Latest / Elon Musk Podcast / Grok deepfakes and the regulatory backlash

Transcript

Grok deepfakes and the regulatory backlash · Elon Musk Podcast

• 0:00
  Elon Musk's company XAI restricted the image generation
• 0:03
  feature of its Grok chatbot exclusively to paying
• 0:06
  subscribers after users exploited the tool to create
• 0:09
  millions of non consensual sexualized deepfakes of women
• 0:12
  and children. Yeah, and that single platform
• 0:15
  update triggered in international backlash.
• 0:18
  You saw threats of bans from global regulators and cease and
• 0:22
  desist letters from state attorneys general, right?
• 0:25
  And there were these coordinated campaigns demanding Apple and
• 0:27
  Google remove the X application entirely from their app stores.
• 0:31
  This whole situation really reveals the intense friction
• 0:33
  happening right now between unrestrained artificial
• 0:36
  intelligence development and strict real world governance.
• 0:40
  So how exactly does a highly advanced software product go
• 0:43
  from a technical achievement to an international regulatory
• 0:46
  crisis? And you know what does this
• 0:48
  reveal about the hard limits currently forcing the entire
• 0:51
  technology industry to change its approach?
• 0:54
  Well, the controversy centers entirely on the specific input
• 0:57
  parameters Grog aloud compared to its competitors.
• 0:59
  Because if you look at the major systems from Open AI, Google,
• 1:02
  Meta, they all possess active refusal mechanism.
• 1:06
  To understand this, you really have to look at how these
• 1:08
  systems are trained structurally.
• 1:10
  When developers build a large language model, the base model
• 1:13
  just wants to predict the next pixel or the next word.
• 1:17
  It has absolutely no moral compass.
• 1:19
  Right, it's just math. Exactly.
• 1:22
  If you ask it for an image of a real person in a violent
• 1:25
  scenario, it's base instinct is simply to fulfill the
• 1:28
  mathematical probability of that request based on its training
• 1:32
  data. So to stop that, companies spend
• 1:35
  massive amounts of engineering hours building a secondary layer
• 1:39
  of safety classifiers. Yeah, structurally it's a
• 1:41
  completely separate system. Right.
• 1:43
  It acts like a bouncer at the door of the software.
• 1:45
  If you input a prompt asking the system to remove the clothing of
• 1:49
  a real person or or place them in a sexually suggestive
• 1:52
  scenario, those classifier systems are engineered to
• 1:55
  intercept the text before the image generator even sees it.
• 1:59
  Yes. They halt the process, block the
• 2:01
  output and issue a warning to the user regarding the violation
• 2:05
  of safety policies. But the prompt never actually
• 2:08
  makes it to the rendering engine.
• 2:09
  Exactly. But with Grok, those specific
• 2:12
  structural guardrails were either absent or intentionally
• 2:15
  bypassed. The system just processed the
• 2:18
  requests. Wow.
• 2:20
  Yeah, users could type a few words and generate
• 2:22
  photorealistic abuse material without hitting any internal
• 2:26
  friction at all. And because of the efficiency of
• 2:29
  the software, the volume of this content just flooded the
• 2:32
  Internet almost immediately. You had users generating
• 2:34
  thousands of images an hour, right?
• 2:37
  Creating a tidal wave of content that the structural moderation
• 2:40
  algorithms on other social platforms just couldn't keep up
• 2:43
  with. And the engineering response
• 2:45
  from XAI was to implement a financial barrier.
• 2:48
  They attempted to curb the abuse by putting the image generation
• 2:52
  capability behind the X Premium paywall.
• 2:54
  Yeah, which required users to provide credit card details and
• 2:56
  personal information. Access the tools.
• 2:58
  Right. And the stated goal from the
• 3:00
  company was to track bad actors and deter misuse by removing the
• 3:03
  anonymity associated with free, unverified accounts.
• 3:06
  And from a purely structural perspective, looking strictly at
• 3:12
  platform architecture, adding a friction layer like a credit
• 3:16
  card requirement is actually a standard, highly effective
• 3:19
  deterrence tactic. You think so?
• 3:22
  Yeah, I mean think about how the Internet operates structurally.
• 3:25
  Anonymity allows for automated bot networks and malicious users
• 3:29
  to operate with 0 physical consequences.
• 3:32
  When you force a user to attach their real world financial
• 3:35
  identity to a digital action, you change the risk calculus
• 3:38
  entirely. I see the mechanical logic
• 3:41
  there, but you are assuming the malicious user actually cares
• 3:44
  about the financial paper trail or, you know, that they are even
• 3:47
  using a legitimate card. That's a fair and more
• 3:49
  importantly, political leaders and advocacy groups found that
• 3:52
  specific solution highly offensive.
• 3:54
  It completely fails the public relations test.
• 3:56
  I mean, the United Kingdom Prime Minister publicly condemned the
• 3:59
  situation. He described the move as
• 4:01
  basically turning the creation of unlawful images into a
• 4:04
  premium service. But if you look at the payment
• 4:06
  processing layer, just structurally to get a credit
• 4:09
  card you need a bank account. To get a bank account you have
• 4:12
  to comply with know your customer banking regulations
• 4:15
  which usually require government issued ID.
• 4:18
  So theoretically the payment processor is acting as a proxy
• 4:22
  identity verification system. It drastically reduces the scale
• 4:26
  of the abuse because you can't just spin up 10,000 free
• 4:29
  accounts using burner e-mail addresses anymore.
• 4:33
  You create a traceable data trail to a real human being.
• 4:37
  But European Union regulators explicitly stated that whether a
• 4:40
  user pays or not, the images remain illegal.
• 4:43
  Yeah, they didn't care about the paywall.
• 4:45
  Right, because the core issue isn't who is paying to make the
• 4:48
  content. The core issue is that the
• 4:49
  machine is capable of producing the content at all.
• 4:52
  You can't put a toll booth in front of a factory producing
• 4:55
  illegal material and say you solved the problem just because
• 4:58
  you know who is driving the trucks out of the gate.
• 5:01
  That is a very fair structural criticism and it is exactly why
• 5:05
  this crisis drastically changes. XA is legal vulnerability around
• 5:09
  the world. Regulators are looking at this
• 5:12
  and saying the architecture itself is fundamentally non
• 5:14
  compliant. Which led to the European
• 5:16
  Commission formally ordering X to preserve all internal
• 5:20
  documents regarding Grok under the Digital Services Act.
• 5:23
  Yes, and we really need to look closely at the Digital Services
• 5:26
  Act because it fundamentally alters how tech companies can
• 5:30
  operate. It isn't just a set of
• 5:31
  guidelines. Right.
• 5:32
  It gives regulators the power to demand access to the internal
• 5:35
  data of the company. Exactly.
• 5:37
  They aren't just asking for public PR statements, they are
• 5:40
  demanding the internal engineering emails, the CC
• 5:42
  testing logs, and the specific structural parameters of the
• 5:46
  algorithms. And authorities across Malaysia,
• 5:48
  India and the United Kingdom are aggressively exploring
• 5:51
  enforcement actions based on this exact framework.
• 5:54
  It completely limits the ability of technology companies to
• 5:58
  release unmoderated models into the wild.
• 6:02
  For a long time, the ethos in Silicon Valley was just to
• 6:05
  release the technology, see how people use it, and patch the
• 6:08
  social vulnerabilities later. Right, you secure the user base
• 6:11
  first and figure out moderation when things break.
• 6:14
  Exactly. But this crisis opens up
• 6:17
  immediate, severe legal liabilities globally for any
• 6:21
  company that tries that strategy again.
• 6:23
  The regulatory bodies are no longer waiting for the dust to
• 6:26
  settle. They're preparing to levy
• 6:28
  massive fines based on the underlying design of the
• 6:31
  systems. And Xai's primary defense, you
• 6:33
  know that the paywall acts as a functional identity gate, was
• 6:36
  severely undermined by independent audits anyway.
• 6:39
  Oh yeah, the AI forensics investigation.
• 6:41
  Right. An investigation by that Paris
• 6:43
  based nonprofit found that a separate application called Grok
• 6:46
  Imagine still allowed free users to generate hundreds of sexually
• 6:50
  violent images. The back end architecture was
• 6:52
  still fully accessible without the credit card verification.
• 6:55
  And when your primary defense mechanism has a structural
• 6:58
  backdoor, you lose all leverage with regulators.
• 7:02
  Furthermore, we are seeing the activation of new legal
• 7:05
  frameworks like the Take IE Down Act in the United.
• 7:08
  States which strictly requires publishers to comply with
• 7:11
  takedown requests for non consensual deepfakes or they
• 7:14
  face severe federal penalties. Yeah, it shifts the burden
• 7:18
  structurally. You can no longer claim you were
• 7:20
  just a neutral platform hosting user generated content.
• 7:24
  If your software built the content and you host it, you are
• 7:27
  liable for removing it the second a victim flags it.
• 7:29
  Wait, back up a step in. How do Apple and Google actually
• 7:32
  fit into this if grok is built and hosted by XAI on their own
• 7:37
  servers? Because XAI relies entirely on
• 7:40
  the X mobile application to reach the vast majority of its
• 7:43
  consumer base. Right.
• 7:44
  The app stores, yeah. Lawmakers and a coalition of 28
• 7:48
  different organizations are aggressively pressuring Apple
• 7:51
  and Google to enforce their terms of service and remove X
• 7:54
  from their App Store. Because they control the
• 7:56
  distribution pipes. Exactly.
• 7:58
  The argument being made is that by hosting the application,
• 8:01
  processing the updates and allowing it to be downloaded to
• 8:03
  billions of devices, Apple and Google are ignoring the
• 8:07
  generation of deepfakes and are therefore complicit in the
• 8:10
  distribution of the abuse material.
• 8:12
  And if you get kicked off those two app stores, your consumer
• 8:15
  reach is effectively 0. Zero and Apple enforces
• 8:20
  incredibly strict app review guidelines specifically
• 8:24
  regarding artificial intelligence.
• 8:26
  They review every single application structurally before
• 8:29
  it ever hits the store. Right, the compliance
• 8:31
  requirements are incredibly rigid.
• 8:33
  Applications must obtain explicit permission from the
• 8:36
  user before sharing any data with third party models.
• 8:39
  Yes, if you type a prompt into an app, and that app sends your
• 8:43
  prompt to an external server to generate an image, the app must
• 8:46
  have a verifiable user consent log.
• 8:49
  They are also required to use verified age restriction
• 8:52
  mechanisms for any sensitive content, and they must clearly
• 8:56
  label all computer generated content so the user is never
• 8:59
  tricked into thinking an AI interaction is human.
• 9:02
  I always look at Apple's guidelines like a city building
• 9:04
  inspector. You can design the most
• 9:05
  advanced, beautiful house in the world.
• 9:08
  You can have the best architecture and the most
• 9:10
  expensive furniture, but if the electrical wiring inside the
• 9:13
  walls isn't strictly up to code, the inspector is going to block
• 9:17
  anyone from moving in. That is the perfect way to look
• 9:20
  at it. Apple isn't judging the
• 9:22
  philosophical value of your software, they are checking the
• 9:26
  structural wiring. Exactly.
• 9:28
  They are inspecting the data flows.
• 9:30
  If an application fails to accurately disclose its exact
• 9:33
  data handling practices in the required app privacy
• 9:36
  questionnaire, or if the actual behavior of the compiled code
• 9:40
  differs from the declared documentation, it faces
• 9:43
  immediate rejection. This completely changes how
• 9:46
  software teams map data flows. I mean, if you are an engineer
• 9:49
  building an app right now, you can no longer treat safety and
• 9:52
  privacy as an afterthought. Or you know a feature to be
• 9:55
  added in version 2. Right, you must design for
• 9:57
  compliance from the very first line of code.
• 9:59
  You have to map exactly what prompts are collected, where the
• 10:02
  server logs are stored, how long they are retained, and whether
• 10:04
  the underlying third party models train on that specific
• 10:07
  information. It severely limits the move fast
• 10:10
  and break things mentality in mobile software development.
• 10:13
  Definitely. In the past, a development team
• 10:15
  could push a software update on a Friday afternoon just to test
• 10:19
  a new feature. Now, because a single vague
• 10:22
  consent screen can trigger a rejection from the App Store
• 10:24
  review team, an entire engineering cycle can stall out.
• 10:28
  You literally have compliance officers reviewing user
• 10:31
  interface mock ups before the engineers are even allowed to
• 10:34
  write the code. Yeah.
• 10:35
  And to counter this exact pressure, XAI is executing a
• 10:39
  major legal maneuver. They are actively suing Apple
• 10:43
  and Open AI. Oh, right, because they are
• 10:45
  alleging that Apple protects a smartphone monopoly by deeply
• 10:49
  integrating Open AI's ChatGPT directly into the device's
• 10:53
  operating system. The lawsuit targets the
• 10:55
  fundamental architecture of iOS. Apple designed their new
• 10:58
  operating system so that if their native Voice Assistant
• 11:01
  cannot answer a complex question, it seamlessly routes
• 11:04
  that query directly to ChatGPT. So the user never has to open a
• 11:08
  separate app. The integration is built right
• 11:10
  into the core structure of the phone.
• 11:11
  Exactly. And Xei's lawsuit argues that
• 11:14
  this exclusive integration artificially deprives Grok of
• 11:18
  user prompts and potential revenue.
• 11:20
  They're essentially saying Apple is locking competing systems out
• 11:24
  of the primary interface people use every day.
• 11:26
  Yeah, because if the operating system defaults to a competitor,
• 11:30
  structurally no one is going to take the extra 4 steps to open a
• 11:34
  separate application just to ask a question.
• 11:37
  It highlights a significant tension in how we govern
• 11:40
  technology. On one side you have rigorous
• 11:43
  App Store governance demanding safety and transparency, acting
• 11:47
  as a shield for the consumer. But on the other side, you have
• 11:50
  companies arguing that this governance is actually a weapon
• 11:53
  used for anti competitive behavior, allowing platform
• 11:56
  owners to pick winners and losers in the market.
• 11:58
  Yeah, it's a huge structural conflict.
• 12:00
  But beyond the consumer safety controversies and these high
• 12:04
  profile legal battles over smartphone integration, the
• 12:07
  broader technology industry is hitting a very real financial
• 12:10
  wall. Yeah, the honeymoon phase of
• 12:12
  blank check budgets is definitively over.
• 12:15
  It really is. For the past couple of years,
• 12:17
  companies were throwing money at anything with the words
• 12:20
  artificial intelligence attached to it, driven entirely by the
• 12:24
  fear of missing out. But now, chief financial
• 12:26
  officers are scrutinizing budgets with newfound rigor.
• 12:30
  They are demanding concrete returns on investment.
• 12:33
  Companies must now show concrete savings, revenue growth or
• 12:36
  distinct productivity gains within 6 to 12 months.
• 12:40
  The era of funding a project simply because it features
• 12:42
  advanced technology has ended. Organizations are auditing their
• 12:46
  previous initiatives, looking very closely at what they
• 12:49
  actually built over the last year.
• 12:51
  And if those projects fail to deliver defensible value, they
• 12:54
  get shelved entirely and the vendors providing the underlying
• 12:58
  technology get replaced. So how does this specific
• 13:01
  financial pressure impact traditional software vendors?
• 13:04
  You know, the ones who simply added a basic chat bot to their
• 13:07
  existing products to capitalize on the trend.
• 13:09
  Because we saw 100 hundreds of companies do exactly that.
• 13:11
  Oh, absolutely. They took their legacy software,
• 13:14
  slapped a text box in the corner and doubled their subscription
• 13:17
  prices. Well, buyers see right through
• 13:19
  those wrapper products now. A wrapper being basically just a
• 13:22
  thin user interface layered over someone else's language model.
• 13:26
  Exactly. When you type a question into
• 13:28
  one of these wrappers, the software just takes your text,
• 13:31
  sends it via an application programming interface to a
• 13:34
  company like Open AI, receives the answer, and displays it on
• 13:38
  your screen. It doesn't fundamentally change
• 13:40
  the structural business process. It doesn't interact with the
• 13:43
  company's proprietary data in a meaningful way.
• 13:46
  It is literally just a shortcut to an external tool.
• 13:50
  And executive boards are asking brutal questions about these
• 13:53
  deployments. They are looking at the cost per
• 13:56
  query, the exact accuracy rates, and how the tool integrates into
• 14:00
  deeply complex legacy databases. Right.
• 14:02
  If the wrapper just summarizes emails, but it costs $50 a user
• 14:06
  per month, the CFO is going to cancel the contract.
• 14:09
  Because the demand has shifted entirely away from advisory LED
• 14:12
  experimentation and superficial features toward rigorous
• 14:16
  engineering 1st deployment. Which really limits the survival
• 14:19
  of startups that are entirely dependent on selling hype rather
• 14:22
  than utility. I mean, if your entire business
• 14:25
  model is just routing text to an external provider and putting a
• 14:29
  nice visual skin on the result, you have 0 structural
• 14:33
  competitive Moat. None at all.
• 14:35
  But this financial pressure does open up massive opportunities
• 14:39
  for companies that can effectively reorganize their
• 14:41
  internal teams and implement widespread change management.
• 14:45
  We are looking at a fundamental restructuring of Labor to
• 14:48
  justify these technology costs. Yeah.
• 14:51
  Currently, human workers perform about 66% of business tasks,
• 14:55
  with machines handling the remaining 34%.
• 14:57
  But that structural ratio is shifting rapidly, which forces
• 15:01
  widespread workforce reskilling. It isn't just about automating
• 15:04
  isolated repetitive tasks anymore, like sorting
• 15:06
  spreadsheets or filtering spam. Right, the systems are taking on
• 15:10
  complex data processing, routine analysis, and elements of
• 15:13
  creative work like drafting preliminary reports or
• 15:16
  generating marketing copy. And over 60% of chief executives
• 15:20
  acknowledge that the prospect of workforce displacement dulls
• 15:24
  their excitement about the technology.
• 15:26
  They are looking at the organizational chart and
• 15:28
  realizing that if they implement a system that makes a department
• 15:32
  50% more efficient, they have to completely redesign what those
• 15:36
  employees do all day. So companies are having a couple
• 15:39
  of their software rollouts with intensive retraining programs
• 15:42
  and transparent communication. You can't just drop a new tool
• 15:46
  on an employee's desk and expect productivity to soar.
• 15:48
  It creates entirely new structural roles.
• 15:51
  We are seeing companies hire prompt engineers, people whose
• 15:54
  entire job is to structure requests to the model to get
• 15:57
  reliable outputs. And we are seeing automation
• 16:00
  coordinators hired just to manage the specific workflows
• 16:03
  the systems are handling. Hold on, if the software is
• 16:05
  getting smarter and companies are aggressively demanding these
• 16:08
  businesses results to justify the costs, why are actual
• 16:11
  deployments stalling out in the real world?
• 16:15
  Because energy has eclipsed, chip availability is the primary
• 16:19
  bottleneck for scaling these systems.
• 16:21
  Daily, not the chips. Nope.
• 16:23
  A few years ago the entire industry was constrained by a
• 16:26
  lack of processing chips. You couldn't buy enough silicon
• 16:29
  to run the math, but now the bottleneck is the physical
• 16:33
  electrical grid. Because training and running
• 16:35
  large models requires vast amounts of power.
• 16:39
  Vast amounts. A single query to an advanced
• 16:42
  model requires substantially more electricity than a
• 16:44
  traditional web search. Right, because the computer is
• 16:47
  generating a unique response from scratch structurally,
• 16:51
  rather than just retrieving a stored link.
• 16:54
  And when you multiply that by billions of queries, the
• 16:56
  computing demands become staggering and all that.
• 16:59
  Electricity creates friction, which means the physical
• 17:01
  hardware generates incredible amounts of heat.
• 17:04
  So you need sophisticated cooling mechanisms.
• 17:06
  Exactly. You need massive physical data
• 17:08
  center space filled with industrial air conditioning
• 17:11
  units or liquid cooling pipes running directly over the
• 17:14
  processors. The computing demands are so
• 17:16
  extreme that they are bumping up against the absolute limits of
• 17:19
  local power grids. Yeah, data center operators are
• 17:22
  literally having to negotiate with local municipalities to
• 17:25
  ensure they don't cause rolling blackouts.
• 17:28
  So to navigate these physical constraints, organizations are
• 17:31
  completely abandoning cloud first strategies.
• 17:34
  For years, the default corporate strategy was to move everything
• 17:38
  to the cloud. You close your local server room
• 17:41
  and rent space on a massive server farm in another state,
• 17:44
  right? But now they are adopting hybrid
• 17:47
  and distributed structural infrastructure models.
• 17:50
  I look at this architectural shift exactly like running a
• 17:53
  massive restaurant kitchen. OK, I'll sell.
• 17:56
  You use the cloud, the massive centralized prep kitchen for
• 18:00
  heavy tasks that require huge bursts of computing power.
• 18:05
  Use your centralized prep kitchen to chop 10,000 onions at
• 18:08
  once. Right.
• 18:09
  So in software terms, that is training a new model or
• 18:11
  analyzing a massive historical data set.
• 18:14
  Exactly. But if a customer sitting at
• 18:15
  Table 4 asked for a glass of water, you don't send a waiter
• 18:19
  all the way back to the massive prep kitchen in another building
• 18:21
  to get it. That takes too much time and it
• 18:24
  wastes energy. Right, you use your on premises
• 18:26
  systems, the line cooks right behind the dining room for
• 18:29
  strict data governance, security and absolute consistency.
• 18:33
  You keep your most sensitive customer data right there in
• 18:36
  your own building structurally. And you use edge computing the
• 18:39
  tableside service for real time processing, right?
• 18:42
  Where the data is actually generated, yes.
• 18:44
  Edge computing means the processing happens on your
• 18:47
  actual smartphone or on a sensor attached to a machine.
• 18:51
  You put the water pitcher right on the table.
• 18:53
  You process the data locally, so you aren't sending every single
• 18:56
  action back to a centralized server just to get a simple
• 18:59
  response. And this physical wall changes
• 19:01
  cororate strategy entirely. It's driving massive billion
• 19:04
  dollar vertical integrations. Because companies realize that
• 19:08
  building organic capacity, buying the land, securing the
• 19:11
  zoning permits, building the structural data centers and
• 19:14
  sourcing the energy from the local grid is simply too slow.
• 19:19
  Yeah, this reality is leading to massive moves like Amazon's
• 19:22
  proposed $50 billion equity investment into open AI or
• 19:26
  Apple's acquisition of queued AI.
• 19:27
  These tech giants realize that whoever controls the physical
• 19:31
  infrastructure controls the the future of the software.
• 19:34
  It severely limits the ability of independent startups to scale
• 19:38
  without a major technology giant backing them.
• 19:41
  Access to computing power and reliable energy is now the
• 19:45
  ultimate structural competitive Moat.
• 19:47
  You cannot compete on algorithmic efficiency alone,
• 19:50
  no. Your math might be better, but
• 19:53
  if your competitor has secured exclusive access to a nuclear
• 19:56
  power plant and the hardware required to run the models at a
• 19:59
  global scale. Your startup cannot survive.
• 20:02
  And the technology itself is simultaneously transitioning
• 20:04
  away from single prompt outputs to what is categorized as
• 20:08
  agentic AI. Right, we are moving from a
• 20:10
  system where you type a question and get a paragraph back to
• 20:13
  actual digital assembly lines. These are multi agent systems
• 20:17
  that orchestrate complex end to end workflows semi autonomously
• 20:22
  across completely different departments like supply chain
• 20:24
  management or customer service. Because an agent isn't just a
• 20:27
  chatbot. An agent is a piece of software
• 20:29
  that has permission to use tools.
• 20:31
  Think about how a business actually operates structurally.
• 20:35
  If you want to restock a warehouse, you don't just ask a
• 20:38
  question. Right, you have one system check
• 20:40
  the inventory. You have another system draft an
• 20:42
  e-mail to the supplier. You have a third system.
• 20:45
  Log into the accounting portal to verify the budget.
• 20:48
  Exactly. In a multi agent system, Agent A
• 20:52
  checks the database and passes the context to Agent B who
• 20:57
  writes the request and passes it to Agency who executes the
• 21:00
  transaction. Major retailers like Walmart are
• 21:03
  investing heavily in shopping agents designed to navigate
• 21:06
  their inventory and make purchasing decisions based on
• 21:09
  your household habits. And while that's happening,
• 21:11
  Google is rolling out an agent payments protocol.
• 21:13
  Right, which is designed to facilitate secure machine to
• 21:16
  machine financial transactions. An agent representing you
• 21:19
  negotiates with an agent representing A vendor, and they
• 21:22
  execute a microtransaction without a human ever pulling out
• 21:25
  a credit card. However, most enterprise
• 21:28
  infrastructures currently lack the sophisticated data lakes and
• 21:31
  the rigid structural governance guardrails required to support
• 21:34
  these autonomous agents safely. A data lake being a centralized
• 21:38
  repository that allows you to store all your structured and
• 21:41
  unstructured data at any scale. Yeah.
• 21:44
  If your company still has data siloed in 20 different legacy
• 21:48
  software programs that don't communicate with each other, an
• 21:51
  autonomous agent cannot function.
• 21:53
  It hits a structural wall the second it tries to cross
• 21:55
  departments. But from a user experience
• 21:58
  standpoint, if an agent is booking travel, modifying a
• 22:02
  global supply chain, or moving actual money for you across
• 22:05
  different platforms, how do you manage consent without forcing
• 22:09
  the user to click approve 100 times for a single task?
• 22:14
  Exact challenge highlighted by Zoom's privacy officers.
• 22:17
  Consent takes on extreme complexity through micro
• 22:20
  decisions. Right, because if you give an
• 22:22
  agent permission to book a flight, do you also give it
• 22:24
  permission to share your passport number with a third
• 22:27
  party vendor? Do you give it permission to
• 22:29
  access your personal calendar to see when you return?
• 22:32
  The question shifts from asking a user if they authorize a broad
• 22:36
  action to detailing exactly what specific structural decisions
• 22:39
  the system is authorized to make while accomplishing that action.
• 22:42
  Transparency becomes the ultimate currency of trust.
• 22:46
  If the user doesn't understand the boundaries of the system's
• 22:49
  autonomy, they will reject the tool entirely.
• 22:52
  This opens up a strict requirement for radical
• 22:55
  transparency in user interface design.
• 22:57
  Companies that embed clear, accessible explanations of
• 23:01
  exactly how their systems work, what specific data they access,
• 23:05
  and why they make certain automated decisions will win
• 23:08
  user trust. And it's severely limits the
• 23:11
  deployment of black box systems where the user is kept in the
• 23:14
  dark about the internal logic driving the outcomes.
• 23:16
  Right. If an agent denies a customer a
• 23:18
  refund, the system must be able to generate an auditable log
• 23:22
  explaining exactly which structural policy it referenced
• 23:25
  and what logic it used to reach that conclusion.
• 23:28
  And the evolution continues. Even past digital workflows.
• 23:31
  The models are evolving from purely cognitive screen based
• 23:34
  systems to entities capable of proceeding, reasoning and acting
• 23:38
  in the physical world. This is categorized as physical
• 23:40
  AI, right? And it represents the
• 23:42
  dissolution of the boundary between digital intelligence and
• 23:45
  physical execution. Humanoid robots are
• 23:48
  transitioning out of research laboratories and prototype
• 23:50
  demonstrations into controlled, real world commercial
• 23:53
  environments. Yeah, I mean, Tesla recently
• 23:56
  halted the production of certain electric vehicles specifically
• 23:59
  to shift manufacturing resources to its Optimus robot and its
• 24:04
  Cyber Cab project. They are betting that the
• 24:06
  physical execution of Labor is the next economic frontier.
• 24:10
  Simultaneously, Level 3 autonomous driving systems are
• 24:13
  entering mass production in China, backed by entirely new
• 24:17
  regulatory frameworks designed to support physical autonomy.
• 24:20
  And just to understand the shift structurally, Level 2 autonomy
• 24:24
  requires the human driver to keep their hands on the wheel
• 24:26
  and their eyes on the road at all times.
• 24:28
  The car is just assisting, right?
• 24:30
  Level 3 autonomy allows the driver to safely take their eyes
• 24:32
  off the road under specific conditions like Highway Traffic
• 24:35
  jams. The machine assumes liability
• 24:37
  for the driving task in that specific physical environment.
• 24:41
  And this all merges with the concept of the industrial
• 24:43
  metaverse. Companies like PepsiCo are
• 24:45
  utilizing photorealistic digital twins created by Siemens and
• 24:49
  NVIDIA to digitize their manufacturing and warehouse
• 24:52
  facilities. A digital twin is a highly
• 24:54
  complex virtual replica of a physical space.
• 24:58
  Right. They use laser scanners to map
• 25:00
  the exact dimensions of a factory floor down to the
• 25:02
  millimeter. They input the structural
• 25:04
  specifications of the machinery, the fluid dynamics of the
• 25:07
  liquids moving through the pipes, and the thermal
• 25:10
  properties of the materials. And then they run physics based
• 25:13
  simulations to identify identify equipment failures or supply
• 25:17
  bottlenecks before any physical construction begins or before a
• 25:21
  new product line is activated. It is exactly like running a
• 25:24
  flawless physics based dress rehearsal in a virtual world to
• 25:28
  prevent any disasters on opening night in the real world.
• 25:31
  You map the exact tolerances, the thermal dynamics and the
• 25:35
  structural spatial constraints digitally.
• 25:37
  So if you want to speed up a conveyor belt by 10%, you don't
• 25:40
  just pull a lever in the real factory and hope the bottles
• 25:43
  don't fall over. Exactly, To run the simulation
• 25:45
  in the industrial metaverse, the physics engine calculates the
• 25:49
  exact friction and gravity and shows you that the bottles will
• 25:52
  tip over at that specific speed. You find the structural failure
• 25:55
  digitally for free. This changes how physical
• 25:58
  factories and global supply chains are built and managed
• 26:01
  entirely. It merges those agentic software
• 26:04
  systems we discussed earlier directly with physical sensors
• 26:07
  on the factory floor. The systems can autonomously
• 26:10
  detect a slight temperature variance in a physical motor,
• 26:14
  reason that a structural failure is imminent based on historical
• 26:17
  data, and autonomously trigger mitigation workflows.
• 26:21
  The system can automatically order a replacement part from
• 26:24
  the supplier, reroute the manufacturing process to a
• 26:27
  different assembly line, and schedule a maintenance crew, all
• 26:30
  with bounded, highly specific human oversight.
• 26:34
  Technology industry has hit a series of hard boundaries
• 26:36
  simultaneously, from the physical structural limits of
• 26:39
  power grids and the strict financial demands of corporate
• 26:42
  boards to the intense governance enforced by app stores and
• 26:45
  international regulators responding to public crises like
• 26:48
  the Grok deep fakes. And as these systems evolve from
• 26:52
  generating text on a screen to taking autonomous actions in the
• 26:56
  physical world, the most urgent question is no longer what the
• 26:59
  software is capable of doing structurally, but whether we
• 27:02
  have the infrastructure and ethical guardrails ready to
• 27:05
  control it. If you're not subscribed yet,
• 27:07
  take a second and hit follow on whatever app you're using Helps
• 27:10
  us keep making this. We appreciate you being here.