Hegseth threatens to blacklist Anthropic over AI-controlled weapons (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) The CEO of one of the biggest AI companies in  the world is meeting with Defense Secretary (00:00:05) Pete Hegsath today as the Pentagon threatens to  essentially blacklist that company, Anthropic, (00:00:10) from lucrative government contracts if the AI  company doesn't lift its restrictions on how (00:00:16) the military can use its technology. The Pentagon  has a $200 million contract with Anthropic. And a (00:00:24) source tells CNN that the company has concerns  over two issues. AI controlled weapons and mass (00:00:29) domestic surveillance of American citizens.  Let's talk about this with Jacob Ward. He is (00:00:35) a technology journalist and host of the RIP  Current podcast. He's also the author of the (00:00:39) book The Loop: How AI is creating a world without  choices and how to fight back. Jacob, the the (00:00:46) Pentagon here wants this company, Anthropic,  to lift its restrictions for how the military (00:00:52) uses its model. Uh they want it to be able to be  used for what's they're calling all lawful use. (00:01:01) What are anthropics concerns with that? Well,  Briana, this is a very unusual situation in my (00:01:09) experience because typically when government  use of technology begins, that means that new (00:01:15) restraints are going to be put onto that company,  the companies typically traditionally just pull (00:01:21) toward greater and greater capabilities. And the  assumption is that the government and democracy (00:01:25) will somehow dampen down the the dangers here.  What Anthropic is discovering is that they don't (00:01:30) live in that world anymore. Suddenly, they're  dealing with a defense department that wants to (00:01:33) be able to use their technology for whatever they  want. And Daario Amadai, the CEO of this company, (00:01:37) who has written very openly about his concerns  that his technology and the industry in general (00:01:43) could be misused for all sorts of authoritarian  purposes, is specifically objecting to the (00:01:48) possibility that we could be using this technology  to surveil Americans on mass or to create weapons (00:01:54) that fire themselves and kill people without human  intervention. Supposedly the trigger here for him (00:01:59) in part was discovering that his work was in  fact used in the capture of Nicholas Maduro (00:02:04) through anthropics partner Palunteer in a way  that he didn't know anything about. He haven't (00:02:09) found out about that the same way we all did. And  it is this effort inside a company that is making (00:02:14) this worldchanging technology to try to impose  some restrictions on that technology that now has (00:02:18) him running a foul of the Pentagon. Can you talk  a little bit about how AI maybe has been used for (00:02:26) say controlling weapons and mass surveillance  and what concerns that usage has raised? Well, (00:02:35) according to the reports we're seeing, Anthropics  Technology, which is the only company by the way (00:02:41) that is allowed inside the most classified systems  in the military, is being used for all sorts of (00:02:46) things that it's supposedly great at, including  things like cyber offense. You need to be able to (00:02:51) exchange digital fire with an opponent in in  nanocond time scales that human beings could (00:02:57) never be part of. That's an incredible use of it.  But at the same time, what AI makes possible is (00:03:02) things like, you know, there's technology that  we've seen, research that we've seen that that (00:03:06) could turn a Wi-Fi router into a an essentially  a radar inside a home that could tell you exactly (00:03:12) where somebody was at any given time. It is the  pattern recognition power of AI. something that (00:03:18) a a team of people, you know, in a vast building  or a bunch of people with headphones on in a van (00:03:24) could never imagine as a surveillance technology.  It is incredible. In fact, it's so incredible that (00:03:29) Dario Amadai actually wrote recently uh he said  here, "Powerful AI looking across billions of (00:03:34) conversations from millions of people could  gauge public sentiment, detect pockets of (00:03:38) disloyalty forming and stamp them out before they  grow." That's the kind of concern that Amadai has (00:03:43) here. Unbelievable. I mean, when you talk about  the possibilities here. So, the way this threat (00:03:49) would kind of operate is Axios is reporting that  the Pentagon would deem that they're threatening (00:03:55) to deem anthropic a what's called a supply chain  risk. That would essentially blacklist Anthropic (00:04:02) uh from having government contracts. It's an  approach that's usually taken with companies that (00:04:07) are extensions of foreign adversaries like China  or Russia. If Anthropic doesn't play ball here, (00:04:12) because as you say, they're unique in that they  kind of plug into this top secret uh nature of of (00:04:17) the US systems, is there another company that  can provide those same AI services a and that (00:04:22) would be willing to capitulate on these potential  human rights and civil liberties concerns? Well, (00:04:29) according to the reporting that we're seeing right  now, Anthropic is the one that they want to use. (00:04:34) def one defense official told Axios uh that the  reason we are talking to these guys is because (00:04:39) they are so good that is the problem that the  defense department is facing but while you say (00:04:44) you know anthropic is unique it's not entirely  unique and it has some very close competitors you (00:04:49) have Google you have meta and you have XAI which  is Elon Musk's company all of them have a you know (00:04:56) are are in line for this possible contract and at  least one of them and according to Axios that's (00:05:01) XAI that's Elon Musk's company has said they  can use it without any safeguards whatsoever. (00:05:07) They won't be bringing any of this uh you know  internal uh hand ringing to it. They'll say you (00:05:12) go ahead and and use it in any way you want.  And it may be that that could cost enthropic (00:05:16) this $200 million contract unless the defense  department figures out that that technology is (00:05:21) irreplaceable truly. But at this point the heat  is on in terms of the competition technologically. (00:05:26) And so it seems to me that this is a a real pickle  that Dario Amodai has gotten into. you are you are (00:05:32) worried about the use of your technology. You're  trying to impose what you consider some ethical (00:05:36) constraints on it and that suddenly could cost  you a $200 million contract. This is that thing (00:05:40) where you know the rubber meets the road when  you try to be an ethical business leader. You (00:05:44) know in the end we're we're talking here just  broadly here Briana about like who controls AI (00:05:50) and under what terms are they going to. All of  these stories are basically one big story about (00:05:55) the future of this technology and who is going to  control it. And I think what could happen next is (00:06:00) you could see somebody with fewer scruples, fewer  ethical qualms stepping up to the mic and saying, (00:06:04) "I'm going to go ahead and give you whatever you  need here." New reporting from Axius details just (00:06:08) how tense this meeting actually could be with  one defense official saying anthropic knows this (00:06:14) is not a get to know you meeting. This is not a  friendly meeting. This is a [ __ ] or get off the (00:06:19) pot meeting. What led to this moment? reporting by  Axius that the Pentagon is threatening to cut ties (00:06:25) with and cut off Anthropic. The company if if the  company continues to say that the military should (00:06:31) not use their AI for mass surveillance or fully  autonomous weapons. The top Democrat on the House (00:06:37) Science and Tech Committee was just on the show.  Here's her take. Anthropic is trying to do the (00:06:44) right thing so far as I can tell and put their own  guard rails in even in the absence of legislation. (00:06:51) I think as best I can tell, the Pentagon is trying  to talk anthropic out of doing the responsible (00:06:58) thing. Joining me now is one of the reporters  behind all of this new detail and reporting, (00:07:03) Dave Lawler, the national security editor for  Axis. It's good to see you again, Dave. Tell me (00:07:07) more about what you're hearing in terms of  what could happen in this meeting and come (00:07:10) out of it. Sure. So the Pentagon is saying it is  unacceptable for Anthropic to put any limitations (00:07:19) on how they use their model. They're saying  that if you're a contractor with the military, (00:07:23) you give us your tool and we use it however  we see fit. But Anthropic is saying we have (00:07:30) some concerns. We you know we we have built our  software uh with some safeguards in it around (00:07:35) things as you mentioned like surveillance or  developing weapons that fire without humans. (00:07:39) We're not willing to drop those guard rails.  The Pentagon is saying, "If you won't play ball, (00:07:45) if you won't come over to our way of thinking, uh,  we will cut off your contract." And there are a (00:07:50) plethora of other penalties that we could apply,  including this supply chain risk designation to (00:07:55) Anthropic to basically say that anyone who does  business with the Pentagon cannot do business with (00:08:02) Anthropic. So, you know, a lot of brinksmanship  leading up into this meeting. We've been texting (00:08:07) with sources on both sides ahead of the meeting.  They're expecting it to be not a particularly (00:08:12) friendly occasion at the Pentagon. It's it should  start in about a half hour, 10:30, uh-ish. And so, (00:08:18) um, you know, I I think that that right now  both sides are very dug in. The Pentagon says (00:08:22) our policy is not going to change. Anthropic  says our policy is not going to change. Uh, (00:08:26) and so we'll see, you know, what happens coming  out of this meeting whether the the relationship (00:08:30) completely breaks down. Yeah. Like the wild card  factor here is like our policy is not going to (00:08:35) change, but like you don't even know. We don't  even know what we're dealing with with AI and (00:08:39) how it develops. Like policy doesn't apply to  the thing that's going that's changing our world (00:08:43) as we speak. I digress. Why is anthropic and  its Claude model so important to the Pentagon? (00:08:53) So Claude is currently the only AI model that's  available in classified systems at the Pentagon. (00:09:00) So if the Pentagon wants to uh do research  and development on weapon systems or wants (00:09:06) to conduct a raid to capture the president of  Venezuela, Claude was actually used as part of (00:09:11) that operation. Claude is currently the only  AI model that's available in that space. Now (00:09:16) we reported last night that Elon Musk's XAI  has signed a deal to move into classified (00:09:22) spaces. They're working with Google and Open  AAI to try to move them in there as well. (00:09:27) Uh but defense officials, even though they're very  mad at anthropic, they admit that Claude is very, (00:09:33) very good. It's very useful for them. And so  they don't want to sort of cut off their nose (00:09:38) despite their face uh by blacklisting Claude.  And yet, like, you know, as we're discussing, (00:09:43) they're so dug in on this position  that you cannot restrict. You know, (00:09:47) we have our own protocols at the Pentagon about  how we do and don't use things. You as the vendor (00:09:52) don't get to tell us what we can't do with your  technology. So, uh, you know, but but again, (00:09:57) part of the problem for them is that Claude is  just very, very useful to the military. Dave, (00:10:02) is there an obvious or compromise or middle ground  here that you're picking up like what the Pentagon (00:10:08) wants and what Anthropic wants to protect, like  where they could meet in the middle on this? (00:10:14) So, I see a potential solution where they keep  clawed in their systems until they feel like (00:10:23) there is a replacement that's ready to roll. And  so, some kind of accommodation where this tension (00:10:31) plays out for a period of months uh and then if  they can't resolve it, they rip out Claude and (00:10:38) they move Google's Gemini or some other system  in. I don't currently see a point where Daario, (00:10:44) the CEO of Anthropic, goes into this meeting and  says, "Fine, have it your way." Uh, you know, (00:10:49) he has written and spoken a lot about his  concerns on these issues. And I also don't (00:10:53) see a point where the Pentagon entirely backs  down. So, there may be a point where neither (00:10:58) side uh is able to declare victory. They  muddle on for a bit uh and then they they (00:11:02) rip clawed out later on. Um, but I don't expect,  you know, handshakes and hugs at the end of this