardrails--this isn't a Huawei or Kaspersky situation where you assume the product will create a backdoor for malign actors.

Given that context, the only reasonable conclusion is that DoD wants to inflict economic harm on Anthropic for "refusing to sell," not to preserve supply chain security.

Fruit of a poison tree. It seems like DoW can't use products that have touched Anthropic's products, so it needs to make sure the people that sell to it don't use Anthropic's products.

If it's such a big deal to Anthropic, they ought to get their ethics in order.

It seems like DoW can't use products that have touched Anthropic's products

Why? Because Pete Hegseth said so in a fit of pique? He hasn’t even bothered to come up with a pretextual argument for how Anthropic’s disputed guardrails pose any operational risk to 99.9% of use cases; the idea that they’re some kind of national security risk is pure bullshit.

Just ask your LLM the following:

1. Can the DoW require suppliers not to use certain types of equipment in the production of goods and services for the DoW.

The answer will be “yes” with wide discretion given to the DoW.

Then:

2. Can the DoW require suppliers “proof” from suppliers that those banned items were not used in production?

Your LLM will say “yes” with wide discretion.

All that is required is for the DoW to crank up the level of “proof.” And there is shockingly little that Anthropic can do about it.

Of course it was written by legal academics. In the real world, Anthropic can win this legal case and still not get any gov't contracts because no gov't program manager is going to risk his job by awarding a contract either to Anthropic or to anyone that lists them as a subcontractor. The PM won't say that's why, and it is next to impossible to know the reason for not winning any single contract.

Predicting elections is hard but spotting Bush Republicans is easy!! I was right about pretty much everything in Trump’s first term including the fact MAGA would end up despising McGahn and Leo because they were clearly Bush loyalists and not MAGA. But what’s hilarious is Bush Republicans saying Dick Cheney is smiling in heaven now that Netanyahu has forced Trump to invade Iran…Cheney feared Iran and it turns out Iran was a paper lion the entire time!! $5 trillion down the toilet!!

4. The problem of shower control is a classic (really THE classic) highly coupled 2x2 dynamic control problem. Both the hot and the cold water supply affect both flow and temperature more or less equally, although how equally depends on the temperature and pressure of the two supplies and the state of the two valves, which will differ according to the specific installation, over time, both short term and long term, and usage of the water system by other users. Hence it is inherently difficult to design a system that adjusts only the flow without affecting the temperature, or adjusts only the temperature without affecting the flow. Inherently difficult problems inspire engineers to invent ever more creative and complex solutions, even if (or let's admit it, particularly if) the simplest solutions are probably just as (partially) effective. Engineers believe in the perfect solution, which is a conceit that can lead us astray. If you add to that the problem of design, where designers want exquisite bathroom fixtures because bathrooms are negatively associated with dirt, smells, nudity, and other forms of chaos that cause stress and must be controlled. The combination of stubbornly challenging engineering problem and archetypical design opportunity lead to the multiplicity of shower controls.

"Thermostatic" valves have solved the temperature/flow control problem in my experience. I have clean control of the flow that is nearly completely independent of temperature, and vice versa. They are several-fold more expensive than "pressure balanced" (barostatic?) valves though.

Excellent set of comments! (Note that I'm the wrong flavor of engineer to have deep thoughts about this.)

I usually run the shower at full blast, which collapses this problem to an uninteresting subspace. But I'm always excited to find a new shower control design - both for the fun little mechanical puzzle of turning it on, and to play and think a bit about some of the control and design tradeoffs. Personally, I welcome the profusion of designs and (atypically I suppose) wish we had more fun little mechanical puzzles integrated into our modern environments!

(Except in cars, that trend is stupid.)

This: "I usually run the shower at full blast". Doesn't everybody? Given the required low-flow showerheads, my showers don't have separate flow and mix controls. They all just turn on full. They were chosen by the builder.

Thanks. I wasn't sure what's typical, so that's interesting feedback.

Regarding flow rate limitations, can't you just get a showerhead that uses a flow restrictor to meet regulations, then remove that? That's what I've always done, and seem to get good flow rates that way. But this is not something I know much about.

#3: Sounds like Rohit would argue that if a dictator tried to take control of AI.

That people argue for oversight does not mean they you can "Gotcha" them when some General puts themselves forward for the job.

Not even a general. Just an a-hole who got fired from Uber for running an illegal pricing scam to hide from regulators.

#4 - Shower controls. So true. The previous owner of our house put in some very high end controls. I made flash cards for each one. At first for myself, but then kept them for our guests. One of the showers, while fancy, is impossible to turn on without getting wet. We've since redone 2 of the showers and used the most basic, common sense controls and locations for them.

I don't travel much now days. By the time I was well-traveled in the Western USA (as distinct (!) from a global traveler, which I'm not), I didn't have any trouble with shower controls. One of the early lessons was you turn it on BEFORE getting in. This, of course, allows you to figure out volume and temperature adjustment. It's probably true that some showers are designed so to prevent this without getting your arm wet. Of course, if you're intending to enter the shower, most likely you will, at some point, want to get wet, so it's a silly issue to raise. (Granted, traveling with children and the less-abled would suggest a potential desire to not get wet myself, but in neither case would I expect to not get wet.) But I didn't start this post to suggest the author was inexperienced. One word: anecdotal. There is absolutely no way that his hotels were randomly selected. I stopped reading at that point, although I'd guess the post isn't about his non-random experience but an attempt to introduce the more fundamental statistical questions. It's just that such an obviously greatly flawed sampling is so annoying that I'm not interested in what he attempts to do with it. Pass.

I did yeah, I did in Columbus; it wasn't all failure. The writer of the note, when I was reading Flying Saucer by Jack London. It was wonderful, then I have learned French, never studying French, whether its Albert Camus; the paintings of Cezanne, the movie Casablanca, and of course, Jack London and Jack Kerouac, took immense interest in French. The idea that WWII can be understood without the paintings of the Versailles and the Louvre; then the Death of Sardinapaulas and the bona-fides of the painter Bonaparte, are widely known. It is Continental Hilarity, almost, then not, and knot, and mottled, and mere, and mare, because English philosophy is analytics. Concise, and foreboding, sick, and diluvian. Steinbeck was analytical I'd say, perhaps. All four, Steinbeck, Kerouac, London, then whom? Then I guess Bob Dylan, but then it was another, were from New York City. And travelled together at 11, to similar places with their families.

The Flying Saucer had such wonderful pictures, I couldn't help but copy, so I showed them to a girl, when I was rich with my timepieces and Vishnu was there as well. I've had success, then.

It must be too much regulation. Get AT on it.

Lawf, I think this article is for you. It's not about shower controls at all. It's about the bogus statistical analyses you love so much!

Using the most rigorous quantitative methods, our determines the horrors of the future explosion of shower control variations:

"And there’s a good chance we could see a large number, possibly >100, given the underestimate bias."

Better call Theta Grungeburg and get her on this. We've got to institute new regulations to protect the planet from the shower control explosion!!

If it is learned that a German train conductor bet on whether his trains would be on time, will he be banned from the train conducting hall of fame?

3. This is so disingenuous. There have been attempts by Democrats to regulate AI. But the current president signed an executive order against AI regulation of any kind. He’s currently going after the governor of Utah over this.

https://www.jdsupra.com/legalnews/white-house-ai-policy-update-utah-ai-9563484/

And just two years ago Marc Andreessen accused the Biden admin of wanting to control AI. He didn’t give any names or dates for these claims, but making allegations of tyranny on a podcast was enough for the tech right to get on board with MAGA.

Trump want federal oversight, not state by state. Not no oversight.

Then perhaps he should propose a law instead of trying to rule by decree.

#2 What is Cowen’s Third Law?

Never reason from a proposition about real interest rates unless it's posted on social media by a twelve year old. Something like that.

...he asks because he's unfamiliar with search engines, any of which would answer him in about 10 seconds (including typing time).

Maybe he thinks Cowen's own answer might be more interesting. Still, if you enjoy being a scold ....

And sometimes a commenter might add supplementary info while answering the question. Such as "it's Tyler's best law" or "Tyler really screwed up in writing his third law".

An object at rest: indeed, that is the American criticism of French Sein. French scientology; or French "Sentimentality." In the Autobiography of Benjamin Franklin it is profoundly put forward, and the American response is dis-interest. Benjamin Franklin was sexual misanthrope, but very successful at foreign diplomacy in France, like the Louisiana Purchase.

Other words for Essentiality, accordingly to myself and Haruki Murakami, are Hikki Komori; indeed his sister is Olivia Wilde, or Nimrata. Nostalgia, Heimweh, Sehnsucht, Pilgrimage, even Weltschmerz, or Schadenfreude, these words do mean Sentiment. Another one is Nirvana. They do lead to suicide. Many French writers, like Lakan, Derrida, and Sartre have written about the topic itself. I taught to a young man wearing Nirvana about that fact.

Even in I Know Why the Caged Bird Sings, when Marguerite, when her mother sends her on a train, when she is left alone with Uncle Willie, she writes of "abandonment", which is a sentiment of self-pity.

Sentiment in the AI age though, is positive, not relative, though elastic, but also scientific. That's a great ontology. Positivity, according to Mr. Chomsky is scientific rationale. Because in the long run, there is full employment, I sure I will get a job in six months.

Normativity, or normative, it means moral. Because in the long, there is full employment, I should have a job in six months.

That is the heartache, right or the "grief-stricken" love of Arjuna for peace, because he has to fight his cousins. He has to fight his cousins, the Druvidians, that is the law of the Ramayana. And his cousins, are able, perhaps, and they can also fly, perhaps, and so he does wear a crown, and so he can only commit to war, or peace, and that is both suicidal ideation.

#5 I fail to see why we should care how liquid the market is for "When will Drake release Iceman;" "Which participants will be eliminated in Episode 1 of the Bachelorette," "Eurovision Winner 2026," "Highest temperature in Oklahoma City today," or "Will the U.S. confirm that aliens exist before 2027."

Don't get me wrong, prediction markets are fun, but also hugely overrated.

#2. The spread won't drop back to 150 - 160bps. the Fed is out of the market and private credit is stretched thin.

"4. There are too many types of shower controls."

Excellent statistics post, an example of applying statistics to an everyday problem. The one disadvantage is that it can't really be assigned reading for intro statistics students, who won't know what maximum likelihood estimation is.

P.S. As for the design question of why so many different controls, I do wonder. Even homeowners are not immune; when I visited a friend they had to stick post-its on the shower door with instructions on how to operate the shower.

(One of the post-its said "ignore this one". Being a person who works mainly with text (and numbers and symbols and figures) I took that literally and thought it was a humorous instruction to "ignore this post-it". In fact the post-it was referring to the physical handle on the other side of the glass door, i.e. the post-it was telling me not to touch that handle.)

"People, there is nothing weird going on here. "

The fuck?

If that's the case, I look forward to you explaining how attempting to destroy a domestic supplier as a supply chain risk worked out all those other times that happened.

I look forward to it.

#4

Definitely. And light bulbs, light switches, faucets, etc.

"6. You can now bet on German train delays."

Two words: Moral hazard. To the max.

Okay, maybe five words. I should create a betting market on how many words my comment will contain.

The OP added a comment immediately after his post:

"It’s a joke!"

I think we are supposed to take the comment literally, because it's such an absurdly bad event to bet on. And thus we are supposed to not take seriously the first comment.

But maybe we are not supposed to take the second comment seriously? Rather than trying to wade through what's serious and what's not, I'm tossing the whole lot and walking away. Yet another example of how this is a waste of time because Twitter sux.

# 5
(1) In the world of Spence (1973) signaling, the Anthropic valuation would go down.
(2) In the world of “black means white”, the Anthropic valuation would go up.
(3) In the world of “I know that you know that I know …”, I don’t know where the Anthropic valuation would go. Sorry.

6. What is wrong with Germany? This seems very unlike them.

Probably they believed their own hype and decided they didn't have to invest properly any more. Happens everywhere.

I was going to suggest that immigrants changed German culture or something and now we have proof.

"This seems very unlike them." Ya reckon? Twenty-five years ago we were disappointed at how ill-managed one of the big Berlin railway stations was. We asked a fellow traveller a question. He said "I don't know either. I'm from Poland; we manage these things better."

It was the East Germans. Communism broke them, turned them into Russians, and then when they got reabsorbed, the Germans got all that foreign Russian culture.

Dare I say 2nd and 3rd generation Turks?

I don't think Matt Yglesias sports betting proposal (as reposted by Scott Alexander and Tyler Cowen) will work as intended.

Say we set a very conservative 10% cap.

Kalshi's 2025 revenues were $263.5 million.
CME Group's 2025 revenues were $6.5 billion.
CME solves for the equilibrium: acquires or competes with Kalshi. Regulatory problem solved.

7. 'Nothing weird going on here'. If anyone told you the day before these events transpired you would have said they were 'kooky' or whatever fifties turn of phrase you enjoy using at the moment to describe extremely consequential world events.

You're a stupid and puffed up little man trying to protect his transparently malicious and also stupid friend

3. AI

Long but interesting background to all this.

For most of its existence, OpenAI maintained a strict "no military" policy. The reasons were both ideological and practical: 
• The Original Ban: Until January 2024, OpenAI’s usage policy explicitly prohibited using its models for "military and warfare." This was rooted in its origin as a non-profit dedicated to "safe" AI that benefits humanity, viewing weaponization as a direct conflict of interest. 
• The Microsoft "Buffer": While OpenAI itself didn't have direct contracts, its partner Microsoft did. Microsoft provides Azure services to the Department of Defense (DoD), and OpenAI technology was available to government agencies through Azure. This allowed OpenAI to stay "hands-off" while their tech still reached the public sector.
• Lack of Infrastructure: Until recently, OpenAI didn't have the dedicated classified "air-gapped" environments required for top-secret defense work. They only recently developed the capacity to host models on the secure servers the military demands.

In 2018, Google signed a contract for Project Maven, a Pentagon initiative using AI to analyze drone footage to identify targets. When the news leaked, it triggered a massive internal revolt that fundamentally changed how AI companies handle the military. 

The Impact on Google
• Mass Resignations: Dozens of senior engineers resigned in protest, and over 4,000 employees signed a petition demanding Google "stay out of the business of war." 
• Contract Cancellation: Google was forced to back out of the contract and promise not to renew it. 
• AI Principles: In response, Google published its "AI Principles," which explicitly banned the development of AI for weapons or "surveillance violating internationally accepted norms." 

Project Maven created a "Standard of Neutrality" across Silicon Valley.

• The "Talent War": OpenAI knew that if they signed a defense contract early on, they would lose their best researchers to competitors. Top AI talent at the time was highly skeptical of military applications.
• Fear of Backlash: For years, OpenAI and Anthropic used Google’s "Maven disaster" as a cautionary tale. They realized that any move toward the military had to be done "quietly" or framed purely as "defensive" (like cybersecurity) rather than "offensive."

The landscape changed in early 2024 when OpenAI quietly deleted the words "military and warfare" from its prohibited use list. 

When Google announced it would not renew the Project Maven contract in 2018, AWS (along with Microsoft) immediately signaled that it was ready to take over.   

• Corporate Philosophy: Unlike Google’s "Don't Be Evil" or OpenAI’s "Safety First" branding, Jeff Bezos famously stated in 2018 that if tech companies turned their backs on the Department of Defense, the country would be in "big trouble."
• Vacuum Filler: After Google exited, AWS reportedly secured subcontracts to continue the work on Project Maven, specifically focusing on "object detection and classification" for drone footage—the exact task Google employees had protested. 

OpenAI couldn't have a defense contract because they didn't have the "Secret" level cloud regions required by the Pentagon. AWS already had them.
• Top Secret Regions: AWS has had a relationship with the CIA since 2013 and holds IL-6 authorization (the highest security level for Top Secret data).   
• The "Host" Strategy: AWS positioned itself as the infrastructure for other people's AI.  Even if a startup like OpenAI or Anthropic was hesitant to sign a direct deal, their models could be hosted on AWS’s secure GovCloud, making the tech available to the military while giving the AI companies "plausible deniability." 

As of 2026, AWS is the primary reason OpenAI's rival, Anthropic, is in the defense spotlight. 
• Claude Gov: Anthropic recently launched a version of its model specifically for AWS’s classified environments.   
• The Fallout: While Anthropic recently made headlines for refusing to remove certain "lethal use" guardrails from a $200 million contract, AWS remains the middleman. AWS provides the "safe house" where the military can run these models without them leaking to the public internet.

If AWS is the "infrastructure" and OpenAI is the "brain," Palantir is the "nervous system."   

While Google was retreating and OpenAI was hesitating, Palantir became the primary beneficiary of the vacuum left by Project Maven. They didn't just take the contract; they built the entire architecture that now allows companies like OpenAI to enter the battlefield.

When Google pulled out of Project Maven in 2018, the Department of Defense (DoD) didn't cancel the project—they handed the keys to Palantir. 
• The "Integrator" Role: Palantir’s Maven Smart System (MSS) became the central interface for the Pentagon.  It takes raw data (drone feeds, satellite imagery, social media) and fuses it into a "single pane of glass" for commanders. 
• Scaling Up: In May 2024, the Army expanded this into a $480 million contract, moving Maven from "hundreds of users" to "thousands."  By late 2025, this grew into a $1.3 billion deal as the system went fully tactical on the battlefield.   

Palantir’s most critical role today is as a "Neutral Platform" (specifically through their AIP—Artificial Intelligence Platform).   
• Model Agnostic: Palantir doesn't care whose AI "brain" is being used. Their software is designed to let the military "plug and play" different models.   
• The Recent Rift: In early 2026, a major rift occurred when Anthropic expressed concern about their Claude model being used in a lethal operation (reportedly a raid in Venezuela).   
• **Palantir as the "Snitch"**: According to reports, Palantir executives were alarmed by Anthropic's "resistance" and reported it to the Pentagon.  This directly led to the Pentagon winding down Anthropic’s contract and opening the door for OpenAI to step in as the more "cooperative" partner. 

As of March 2026, Palantir is the "Prime Contractor" for the Pentagon's AI efforts. They provide the interface (Gotham/AIP) that OpenAI’s models now run on.  Essentially, OpenAI provides the intelligence, but Palantir provides the targeting and deployment tools.

Real world results:

Hosted by the XVIII Airborne Corps, Scarlet Dragon is a series of exercises designed to test how fast the military can move from "Sensor to Shooter." In the past, identifying a target and getting permission to strike it could take hours.
• The Goal: Use AI to scan thousands of square kilometers of satellite imagery and drone feeds, identify targets, and pass them to a human commander in seconds.
• The "Hyper-War" Result: In the recent Scarlet Dragon 26-1 (held in December 2025), ****a small team of 20 soldiers managed a targeting workload that previously required 2,000 personnel****. They were processing over 1,000 targets per hour.

Palantir is the primary software provider for these exercises. Their Maven Smart System (MSS) acts as the operating system for the entire battlefield.
• The "Single Pane of Glass": Palantir takes data from every branch (Army, Navy, Air Force) and puts it on one screen.
• The Integration: During the 2025 exercises, Palantir’s platform successfully integrated HIMARS (rocket systems) with live AI data. This allowed the rockets to automatically calculate their coordinates the moment they were unloaded from a plane, based on AI targeting data received mid-flight.

Seems quite impressive/terrifying for a “fancy autocomplete” toy.

#5

Clicking through to the Ventuals prediction market that is referenced, the contract on Anthropic's IPO valuation shows open interest of $979k and latest 24 hour volume of $169k. (These appear to be actual $ amounts. I don't see any indication that some zeroes should be added to these amounts.)

These $ amounts are, of course, tiny relative to the predicted IPO valuation or to Anthropic's funding rounds. As for the latter, quick Googling tells me that Anthropic has raised a total of $60+ billion, including a latest Series G round of $30 billion (at a $380 billion valuation).

Given the very small size of this Ventuals contract, it seems fairly trivial for an interested party (e.g., an Anthropic investor) to put enough money into this prediction market to support the value of this contract with the goal of heading off any negative press that might come from a steep decline in its value. I'm therefore highly dubious how much information we can actually glean from this prediction contract.

4. It used to be that every hotel shower worked differently from every other hotel shower, but I think hotel designers are finally relenting. Now a lot of hotels built or remodeled in the last few years seem to have the same setup: a handle with a label showing off at 0 degrees and hot at 90 degrees. To get hot water, you turn it 180-270 degrees.

#2. You posted the wrong link. Here:

https://share.google/w8nCsjUhBshCzZeJF

Imagine you are an AI doomer. If you actually believe that AGI is possible, that it would be able to subvert any human system it touches, and that Anthropic might get there first, then Anthropic really *is* a supply chain risk to the US military, isn't it?

Yet the doomers are not saying, Hegseth is correct, Anthropic is in fact a supply chain risk, every AI company is a supply chain risk. It seems like many doomers are in fact hoping for a world where Anthropic is a supply chain risk, and takes advantage of that to subvert the US military, and in fact replaces the entire US government with AI control.

Credit where credit is due, one of the intellectually honest doomers here is Eliezer Yudkowsky. I note that he has not objected to the supply chain risk designation, only claiming that other companies should be considered even more of a supply chain risk.

Hegseth is very much not saying "every AI company is a supply chain risk" and then rejecting the Department of War's use of AI technology.

So how could "AI doomers" say Hegseth is correct on such a basis?

3. Simply put when it comes to transparency and self-awareness, Anthropic has put every Presidency in US History to shame. And if you randomly selected corporate earnings calls and compared them to randomly selected presidential speeches by any President, our political system would emerge as a hideous joke in comparison. It's not even close. Even the most crooked corporate earnings call crushes just about all political speeches in terms of honesty.

The difference being that no one at Anthropic has to get elected to their position of leadership. :)

Let Amodei or any Anthropic management run for office. Then we'll compare how they do on the "transparency" front vs. the "winning elections" front.

5. Some folks are taking offense about the war predictions in prediction markets. Demanding that prediction markets be regulated. War brings out the worst. Trump insiders (e.g., Junior) are being accused of placing bets on the war. Why would anybody believe such a nice cast of characters would do that? Trump insiders are even giving prediction markets a bad name. Is there anything that Trump doesn't screw up? Now, do I believe that Trump started the war so Junior could place bets in prediction markets? Only if Junior agreed to share the loot.

Comments for this post are closed