Latest / Elon Musk Podcast / AI UPDATE: How a Drone Strike Crashed Claude

Transcript

AI UPDATE: How a Drone Strike Crashed Claude · Elon Musk Podcast

• 0:00
  A physical drone strike on a server farm in the desert caused
• 0:03
  millions of programmers and businesses worldwide to
• 0:06
  instantly lose access to their primary AI coding tools.
• 0:09
  Yeah, the sheer scale of that disruption is staggering,
• 0:12
  honestly. A localized conflict in the
• 0:15
  Middle East collided directly with a massive political protest
• 0:18
  against a major tech company, right?
• 0:20
  And that created a perfect storm that just completely took down
• 0:23
  Anthropics clawed AI. So how did a public boycott over
• 0:28
  military AI use end up paralyzing global tech
• 0:31
  workflows? And what does this physical
• 0:33
  vulnerability mean for everyone relying on these systems?
• 0:37
  Well, the catalyst for all of this was Open AI striking a deal
• 0:41
  with the US Department of Defense, OK.
• 0:43
  And this deal allow their technology to be integrated into
• 0:46
  systems used for mass surveillance and autonomous
• 0:48
  weapons. Which is a massive contrast to
• 0:50
  Anthropic because they actively refused a similar deal.
• 0:54
  Yeah, they did. They actually cited their
• 0:55
  responsible scaling policy. Specifically, they pointed to
• 0:58
  their. Their ASL 3 standards.
• 1:00
  Right, exactly. Meaning the AI is sophisticated
• 1:03
  enough that it could potentially help design cyber attacks or
• 1:06
  even biological and chemical weapons.
• 1:09
  Yeah, so Anthropic's policy basically dictates that once
• 1:12
  their models hit that specific level, they completely refuse to
• 1:16
  let the tech anywhere near military applications.
• 1:19
  And because of that refusal, the US government labeled Anthropic
• 1:22
  a supply chain risk. They actually ordered federal
• 1:26
  agencies to stop using their products entirely.
• 1:29
  Wait back up a supply chain risk.
• 1:31
  Yeah, that exact label. But that designation is
• 1:33
  completely unprecedented for an American company.
• 1:36
  It really is. Historically, you only see the
• 1:38
  supply chain risk label reserved strictly for foreign
• 1:41
  adversaries, like foreign telecommunications companies or
• 1:44
  overseas hardware manufacturers where there's a fear of
• 1:47
  espionage. Applying that to a domestic tech
• 1:49
  developer operating right out of San Francisco is a massive
• 1:52
  escalation by the government. So that government action
• 1:55
  triggered a huge public backlash against Open AI.
• 1:58
  Users flooded away from ChatGPT and moved directly over to
• 2:01
  Claude in protest. Pushing the quad app to the
• 2:04
  number one spot on the App Store overnight.
• 2:07
  Exactly. And to capitalize on the
• 2:09
  boycott, Anthropic released a free memory import tool.
• 2:13
  Oh yeah, the migration tool. Right, so it allowed users to
• 2:16
  easily port their entire chat histories, their custom
• 2:19
  instructions and preferences directly over from ChatGPT.
• 2:23
  You could literally just use a specific prompt to extract all
• 2:26
  your stored memories from Open AI into a single massive code
• 2:30
  block. Just copy that block, paste it
• 2:32
  right into Claude. And instantly the new system
• 2:34
  knew your workflow. But that massive influx of new
• 2:37
  users completely overwhelmed Claude's authentication systems.
• 2:41
  Yeah, people trying to login were suddenly hit with HTTP50529
• 2:46
  errors. A 529 error, basically meaning
• 2:48
  the system is completely overloaded and cannot process
• 2:51
  another request. Freezing coding workflows
• 2:54
  everywhere. You just had developers staring
• 2:56
  at a blank screen with a tiny message saying Claude will
• 2:58
  return soon. I do want to push back slightly
• 3:00
  on the severity of the failure though.
• 3:02
  OK. How so?
• 3:03
  Because while the consumer facing web interface and the
• 3:06
  Claude code app completely broke down, the core API actually
• 3:09
  remained fully operational. I see.
• 3:12
  Yeah, and the status page showed that Claude for Government was
• 3:16
  completely unaffected. It was running normally.
• 3:18
  So that limits the scope of the technical failure considerably,
• 3:21
  right? It wasn't the actual underlying
• 3:23
  AI model that broke under the pressure.
• 3:25
  The core engine was fine. It was just the front door login
• 3:28
  system struggling under the weight of unexpected consumer
• 3:31
  demand. Exactly.
• 3:33
  The side doors. The APIs were still operating
• 3:35
  perfectly, but ironically, at the exact same time as that
• 3:39
  digital user surge, AWS data centers in the United Arab
• 3:43
  Emirates and Bahrain were hit by drone strikes and caught fire.
• 3:47
  Wow. We talk about the cloud like
• 3:49
  it's this invisible magical entity floating above us, but
• 3:53
  it's incredibly physical. It's just endless rows of
• 3:55
  hardware sitting in massive warehouses.
• 3:58
  Yeah, and the fire department response really highlights that
• 4:01
  physical vulnerability because when a server farm catches fire,
• 4:04
  you cannot just spray water on incredibly high voltage
• 4:07
  equipment, right? The local fire department had to
• 4:09
  completely cut the power to the facility.
• 4:12
  They forced the operators to shut off the backup generators
• 4:14
  just to manage the blaze safely. And the tech company couldn't do
• 4:18
  any anything to override it. They literally had to wait for
• 4:21
  official permission from local authorities to restore power to
• 4:25
  those availability zones. So this physical destruction
• 4:28
  combined with the digital user search created an absolute
• 4:31
  infrastructure crisis. You.
• 4:32
  Have a physical explosive traveling thousands of
• 4:35
  kilometers down fiber optic cables to essentially puncture
• 4:38
  the monitors in front of programmers.
• 4:40
  Turning them into digital refugees overnight.
• 4:42
  And that changes how enterprises view cloud reliance forever.
• 4:47
  It limits the perceived security of centralized AI architecture,
• 4:51
  opening up an urgent demand for decentralized, sovereign and
• 4:54
  physically hardened AI environments.
• 4:56
  If you've got a decent mic and laptop and some free time,
• 4:59
  Babble Audio is paying people to record speech data and annotate
• 5:03
  audio for AI training. No minimums, no fixed hours.
• 5:06
  You work when you want and get paid weekly via PayPal, Venmo,
• 5:09
  or bank transfer. They pay per recorded or
• 5:11
  annotated hour, plus bonus challenges for hitting weekly
• 5:14
  goals. And if you sign up through our
• 5:16
  link, you get priority processing and a $15 bonus.
• 5:19
  Links in the show notes. Shifting focus a bit, the reason
• 5:22
  the loss of Claude was so paralyzing for businesses in the
• 5:25
  1st place is that Anthropic had just released Claude Opus 4.6
• 5:29
  and Son at 4.6 and this massive launch was backed by a $30
• 5:34
  billion funding round that value the entire company at $380
• 5:39
  billion. What exactly makes 4.6 different
• 5:43
  from the previous versions? Why were businesses so dependent
• 5:46
  on this specific update? They introduced a feature called
• 5:48
  adaptive thinking. The model dynamically decides
• 5:51
  when to think deeply based on an adjustable effort parameter.
• 5:55
  OK, so you can control how hard it works on a prompt.
• 5:58
  Exactly. If you ask a simple question, it
• 6:00
  uses low effort and answers instantly, but if you ask it to
• 6:03
  restructure a massive database, it cranks the effort to Max and
• 6:07
  spends minutes calculating before responding.
• 6:09
  Interesting. There's also a major feature
• 6:11
  called Agent Teams, allowing multiple AI agents to work in
• 6:14
  parallel on complex code bases. So it's essentially like hiring
• 6:18
  an entire team of junior developers who can message each
• 6:21
  other directly to coordinate tasks.
• 6:23
  Exactly, and Opus 4.6 features a massive 1,000,000 token context
• 6:28
  window. Which means the system can hold
• 6:30
  roughly the equivalent of several thick books of
• 6:33
  information in its active memory all at once.
• 6:35
  Right. It can read through through
• 6:37
  hundreds of PDFs or an entire software repository in seconds
• 6:42
  without forgetting the beginning by the time it reaches the end.
• 6:44
  Not only can it read that much, but it can also generate 128,000
• 6:49
  output tokens in a single response.
• 6:51
  Enough capacity to write complete technical documentation
• 6:55
  or execute full code base migrations in a single prompt.
• 6:59
  And to handle those massive memory loads without crashing
• 7:02
  the servers, Anthropic introduced Context Compaction.
• 7:05
  Right, the API that automatically summarizes older
• 7:08
  parts of the conversation. It compresses the old
• 7:10
  information so it doesn't lose the foundation of the
• 7:12
  discussion, Which. Changes the technology from a
• 7:15
  simple chat bot into a system capable of effectively infinite
• 7:18
  conversations. It makes businesses incredibly
• 7:21
  dependent on its continuous operation because they can
• 7:23
  maintain context over multi month projects without ever
• 7:27
  losing the threat. Which brings us right back to
• 7:29
  the vulnerability. Relying entirely on centralized
• 7:32
  AI infrastructure means that a localized geopolitical conflict
• 7:37
  or a sudden surge from a user boycott can instantly halt
• 7:41
  global productivity. If a single outage can instantly
• 7:44
  turn you into a digital refugee, how long until businesses stop
• 7:48
  renting intelligence from the cloud and start demanding
• 7:50
  physical control over their own AI?
• 7:52
  If you're not subscribed yet, take a second and hit follow on
• 7:55
  whatever app you're using. It helps us keep making this.
• 7:57
  We appreciate you being here.