Latest / Elon Musk Podcast / Military AI partnership triggers record ChatGPT uninstalls

Transcript

Military AI partnership triggers record ChatGPT uninstalls · Elon Musk Podcast

• 0:00
  ChatGPT app uninstalls just surged by 295%, accompanied by a
• 0:05
  775% spike in one star reviews. Yeah, the sheer volume of that
• 0:10
  exodus is staggering. I mean, you almost never see an
• 0:13
  established dominant platform experience a mass abandonment of
• 0:17
  that scale so quickly. And this massive shift in user
• 0:20
  behavior, It all stems directly from.
• 0:22
  Open AI signing a military partnership with the US
• 0:25
  Department of Defense. People are leaving in droves,
• 0:28
  and meanwhile their main rival Chat bought Claude, just
• 0:31
  rocketed straight to #1 on the App Store.
• 0:33
  So we are exploring exactly how a massive government contract
• 0:36
  essentially fractured the artificial intelligence industry
• 0:39
  and completely shifted consumer loyalties.
• 0:41
  Right, because it leads directly to the core issue we're looking
• 0:44
  at today. When commercial models become
• 0:46
  battlefield brains, who ultimately holds the kill
• 0:49
  switch? So opening I partnered with the
• 0:51
  Department of Defense under the Trump administration, and they
• 0:53
  simultaneously closed $110 billion funding round.
• 0:57
  That move effectively shifts them from functioning as a
• 0:59
  consumer research lab into operating as critical national
• 1:03
  infrastructure. $110 billion, we really have to just sit with
• 1:07
  that number for a second. We are looking at a fundamental
• 1:10
  reclassification of what this company actually is.
• 1:13
  Right, they aren't just an app on your phone anymore.
• 1:15
  Exactly. They are no longer just building
• 1:17
  a helpful digital assistant to draft your emails or, you know,
• 1:22
  brainstorm dinner recipes. They have embedded themselves
• 1:25
  into the machinery of national security.
• 1:29
  But what does that actually look like in practice?
• 1:31
  Because I think a lot of people hear the phrase military
• 1:33
  artificial intelligence and the immediately picture autonomous
• 1:37
  robot soldiers marching across a battlefield.
• 1:40
  Yeah, the reality is less cinematic, but it is far more
• 1:42
  pervasive. It is primarily about processing
• 1:45
  unimaginable amounts of information O Imagine the model
• 1:48
  you use to summarize a long PDF document.
• 1:51
  Sure, I use it all the time. Right.
• 1:52
  So now, instead of a DF, that system is ingesting thousands of
• 1:55
  interceted radio transmissions and satellite imagery and drone
• 1:59
  video feeds in real time. It parses all of that raw noise
• 2:03
  to suggest troop movements or identify specific targets or
• 2:07
  even synthesize complex intelligence reports for a
• 2:10
  commanding officer. That is the technological leap
• 2:13
  we are talking about here. And that reality changes the
• 2:15
  entire public perception of the tool you might have sitting on
• 2:18
  your phone right now. When you pull out your phone to
• 2:21
  have the app generate a workout routine, you have to realize
• 2:25
  you're interacting with the exact same underlying
• 2:27
  architecture that might be analyzing drone footage in a war
• 2:31
  zone that. Dual use nature creates A
• 2:33
  profound discomfort for the user.
• 2:35
  I mean, there are intense public fears right now about the
• 2:37
  technology being utilized for domestic surveillance and
• 2:40
  autonomous weapons. Which absolutely limits Open
• 2:43
  AI's consumer goodwill. You can trace a Direct Line from
• 2:46
  those surveillance fears to the massive uninstall rate and that
• 2:49
  50% drop in five star ratings they just experienced.
• 2:52
  People are actively deciding they do not want a military
• 2:55
  contractor operating in their pocket.
• 2:57
  It alters the implied contract with the user.
• 3:00
  People originally adopted the tool because it felt like a
• 3:03
  creative sandbox. You could ask it funny questions
• 3:06
  or have it write a poem. Now the context is entirely
• 3:09
  shifted. The server farm processing your
• 3:11
  casual chat is owned by an entity building tools for the
• 3:14
  Defense Department. Hold on, wait back up.
• 3:16
  Let's look at how the competition handled this,
• 3:18
  because it is essentially the exact opposite approach.
• 3:21
  Yeah, Anthropic, the company behind the Claude chatbot, took
• 3:24
  a completely different path. They actually walked away from a
• 3:26
  $200 million deal. 200 million just left on the table.
• 3:30
  Just walked away. Their CEO, Dario Amede,
• 3:33
  established firm boundaries against mass surveillance and
• 3:36
  fully autonomous weapons. He stated clearly that the
• 3:39
  technology is simply not reliable enough to be trusted
• 3:42
  with lethal force. Refusing that kind of money
• 3:45
  absolutely limits Anthropic's immediate government revenue,
• 3:49
  but it opens up a massive consumer trust advantage.
• 3:52
  And we really should talk about why a leading tech CEO would
• 3:56
  explicitly say his own product is not reliable.
• 3:58
  It comes down to a technical reality called hallucination.
• 4:02
  These models do not know facts the way a human does.
• 4:06
  They are essentially highly advanced prediction engines,
• 4:09
  guessing the next most likely word in a sequence based on
• 4:12
  their training data. Sometimes they confidently
• 4:15
  invent information that sounds entirely plausible but is
• 4:18
  completely false. Which is funny if you ask it for
• 4:21
  a historical fact and it makes up a fake battle.
• 4:23
  Exactly. If a chat bot hallucinates a
• 4:26
  legal citation in a brief batting, it is an embarrassing
• 4:29
  professional mistake. But if the system hallucinates a
• 4:33
  target identification in a lethal military strike, like
• 4:36
  confidently identifying a civilian convoy as a hostile
• 4:39
  threat. Then it is an absolute
• 4:40
  catastrophe. Exactly.
• 4:42
  That makes perfect sense from an engineering perspective, but
• 4:45
  drawing that ethical line radically changes anthropic
• 4:48
  standing with the federal government. the US actually
• 4:50
  designated anthropic as a supply chain.
• 4:52
  Risk that. Is an incredibly severe label.
• 4:55
  We usually see that designation reserved for hostile foreign
• 4:58
  entities, like the restrictions placed on the telecom company
• 5:01
  Huawei. So what does that designation
• 5:03
  actually mean for an American company day-to-day?
• 5:05
  Practically, being labeled a supply chain risk means you get
• 5:08
  locked out of lucrative government grants.
• 5:11
  It can trigger strict export restrictions, making it hard to
• 5:14
  sell your software internationally.
• 5:16
  It also makes it incredibly difficult to partner with any
• 5:19
  university or research lab that receives federal funding.
• 5:22
  Because those institutions cannot risk their own funding by
• 5:25
  working with a blacklisted entity, right?
• 5:27
  The friction there is fascinating.
• 5:29
  You have an American company building state-of-the-art
• 5:31
  technology, and because they draw a firm ethical line
• 5:35
  regarding how that technology is deployed, they get categorized
• 5:39
  alongside foreign adversaries. It places them in a highly
• 5:42
  adversarial relationship with the very defense establishment
• 5:46
  Open AI just embraced. It brings up a massive
• 5:49
  philosophical divide between the leaders of these two companies.
• 5:52
  Sam Altman's stance is that private companies should not
• 5:54
  dictate global ethics. He argues that elected
• 5:57
  government should be the ones making these massive decisions
• 5:59
  about national security and military deployment, rather than
• 6:02
  tech executives. I actually agree with Altman's
• 6:04
  pragmatism here. Elected officials should be the
• 6:07
  ones controlling national security tools rather than tech
• 6:11
  CE OS. See, I completely disagree with
• 6:13
  that. In theory, sure, but the flaw
• 6:16
  there is assuming a general or a senator actually understands the
• 6:20
  black box mechanics of a neural network.
• 6:22
  Well, they aren't software engineers.
• 6:24
  Exactly. If you build a bridge and you
• 6:25
  know the steel is faulty, you don't let the government drive
• 6:28
  tanks over it just because they are the government.
• 6:31
  Amede is looking at the raw capabilities of the models and
• 6:34
  saying they are inherently flawed.
• 6:36
  They are not ready for life or death scenarios.
• 6:39
  Accountability after the fact does not fix a misfired
• 6:43
  autonomous weapon. I hear that, but government
• 6:46
  defense agencies test and validate their own equipment
• 6:49
  extensively. If the military decides the
• 6:51
  technology provides A strategic advantage, that is a decision
• 6:55
  for the Defense Department to make.
• 6:56
  By taking the stance Almond did, you put the tools in the hands
• 6:59
  of democratic institution? But passing a buck to an elected
• 7:03
  official doesn't suddenly make the underlying code more stable
• 7:06
  in a legal scenario. If you know the tool
• 7:09
  hallucinates, handing it over for lethal purposes is a
• 7:12
  fundamental failure of responsibility.
• 7:14
  But you ensure there is public oversight through the electoral
• 7:18
  process, rather than leaving global ethics entirely up to the
• 7:22
  whims of unelected tech CEO's sitting in Silicon Valley
• 7:26
  boardrooms. If a democratic government
• 7:29
  determines they need advanced technology for defense, private
• 7:32
  citizens running a software company shouldn't have the power
• 7:36
  to just veto National Defense strategy.
• 7:38
  That assumes the oversight is capable of understanding the
• 7:40
  technology in real time, which is a huge leap of faith.
• 7:43
  The core issue remains the reliability of the system
• 7:46
  itself. The CEO knows the actual
• 7:48
  limitations of the software better than a military general
• 7:51
  does. We.
• 7:51
  Are going to get into how users are actively taking sides in
• 7:54
  this debate in a second, but first, Speaking of artificial
• 7:56
  intelligence training data. If you've got a decent mic and
• 7:59
  laptop and some free time, Babble Audio is paying people to
• 8:02
  record speech data and annotate audio for AI training.
• 8:06
  No minimums, no fixed hours. You work when you want and get
• 8:09
  paid weekly via PayPal, Venmo, or bank transfer.
• 8:12
  They pay per recorded or annotated hour, plus bonus
• 8:15
  challenges for hitting weekly goals.
• 8:17
  And if you sign up through our link, you get priority
• 8:19
  processing and a $15 bonus. Links in the show notes.
• 8:23
  So while all of this high level ethical maneuvering is
• 8:26
  happening, we are also watching a master class in capitalizing
• 8:29
  on a competitor's vulnerability. Anthropic fully capitalized on
• 8:33
  the public outrage by actively courting the defectors leaving
• 8:37
  open AI. They absolutely did.
• 8:38
  They enabled free context recall for their users.
• 8:41
  We should explain what context recall is, because a layperson
• 8:44
  might not realize why that matters or why it usually cost
• 8:47
  money. Context recall basically means
• 8:49
  the artificial intelligence can remember specific details from
• 8:52
  way earlier in your conversation without losing the threat of the
• 8:55
  discussion. Like remembering a specific
• 8:57
  formatting instruction. You gave it 30 messages prior.
• 9:00
  Exactly. Maintaining that kind of active
• 9:03
  memory requires immense computational processing power.
• 9:06
  It is incredibly expensive to run on the server side, so
• 9:09
  giving that capability away for free was a massively expensive,
• 9:12
  highly calculated move to attract power users.
• 9:15
  They also heavily promoted a tool that allows you to
• 9:17
  seamlessly import your chat histories and memories from
• 9:21
  competitors right into Claude. The user behavior around this
• 9:25
  migration has been incredible to watch.
• 9:27
  People are actively seeking advice online asking for step by
• 9:30
  step guides on how to migrate their data out of open a eyes
• 9:33
  ecosystem. They're explicitly talking about
• 9:35
  wanting to completely delete ChatGPT to, in their exact
• 9:38
  words, stick it to the man. It is a highly motivated, almost
• 9:42
  ideological migration. Think about how hard it usually
• 9:44
  is to get people to switch apps. Getting someone to switch from
• 9:47
  Spotify to Apple Music, or from an iPhone to an Android device
• 9:50
  is nearly impossible because of the friction.
• 9:53
  People hate losing their playlists, their photos, their
• 9:55
  save data. People stay in software
• 9:57
  ecosystems they dislike simply because the idea of manually
• 10:00
  copying over hundreds of saved prompts or chat logs is
• 10:03
  exhausting. And providing the import tool
• 10:06
  entirely eliminates that friction point for consumers.
• 10:09
  By removing the barrier of lost data, Anthropic opens the
• 10:14
  floodgates for mass migration. That single feature directly
• 10:18
  fueled Claude's 51% spike in downloads and pushed it straight
• 10:22
  to the top of the App Store charts.
• 10:24
  It's entirely strategic. They saw an opening created by a
• 10:27
  competitors public relations crisis and they immediately
• 10:30
  provided the exact technical bridge users needed to abandon
• 10:34
  ship without losing their personal archives.
• 10:36
  Which forced Open AI into a very uncomfortable position publicly.
• 10:40
  Sam Altman actually had to publicly admit that the rollout
• 10:43
  of the defense deal looked opportunistic and sloppy.
• 10:46
  Open AI then quickly amended its contract to explicitly prohibit
• 10:49
  domestic surveillance under U.S. law.
• 10:51
  That contract amendment legally limits Open AI's involvement in
• 10:55
  domestic spying. It is a direct response to the
• 10:58
  major public fears we have been discussing.
• 11:00
  However, defining surveillance in a legal context is incredibly
• 11:03
  murky. Is analyzing public social media
• 11:05
  feeds considered surveillance? Is processing commercially
• 11:08
  purchased cell phone location data considered surveillance?
• 11:11
  The technical loopholes are vast.
• 11:12
  Exactly. And the amendment exposes a
• 11:15
  company operating in pure damage control mode.
• 11:18
  They are reacting defensively as the uninstalls continue to
• 11:20
  climb, trying to put the toothpaste back in the tube.
• 11:23
  Once consumer trust is fractured on heavy issues like domestic
• 11:26
  surveillance and military alignment, a simple contract
• 11:29
  revision rarely brings users back immediately.
• 11:32
  The perception is already set in the public consciousness.
• 11:36
  You can't just release a legal addendum and expect users to
• 11:39
  forget that they felt uncomfortable with the
• 11:41
  application on their phone. No, they have already made-up
• 11:44
  their minds. So the competition between these
• 11:46
  two tech giants is now entirely driven by consumer trust and
• 11:50
  military alignment. If opting out of defense
• 11:52
  contracts gets a tech company labeled a national security
• 11:55
  risk, is it even possible to build neutral technology
• 11:58
  anymore? If you're not subscribed yet,
• 12:00
  take a second and hit follow on whatever app you're using.
• 12:02
  It hels us Kee making this. We appreciate you being here.