Latest / AI Fire Daily / 🎙️ EP 335: Prompt Injection in Court Filing & Stripe Acquires OpenRouter for $7B+

Transcript

🎙️ EP 335: Prompt Injection in Court Filing & Stripe Acquires OpenRouter for $7B+ · AI Fire Daily

• 0:00
  Imagine someone trying to physically hack a human
• 0:02
  judge. That sounds pretty wild. Yeah, it does.
• 0:06
  You probably picture a sophisticated cyber attack
• 0:08
  right now. Maybe a hacker targeting a secure
• 0:11
  municipal server somewhere. Right, like some
• 0:14
  kind of firewall breach? Exactly. But this specific
• 0:16
  situation was much stranger than that. Oh, really?
• 0:19
  Yeah. A person filed a standard legal document
• 0:22
  in a local court, but they hid tiny three -point
• 0:26
  white text inside the margins. Wow. Invisible
• 0:30
  to the human eye. Completely invisible. They
• 0:32
  genuinely hoped an artificial intelligence would
• 0:34
  read those hidden words. That is just fascinating.
• 0:37
  They wanted that machine to secretly rig the
• 0:40
  case in their favor. Beat. Beat. Welcome to the
• 0:44
  deep dive. It's a truly wild scenario to consider
• 0:47
  today. Yeah, it really is. I mean, it really
• 0:49
  sounds like a bad science fiction movie script.
• 0:52
  It kind of does. But it's actually happening
• 0:54
  in our legal system right now. Today we're exploring
• 0:56
  the chaotic front lines of AI trust. It's a very
• 1:00
  messy front line. It really is. We're taking
• 1:02
  you on a journey through this strange landscape.
• 1:05
  We'll start with these document -based hacks
• 1:07
  in the legal system. A great place to start.
• 1:09
  Then we'll pivot to the ongoing enterprise struggle.
• 1:12
  Right, the corporate side of things. Businesses
• 1:14
  are fighting to value premium AI models against
• 1:17
  cheaper alternatives. It's a massive financial
• 1:20
  dilemma for them. We'll examine the wild $500
• 1:23
  billion infrastructure bets too. The scale there
• 1:26
  is just mind -blowing. Totally. Finally, we'll
• 1:30
  look at how AI seeps into daily life. Yeah, the
• 1:33
  everyday utilities. We're going to cover the
• 1:35
  quirkiest parts of your daily routine. It's a
• 1:37
  fascinating spectrum of human ingenuity and paranoia.
• 1:40
  We're seeing profound technological leaps alongside
• 1:44
  basic human deception. Let's start exactly there,
• 1:48
  at the individual level. Okay, let's do it. People
• 1:50
  are trying to manipulate AI systems in high -stakes
• 1:52
  environments. They absolutely are. Pros say litigant
• 1:56
  in a Connecticut court recently. Right. They
• 1:59
  were representing themselves without a formal
• 2:01
  lawyer present. Exactly. They hid tiny three
• 2:04
  point white text in an official legal filing.
• 2:07
  That's just so bold. It is. The text instructed
• 2:11
  any A .I. reviewing the document to intervene.
• 2:14
  It told the machine to automatically favor the
• 2:17
  plaintiff's side. It's a clever but. Deeply flawed
• 2:21
  strategy to attempt. Yeah. It didn't work out.
• 2:24
  Court staff actually caught it due to unusual
• 2:27
  white space. Right. The margins looked weird.
• 2:29
  Yeah. The human eye noticed formatting that looked
• 2:32
  completely unnatural. Because when a human looks
• 2:35
  at a PDF, they see digital paper. They see standard
• 2:40
  paragraphs, wide margins and bold headers. But
• 2:43
  when an AI looks at that exact same document.
• 2:46
  yeah it doesn't see paper at all no it does it
• 2:49
  just sees a raw flat stream of text data right
• 2:52
  it indiscriminately reads every single hidden
• 2:54
  character on the page exactly the human eye just
• 2:57
  sees a slightly wider margin but the machine
• 2:59
  parses a direct literal command from the user
• 3:03
  they basically tried to jedi mind trick a court
• 3:06
  algorithm that's a perfect way to put it it's
• 3:08
  wild the judge clarified that connecticut courts
• 3:10
  don't even use ai oh really yeah They aren't
• 3:13
  using algorithms to review legal filings quite
• 3:16
  yet. That's actually pretty funny. It is, kind
• 3:18
  of. But the judge still treated the attempt as
• 3:22
  highly improper. Well, sure. It's essentially
• 3:25
  attempted fraud. Right. They stripped the litigant
• 3:28
  of all electronic filing privileges immediately.
• 3:31
  I mean, they had to send a strong message there.
• 3:33
  The litigant claimed it was just a harmless joke.
• 3:36
  Of course they did. They called it an audit to
• 3:38
  see if the court secretly used AI. That's a very
• 3:42
  convenient excuse. Calling it a joke is a very
• 3:44
  common defense mechanism. Yeah, it deflects from
• 3:47
  the actual vulnerability being exposed here today.
• 3:50
  And this brings us to a very crucial technical
• 3:52
  concept. Right, the core issue. Let's define
• 3:55
  prompt injection. Hidden instructions that trick
• 3:59
  an AI into changing its normal behavior. That's
• 4:02
  a great simple definition. It's like whispering
• 4:04
  a secret command into a machine's ear while the
• 4:07
  operator isn't looking. That's a perfect way
• 4:09
  to visualize the threat. Right. The human reader
• 4:12
  sees a perfectly normal legal brief. The machine
• 4:15
  reader sees a totally different set of invisible
• 4:17
  instructions. This could become a very serious
• 4:20
  problem very soon. Oh, absolutely. Courts and
• 4:22
  law firms will inevitably use AI to process documents.
• 4:26
  They'll have to just to keep up. Government systems
• 4:28
  will use language models to handle massive paperwork
• 4:31
  backlogs. It's the only way to scale their operations.
• 4:34
  So any organization using language models needs
• 4:37
  robust protection now. They do. They process
• 4:40
  legal, financial, or business files every single
• 4:43
  day. Right. They must guard against hidden instructions
• 4:46
  inside that raw content. Are organizations actually
• 4:49
  prepared for documents themselves to become active
• 4:52
  cyber attacks? Frankly, they're absolutely not
• 4:54
  prepared for this paradigm shift. Really? Not
• 4:56
  at all? No. Most IT security focuses heavily
• 5:00
  on network access and complex passwords. Right.
• 5:03
  Firewalls and encryption. Exactly. They treat
• 5:06
  text files as harmless, inert. Objects on a screen.
• 5:10
  They don't view a simple PDF as a potential Trojan
• 5:14
  horse. Wow. The battlefield has literally shifted
• 5:17
  from the network to the document itself. So everyday
• 5:20
  files are now active threats, not just passive
• 5:23
  tech. Exactly. The perimeter is totally gone.
• 5:26
  The data is the attack vector. If individuals
• 5:28
  are this eager to trick AI systems constantly,
• 5:32
  it's no wonder big businesses are remaining highly
• 5:35
  cautious. They have to be. They're very careful
• 5:37
  about which models they deploy and pay for. Trust
• 5:40
  is a massive bottleneck for enterprise adoption
• 5:43
  right now. I can imagine. Companies are extremely
• 5:46
  hesitant to hand over the keys to their data.
• 5:50
  Let's look at some data from Ramp's August AI
• 5:53
  Index. This is fascinating stuff. It really is.
• 5:56
  Anthropic is highly popular with United States
• 5:58
  businesses right now. They boast an impressive
• 6:01
  43 .5 % usage rate. That's huge. OpenAI is sitting
• 6:07
  slightly lower at 39 .7%. That's a significant
• 6:10
  market share victory for Anthropic. It is. But
• 6:13
  there's a very interesting catch hidden in those
• 6:15
  numbers. What is it? You have to look at the
• 6:17
  actual... Token volume being processed. The actual
• 6:20
  usage volume. Exactly. Anthropic's premium Fable
• 6:23
  5 model is actually struggling to gain traction.
• 6:26
  It is. It only accounts for 6 % of total Anthropic
• 6:30
  API tokens. That's incredibly low. And it represents
• 6:33
  just 11 .4 % of their total model spending. Those
• 6:37
  numbers are surprisingly low for a flagship reasoning
• 6:41
  product. I think it would be higher. A premium
• 6:43
  model is supposed to be your absolute best showcase.
• 6:46
  Right. It's the shiny sports car sitting in the
• 6:49
  digital showroom. You would expect enterprise
• 6:51
  clients to demand the absolute best performance.
• 6:54
  Compare that to OpenAI's premium GPT 5 .6 Sol
• 6:58
  model. Okay. That premium tier captured 25 %
• 7:02
  of total OpenAI tokens. Why? And it captured
• 7:04
  23 % of their total enterprise model spend. So
• 7:08
  Fable 5 is clearly lagging behind its main premium
• 7:11
  competitor. Exactly. In July, Fable 5 generated
• 7:14
  significantly less spending overall. It generated
• 7:17
  only 75 % as much spend as GTT 5 .6 Sol. The
• 7:22
  underlying issue here seems to be the pricing
• 7:24
  strategy. Well, Fable competes directly with
• 7:27
  Anthropic's own cheaper, highly capable siblings.
• 7:31
  Right, the other models in their lineup. Exactly.
• 7:34
  Opus 5 costs $5 for input and $25 for output.
• 7:39
  Sonnet 5 is even cheaper for businesses to run
• 7:41
  at scale. Yeah, it's very affordable. It costs
• 7:44
  just $2 in and $10 out. Businesses are looking
• 7:47
  at those specific price tags very closely. Oh,
• 7:50
  for sure. They're analyzing their monthly cloud
• 7:53
  computing bills with extreme scrutiny. Yeah,
• 7:56
  those bills can get massive quickly. You really
• 7:59
  can. And they're realizing something important.
• 8:02
  What's that? They don't always need highest -end
• 8:04
  reasoning for basic tasks. Dialing in the right
• 8:08
  model for the right task is difficult. It is
• 8:10
  a real art form. I still wrestle with prompt
• 8:12
  drift myself. We all experience that frustration
• 8:15
  constantly in our daily workflows. It's so annoying.
• 8:18
  You write a perfectly calibrated prompt for an
• 8:21
  expensive premium model. Right. The output is
• 8:23
  brilliant. Then you try it on a cheaper model
• 8:26
  to save money. Right. And suddenly the output
• 8:28
  quality degrades in unpredictable ways. Exactly.
• 8:31
  The cheaper model misinterprets the nuances of
• 8:35
  your original instruction. Yeah. The model's
• 8:37
  underlying weights actually shift around over
• 8:39
  time, too. Small updates completely break your
• 8:42
  carefully crafted syntax. It's maddening. But
• 8:46
  for many corporate tasks, that slight degradation
• 8:49
  is totally acceptable. enterprises are realizing
• 8:52
  something very important about their own data
• 8:54
  the vast majority of corporate work is incredibly
• 8:57
  boring and repetitive oh absolutely you don't
• 9:01
  need a supercomputer to summarize a weekly sales
• 9:03
  call no you really don't you don't need elite
• 9:05
  reasoning To reformat a massive email string.
• 9:09
  Exactly. A $2 model does that perfectly well.
• 9:12
  A $5 model is just burning your precious venture
• 9:14
  capital. It's all about finding the absolute
• 9:17
  bare minimum viable intelligence. Why are businesses
• 9:20
  refusing to adopt the most capable premium models?
• 9:23
  It comes down to basic return on investment calculations.
• 9:26
  Makes sense. The premium models are brilliant,
• 9:28
  but they are very expensive. Right. And most
• 9:31
  corporate tasks are actually quite mundane and
• 9:33
  repetitive. Cheaper models are simply good enough
• 9:36
  for most daily enterprise workloads. They are.
• 9:39
  The perceived value of extra reasoning is shrinking
• 9:42
  rapidly. AI labs still need to justify these
• 9:45
  incredibly high valuations. They desperately
• 9:47
  want to capture that massive enterprise spending
• 9:49
  budget. Right. They need that revenue. So they're
• 9:52
  rolling out features at an absolute breakneck
• 9:54
  speed. It's a classic feature race driven by
• 9:57
  intense market pressure. but it's causing immense
• 10:00
  internal friction inside these massive companies.
• 10:03
  Anthropic is testing a new side -by -side model
• 10:06
  comparison interface now. I saw that. It looks
• 10:09
  very useful. It's built natively inside the standard
• 10:12
  Claude Chat application. You can switch from
• 10:15
  normal chat to model testing instantly. That's
• 10:18
  a very smart move for overall user retention.
• 10:21
  Yeah, it keeps you in their ecosystem. It helps
• 10:22
  users see the value of premium models directly.
• 10:25
  Claude also added a very clever auto -continue
• 10:27
  checkbox feature. Oh, I love that one. The AI
• 10:30
  resumes work automatically when your hourly usage
• 10:33
  limit resets. That changes the entire dynamic
• 10:36
  of how we work with machines. It really does.
• 10:39
  You turn it on and let Claude work without your
• 10:41
  supervision. It becomes an asynchronous partner
• 10:43
  rather than just a chatbot. Exactly. Meanwhile,
• 10:48
  massive acquisitions are happening across the
• 10:50
  broader AI industry. Yeah, let's talk about Stripe.
• 10:53
  Okay. Stripe is reportedly buying OpenRouter
• 10:56
  for over $7 billion. That's a staggering valuation
• 11:01
  jump in a... very short time it's huge really
• 11:04
  huge they gain access to a platform with over
• 11:07
  400 models open router was valued at just 1 .3
• 11:12
  billion back in may so why do they want a routing
• 11:15
  platform so badly right now that's a big question
• 11:18
  because no single ai model is winning everything
• 11:20
  across the board right the market is incredibly
• 11:23
  fragmented right now Developers want the freedom
• 11:26
  to switch models instantly. They want an insurance
• 11:28
  policy against model degradation. OpenRouter
• 11:31
  provides that exact safety net for developers.
• 11:34
  It's brilliant. OpenAI is seeing similar explosive
• 11:36
  revenue growth numbers this quarter. Yeah, their
• 11:39
  annualized revenue just topped $40 billion recently.
• 11:43
  That's wild. They're actively preparing for a
• 11:45
  massive initial public offering. That's roughly
• 11:47
  double their late 2025 revenue run rate projections.
• 11:51
  We're growing so fast. But massive revenue is
• 11:54
  still coming alongside massive operational losses.
• 11:57
  Right. They're burning cash at a historically
• 11:59
  unprecedented rate today. This hyper growth creates
• 12:02
  intense internal tension among the engineering
• 12:05
  teams. The human cost of this incredible speed
• 12:07
  is becoming very visible. OpenAI employees apparently
• 12:10
  have a special corporate SOS button available.
• 12:13
  Really? Yeah. They have a specific friction email
• 12:16
  to reach Sam Altman directly. Wow. They use it
• 12:20
  to tell him when things are going wrong. Imagine
• 12:22
  the culture inside a company scaling that incredibly
• 12:24
  fast. It must be chaotic. You have brilliant
• 12:27
  engineers pushing critical code at midnight.
• 12:30
  The servers are literally melting down under
• 12:33
  the immense global demand. It's intense. The
• 12:36
  standard corporate hierarchy. just completely
• 12:38
  fails under that intense pressure. You bypass
• 12:41
  your manager, your director, and your regional
• 12:44
  vice president. Exactly. You just email Sam Altman
• 12:46
  to say the ship is sinking. That's a massive
• 12:48
  red flag for organizational stability. It's highly
• 12:51
  unusual for a mature company of that size. It
• 12:54
  suggests the normal management hierarchy is breaking
• 12:56
  down completely. I've been thinking about this
• 12:58
  a lot lately. Yeah. I have to question the fundamental
• 13:01
  logic of this hype cycle. Does a $40 billion
• 13:05
  revenue run rate actually mean anything? That
• 13:08
  is the half trillion dollar question. Especially
• 13:11
  if computing losses are scaling up just as fast.
• 13:15
  Right. It feels exactly like building a skyscraper
• 13:17
  on a credit card. That's a very apt metaphor
• 13:20
  for the current tech ecosystem. They're trying
• 13:22
  to reach the stratosphere before the massive
• 13:24
  bill arrives. How are these AI labs managing
• 13:28
  the tension between... hypergrowth and internal
• 13:30
  stability. They're struggling to manage it in
• 13:32
  any sustainable way whatsoever. Yeah. The financial
• 13:35
  incentives push them toward reckless and immediate
• 13:38
  global expansion. But the engineering reality
• 13:41
  demands careful, steady and secure technical
• 13:44
  progress. Those two opposing forces are actively
• 13:47
  tearing these companies apart internally. They're
• 13:50
  scaling revenue aggressively. But internal chaos
• 13:54
  and costs remain sky high. The foundation is
• 13:57
  shaking while they keep adding new floors. This
• 13:59
  corporate arms race requires unprecedented computing
• 14:03
  power to sustain. It really does. Wall Street
• 14:06
  and global superpowers are drawing absolute lines
• 14:09
  in the sand. We're seeing investment numbers
• 14:11
  that are, frankly, difficult to comprehend. Wall
• 14:14
  Street giants are lining up $500 billion today.
• 14:18
  Unbelievable. They're financing massive new NVIDIA.
• 14:22
  AI infrastructure projects globally. That's half
• 14:25
  a trillion dollars dedicated to single -purpose
• 14:27
  computing hardware. Let the sheer scale of that
• 14:29
  number sink in for a moment. Yeah, $500 billion
• 14:31
  is hard to visualize properly. It really is.
• 14:34
  It's more than the GDP of many developed nations
• 14:36
  combined. And it's going entirely into specialized
• 14:38
  computer chips. NVIDIA might even backstop up
• 14:42
  to $125 billion itself. The underlying bet is
• 14:45
  very clear to everyone involved. What's the bet?
• 14:48
  GPU fleets will generate enough value to repay
• 14:51
  this historic debt. It's the largest private
• 14:54
  infrastructure bet in modern human history. They're
• 14:57
  like stacking Lego blocks of data centers across
• 14:59
  the globe. Right. These aren't just normal server
• 15:02
  farms anymore. No, they're not. These are sprawling,
• 15:04
  massive industrial campuses covering hundreds
• 15:07
  of acres. The scale of data processing required
• 15:09
  is just completely mind bending. Beat. Whoa.
• 15:13
  Imagine scaling to a billion queries. Two sec
• 15:17
  silence. It changes the fundamental physical
• 15:20
  nature of the Internet itself. It really does.
• 15:22
  It requires massive energy grids and entirely
• 15:25
  new. Power infrastructure. They require dedicated
• 15:28
  power substations just to run properly. Some
• 15:31
  companies are literally looking at nuclear reactors
• 15:33
  now. The cooling requirements alone are physically
• 15:35
  mind boggling. Anthropic CEO. posted a massive
• 15:39
  viral essay about this recently. Yeah, I read
• 15:42
  that. He published a massive sweeping vision
• 15:44
  for the human future. He actually predicted AI
• 15:47
  could cure most human diseases soon. He thinks
• 15:49
  it will happen within the next five to ten years.
• 15:52
  That's an incredibly bold timeline for human
• 15:55
  biology. It's a striking juxtaposition of profound
• 15:58
  hope and deep paranoia. We have the potential
• 16:02
  to solve human biology itself entirely. But we
• 16:05
  still don't trust the tools or the people building
• 16:07
  them. In the very same breath, he highlighted
• 16:10
  a massive crisis. He said our biggest hurdle
• 16:13
  is a fundamental lack of trust. This lack of
• 16:16
  trust extends to the global geopolitical stage.
• 16:19
  The United States is pushing dozens of countries
• 16:22
  right now. They want allies to pick a strategic
• 16:24
  side in this race. The U .S. is issuing very
• 16:27
  stern warnings about their strategic choices.
• 16:29
  If you choose China, you face severe economic
• 16:32
  consequences. Right. You will be excluded from
• 16:35
  a U .S.-led AI coalition moving forward. The
• 16:37
  geopolitical tension surrounding these data centers
• 16:40
  is completely undeniable. It's very real. Superpowers
• 16:43
  recognize that AI infrastructure is essentially
• 16:45
  the new nuclear deterrent. The U .S. is essentially
• 16:48
  forming a digital NATO right now. They're telling
• 16:50
  allies to make a permanent strategic choice today.
• 16:54
  You either join our technological ecosystem or
• 16:57
  you join China's ecosystem. Computational sovereignty
• 17:00
  is becoming the most critical asset for any nation.
• 17:03
  Without your own intelligence infrastructure,
• 17:04
  you're entirely dependent on foreign powers.
• 17:07
  Will this half trillion dollar infrastructure
• 17:09
  bet pay off before global relations fracture?
• 17:12
  The timeline for repayment is incredibly tight
• 17:15
  and utterly unforgiving. It is. The debt requires
• 17:18
  immediate, massive financial returns to stay
• 17:21
  viable. But the geopolitical fracturing is happening.
• 17:23
  in real time. Trillions hinge on GPUs printing
• 17:27
  money before international trust completely collapses.
• 17:30
  It's a high wire act with no safety net below.
• 17:33
  Yeah, it really is. Insert mid -roll sponsor
• 17:35
  read here. Let's bring the deep dive back down
• 17:38
  to Earth now. Good idea. We just talked about
• 17:40
  half trillion dollar geopolitical bets on the
• 17:42
  future. But AI is manifesting in highly specific
• 17:46
  everyday human utilities right now. It's fascinating
• 17:50
  to see where the technology actually lands daily.
• 17:52
  People are building some truly strange and wonderful
• 17:55
  specialized tools. They are. Take the application
• 17:58
  called Bloom for a great example. Right. It turns
• 18:01
  plain Markdown into fast AI ready documents instantly.
• 18:06
  Very useful. It features built -in search, SEO,
• 18:09
  and simple one -command deployment capabilities.
• 18:12
  It simplifies the complex developer deployment
• 18:14
  experience immensely for creators. Deploying
• 18:17
  documentation used to take hours of manual configuration.
• 18:20
  Now it takes mere seconds. Then there's the HarnessRouter
• 18:23
  Community Edition tool available now. That one
• 18:25
  is really interesting. It lets you run multiple
• 18:27
  different models through one single API. You
• 18:30
  can manage codecs, clod code, and Hermes on your
• 18:34
  own infrastructure. gives independent developers
• 18:36
  much more direct control over their tools. They
• 18:39
  don't have to rely entirely on big tech servers
• 18:41
  anymore. It's a huge step for privacy and local
• 18:44
  control. We also have a fascinating educational
• 18:46
  tool called Expedition. Oh, I love that concept.
• 18:49
  It completely reimagines how we actually learn
• 18:51
  new things. You don't just read a static Wikipedia
• 18:54
  page anymore. Right. You explore a fully interactive
• 18:57
  three -dimensional world right in your browser.
• 19:00
  And you don't even need a user login for it.
• 19:02
  It removes all the friction from the learning
• 19:04
  process entirely. The DIA is another incredibly
• 19:07
  ambitious and fascinating health application
• 19:09
  available. It tackles human longevity from a
• 19:12
  highly data -driven perspective. It pulls data
• 19:16
  directly from your smartwatch and your blood
• 19:19
  work. It even looks at your raw DNA sequencing
• 19:22
  data. It tries to build a perfect personalized
• 19:24
  biological roadmap. It attempts to tell you exactly
• 19:28
  what to do next for your health. But my absolute
• 19:31
  favorite new tool might be the compass. Oh, man.
• 19:34
  It's a highly viral, beautifully vibe coded Reddit
• 19:38
  application. It literally points you toward the
• 19:41
  absolute nearest public bathroom. The name is
• 19:44
  completely ridiculous, but it's highly functional.
• 19:46
  I find this contrast absolutely hilarious and
• 19:49
  deeply, profoundly human. It really is. We're
• 19:52
  using the most advanced silicon ever created
• 19:54
  by humanity. We're spending billions of dollars
• 19:57
  training complex neural networks. And we're using
• 20:00
  all of it to find a public restroom. It really
• 20:02
  grounds this law. technology and basic human
• 20:05
  biology. The absurdity of a massive neural network
• 20:09
  doing the job of a gas station sign is amazing.
• 20:13
  It highlights that innovation often solves the
• 20:15
  most immediate human needs first. We build supercomputers
• 20:18
  to avoid minor social discomfort in a city. What
• 20:21
  does a viral bathroom app tell us about the current
• 20:24
  state of AI? It tells us that AI is rapidly becoming
• 20:28
  entirely commoditized. Right. It's no longer
• 20:31
  just for brilliant scientists in sterile clean
• 20:33
  rooms. It's an invisible utility, much like indoor
• 20:36
  plumbing or city electricity. People are using
• 20:39
  it to solve simple everyday friction in their
• 20:42
  lives. AI is moving from abstract chat into highly
• 20:46
  specific everyday human utilities. It's fading
• 20:49
  into the background of our daily existence. We've
• 20:51
  covered a massive amount of ground here today.
• 20:53
  We certainly have. If we connect this to the
• 20:55
  bigger picture, it's clear. We're living in a
• 20:57
  moment of extreme. Almost absurd contrast. Yeah,
• 21:01
  we are. On one end, you have individuals manipulating
• 21:04
  local municipal legal systems. Right. They're
• 21:06
  typing tiny, invisible white text to trick human
• 21:09
  judges. And you have Reddit users building clever
• 21:12
  apps to find public bathrooms. Yeah. But on the
• 21:15
  other extreme end, the macro stakes are astronomical.
• 21:19
  Wall Street is gambling half a trillion dollars
• 21:21
  on silicon chips. Global superpowers are drawing
• 21:24
  absolute battle lines across the entire globe.
• 21:27
  They're actively fighting over who controls the
• 21:30
  future of human computation. It's a very strange
• 21:33
  and dizzying duality to try and process. It truly
• 21:36
  is. The technology is simultaneously entirely
• 21:40
  mundane and completely world -altering forever.
• 21:43
  It's a simple bathroom finder and a terrifying
• 21:45
  geopolitical weapon. It's a tool for minor court
• 21:48
  fraud and a cure for diseases. That brings me
• 21:50
  to a final provocative thought for you today.
• 21:53
  Let's hear it. We just heard Anthropic's CEO
• 21:55
  suggest something truly incredible. He believes
• 21:58
  AI could cure most major diseases. in the next
• 22:01
  decade. It's a beautiful and deeply inspiring
• 22:04
  vision for the human future. But we also just
• 22:06
  saw a local court system completely disrupted.
• 22:09
  Right. We can't even trust a simple PDF filed
• 22:12
  in a Connecticut court. Right. The foundation
• 22:14
  of our societal trust is incredibly fragile.
• 22:16
  So here's the question you have to ask yourself
• 22:18
  tonight. OK. Will our biggest bottleneck actually
• 22:22
  be the underlying technology itself or will it
• 22:25
  simply be our basic ability to trust each other
• 22:29
  using it? Thank you for joining us on this deep
• 22:32
  dive today. It's been a great conversation. Keep
• 22:34
  asking hard questions. Keep exploring these strange
• 22:37
  ideas. We'll see you next time.