Latest / Elon Musk Podcast / OpenAI signs the military deal Anthropic refused

Transcript

OpenAI signs the military deal Anthropic refused · Elon Musk Podcast

• 0:00
  Open AI signed a contract to put their AI models into classified
• 0:03
  military networks just hours after the US government
• 0:07
  blacklisted their rival Anthropic for refusing that
• 0:10
  exact same deal over ethical red lines.
• 0:13
  Yeah. The sheer speed of this sequence
• 0:15
  is staggering. Frankly, the immediate fallout
• 0:18
  resulted in open AICEO Sam Altman publicly admitting the
• 0:21
  deal was opportunistic and sloppy, while consumers just
• 0:25
  staged this impromptu boycott of ChatGPT.
• 0:28
  So how does a tech giant secure a massive military contract,
• 0:32
  claimed to have kept their ethical standards intact and
• 0:35
  instantly face an industry wide revolt?
• 0:37
  Well, you really have to look at anthropics hard limits first.
• 0:40
  They established 2 absolute, nonnegotiable boundaries for the
• 0:45
  military use of their technology, right?
• 0:47
  These were hardlines they brought to the negotiation table
• 0:49
  and they made it clear they would not proceed without.
• 0:51
  Them right. And these limits are highly
• 0:53
  specific. First, they demanded strict
• 0:55
  prohibitions against any mass surveillance of Americans.
• 0:58
  Second, they stipulated there could be no lethal autonomous
• 1:00
  weapons capable of striking targets without human oversight.
• 1:03
  They were entirely unwilling to budget on those two.
• 1:06
  Which provoked an extraordinarily severe response
• 1:09
  in the government. The Department of War, which is
• 1:11
  the Trump administrations preferred name for the Defense
• 1:14
  Department, completely rejected these terms.
• 1:17
  They didn't just walk away from the negotiation table.
• 1:19
  They officially blacklisted Anthropic.
• 1:22
  And the rhetoric surrounding this rejection was intense.
• 1:26
  Pete Hegseth promised a scorched earth campaign against the
• 1:29
  company. He took it a profound step
• 1:32
  further by declaring Anthropic Assume.
• 1:35
  A supply chain risk. The administrative weight of
• 1:37
  that specific designation is incredibly heavy.
• 1:40
  Labeling a software company. A supply chain risk legally
• 1:44
  forbids any military contractor or partner from conducting
• 1:47
  commercial activity with Anthropic.
• 1:49
  Wait, hold on back up. A supply chain risk designation
• 1:52
  means they are entirely cut off from doing commercial business
• 1:54
  with anyone who works the military.
• 1:56
  Exactly. Think about the modern tech
• 1:58
  ecosystem you interact with daily.
• 2:00
  Almost every major cloud provider, hardware manufacturer,
• 2:04
  and enterprise software company holds some form of government or
• 2:07
  military contract. If you are a military partner,
• 2:11
  this mandate dictates that you cannot do commercial business
• 2:14
  with a blacklisted entity. It creates this this radioactive
• 2:18
  zone around Anthropic. That is severe.
• 2:20
  It threatens to sever them from the entire technology
• 2:23
  infrastructure they need to survive.
• 2:25
  I mean, if you are a contractor building anything for the
• 2:28
  government and you use anthropic software to help write your code
• 2:32
  or draft your reports, you are suddenly in violation of federal
• 2:35
  rules. You have to RIP their software
• 2:37
  out of your entire infrastructure.
• 2:38
  Anthropic is threatening to sue over this, arguing the move is
• 2:41
  legally questionable. And in a surprising twist, Open
• 2:44
  AI actually came out and openly opposed this punishment against
• 2:48
  their rival. Yeah, they avoid support, but
• 2:50
  the timeline compresses incredibly fast.
• 2:52
  Right after that, Open AI stepped directly into the void
• 2:55
  Anthropic left behind. They successfully negotiated
• 2:57
  terms for the military contract almost immediately after
• 3:00
  Anthropic was ousted. And then Sam Altman claimed
• 3:04
  their newly signed agreement actually included the exact same
• 3:07
  safety principles Anthropic had demanded.
• 3:10
  He specifically stated the contract prohibits domestic mass
• 3:13
  surveillance and requires human responsibility for the use of
• 3:16
  force. Which presents us with a glaring
• 3:18
  contradiction here. The Pentagon definitively
• 3:21
  rejected these exact ethical boundaries from one company,
• 3:25
  penalizing them with a devastating blacklist.
• 3:28
  Then, almost in the same breath, they supposedly agreed to those
• 3:32
  identical boundaries with another company.
• 3:34
  The logic just does not track to me.
• 3:36
  Why would the Department of War fiercely guard their right to
• 3:39
  mass surveillance and autonomous strikes, wages scorched earth
• 3:42
  campaign to defend that right, and then instantly sign a
• 3:45
  contract with open AI bound by those exact same rules?
• 3:47
  They clearly just preferred Open AI for other reasons and use the
• 3:51
  red lines as an excuse to clear Anthropic off the board.
• 3:54
  I actually see it differently. You have to consider the
• 3:56
  enforcement mechanisms behind these kinds of agreements.
• 3:59
  Contracts are highly standardized, heavily
• 4:02
  scrutinized documents. Anthropic likely demanded strict
• 4:05
  software level locks preventing those specific uses.
• 4:09
  They wanted the code itself to refuse the command if asked to
• 4:12
  conduct surveillance on Americans.
• 4:14
  Open AI, on the other hand, might have simply accepted a
• 4:17
  written promise from the government that they wouldn't
• 4:19
  use the tool that way. So basically relying on the
• 4:22
  honor system for lethal military operations.
• 4:24
  Precisely. And when you are dealing with
• 4:26
  classified operations, visibility into how a tool is
• 4:29
  actually being used vanishes the moment it crosses into secure
• 4:33
  networks. The company providing the
• 4:35
  software loses all oversight. If you've got a decent mic and
• 4:38
  laptop and some free time, Babble Audio is paying people to
• 4:41
  record speech data and annotate audio for AI training.
• 4:45
  No minimums, no fixed hours. You work when you want and get
• 4:48
  paid weekly via PayPal, Venmo or bank transfer.
• 4:51
  They pay per recorded or annotated hour, plus bonus
• 4:54
  challenges for hitting weekly goals.
• 4:56
  And if you sign up through our link, you get priority
• 4:58
  processing and a $15 bonus. Links in the show notes.
• 5:02
  O While Open AI was busy inking that deal, the public was
• 5:05
  definitely not having it. The reaction to Open AI's
• 5:09
  maneuver was instantaneous. The Internet voted with its
• 5:12
  downloads. Consumers rallied behind
• 5:14
  Anthropic in a sweeping way. Users began A spontaneous
• 5:18
  boycott of ChatGPT, abandoning the platform and flocking to
• 5:21
  Anthropics Alternative Claude. The scale of that migration was
• 5:25
  incredible. Claude surged past ChatGPT to
• 5:28
  become the top app in the Apple App Store.
• 5:30
  The influx of new traffic was so sudden and overwhelming, the
• 5:33
  Claude actually crashed briefly under the weight of it all.
• 5:35
  We see consumer boycotts over physical goods or streaming
• 5:38
  services frequently, but watching 1 unfold in real time
• 5:42
  between enterprise level artificial intelligence giants
• 5:45
  is wild. People clearly felt a strong
• 5:48
  sense of loyalty to Anthropic for taking an immense financial
• 5:51
  hit to stand by their ethics, and they altered their daily
• 5:54
  tech habits to show support. This public relations disaster
• 5:58
  forced a very unusual response from Sam Altman.
• 6:02
  He took to social media and openly admitted the deal was
• 6:05
  rushed. He used the exact words,
• 6:08
  opportunistic and sloppy, to describe his own company's
• 6:11
  handling of the situation. Aceo of that magnitude calling
• 6:14
  his own major military contract sloppy is definitely not
• 6:17
  something you hear often. He was entirely in damage
• 6:19
  control mode. He prompts immediate revisions
• 6:22
  to the agreement. Specifically, he highlighted new
• 6:24
  language stating the AI shall not be intentionally used for
• 6:27
  domestic surveillance of US persons and nationals.
• 6:30
  That word intentionally is doing a tremendous amount of heavy
• 6:33
  lifting in that sentence. Right creates A buffer for
• 6:36
  accidental or incidental surveillance.
• 6:38
  If a system happens to sweep up data on U.S. citizens while
• 6:42
  conducting A broader operation, the intentional clause might
• 6:45
  protect the deployment. Legally, it provides a degree of
• 6:48
  cover that a strict absolute prohibition does not.
• 6:52
  Think about the internal politics of this.
• 6:54
  If you work there, Open AI spends enormous sums of money
• 6:57
  acquiring the top AI talent in the world.
• 7:00
  These researchers and engineers are historically very protective
• 7:03
  of how their creations are used. If you are one of those
• 7:06
  engineers watching this unfold, you are looking at a management
• 7:08
  team that rushed a contract, faced a public backlash, and
• 7:11
  then tried to patch the ethical holes with highly specific legal
• 7:15
  phrasing. The internal risk for open AI is
• 7:17
  severe. The engineering culture within
• 7:20
  AI labs is deeply focused on alignment, meaning ensuring the
• 7:23
  AI's goals perfectly match human values and safety constraints.
• 7:27
  Highly critical staff members may view these newly added
• 7:30
  loopholes and the handling of the contract as unforgivable.
• 7:33
  It absolutely opens up the possibility of a major employee
• 7:37
  revolt. If the talent building the
• 7:39
  technology decides this compromise crosses a line, they
• 7:42
  could easily walk out the door. And in this industry, losing
• 7:46
  your top engineers is a catastrophic blow to your
• 7:49
  development pipeline. Especially when they could walk
• 7:51
  straight over to Anthropic, a company that just very publicly
• 7:55
  martyred itself for those exact safety principles.
• 7:59
  The talent war and artificial intelligence is fierce, and
• 8:02
  securing the moral high ground is a highly effective recruiting
• 8:05
  tool when you were trying to hire people who care deeply
• 8:08
  about safety. We have to talk about the
• 8:09
  physical reality on the ground too, because this isn't just a
• 8:12
  philosophical debate happening in corporate boardrooms.
• 8:15
  Anthropics Claude is currently the only AI model the Pentagon
• 8:18
  actively uses in classified operations.
• 8:21
  That is a crucial operational detail.
• 8:24
  The military relies exclusively on Claude for the secure
• 8:27
  environments. They are actively using it in
• 8:29
  operations right now, including highly sensitive missions in
• 8:32
  Venezuela. So the government just banned
• 8:34
  the commercial ecosystem from touching the one software tool
• 8:38
  their own military relies on for active operations.
• 8:41
  Yes, and Higgseth granted the military a six month window to
• 8:44
  solve this problem. They have six months to
• 8:47
  completely phase out clawed and phase in replacement models from
• 8:50
  Open AI as well as models from Elon Musk's X AI.
• 8:54
  I cannot even begin to imagine the logistical nightmare of
• 8:58
  upping out the foundational intelligence engine of a
• 9:00
  classified military operation while that operation is actively
• 9:04
  running. It forces a highly complex and
• 9:06
  risky technical transition. If you swap out a standard word
• 9:09
  processor for a different brand, the text remains the same.
• 9:13
  But an AI model is not a simple software update.
• 9:16
  Different models interpret data differently.
• 9:18
  They hallucinate differently. They rioritize information in
• 9:21
  completely distinct ways. Think about the software
• 9:23
  architecture you use at work. Every day you build entire
• 9:26
  workflows and structures around the specific quirks of the tools
• 9:29
  you are using. If troops in the field are
• 9:32
  relying on how Claude processes intelligence and suddenly you
• 9:35
  swap the brain out for ChatGPT or XAI, the entire output
• 9:38
  changes. Now imagine someone telling you
• 9:40
  that you have to swap out that core operating system while
• 9:43
  you're in the middle of a critical project, just with
• 9:45
  actual lives on the line. In a classified military
• 9:48
  setting, unpredictability is dangerous.
• 9:51
  Introducing new unvetted models into a live operational
• 9:55
  environment carries immense risk.
• 9:57
  You have to retrain personnel, rewrite your prompts, and
• 10:00
  constantly verify that the new models are not misinterpreting
• 10:03
  critical data or fabricating intelligence.
• 10:06
  A rushed government contract left one company fighting an
• 10:09
  existential threat for sticking to its ethics while forcing the
• 10:12
  other to scramble and rewrite the rules of its own agreement
• 10:15
  amidst public backlash. As the military spends the next
• 10:18
  six months swapping out the only AI it currently trusts for
• 10:21
  classified operations, will open AI's newly romised safety
• 10:25
  language hold U under the pressure of actual deployment.
• 10:28
  If you're not subscribed yet, take a second and hit follow on
• 10:31
  whatever a you're using. It hels us Kee making this.
• 10:33
  We appreciate you being here.