It's too bad to see the old blue/red framing infecting everything. What's being done would be a bad thing to do regardless of the administration in power, and I can all but guarantee that even if the Bad Orange Man weren't in, someone from the other party would be busy misusing LLMs in a creepy way. It's just the way of the world.
I worked on AI back in the late 80s (natural language processing). The current AI bears as much resemblance to the kind I worked on as a 787 does to a Lilienthal glider. And while people were experimenting with neural nets back then, those nets were even less capable than the rule-based system I worked on.
So no, (this) AI is a new strategic variable, as Oleg said.
If you want to read about the truth regarding ai in the 1980’s at research labs you need to do a LLM to explain Roger Penrose’s book The Emperor’s New Mind. In it he details the state of the 1980’s and all of what is developed today was well known in the physics departments. Written in 1985.
I'd like to believe that the military will "ensure they thorough understand what it [AI] can and cannot do." But Hegseth and Trump have fired so many top generals and admirals that I have to worry about the ones that they've let remain.
READ THIS DEEP DIVE on the Pentagon's bullying of Anthropic. How successful will the government be in enslaving smart people and demanding their intellectual cooperation?
He also had perceptive comments on Yudkowsky & Soares' (2025) book "If Anyone Builds It, Everyone Dies":
"While I'm skeptical that the current trajectory of AI development will lead to human extinction, I acknowledge that this view may reflect a failure of imagination on my part. Regardless of where one stands on Yudkowsky and Soares's central argument, this book makes a valuable contribution by reminding us, once again, that every transformative technology in history has carried both promise and peril. Given AI's exponential pace of change, there's no better time to take prudent steps to guard against worst-case outcomes. The authors offer important proposals for global guardrails and risk mitigation that deserve serious consideration."
— Lieutenant General John N.T. "Jack" Shanahan (USAF, Ret.),
Inaugural Director, Department of Defense Joint AI Center
I seriously doubt that anyone would be involved in this real-live doomsday discussion if we had an administration we might disagree with, but that, in this broader field, "reasonable and sane" people could trust to be discerning about what's actually good for everyone.
As it is, it's a "One Flew Over the Coo-Coo's Nest" scenario ramped up to warp speed by people who are making trouble for trouble's sake. My guess is (and it's only a guess because I do not know these people) that the people calling the shots at Anthropic were/are speaking to the Trump Train that (it takes no genius to realize) is barreling down the tracks at us. Otherwise, problems would have been worked out differently and maybe more quietly.
As it is, I'm glad it is "going viral." The public deserves to know what's going on IN THIS CASE and with THESE PEOPLE who presently are in power.
And "excellent vindictive skills" or "extortionist-bully" are not on my ideas of bullet points to put on a resume for President of the United States or head of the so-called "war department."
The quote you extracted for the subtitle says it. Large language models are just that language models. The only thing that they know is how to predict the next token. The problem is not just that they get things wrong, it is that their "decision" basis is almost orthogonal to the factors that (should) go into making lethal decisions. It is a fundamental category error because language models have no way to represent the factors that must go into making these decisions.
Amodei does himself and his company a disservice when he claims that his models will backmail you or try to kill you. He harms his cause when he claims that he cannot determine whether his models are conscious. He hoisted himself on his own petard. He backed himself into a corner where he has to admit that he lied about how competent his models are or accept using them for purposes for which they are totally unsuited. There is a lot of ludicrous to go around.
I love Ai, I build with it 14-16 hours a day using all the big models, literally thousands of hours of screen time with them… the General is absolutely right they are not even close to ready for fully autonomous weapons. The idea makes me sick in the first place, but even if it didn’t, this is the tech equivalent of letting a toddler juggle loaded .357’s … and when it goes badly they’ll absolutely blame the Ai 🤦♂️
Gary, thank you for amplifying General Shanahan's voice here — a retired senior military officer with direct AI oversight experience saying "it's ludicrous even to suggest it" is not a fringe perspective. It's operational wisdom from inside the institution making the demands.
The Tension Transformation Framework reveals the precise paradox this creates: the Pentagon is exhibiting Victim-identity behavior — demanding unconstrained control to eliminate the feeling of vulnerability — while simultaneously ignoring the counsel of its own most experienced AI practitioners, which would actually reduce its vulnerability.
General Shanahan's statement isn't an argument against military AI. It's an Architect-identity position: understand what the technology can reliably do before deploying it in irreversible contexts. That's not weakness. That's sound systems design.
The deeper irony is that the Pentagon's maximalist demand — remove all guardrails now — will likely produce less usable military AI, not more. Coerced retraining of a reluctant lab, producing a model its best researchers didn't build carefully, deployed in high-stakes contexts its architecture wasn't designed for, is not a national security strategy. It's a Maladaptive response generating the exact fragility it claims to be preventing.
When the people who built the institution's AI oversight infrastructure are publicly saying "this is ludicrous," the tension has stopped being about Anthropic's contract terms. It's become a diagnostic of identity.
The below is from "Fortune" magazine this morning but most of it is behind a paywall and I don't have a subscription. However, it's so Trumpian I almost laughed: What happened to Congress? And this is hardly "the military or the Pentagon. It's Hegseth's ignorance in full display.
ALL COPIED BELOW/SNIPS
Anthropic/The Pentagon brands Anthropic’s CEO a ‘liar’ with a ‘God-complex’ as deadline looms over AI use in weapons and surveillance/By Beatrice Nolan/Tech Reporter
February 27, 2026, 10:52 AM ET
AI company Anthropic said it could not accept the Pentagon’s “best and final” offer to resolve a dispute over restrictions the company has in place on how the U.S. military can use its AI models. With just hours left before a Friday deadline to comply with the Pentagon’s demands or face actions that could see Anthropic barred from doing business with any company that also does business with the U.S. military, the dispute turned increasingly ugly.
END COPIED MATERIAL
This is just more mob-boss intimidation and extortion tactics--they seem to be threatening companies **already under contract* with the United States military by reneging on their contracts if they do business with Anthropic?
The general is extra-retired if he forgets that no ethical limitations ever bridled the military to employ technologies looking to kill and win. Only rival force can counter them. It was the understanding that directed the technology change between the American and Soviet nuclear and missile scientists at the beginning of the Cold War. Look not for self-restrictions but the rival contender
The general is extra-retired if he forgets that no ethical limitations ever bridled the military to employ technologies looking to kill and win. Only rival force can counter them. It was the understanding that directed the technology change between the American and Soviet nuclear and missile scientists at the beginning of the Cold War. Look not for self-restrictions but the rival contender
This is an addendum to a note to OLEG repeated below in that conversation about the military and Anthropic.
About "sane and reasonable," we don't have to go to some obscure philosophical, ethical, or spiritual point to refer to the idea that what is sane and reasonable is not arbitrary, or to understand the great difference between what is "sane and reasonable" in an authoritarian state and then in a democratic political system. We have the U. S. Constitution that holds those principles together for us under a defined political order, a tri-part system, and a genuine identity of "The People" with the rule of law.
It IS complex and sometimes difficult (duh) but, for instance, the two ethical standards that Anthropic cited and wants to stand by are easily related to the U.S. Constitution and the tri-part system where the power, though remote because of the utter size of the nation and the diversity of its electorate, still resides in "The People."
If Donald Trump has done one good thing for us all, it's that he makes it imperative that we all know and regularly consult the U.S. Constitution and the laws that are mediated through our tri-part system and as that system also exists in some form in individual States.
Trump is having a hard time tearing the country apart precisely because people are holding to the order that I used to think everyone loved and realized the value of as much as I do.
The military, however, is UNDER that (sane and reasonable) democratic political order, and not in charge of it, even as they have abused that order from time to time, as CEO's and corporations have and still do, and as Trump et all are doing as we speak. However, for instance, isn't MILITARY surveillance of U. S. citizens within our borders a No-No? Though "we" do have to work out how such ideas (that come from such massive changes in the culture) get played out by state and local police.
I don't think the word "ethics" is actually in the Constitution. However, and similar to the term education, these ideas are still woven deep into The U. S. Constitution as clearly as I write this note now and as you are reading it. Catherine Blanche King
This statement in his letter is very important, “Why not work on what kind of new governance is needed to ensure secure, reliable, predictable use of all frontier models, from all companies? This is a shared government-industry challenge, demanding a shared government-
“Let reason and sanity prevail“ - you're asking way too much of the Trump administration there.
It's too bad to see the old blue/red framing infecting everything. What's being done would be a bad thing to do regardless of the administration in power, and I can all but guarantee that even if the Bad Orange Man weren't in, someone from the other party would be busy misusing LLMs in a creepy way. It's just the way of the world.
It is not new. It has been in development since the 1940’s
Ai has been in development for decades
I worked on AI back in the late 80s (natural language processing). The current AI bears as much resemblance to the kind I worked on as a 787 does to a Lilienthal glider. And while people were experimenting with neural nets back then, those nets were even less capable than the rule-based system I worked on.
So no, (this) AI is a new strategic variable, as Oleg said.
If you want to read about the truth regarding ai in the 1980’s at research labs you need to do a LLM to explain Roger Penrose’s book The Emperor’s New Mind. In it he details the state of the 1980’s and all of what is developed today was well known in the physics departments. Written in 1985.
And I beg of you sir, don’t badmouth Sir Penrose!
I'd like to believe that the military will "ensure they thorough understand what it [AI] can and cannot do." But Hegseth and Trump have fired so many top generals and admirals that I have to worry about the ones that they've let remain.
Disclaimer: I was once in the Navy.
READ THIS DEEP DIVE on the Pentagon's bullying of Anthropic. How successful will the government be in enslaving smart people and demanding their intellectual cooperation?
https://www.ifyoucankeepit.org/p/a-steel-mill-the-president-might
He also had perceptive comments on Yudkowsky & Soares' (2025) book "If Anyone Builds It, Everyone Dies":
"While I'm skeptical that the current trajectory of AI development will lead to human extinction, I acknowledge that this view may reflect a failure of imagination on my part. Regardless of where one stands on Yudkowsky and Soares's central argument, this book makes a valuable contribution by reminding us, once again, that every transformative technology in history has carried both promise and peril. Given AI's exponential pace of change, there's no better time to take prudent steps to guard against worst-case outcomes. The authors offer important proposals for global guardrails and risk mitigation that deserve serious consideration."
— Lieutenant General John N.T. "Jack" Shanahan (USAF, Ret.),
Inaugural Director, Department of Defense Joint AI Center
I seriously doubt that anyone would be involved in this real-live doomsday discussion if we had an administration we might disagree with, but that, in this broader field, "reasonable and sane" people could trust to be discerning about what's actually good for everyone.
As it is, it's a "One Flew Over the Coo-Coo's Nest" scenario ramped up to warp speed by people who are making trouble for trouble's sake. My guess is (and it's only a guess because I do not know these people) that the people calling the shots at Anthropic were/are speaking to the Trump Train that (it takes no genius to realize) is barreling down the tracks at us. Otherwise, problems would have been worked out differently and maybe more quietly.
As it is, I'm glad it is "going viral." The public deserves to know what's going on IN THIS CASE and with THESE PEOPLE who presently are in power.
And "excellent vindictive skills" or "extortionist-bully" are not on my ideas of bullet points to put on a resume for President of the United States or head of the so-called "war department."
The quote you extracted for the subtitle says it. Large language models are just that language models. The only thing that they know is how to predict the next token. The problem is not just that they get things wrong, it is that their "decision" basis is almost orthogonal to the factors that (should) go into making lethal decisions. It is a fundamental category error because language models have no way to represent the factors that must go into making these decisions.
Amodei does himself and his company a disservice when he claims that his models will backmail you or try to kill you. He harms his cause when he claims that he cannot determine whether his models are conscious. He hoisted himself on his own petard. He backed himself into a corner where he has to admit that he lied about how competent his models are or accept using them for purposes for which they are totally unsuited. There is a lot of ludicrous to go around.
I love Ai, I build with it 14-16 hours a day using all the big models, literally thousands of hours of screen time with them… the General is absolutely right they are not even close to ready for fully autonomous weapons. The idea makes me sick in the first place, but even if it didn’t, this is the tech equivalent of letting a toddler juggle loaded .357’s … and when it goes badly they’ll absolutely blame the Ai 🤦♂️
Gary, thank you for amplifying General Shanahan's voice here — a retired senior military officer with direct AI oversight experience saying "it's ludicrous even to suggest it" is not a fringe perspective. It's operational wisdom from inside the institution making the demands.
The Tension Transformation Framework reveals the precise paradox this creates: the Pentagon is exhibiting Victim-identity behavior — demanding unconstrained control to eliminate the feeling of vulnerability — while simultaneously ignoring the counsel of its own most experienced AI practitioners, which would actually reduce its vulnerability.
General Shanahan's statement isn't an argument against military AI. It's an Architect-identity position: understand what the technology can reliably do before deploying it in irreversible contexts. That's not weakness. That's sound systems design.
The deeper irony is that the Pentagon's maximalist demand — remove all guardrails now — will likely produce less usable military AI, not more. Coerced retraining of a reluctant lab, producing a model its best researchers didn't build carefully, deployed in high-stakes contexts its architecture wasn't designed for, is not a national security strategy. It's a Maladaptive response generating the exact fragility it claims to be preventing.
When the people who built the institution's AI oversight infrastructure are publicly saying "this is ludicrous," the tension has stopped being about Anthropic's contract terms. It's become a diagnostic of identity.
Remember that people in the administration make decisions on AI when they can't create a group in Signal.
Scratching his head? Does he know nothing about Pete Hegseth? C'mon.
Breaking. In "Conspiracy Today": OpenAI, desperate for a break, was instrumental in getting Hegseth to attack Anthrop\c.
In brilliant conspiracy counterstrike, Amodei *manipulates* OAI into getting Hegseth to attack Anthropic causing Anthropic name recognition to soar.
Thanks General, for this sane perspective. All the rest, well, seemed rather insane to me.
The below is from "Fortune" magazine this morning but most of it is behind a paywall and I don't have a subscription. However, it's so Trumpian I almost laughed: What happened to Congress? And this is hardly "the military or the Pentagon. It's Hegseth's ignorance in full display.
ALL COPIED BELOW/SNIPS
Anthropic/The Pentagon brands Anthropic’s CEO a ‘liar’ with a ‘God-complex’ as deadline looms over AI use in weapons and surveillance/By Beatrice Nolan/Tech Reporter
February 27, 2026, 10:52 AM ET
AI company Anthropic said it could not accept the Pentagon’s “best and final” offer to resolve a dispute over restrictions the company has in place on how the U.S. military can use its AI models. With just hours left before a Friday deadline to comply with the Pentagon’s demands or face actions that could see Anthropic barred from doing business with any company that also does business with the U.S. military, the dispute turned increasingly ugly.
END COPIED MATERIAL
This is just more mob-boss intimidation and extortion tactics--they seem to be threatening companies **already under contract* with the United States military by reneging on their contracts if they do business with Anthropic?
FORWARDED FYI from CNN Newsfeed 02-27-26/Friday
https://www.msn.com/en-us/news/us/sam-altman-shares-anthropic-s-concerns-when-it-comes-to-working-with-the-pentagon/ar-AA1XdjQs?ocid=winp2fp&cvid=69a17e68601a4f8495f7bac56e7d41d9&cvpid=69a1c092875a4c4a85d20b7b25ce3d0f&ei=57
The general is extra-retired if he forgets that no ethical limitations ever bridled the military to employ technologies looking to kill and win. Only rival force can counter them. It was the understanding that directed the technology change between the American and Soviet nuclear and missile scientists at the beginning of the Cold War. Look not for self-restrictions but the rival contender
The general is extra-retired if he forgets that no ethical limitations ever bridled the military to employ technologies looking to kill and win. Only rival force can counter them. It was the understanding that directed the technology change between the American and Soviet nuclear and missile scientists at the beginning of the Cold War. Look not for self-restrictions but the rival contender
This is an addendum to a note to OLEG repeated below in that conversation about the military and Anthropic.
About "sane and reasonable," we don't have to go to some obscure philosophical, ethical, or spiritual point to refer to the idea that what is sane and reasonable is not arbitrary, or to understand the great difference between what is "sane and reasonable" in an authoritarian state and then in a democratic political system. We have the U. S. Constitution that holds those principles together for us under a defined political order, a tri-part system, and a genuine identity of "The People" with the rule of law.
It IS complex and sometimes difficult (duh) but, for instance, the two ethical standards that Anthropic cited and wants to stand by are easily related to the U.S. Constitution and the tri-part system where the power, though remote because of the utter size of the nation and the diversity of its electorate, still resides in "The People."
If Donald Trump has done one good thing for us all, it's that he makes it imperative that we all know and regularly consult the U.S. Constitution and the laws that are mediated through our tri-part system and as that system also exists in some form in individual States.
Trump is having a hard time tearing the country apart precisely because people are holding to the order that I used to think everyone loved and realized the value of as much as I do.
The military, however, is UNDER that (sane and reasonable) democratic political order, and not in charge of it, even as they have abused that order from time to time, as CEO's and corporations have and still do, and as Trump et all are doing as we speak. However, for instance, isn't MILITARY surveillance of U. S. citizens within our borders a No-No? Though "we" do have to work out how such ideas (that come from such massive changes in the culture) get played out by state and local police.
I don't think the word "ethics" is actually in the Constitution. However, and similar to the term education, these ideas are still woven deep into The U. S. Constitution as clearly as I write this note now and as you are reading it. Catherine Blanche King
This statement in his letter is very important, “Why not work on what kind of new governance is needed to ensure secure, reliable, predictable use of all frontier models, from all companies? This is a shared government-industry challenge, demanding a shared government-
industry (+ academia) solution.”