The conflict between the Department of War and Anthropic has quieted somewhat, but nothing has been resolved and a catastrophic outcome is still entirely possible. Regardless of what happens next, two things are very clear.
This is the least political that AI will ever be. Politicians are finally waking up to the fact that AI is a big deal. Even though most of them don’t understand why it’s a big deal, you can safely assume they will have an increasing appetite for government intervention. The DoW incident is a preview, not an aberration.
This is the least stressful that AI will ever be. The last two weeks have been brutal: I notice several of the writers and thinkers that I most respect have been publicly struggling and in some cases decompensating. I’m afraid the pace is only going to get faster, and the stakes are only going to get higher. Pace yourselves.
In the spirit of pacing ourselves, we’ll cover what we need to cover about DoW, then put it down and move on to happier topics.
For years now, questions about A.I. have taken the form of “what happens if?” […]
This year, the A.I. questions have taken a new form, “what happens now?”
Ezra Klein’s opinion piece in NY Times ($) is nominally about the conflict between the Department of War and Anthropic and his analysis of that situation is spot-on: this is possibly the best short piece on that topic. But that conflict is a symptom of a much deeper problem: we’ve gone from being unprepared for AI capabilities that are coming soon to being unprepared for AI capabilities that have now arrived.
AI profoundly changes the nature of government surveillance—it’s now possible to intensively surveil every single American in a way that was previously (sort of) legal but completely impractical. In a sane world, the US Congress would carefully consider the implications of that change and pass appropriate legislation that codifies a reasonable balance between security and privacy.
Lamentably, we don’t seem to live in that world. Plan accordingly.
Zvi reports on Gemini 3.1. It’s a great model, but Google DeepMind just isn’t quite keeping up with Anthropic and OpenAI. Image generation is state of the art, but aside from that there’s no good reason for most people to pick Gemini as their daily driver.
Department of War vs Anthropic, part 1
Let’s start with some of the most interesting pieces from the past week.
Obviously a conversation between Ezra Klein and Dean Ball ($) is going to be good, and this one exceeds expectations. Dean is both highly-informed about the political situation and deeply thoughtful about the deeper implications of what’s happening here.
There’s been a lot of discussion about what the contracts between DoW and Anthropic / OpenAI actually mean. If you want to go down that rabbit hole, Zvi does a great job of breaking down what we currently know. See also Tom Smith’s analysis.
I’m glad people are doing the important work of scrutinizing these contracts and doing their best to ensure that they establish clear legal boundaries. But ultimately, legal documents can only do so much. If you don’t trust the three letter agencies not to spy on you in the first place, you probably shouldn’t trust them to honor a contract.
Much of the AI world has been highly critical of DoW’s recent actions, for obvious reasons. Pirate Wires’ conversation with DoW’s Emil Michael (partial $) is the best piece I’ve found in support of DoW’s position—there’s a lot I don’t agree with, but it’s more reasonable and coherent than many of the straw men being tilted at online.
Department of War vs Anthropic, part 2
The immediate consequences of the situation are bad enough, but the long-term collateral damage will be even worse. A lot of individuals, companies, and countries are going to look at the events of the last two weeks and start quietly making contingency plans that ultimately weaken both America and the entire AI industry. Nobody is well-served by any of this, and the longer the situation drags on the worse the fallout will be.
Here are two early examples—I’m certain many similar conversations are happening behind closed doors.
In the wake of the conflict between DoW and Anthropic, Anton Leicht considers whether it’s feasible for one of the middle powers to “poach” a frontier lab. He concludes it isn’t realistic to outright move one of the big labs outside the US, but proposes some intermediate strategies:
Stepwise and subtle, however, is a possible way to do this: understand the project of ‘poaching’ a frontier lab not as an attempt to extract value from the U.S., but to diversify the Western stack to make it more resilient to transient political trends and disruptions. My broader claim here is simple: it would be good for the world if a sizeable minority of American developers’ compute, business activity, and government cooperation were located in allied democracies. That could be about Anthropic, but I’d be just as happy with OpenAI or Google DeepMind. In a pinch, I might even take Meta. That outcome is eminently reachable and obviously beneficial in the aftermath of the Anthropic/Pentagon saga—and it’s never been more clear to the frontier developers that some hedging might be in their very best interest.
The DoW / Anthropic dispute has rekindled serious discussion about the US government nationalizing frontier AI development. Much of that discussion has focused on legal, political, and philosophical questions, but there hasn’t been much serious discussion of the practicalities.
John Allard dives into the nuts and bolts of nationalization, considering what strategies the government might use and whether those strategies would actually work. He isn’t optimistic about the outcome (which doesn’t mean it wouldn’t happen anyway):
It was always an inevitability that the government would try to exert control over frontier AI. The problems arise when the government begins exerting control without understanding that the frontier is a living process, not an asset. At some point the frontier may commoditize enough that tacit knowledge stops mattering and the government can brute-force its way to capability. But we’re not there yet. And until someone can answer the harder question — whether the US is better off accepting less control in exchange for maintaining its lead — the risk is that every attempt to capture the frontier is what finally kills it.
AI Village is the sensible, grownup version of Moltbook. A team of frontier AIs are assigned a group project and attempt to tackle it in full view of an amused world. Recent projects have included fundraising for charity and writing a blog. While there are elements of robot reality TV here, it’s an interesting way of exploring agent capabilities in the real world. Of particular note, it gives us information about how well a diverse group of frontier agents can work together (that’s going to be a big deal by the end of this year).
In the AI Village, we’ve observed substantial improvement in agent capabilities over the span of months. Early 2025 agents often fabricated information, got stuck, or became easily distracted in a few minutes to hours. Late 2025 agents tend to be more truthful and stay on task longer (though their effectiveness often drops off once the most obvious tasks are done).
Jason Crawford reflects on recent progress in agentic coding. There aren’t a lot of novel insights here, but it’s a great overview and a strong choice for sharing with people who haven’t been following AI closely.
Robots as art directors
2025: why would I do work when I can tell a robot to do it for me?
2026: why would I tell a robot to do work when I can have a robot tell it for me?
I’ve recently needed artwork for a couple of personal projects, and I’ve found that SOTA models aren’t just capable artists—they’re also quite good art directors. My current workflow goes like this:
Discuss the style and content of the image with Claude, who has a much better understanding of art terminology and styles than I do.
Once we’ve figured out the goal, Claude writes a detailed prompt.
The prompt goes to Gemini for rendering.
Back to Claude, who assesses the image and makes changes to the prompt (sometimes but not always with my feedback).
Iterate until I’m satisfied with the result.
Claude is surprisingly good at looking at an image and finding areas for improvement in everything from line style to facial expressions. The results can’t (yet) compete with professional work, but they’re getting very good. And from a process perspective, the AI is light years better: I can experiment with multiple directions and styles within minutes, and the robots never get frustrated when I change my mind seven times in half an hour for no good reason.
Along with mass domestic surveillance, autonomous weapons are one of the red lines in the Anthropic / DoW dispute. Policy and ethical considerations aside, it’s surprisingly hard to define what “autonomous weapons” actually means. We have well-defined autonomy levels for cars, but no similar concept for weapons (yet). Autonomous missile defenses have been deployed since the 1980s, but that feels very different from a system that can autonomously identify and engage individual soldiers.
Transformer explores some of the technical and legal questions, and looks at what’s currently on the battlefield in Ukraine.
New communications technologies often transform how the public gets information and forms opinions. The printing press democratized the spread of information, weakening the control of the church and monarchy. Social media is a breeding ground for outrage, tribalism, and conspiracy theories. How might AI affect public discourse?
Dan Williams argues that AI might be a force for good, nudging us closer to a consensus view of reality based on expert understanding and strong epistemics. We don’t have much data yet, but he cites some promising early research suggesting that LLMs are surprisingly effective at getting people to change their minds.
His arguments sound plausible, although I note that many of us initially expected social media to be a force for good.
AI Frontiers explores how AI might affect workers, arguing that if AI is much better than humans at many but not all jobs, human wages might actually rise.
That counter-intuitive result follows from basic economics, which the article does a good job of explaining. It’s a solid piece, and a good introduction to some of the relevant economics if you’re not already familiar with them. But note that this whole analysis only applies if AI is powerful but not superhuman. Without careful intervention, everything falls apart in a world with superhuman AI:
If machines do everything, then those who own the machines will capture all this value. Products and services would become very cheap, but workers, outcompeted by machines in all tasks, would end up with a vanishingly small share of the economy’s income.
We can flourish alongside superintelligent AI, but only if we make smart choices.
Eleos AI Research is a small nonprofit dedicated to studying AI sentience and wellbeing, a topic which until very recently has largely been ignored. Executive Director Robert Long goes on the 80,000 Hours podcast to discuss their work and some of the big open questions they’re tackling.
Good interviews answer the questions you wanted to learn about, but great interviews raise (and occasionally answer) questions you hadn’t realized you ought to be asking. I came out of this one with new questions about the ethics of creating sentient AI that wants to be subservient to humans and about AI consciousness that is as meaningful as ours but unrecognizably different.
As the subtitle implies, Nicholas Carlini has opinions about how to write papers good enough to win best paper awards—and more generally, how to do good research. It’s a dauntingly long piece but very good: even though I’m not a researcher, I found multiple insights that I’m excited to put to use in my own work.
The conflict between the Department of War and Anthropic has quieted somewhat, but nothing has been resolved and a catastrophic outcome is still entirely possible. Regardless of what happens next, two things are very clear.
This is the least political that AI will ever be. Politicians are finally waking up to the fact that AI is a big deal. Even though most of them don’t understand why it’s a big deal, you can safely assume they will have an increasing appetite for government intervention. The DoW incident is a preview, not an aberration.
This is the least stressful that AI will ever be. The last two weeks have been brutal: I notice several of the writers and thinkers that I most respect have been publicly struggling and in some cases decompensating. I’m afraid the pace is only going to get faster, and the stakes are only going to get higher. Pace yourselves.
In the spirit of pacing ourselves, we’ll cover what we need to cover about DoW, then put it down and move on to happier topics.
Top pick
The Future We Feared Is Already Here
Ezra Klein’s opinion piece in NY Times ($) is nominally about the conflict between the Department of War and Anthropic and his analysis of that situation is spot-on: this is possibly the best short piece on that topic. But that conflict is a symptom of a much deeper problem: we’ve gone from being unprepared for AI capabilities that are coming soon to being unprepared for AI capabilities that have now arrived.
AI profoundly changes the nature of government surveillance—it’s now possible to intensively surveil every single American in a way that was previously (sort of) legal but completely impractical. In a sane world, the US Congress would carefully consider the implications of that change and pass appropriate legislation that codifies a reasonable balance between security and privacy.
Lamentably, we don’t seem to live in that world. Plan accordingly.
New releases
Gemini 3.1
Zvi reports on Gemini 3.1. It’s a great model, but Google DeepMind just isn’t quite keeping up with Anthropic and OpenAI. Image generation is state of the art, but aside from that there’s no good reason for most people to pick Gemini as their daily driver.
Department of War vs Anthropic, part 1
Let’s start with some of the most interesting pieces from the past week.
Ezra Klein interviews Dean Ball
Obviously a conversation between Ezra Klein and Dean Ball ($) is going to be good, and this one exceeds expectations. Dean is both highly-informed about the political situation and deeply thoughtful about the deeper implications of what’s happening here.
Zvi reviews the situation
Zvi summarizes the state of play as of March 6.
Zvi: A Tale of Three Contracts
There’s been a lot of discussion about what the contracts between DoW and Anthropic / OpenAI actually mean. If you want to go down that rabbit hole, Zvi does a great job of breaking down what we currently know. See also Tom Smith’s analysis.
I’m glad people are doing the important work of scrutinizing these contracts and doing their best to ensure that they establish clear legal boundaries. But ultimately, legal documents can only do so much. If you don’t trust the three letter agencies not to spy on you in the first place, you probably shouldn’t trust them to honor a contract.
Pirate Wires talks with Emil Michael
Much of the AI world has been highly critical of DoW’s recent actions, for obvious reasons. Pirate Wires’ conversation with DoW’s Emil Michael (partial $) is the best piece I’ve found in support of DoW’s position—there’s a lot I don’t agree with, but it’s more reasonable and coherent than many of the straw men being tilted at online.
Department of War vs Anthropic, part 2
The immediate consequences of the situation are bad enough, but the long-term collateral damage will be even worse. A lot of individuals, companies, and countries are going to look at the events of the last two weeks and start quietly making contingency plans that ultimately weaken both America and the entire AI industry. Nobody is well-served by any of this, and the longer the situation drags on the worse the fallout will be.
Here are two early examples—I’m certain many similar conversations are happening behind closed doors.
Can You Poach A Frontier Lab?
In the wake of the conflict between DoW and Anthropic, Anton Leicht considers whether it’s feasible for one of the middle powers to “poach” a frontier lab. He concludes it isn’t realistic to outright move one of the big labs outside the US, but proposes some intermediate strategies:
Can you nationalize a frontier AI lab?
The DoW / Anthropic dispute has rekindled serious discussion about the US government nationalizing frontier AI development. Much of that discussion has focused on legal, political, and philosophical questions, but there hasn’t been much serious discussion of the practicalities.
John Allard dives into the nuts and bolts of nationalization, considering what strategies the government might use and whether those strategies would actually work. He isn’t optimistic about the outcome (which doesn’t mean it wouldn’t happen anyway):
Agents!
What did we learn from the AI Village in 2025?
AI Village is the sensible, grownup version of Moltbook. A team of frontier AIs are assigned a group project and attempt to tackle it in full view of an amused world. Recent projects have included fundraising for charity and writing a blog. While there are elements of robot reality TV here, it’s an interesting way of exploring agent capabilities in the real world. Of particular note, it gives us information about how well a diverse group of frontier agents can work together (that’s going to be a big deal by the end of this year).
As you might expect, the agents made a lot of progress last year:
As we may vibe
Jason Crawford reflects on recent progress in agentic coding. There aren’t a lot of novel insights here, but it’s a great overview and a strong choice for sharing with people who haven’t been following AI closely.
Robots as art directors
2025: why would I do work when I can tell a robot to do it for me?
2026: why would I tell a robot to do work when I can have a robot tell it for me?
I’ve recently needed artwork for a couple of personal projects, and I’ve found that SOTA models aren’t just capable artists—they’re also quite good art directors. My current workflow goes like this:
Claude is surprisingly good at looking at an image and finding areas for improvement in everything from line style to facial expressions. The results can’t (yet) compete with professional work, but they’re getting very good. And from a process perspective, the AI is light years better: I can experiment with multiple directions and styles within minutes, and the robots never get frustrated when I change my mind seven times in half an hour for no good reason.
AI in the real world
What you need to know about autonomous weapons
Along with mass domestic surveillance, autonomous weapons are one of the red lines in the Anthropic / DoW dispute. Policy and ethical considerations aside, it’s surprisingly hard to define what “autonomous weapons” actually means. We have well-defined autonomy levels for cars, but no similar concept for weapons (yet). Autonomous missile defenses have been deployed since the 1980s, but that feels very different from a system that can autonomously identify and engage individual soldiers.
Transformer explores some of the technical and legal questions, and looks at what’s currently on the battlefield in Ukraine.
How AI Will Reshape Public Opinion
New communications technologies often transform how the public gets information and forms opinions. The printing press democratized the spread of information, weakening the control of the church and monarchy. Social media is a breeding ground for outrage, tribalism, and conspiracy theories. How might AI affect public discourse?
Dan Williams argues that AI might be a force for good, nudging us closer to a consensus view of reality based on expert understanding and strong epistemics. We don’t have much data yet, but he cites some promising early research suggesting that LLMs are surprisingly effective at getting people to change their minds.
His arguments sound plausible, although I note that many of us initially expected social media to be a force for good.
Jobs and the economy
How AI Could Benefit the Workers it Displaces
AI Frontiers explores how AI might affect workers, arguing that if AI is much better than humans at many but not all jobs, human wages might actually rise.
That counter-intuitive result follows from basic economics, which the article does a good job of explaining. It’s a solid piece, and a good introduction to some of the relevant economics if you’re not already familiar with them. But note that this whole analysis only applies if AI is powerful but not superhuman. Without careful intervention, everything falls apart in a world with superhuman AI:
We can flourish alongside superintelligent AI, but only if we make smart choices.
AI psychology
Robert Long on AI consciousness and wellbeing
Eleos AI Research is a small nonprofit dedicated to studying AI sentience and wellbeing, a topic which until very recently has largely been ignored. Executive Director Robert Long goes on the 80,000 Hours podcast to discuss their work and some of the big open questions they’re tackling.
Good interviews answer the questions you wanted to learn about, but great interviews raise (and occasionally answer) questions you hadn’t realized you ought to be asking. I came out of this one with new questions about the ethics of creating sentient AI that wants to be subservient to humans and about AI consciousness that is as meaningful as ours but unrecognizably different.
Other interests
How to win a best paper award
(or, an opinionated take on how to do important research that matters)
As the subtitle implies, Nicholas Carlini has opinions about how to write papers good enough to win best paper awards—and more generally, how to do good research. It’s a dauntingly long piece but very good: even though I’m not a researcher, I found multiple insights that I’m excited to put to use in my own work.
Something frivolous?
A very short story
Sam Altman: