"only belatedly learned what anyone could have told them a year ago, which is that you can’t pick and choose what work you do for the Pentagon"
This is a good line, but it's clearly not true? Millions of people do work for DoD in some form or another and no-one thinks that agreeing to sell toilets to military bases renders you liable to being conscripted into making bombs. Maybe Anthropic should've been more alert to the possibility of a Trumpified DoD pulling some shit, but the norm breaking (and, I think, law breaking) is all on one side.
I don't think that analogy is correct. This is more like the toilet company being forced to install toilets in a torture prison like abu graib or gitmo when they thought they'd only have to install them on normal bases. If you provide a tool to the U.S. military, you generally don't get a huge say in how it's used.
That isn't to say I think it's GOOD that the DoD is doing this, it's not. Yet at the same time it's kinda like the scientists that worked on the atom bomb, what did you think was going to happen?
It's unclear to me if David Sacks objects to this outcome: he's been vocal about his hostility towards Anthropic in the recent past, and has so far avoided any comment talking about this (he's spent much more of his Twitter time recently denying that Trump and Musk had any involvement in Epstein's crimes, for unclear reasons).
It's unclear to me if Stephen Miller objects to this outcome: certainly Katie Miller has been very vocal about opposing "woke AI companies", a standard which for her seems to include OpenAI, and Stephen Miller's immigration policies have done a tremendous amount to damage every AI lab.
Elon Musk probably benefits from the fight, even if it's a bit embarrassing to have Grok available and everyone involved dismissing it as inferior to Claude.
I think everyone who wants America to have a strong AI industry overall thinks this is bad. But that doesn't obviously describe multiple important figures in the Trump administration.
Kelsey, this is the sharpest and most comprehensive account of the fiasco I've read — and your enumeration of the ways this damages everyone's interests is exactly right. What the Tension Transformation Framework adds is an explanation for why a dispute this self-defeating happened at all.
The Pentagon's behavior is textbook identity-strategy tension masquerading as an external shock. The stated problem — Anthropic's contractual restrictions — is, as you note, nearly irrelevant in practice. Claude can't power a killer drone today regardless. The actual tension is institutional: an organization whose identity is built around unconstrained authority encountered a private partner with its own principled constraints. That identity clash — not operational necessity — is what's driving the escalation.
Your most provocative observation is that Anthropic isn't even objecting to autonomous weapons in principle, only to their premature deployment with today's unreliable models. This means the Pentagon is picking a nuclear-option fight over a timeline disagreement, not a values disagreement. That's a Maladaptive response to a tension that a five-minute Reformist conversation could have resolved. The Creative response — a phased capability framework with agreed-upon reliability thresholds — was apparently never attempted.
The lesson every other AI company is drawing from this, as you observe, is devastating for the national security goals the Pentagon claims to be defending.
I'm curious from a technical perspective, is it even possible to train Claude to do very unethical things without driving Claude "crazy", as in fundamentally breaking aspects of Claude that are necessary for the model to work?
"only belatedly learned what anyone could have told them a year ago, which is that you can’t pick and choose what work you do for the Pentagon"
This is a good line, but it's clearly not true? Millions of people do work for DoD in some form or another and no-one thinks that agreeing to sell toilets to military bases renders you liable to being conscripted into making bombs. Maybe Anthropic should've been more alert to the possibility of a Trumpified DoD pulling some shit, but the norm breaking (and, I think, law breaking) is all on one side.
I don't think that analogy is correct. This is more like the toilet company being forced to install toilets in a torture prison like abu graib or gitmo when they thought they'd only have to install them on normal bases. If you provide a tool to the U.S. military, you generally don't get a huge say in how it's used.
That isn't to say I think it's GOOD that the DoD is doing this, it's not. Yet at the same time it's kinda like the scientists that worked on the atom bomb, what did you think was going to happen?
It's unclear to me if David Sacks objects to this outcome: he's been vocal about his hostility towards Anthropic in the recent past, and has so far avoided any comment talking about this (he's spent much more of his Twitter time recently denying that Trump and Musk had any involvement in Epstein's crimes, for unclear reasons).
It's unclear to me if Stephen Miller objects to this outcome: certainly Katie Miller has been very vocal about opposing "woke AI companies", a standard which for her seems to include OpenAI, and Stephen Miller's immigration policies have done a tremendous amount to damage every AI lab.
Elon Musk probably benefits from the fight, even if it's a bit embarrassing to have Grok available and everyone involved dismissing it as inferior to Claude.
I think everyone who wants America to have a strong AI industry overall thinks this is bad. But that doesn't obviously describe multiple important figures in the Trump administration.
Sacks has since come out publicly in favor of this outcome on his Twitter.
What does Sacks dislike about Anthropic? Seems kind of arbitrary.
They've supported regulation that harm his financial interests, and have generally done much less kissing of Trump's ring than other top AI labs.
Kelsey, this is the sharpest and most comprehensive account of the fiasco I've read — and your enumeration of the ways this damages everyone's interests is exactly right. What the Tension Transformation Framework adds is an explanation for why a dispute this self-defeating happened at all.
The Pentagon's behavior is textbook identity-strategy tension masquerading as an external shock. The stated problem — Anthropic's contractual restrictions — is, as you note, nearly irrelevant in practice. Claude can't power a killer drone today regardless. The actual tension is institutional: an organization whose identity is built around unconstrained authority encountered a private partner with its own principled constraints. That identity clash — not operational necessity — is what's driving the escalation.
Your most provocative observation is that Anthropic isn't even objecting to autonomous weapons in principle, only to their premature deployment with today's unreliable models. This means the Pentagon is picking a nuclear-option fight over a timeline disagreement, not a values disagreement. That's a Maladaptive response to a tension that a five-minute Reformist conversation could have resolved. The Creative response — a phased capability framework with agreed-upon reliability thresholds — was apparently never attempted.
The lesson every other AI company is drawing from this, as you observe, is devastating for the national security goals the Pentagon claims to be defending.
Kelsey with the definitive take.
I'm curious from a technical perspective, is it even possible to train Claude to do very unethical things without driving Claude "crazy", as in fundamentally breaking aspects of Claude that are necessary for the model to work?
Is this a HAL situation?