Anthropic finalized a classified contract with the National Security Agency and Defense Department permitting AI model deployment on classified systems while omitting standard any lawful use language and barring use on Americans' data
White House approved $9 billion for spy agencies to buy advanced chips.
@NathanpmYoung OpenAI has not signed a contract with NSA at this point iirc, which is what this is?
Nathan is in Berkeley 🔎@NathanpmYoungSeems like DeepMind and OpenAI could just have got better contracts if they hadn't caved. Good for Anthropic.
@NathanpmYoung But by this logic OpenAI is still “holding out”? Anthropic did not get these terms with the rest of DoW
Nathan is in Berkeley 🔎@NathanpmYoung@yonashav Yes, that's fair. Though I think it suggests that holding out is possible, though I think I was too strong.
@NathanpmYoung Come on Nathan
Nathan is in Berkeley 🔎@NathanpmYoungSeems like DeepMind and OpenAI could just have got better contracts if they hadn't caved. Good for Anthropic.
@NathanpmYoung This is from the blog they put out. I see zero reason to distrust this disproportionately based on FUD.
Nathan is in Berkeley 🔎@NathanpmYoung@krishnanrohit I roll to doubt. They agreed all lawful use didn't they?
@NathanpmYoung If neither tor us are lawyers and neither contract is public what makes you so confident a) this contract is meaningfully different, b) oai lied in their blog baldfaced, and and c) oai/gdm could do better? It's not a credible update from evidence, which is my point.
Nathan is in Berkeley 🔎@NathanpmYoung@krishnanrohit Seems pretty plausible the DoD can still use it for surveilling Americans. Neither of us are lawyers but the consensus is not that this prohibits that, which I think it was for the Anthropic red line.
#1220rohit@KRISHNANROHIT@NathanpmYoung @bradrcarson See from above. I'm still saying the exact same thing and I think you have trapped priors here, but its better to stop at this point.
rohit@krishnanrohit@NathanpmYoung If neither tor us are lawyers and neither contract is public what makes you so confident a) this contract is meaningfully different, b) oai lied in their blog baldfaced, and and c) oai/gdm could do better? It's not a credible update from evidence, which is my point.
No it is not. Language like this is highly specific. My pov is that all labs have same-ish language for what won't get done and the govt has loopholes to drive through it that they can do surveillance.
Here, however, my problem is with differential interpretation of the same statements since it's terrible bayesian updating. Implicitly you're trusting Anthropic's PR posts that they will do no surveillance, despite them being used by DoD for years now, and Openai against their actual posts that they will do surveillance. This is irrational.
Nathan is in Berkeley 🔎@NathanpmYoung@krishnanrohit @bradrcarson This is in direct conflict to the notion that they won’t do surveillance. They may do any lawful use.
@NathanpmYoung @bradrcarson They were supposedly a couple words away from signing the agreement, you really think it was some major fundamental philosophical difference? Or getting upset at Pete Hegseth due to asking questions about Maduro and that spiralling?
My pov is here https://www.strangeloopcanon.com/p/aligning-anthropic
Nathan is in Berkeley 🔎@NathanpmYoung@krishnanrohit @bradrcarson What do you think they got into a huge fight with the DoD over? This seems a very costly signal that they don’t do surveillance of US citizens.
#1402bilal@BILALTWOVECnice
Seems like DeepMind and OpenAI could just have got better contracts if they hadn't caved.
Good for Anthropic.
Images from here:
@krishnanrohit I roll to doubt. They agreed all lawful use didn't they?
rohit@krishnanrohit@NathanpmYoung Come on Nathan
@krishnanrohit Seems pretty plausible the DoD can still use it for surveilling Americans. Neither of us are lawyers but the consensus is not that this prohibits that, which I think it was for the Anthropic red line.
Link: https://www.nytimes.com/2026/05/22/us/politics/spy-agencies-ai-chips-shortage.html
Nathan is in Berkeley 🔎@NathanpmYoungImages from here:
@yonashav Yes, that's fair. Though I think it suggests that holding out is possible, though I think I was too strong.
Yo Shavit@yonashav@NathanpmYoung OpenAI has not signed a contract with NSA at this point iirc, which is what this is?
@krishnanrohit @bradrcarson What do you think they got into a huge fight with the DoD over? This seems a very costly signal that they don’t do surveillance of US citizens.
rohit@krishnanrohitNo it is not. Language like this is highly specific. My pov is that all labs have same-ish language for what won't get done and the govt has loopholes to drive through it that they can do surveillance. Here, however, my problem is with differential interpretation of the same statements since it's terrible bayesian updating. Implicitly you're trusting Anthropic's PR posts that they will do no surveillance, despite them being used by DoD for years now, and Openai against their actual posts that they will do surveillance. This is irrational.