Latest / Elon Musk Podcast / Anthropic rejects latest Pentagon offer: ‘We cannot in good conscience accede to their request’

Transcript

Anthropic rejects latest Pentagon offer: ‘We cannot in good conscience accede to their request’ · Elon Musk Podcast

• 0:00
  The United States government has officially designated a leading
• 0:03
  American artificial intelligence company a supply chain risk to
• 0:06
  national security. Which is a label they usually
• 0:10
  save for foreign adversaries. We are talking about the exact
• 0:13
  same classification used for Huawei and ZTE.
• 0:16
  The Department of War has effectively placed a Do not
• 0:19
  touch order on Anthropic. That sounds like a lot more than
• 0:22
  just a cancelled contract. Oh, it is.
• 0:24
  It's an absolute blacklist. It prohibits military
• 0:27
  contractors, suppliers, basically any partners from
• 0:31
  conducting commercial activity with the company.
• 0:33
  If the government can blacklist a domestic company for refusing
• 0:36
  to remove safety guardrails from its software, does the private
• 0:39
  sector actually retain any control over how its technology
• 0:43
  is used in warfare? That is the multibillion dollar
• 0:46
  question here. And this whole thing didn't
• 0:48
  start in some corporate boardroom.
• 0:50
  It started with a specific military operation, specifically
• 0:53
  the raid in Caracas to capture Nicolas Maduro, right?
• 0:57
  Operation absolute resolve. Exactly.
• 1:00
  According to reports, the US military utilized Anthropic's
• 1:03
  model CLAWED during the planning and execution of this operation.
• 1:07
  Now Anthropic doesn't have a direct contract to sit in the
• 1:11
  situation room. The integration happened through
• 1:14
  Palantir. We should probably clarify
• 1:17
  Palantir's role here, because they aren't just, you know, a
• 1:20
  simple database company. No, no.
• 1:22
  Think of Palantir as as the operating system for Modern
• 1:25
  Warfare. They have this platform called
• 1:26
  the Maven Smart system. It ingests satellite imagery,
• 1:30
  drone feeds, intercepted communications, basically all
• 1:34
  the messy data of war. And it synthesizes it into a
• 1:37
  coherent picture for the commanders on the ground.
• 1:40
  Right. And they were running clawed
• 1:42
  inside that ecosystem to help process all that information.
• 1:44
  The mission itself was technically a success.
• 1:46
  I mean, the administration got what they wanted, but the
• 1:49
  friction actually started afterward.
• 1:51
  Yeah, it started because Anthropic asked a question.
• 1:54
  Following the operation, company leadership allegedly inquired
• 1:57
  whether their technology had been utilized in the raid, and
• 2:00
  that inquiry triggered an immediate alarm at the Pentagon.
• 2:03
  Why though? I mean, if I sell a product,
• 2:06
  asking how it was used seems pretty standard.
• 2:09
  Not when you were dealing with classified operations.
• 2:12
  But Anthropic felt they had to ask because of constitutional
• 2:17
  AI. Let's use plain English for
• 2:19
  that. What exactly is Constitutional
• 2:21
  AI? Sure, so most AI models are
• 2:23
  trained on massive amounts of Internet text and then just fine
• 2:27
  tune to be helpful. Anthropic takes a totally
• 2:29
  different approach. They give the model a
• 2:32
  constitution, a specific set of ethical principles it has to
• 2:35
  follow. Like rules against generating
• 2:36
  harm or things like that. Exactly.
• 2:38
  Rules like do not help kill people, do not violate privacy,
• 2:42
  do not create hate speech. The model essentially checks its
• 2:45
  own output against this Constitution before it even
• 2:48
  responds. So to know if the model actually
• 2:51
  followed its own constitution, the company needs to know what
• 2:54
  it was doing. Yes, they need a degree of
• 2:56
  transparency to ensure their ethical guidelines aren't being
• 2:58
  violated, and the Pentagon viewed this request for
• 3:01
  transparency as a massive liability.
• 3:04
  Because it's an operational security issue for them.
• 3:06
  Right. They saw a private vendor in San
• 3:08
  Francisco claiming they had the right to audit a military
• 3:11
  operation. To the generals, that just looks
• 3:14
  like a vendor trying to exercise veto power over national
• 3:17
  security decisions. They see a tool they bought and
• 3:19
  the tool maker wants to verify if it was used to hurt someone.
• 3:23
  And that friction LED directly to the ultimatum.
• 3:26
  The Pentagon, led by Secretary Pete Hegseth, issued a directive
• 3:30
  to all AI vendors. The rule was very simple.
• 3:34
  AI vendors must allow their tools to be used for all lawful
• 3:38
  purposes. All lawful purposes.
• 3:40
  That phrase does a lot of heavy lifting, but Anthropic refused
• 3:44
  to agree to that broad language. They did.
• 3:46
  They insisted on maintaining 2 very specific red lines in their
• 3:50
  contract, regardless of whether the government called it lawful
• 3:53
  or not. What were the 2 lines?
• 3:55
  1st, The AI cannot be used for mass domestic surveillance of
• 3:58
  American citizens. OK, that seems straightforward.
• 4:01
  And the second. Second, it cannot be used for
• 4:03
  fully autonomous weapons. Wait, hold on, let me just make
• 4:05
  sure I have this straight. What is the specific definition
• 4:07
  of an autonomous weapon in this context?
• 4:09
  We are talking about systems that select and engage targets
• 4:13
  without human intervention. In military terminology, it's
• 4:16
  lethal autonomous weapons systems.
• 4:19
  Meaning the software processes the sensor data, identifies a
• 4:23
  human being as a target, and authorizes the strike.
• 4:26
  Yes, all without a person ever pressing a button.
• 4:29
  So Anthropic basically said we will not build the software that
• 4:32
  decides who dies. Correct.
• 4:34
  The government's counter argument is that lawful use
• 4:37
  already covers those concerns. They argue that if an action is
• 4:40
  legal under the Constitution, a private company has no right to
• 4:43
  block the government from using its procured tools to execute
• 4:46
  it. But we if the Pentagon says
• 4:49
  lawful use is the only standard they will accept, and they
• 4:51
  refuse to sign a ban on autonomous weapons, are they
• 4:54
  admitting they plan to use them? That is the real Gray area, the
• 4:58
  Pentagon spokesperson stated. They have no interest in mass
• 5:01
  surveillance or autonomous weapons without human oversight
• 5:04
  right now. They actually called those fears
• 5:06
  fake. However, they refuse to sign a
• 5:08
  contract that explicitly forbade it.
• 5:11
  Exactly. Their position is entirely about
• 5:14
  operational flexibility. They on a you they cannot have a
• 5:18
  Terms of Service agreement overriding a commander's
• 5:20
  decision in the field 5 or 10 years from now.
• 5:23
  And Anthropic is saying the technology just isn't safe
• 5:26
  enough for that. That is their core argument.
• 5:28
  They point out that current large language models
• 5:31
  hallucinate. They make up facts.
• 5:33
  They misinterpret context entirely.
• 5:35
  Which is a known issue across the industry.
• 5:37
  Yes. So Anthropic believes relying on
• 5:40
  them for autonomous killing is reckless and mass surveillance
• 5:44
  violates fundamental rights. They view these strictly as
• 5:47
  safety issues. While the administration argues
• 5:50
  that lawful is the only standard that matters and what corporate
• 5:53
  policies cannot constrain the military?
• 5:55
  Exactly. OK, let's pause here to reset
• 5:58
  the pace. Usually when a company says no
• 6:00
  to the government, they just lose the contract.
• 6:02
  They don't get the money. Right.
• 6:04
  You walk away, you lose the revenue, maybe your stock dips a
• 6:06
  bit, but you just continue doing business with everyone else.
• 6:10
  But the administration didn't just cancel the contract in this
• 6:12
  case, they went much further. They absolutely did.
• 6:16
  President Trump ordered all federal agencies to immediately
• 6:19
  cease using Anthropics technology, and then Secretary
• 6:23
  Hegseth followed up by designating the company a supply
• 6:25
  chain risk under 10US Code Section 3252.
• 6:29
  I want to focus on that specific statute, NUS Code section 3252.
• 6:33
  What does that actually mean? This is a statute designed
• 6:36
  specifically to prevent espionage and sabotage by
• 6:39
  foreign enemies. It is the exact legal tool used
• 6:43
  to ban Huawei and ZTE because of fears they were funneling data
• 6:47
  straight to the Chinese Communist Party.
• 6:48
  So it implies the company is an active threat to the integrity
• 6:51
  of the nation's defense. Yes, and applying it to an
• 6:54
  American firm for a contract dispute is entirely
• 6:56
  unprecedented. It acts as a corporate death
• 6:58
  penalty in the federal sector because.
• 7:00
  It's not just the Pentagon that has to top using them now.
• 7:02
  Exactly. It effectively forces any
• 7:04
  company that wants to work with the military.
• 7:06
  Companies like Boeing, Lockheed Martin or Palantir to strip
• 7:10
  Anthropic out of their own internal workflows.
• 7:12
  So if I'm an engineer at Lockheed Martin and I use Claude
• 7:15
  to write code or summarize technical documents for my daily
• 7:19
  work. You are now a liability.
• 7:21
  Lockheed Martin cannot risk its massive government contracts by
• 7:26
  harboring a designated supply chain risk in its software
• 7:29
  stack. They have to RIP it completely
• 7:31
  out. There's a massive contradiction
• 7:34
  in the order itself though. It claims Anthropic is a
• 7:37
  security risk, but simultaneously mandates they
• 7:40
  continue providing services for a six month transition period.
• 7:43
  Which really highlights that this isn't about espionage.
• 7:47
  Think about it, if Huawei was actively spying on the Pentagon,
• 7:50
  you wouldn't say, OK, keep the routers plugged in for six more
• 7:54
  months while we find a replacement.
• 7:55
  Right, you would cut the line immediately.
• 7:57
  You would sever it that second. So that tells us they need the
• 7:59
  tech. They absolutely need the tech,
• 8:01
  they just don't want the rules that come with it.
• 8:03
  Keeping them on for six months admits that the government is
• 8:06
  highly dependent on these models.
• 8:08
  So this designation is just a power play.
• 8:10
  Completely. It threatens Anthropics entire
• 8:13
  enterprise ecosystem and its projected IPO.
• 8:16
  It signals to the rest of Silicon Valley that
• 8:18
  noncompliance comes with an existential cost.
• 8:21
  And almost immediately, another company stepped in to take
• 8:24
  advantage. Yes, hours after the ban on
• 8:27
  Anthropic was announced, open AICEO, Sam Altman announced a
• 8:32
  new agreement. The pivot was instant open AI
• 8:35
  stepped right into the vacuum. They announced an agreement to
• 8:38
  deploy open AI models on the Department of War's classified
• 8:41
  networks. Here's the part that is really
• 8:43
  confusing to me. Open AI claims they have the
• 8:46
  exact same red lines as Anthropic regarding surveillance
• 8:50
  and autonomous weapons. Publicly, yes, their safety
• 8:53
  guidelines prohibit the exact same things.
• 8:55
  So if they have the same red lines, why did the Pentagon ban
• 8:59
  one and sign the other? Why is Anthropic a security risk
• 9:02
  and Open AIA trusted partner? It comes down to the mechanism
• 9:05
  of enforcement. Anthropic wanted strict
• 9:08
  contractual prohibitions. They wanted a written veto in
• 9:11
  the deal that they could enforce legally.
• 9:14
  Open AI adopted A layered approach instead.
• 9:16
  What does a layered approach actually mean in this context?
• 9:19
  It means they are relying on a cloud only deployment structure
• 9:22
  and the government's own interpretation of the law.
• 9:25
  Open AI agreed to the All Lawful Use standard.
• 9:28
  So they are effectively betting that the government won't define
• 9:31
  mass surveillance or autonomous weapons as lawful.
• 9:34
  Correct. Altman stated that the
• 9:35
  department agrees domestic surveillance is illegal, so they
• 9:38
  didn't need to fight over it. Open AI integrated their safety
• 9:42
  stack into the deployment rather than demanding external
• 9:45
  oversight of specific missions. So Anthropic wanted the
• 9:48
  government to sign a paper saying we promise not to do X,
• 9:52
  Open AI said. We know X is illegal, so we
• 9:55
  don't need you to sign a paper. Essentially, yes.
• 9:58
  They gave the Pentagon the lawful use language they
• 10:00
  demanded while assuring the public that their technical
• 10:03
  architecture prevents the bad stuff.
• 10:05
  It sounds like Open AI gave the Pentagon an optical win.
• 10:08
  And a contractual 1 The Pentagon gets to say we don't bow to
• 10:11
  terms of service and Open AI gets the massive contract.
• 10:14
  During this entire fallout, the administration publicly attacked
• 10:18
  Anthropic as a radical left and woke company.
• 10:21
  The record really contradicts that characterization.
• 10:24
  Though Anthropic is funded by major corporate players like
• 10:27
  Amazon and Google, their CEO Dario Amadei has publicly stated
• 10:31
  that AI is existentially important for National Defense.
• 10:35
  He's not exactly a pacifist. Far from it.
• 10:37
  He has explicitly stated he supports helping the US military
• 10:41
  defeat autocratic adversaries. He even admitted in interviews
• 10:45
  that autonomous weapons might eventually be necessary.
• 10:48
  Wait, really? So what is his actual objection?
• 10:51
  His objection is entirely that the current technology isn't
• 10:54
  safe enough yet. He argues that LLMS hallucinate
• 10:57
  and make unpredictable errors and therefore shouldn't be
• 11:01
  making keel decisions today. That seems like a technical
• 11:04
  argument, not a political one. Exactly.
• 11:06
  It suggests the woke label is just political cover.
• 11:09
  The real issue here is the sovereign AI doctrine.
• 11:12
  Meaning what? The administration is
• 11:14
  establishing that the state, not private labs, is the final
• 11:17
  arbiter of how technology is deployed.
• 11:20
  It is a fundamental shift from a safety first model to a
• 11:22
  deployment first mandate. They are worried about falling
• 11:25
  behind. They are looking straight at
• 11:27
  China. The argument is that if American
• 11:29
  companies are permitted to constrain the military with
• 11:31
  safety guardrails, the United States will lose the advantage
• 11:34
  to an adversary that operates without any such constraints.
• 11:38
  They want the tech to be totally subservient to national security
• 11:41
  objectives. This supply chain risk, does it
• 11:44
  stop at the Pentagon? No, the General Services
• 11:48
  Administration has already terminated Anthropics One Gov
• 11:51
  deal. That means agencies totally
• 11:54
  outside the military, like the Department of Energy or the EPA,
• 11:58
  might have to stop using Quad. It it effectively chills the
• 12:01
  entire market for them. It forces every company to make
• 12:04
  a hard choice. If you want to do business with
• 12:06
  the US government in any capacity, you cannot use the
• 12:09
  Risk software. It consolidated power strictly
• 12:12
  around the vendors who agree to the government's terms.
• 12:15
  To wrap this up, Anthropic held the line on their terms of
• 12:18
  service and was designated a National Security Risk Open AI
• 12:22
  aligned with the government's legal framework and became the
• 12:24
  primary partner for the military.
• 12:26
  The main take away here is that the Silicon Valley consensus,
• 12:29
  this era where companies felt they could dictate ethical terms
• 12:32
  to the government, is effectively dead.
• 12:34
  The government has demonstrated it is entirely willing to use
• 12:37
  the full weight of executive power to ensure AI companies are
• 12:40
  subservient to national security objectives.
• 12:43
  If you're not subscribed yet, take a second and hit follow on.
• 12:47
  Whatever a you're using, it helps us keep making this.
• 12:49
  We appreciate you being here.