Latest / AI Newspaper / Companies Rushed Into AI. Now Meta, Microsoft and Others Are Pulling Back

Transcript

Companies Rushed Into AI. Now Meta, Microsoft and Others Are Pulling Back · AI Newspaper

• 0:00
  Welcome to the AI Newspaper, your daily update
• 0:02
  on everything happening with AI. Visit AINewspaper
• 0:05
  .org to see today's stories and sign up for email
• 0:07
  alerts. So imagine buying a multi -million dollar
• 0:12
  Formula One car. Okay. And then you just like
• 0:15
  fire your driver. Right. You just assume the
• 0:18
  machine's going to navigate the track and, you
• 0:20
  know, win the race entirely on its own. Which
• 0:22
  is obviously a terrible idea. Exactly. But that
• 0:26
  is essentially what the biggest tech companies
• 0:28
  in the world has spent the last two years doing
• 0:30
  with artificial intelligence. And right now,
• 0:33
  we are watching them crash in slow motion. Yeah,
• 0:36
  it is a massive reality check. I mean, we've
• 0:38
  spent the last couple of years trapped in this
• 0:41
  echo chamber where the industry mandate was just
• 0:43
  so simple. Adopt AI at all costs, right? Right.
• 0:46
  Adopt AI, cut your human staff, and just watch
• 0:49
  productivity soar. But it didn't happen. No,
• 0:53
  it really didn't. Now employers from Meta to
• 0:56
  Microsoft are quietly realizing that the reality
• 0:58
  is much darker. The massive amount of cleanup
• 1:02
  work required, the terrifying data security exposure,
• 1:05
  and just the sheer rising costs of running these
• 1:08
  models are completely wiping out the promised
• 1:11
  productivity gains. So we are witnessing this
• 1:14
  massive industry pivot away from that adoption
• 1:16
  at any cost mentality toward a much slower, highly
• 1:20
  human supervised approach. Exactly. To understand
• 1:25
  why this pivot is happening, we have to look
• 1:26
  at just how extreme that initial corporate gold
• 1:29
  rush really was. I mean, there have been more
• 1:31
  than 112 ,000 job cuts this year alone, where
• 1:35
  companies explicitly cited AI implementation
• 1:37
  as the primary reason. Yeah, the C -suite pressure
• 1:40
  was enormous. The underlying assumption was,
• 1:43
  well, if a software program can generate a line
• 1:45
  of code or draft an email, you simply no longer
• 1:48
  need the human being who used to do those things.
• 1:50
  Right. You're just swapping out a biological
• 1:52
  worker for a digital one. Exactly. It's a very
• 1:54
  spreadsheet -driven way of thinking. And meta
• 1:57
  is actually the perfect poster child for how
• 2:00
  that theory crashed into reality. So earlier
• 2:02
  this year, Mark Zuckerberg initiated this massive
• 2:05
  internal plan, reportedly codenamed Project OT,
• 2:09
  which stood for Organization Transformation.
• 2:11
  Right. And the stated goal was to slash certain
• 2:14
  internal teams by up to 60%. Wow. 60%. Yeah.
• 2:19
  They wanted to rip out... entire layers of middle
• 2:22
  management to become what they called an AI native
• 2:24
  company. And they actually started doing it,
• 2:27
  right? They did. They went ahead and executed
• 2:29
  the first wave of those layoffs in May, but then
• 2:32
  completely out of nowhere, they canceled the
• 2:34
  second wave that was scheduled for November.
• 2:36
  And now they're actually expanding their managerial
• 2:39
  ranks again. I look at that reversal and honestly,
• 2:42
  I see the ultimate collision between spreadsheet
• 2:44
  theory and operational reality. I hope so. Well,
• 2:48
  on paper, an AI native company looks incredibly
• 2:51
  lean and profitable. You replace a complex human
• 2:54
  workflow with automated digital agents. But the
• 2:57
  CEO of the Society for Human Resource Management
• 3:00
  had a brilliant way of describing what actually
• 3:02
  happens when you do that. What did he say? He
• 3:05
  said companies are finding out they are trying
• 3:07
  to drive the truck and build the truck at the
• 3:10
  exact same time. Oh, that's good. Yeah. When
• 3:12
  you aggressively strip away the human expertise
• 3:14
  that actually keeps your operations running,
• 3:17
  you introduce a devastating amount of hidden
• 3:19
  risk into your business. Let's unpack the mechanics
• 3:22
  of that risk, because there is this explosive
• 3:25
  courtroom anecdote in the reporting today that
• 3:28
  perfectly illustrates what happens when you strip
• 3:30
  away that human expertise. Oh, the 3M case? Yeah,
• 3:33
  this is wild. It really is. So there was a major
• 3:36
  lawsuit over a fatal 2020 gas explosion. at a
• 3:40
  3M manufacturing plant in Houston. We are talking
• 3:42
  about a catastrophic event. Lives were lost.
• 3:45
  Neighboring homes were destroyed. And in a case
• 3:48
  like that, the litigation is going to rely on
• 3:50
  incredibly dense, highly technical engineering
• 3:52
  data to determine liability. It all comes down
• 3:55
  to the forensic analysis of the machinery. Right.
• 3:57
  So the defense team hires an independent expert
• 4:00
  witness to write a definitive report proving
• 4:02
  the company wasn't liable. And this expert eventually
• 4:05
  bills them. $90 ,000 for his work. 90 grand.
• 4:10
  90 grand. But the night before the trial begins,
• 4:13
  the opposing attorney is reviewing this massive
• 4:15
  14 ,500 -page report on his iPad. That's a lot
• 4:19
  of pages. Right. And he starts noticing all these
• 4:21
  bizarre red flags, like the fonts keep changing
• 4:24
  randomly. Oh, no. The formatting is completely
• 4:26
  disjointed. And worst of all, the digital timestamps
• 4:30
  indicate that this 14 ,000 -page highly technical
• 4:33
  document was... Essentially assembled in just
• 4:37
  a few hours. Which should instantly trigger alarm
• 4:40
  bells. It is physically impossible for a human
• 4:43
  engineer to synthesize that volume of complex
• 4:46
  technical data in that time frame. Exactly. So
• 4:49
  during the cross -examination on the stand, the
• 4:51
  lawyers actually pull the experts' chat logs.
• 4:54
  And it turns out this supposed expert never even
• 4:58
  traveled to the site of the explosion. Are you
• 5:00
  kidding me? Not once. He never physically inspected
• 5:03
  the gas detector in question. He literally took
• 5:06
  a two -dimensional photograph of the machine,
• 5:08
  uploaded it to ChatGPT, and prompted it with
• 5:11
  the question, wait for it, what am I looking
• 5:14
  at? That is just wow. And from there, he just
• 5:18
  let the AI hallucinate a 14 ,500 -page defense.
• 5:22
  The total abdication of professional judgment
• 5:24
  there is just stunning. And, you know, it perfectly
• 5:28
  highlights a fundamental misunderstanding of
• 5:30
  how large language models actually work. Because
• 5:32
  it's not an engineer. Exactly. An LLM doesn't
• 5:35
  have physical intuition. When you upload a photo
• 5:37
  of a gas detector, the AI doesn't see a machine
• 5:39
  with working parts. It translates the pixels
• 5:42
  into data tokens, searches its vast training
• 5:44
  data for statistical correlations to those tokens,
• 5:47
  and then generates text that predicts what a
• 5:50
  human might say about that image. It's a text
• 5:52
  prediction engine. Yes, it is a text prediction
• 5:54
  engine, not a mechanical engineer. Well, the
• 5:56
  jury was in utter disbelief. They ended up awarding
• 5:59
  the plaintiffs $61 .5 million. Which makes sense.
• 6:02
  But beyond the legal drama, this is the exact
• 6:05
  same dynamic playing out inside the Fortune 500
• 6:07
  right now. These companies are paying for expensive
• 6:10
  AI tools, firing the drivers, the employees who
• 6:14
  actually understand the nuances of the business,
• 6:16
  and letting the text prediction engine steer
• 6:19
  the company. And the macroeconomic data is finally
• 6:22
  starting to reflect that reality. Researchers
• 6:24
  at the University of Pittsburgh just released
• 6:26
  a study looking at this specific corporate behavior.
• 6:29
  And what did they find? They found that companies
• 6:31
  investing heavily in AI while simultaneously
• 6:34
  laying off workers saw absolutely zero actual
• 6:38
  rise in their productivity. Zero. Zero. The optimism
• 6:41
  of the executive board simply does not translate
• 6:44
  into better performance on the ground. And in
• 6:46
  fact, Wall Street is catching on. Yeah. Yeah,
• 6:49
  the stock market is now actively punishing companies
• 6:51
  that announce these specific AI -driven layoffs
• 6:54
  because investors have realized that stripping
• 6:57
  out your human capital in favor of an unfinished
• 6:59
  technology is a massive liability. I want to
• 7:02
  dig into the day -to -day reality of that liability.
• 7:05
  What is actually happening inside the offices
• 7:07
  of a company when they hand the keys over to
• 7:09
  the bots? Because if we look at internal communications
• 7:12
  from Meta, we can see exactly why they had to
• 7:14
  cancel that second wave of layoffs. Right. The
• 7:17
  internal systems were chaotic. Exactly. Their
• 7:19
  internal AI agents were making what management
• 7:21
  called large -scale disruptive actions. And this
• 7:25
  is the hidden cleanup crisis that the industry
• 7:27
  is trying to keep quiet. The operational failures
• 7:30
  of these models are actively wiping out the promised
• 7:32
  gains. Major technical and security incidents
• 7:35
  on Meta's internal platforms spiked by 40%. compared
• 7:39
  to the prior year but here is the truly incredible
• 7:42
  statistic the amount of human employee time spent
• 7:46
  resolving these ai created problems surged by
• 7:49
  70 just think about the underlying math of that
• 7:52
  you implement an expensive ai system to save
• 7:55
  your engineers time right but instead the system
• 7:59
  generates such a massive technical debt of errors
• 8:02
  that your highly paid human engineers have to
• 8:04
  stop building new profitable features. They are
• 8:07
  literally reduced to acting as full -time bot
• 8:10
  babysitters just to keep the platform from crashing.
• 8:13
  Andrew Bosworth met at CTO. He put a spotlight
• 8:16
  on this in a recent internal memo. He pointed
• 8:19
  out that the sheer volume of code changes made
• 8:21
  to their internal platforms skyrocketed by 220
• 8:24
  % year over year. Because the AI is just churning
• 8:28
  it out. Right. It's churning out a massive volume
• 8:30
  of code. But the actual new features that successfully
• 8:34
  reached users, that number only went up by 36%.
• 8:38
  We need to explain why this babysitting tax actually
• 8:40
  exists. It comes back to the probabilistic nature
• 8:44
  of the technology. AI is incredibly adept at
• 8:47
  producing what we call plausible looking output.
• 8:50
  Plausible looking. Yeah. Because it is predicting
• 8:53
  the most likely next sequence of words or code,
• 8:55
  it generates a legal report or a software module
• 8:58
  that looks perfectly correct at a quick glance.
• 9:01
  But baked deep inside that plausible output are
• 9:04
  hidden flaws. Like what? Like hallucinated variables
• 9:07
  or logic chains that make absolutely no sense
• 9:10
  to a human developer. And because it looks so
• 9:12
  authoritative, people just accept it and move
• 9:14
  on. Yes. And that is where the danger multiplies.
• 9:17
  When those plausible errors are left unsupervised,
• 9:20
  they snowball. A subtle error in a piece of generated
• 9:23
  code gets automatically integrated into a larger
• 9:25
  software build. And then that build interacts
• 9:28
  with another system. Exactly. So by the time
• 9:31
  the code finally breaks in a live production
• 9:33
  environment, untangling that massive web of synthetic
• 9:36
  logic takes a human engineer days of forensic
• 9:39
  cleanup. You are building a house with bricks
• 9:41
  that look solid but dissolve in the rain. Think
• 9:44
  about your own workday for a second. Are you
• 9:46
  actually saving time with AI? Or are you just
• 9:49
  replacing the manual labor of doing the work
• 9:51
  with the new manual labor of babysitting a bot
• 9:54
  that does the work badly? That's a great question.
• 9:57
  Because this isn't just a coding problem. It's
• 9:59
  infecting basic corporate communication. Ryan
• 10:01
  Roslansky, an executive at Microsoft, recently
• 10:04
  warned about a phenomenon he calls the doom loop.
• 10:07
  Oh, the memo story. Yeah. He received a lengthy
• 10:09
  memo from a coworker that was clearly generated
• 10:11
  by an AI. And not having the time to read it,
• 10:14
  he used his own AI to summarize the AI -generated
• 10:18
  memo. And realized he learned absolutely nothing
• 10:21
  new from the exchange. Exactly. He warned that
• 10:24
  we are sinking into an emerging ocean of sameness
• 10:26
  in corporate work. It is a closed loop of zero
• 10:29
  -value creation. The AI is doing the reading,
• 10:31
  the AI is doing the writing, and the human beings
• 10:33
  are completely disconnected from the actual critical
• 10:35
  thinking process that moves a business forward.
• 10:38
  And it gets so much worse when you step outside
• 10:40
  the white collar office and apply this tech to
• 10:42
  physical labor. Like at Walmart? Right. At Walmart,
• 10:46
  they rolled out bots to assign floor tasks, monitor
• 10:49
  performance, and plot fulfillment rules for workers.
• 10:51
  And the human managers are actively reporting
• 10:54
  that these bots are setting entirely impossible
• 10:56
  standards. Well, spatial reasoning is incredibly
• 10:59
  difficult for these models. The AI routing algorithm
• 11:02
  is optimizing for a mathematical ideal. It basically
• 11:04
  assumes a human worker is a frictionless point
• 11:07
  moving through a perfect grid. Right, with zero
• 11:10
  fatigue every single second of an eight -hour
• 11:12
  shift. Exactly. The bot looks at a 2D floor plan
• 11:15
  of the warehouse and plots a route that makes
• 11:17
  geometric sense. But it completely ignores the
• 11:20
  physical reality of a broken pallet blocking
• 11:22
  aisle four or a spill in aisle six. And the result
• 11:25
  is completely exhausted employees. cutting corners
• 11:27
  on safety protocols just to keep up with the
• 11:29
  algorithm's impossible pace. And then what happens?
• 11:32
  The system starts pumping out useless, false
• 11:34
  safety alerts because the humans aren't following
• 11:36
  the mathematical ideal. The human managers are
• 11:40
  literally having to step in and fight the software
• 11:42
  just to protect their workers. The takeaway here
• 11:44
  is that rapid, frictionless deployment without
• 11:47
  deep human supervision isn't just a little inefficient.
• 11:50
  It is actively destructive to the core operations
• 11:53
  of a business. But the operational mess is really
• 11:56
  only half the story here. Because unchecked AI
• 11:59
  adoption is also turning into a terrifying data
• 12:02
  security vulnerability and, frankly, a financial
• 12:05
  sinkhole. We are suddenly seeing a massive corporate
• 12:08
  backlash against the major third -party models
• 12:10
  from companies like Anthropic and OpenAI. We
• 12:14
  are looking at a seismic shift in enterprise
• 12:17
  strategy. I mean, for the last year, those third
• 12:19
  -party frontier models were the absolute gold
• 12:22
  standard. Every Fortune 500 company was... desperate
• 12:25
  for access. Right. And now the biggest players
• 12:28
  in the industry are slamming the brakes. Yeah.
• 12:31
  Companies like Palantir, Booz Allen Hamilton,
• 12:34
  and NVIDIA are actively restricting their own
• 12:37
  employees from using these advanced models because
• 12:39
  of intense fears over data retention. This all
• 12:42
  boiled over recently when Anthropic implemented
• 12:45
  a 30 -day data retention policy for their new
• 12:47
  Fable 5 model. And the mechanics of this are
• 12:51
  important to understand. Anthropic argued that
• 12:53
  they needed to retain user query logs for 30
• 12:56
  days to detect complex, novel security threats.
• 13:00
  Okay, so why did they need 30 days? Well, think
• 13:04
  of their Fable 5 model, like a highly complex
• 13:07
  digital vault. Hackers aren't just trying to
• 13:09
  pick the lock in one single attempt. They are
• 13:12
  testing one vulnerability on a Monday, probing
• 13:15
  a different weakness on a Thursday, and trying
• 13:18
  to execute a complex prompt injection attack
• 13:20
  over weeks. Right. So Anthropic needs a 30 -day
• 13:23
  window of log data to identify those long -term
• 13:26
  attack patterns. But from the perspective of
• 13:28
  an enterprise customer, that explanation is a
• 13:31
  massive red flag. Oh, absolutely. If you are
• 13:34
  Palantir and your employees are feeding highly
• 13:36
  classified defense data into the model, or you're
• 13:40
  Booz Allen doing proprietary cybersecurity analysis.
• 13:43
  You cannot allow a third -party AI lab to hold
• 13:46
  onto your most sensitive queries on their servers
• 13:48
  for a month. No, you can't. And the enterprise
• 13:51
  revolt was immediate. Palantir is now demanding
• 13:54
  irrevocable, absolute zero data retention guarantees
• 13:57
  before they will allow Anthropic anywhere near
• 14:00
  their internal software. And Booz Allen went
• 14:02
  even further, right? Yeah, they completely banned
• 14:05
  Anthropic's commercial model for any of their
• 14:07
  cybersecurity work. NVIDIA is restricting the
• 14:10
  use of third -party models to strictly non -sensitive
• 14:13
  tasks, and they're forcing their staff to use
• 14:14
  their own in -house Nematron models instead.
• 14:17
  Because it's localized. Exactly. A model like
• 14:20
  Nematron is smaller and highly localized, meaning
• 14:23
  it runs entirely on NVIDIA's own secure servers.
• 14:26
  So the proprietary data never actually leaves
• 14:28
  the building. And, you know, it isn't just about
• 14:30
  the security fears. The sheer cost is causing
• 14:33
  executives to panic. Microsoft was reportedly
• 14:36
  on track to spend an estimated $1 billion this
• 14:39
  year just on internal cloud usage. One billion.
• 14:43
  That is $1 billion just paying for their own
• 14:47
  employees to use a third -party chatbot. When
• 14:50
  you combine the hidden compounding costs of human
• 14:53
  cleanup time that we discussed earlier with a
• 14:56
  billion -dollar licensing bill, the economics
• 14:58
  of this technology suddenly look terrifying for
• 15:01
  a corporate finance department. Oh, yeah. So
• 15:04
  Microsoft is now actively instructing their employees
• 15:06
  to stop using Claude and forcing them to use
• 15:09
  their own in -house copilot system instead. And
• 15:11
  Meta is making the exact same defensive move.
• 15:13
  Right, with Metacode. Yeah, they just slashed
• 15:16
  their internal Claude licenses from 60 ,000 seats
• 15:18
  down to 30 ,000. And they are pushing their engineers
• 15:21
  onto their homegrown alternative called Metacode.
• 15:24
  The ripple effects of this enterprise backlash
• 15:25
  are existential for the AI labs. Anthropic was
• 15:29
  recently labeled a supply chain risk by the U
• 15:32
  .S. Department of Defense. Ouch. Yeah. For a
• 15:35
  startup that is aiming for a massive $2 trillion
• 15:37
  IPO valuation, being designated a national security
• 15:41
  risk and losing your biggest, most lucrative
• 15:44
  enterprise customers is devastating. So what
• 15:46
  do they do? It forced them to rush out a new
• 15:49
  system called Enterprise Frontier Safeguards.
• 15:51
  This system finally allows companies to store
• 15:54
  their AI activity data on their own private cloud
• 15:57
  infrastructure, locked behind their own encryption
• 15:59
  keys. Think about what this means for you, the
• 16:02
  end user. The biggest, most powerful tech companies
• 16:05
  in the world are looking at the landscape and
• 16:08
  telling their own employees that third -party
• 16:10
  AI is too expensive and far too risky to use
• 16:13
  with any proprietary data. It is the ultimate
• 16:16
  case of AI for thee, but not for me. And what
• 16:19
  is truly brilliant from a ruthless, opportunistic
• 16:22
  business perspective is how Microsoft is capitalizing
• 16:25
  on this exact anxiety. While they are rapidly
• 16:28
  reeling in their own spending on third -party
• 16:30
  models, they are aggressively marketing their
• 16:33
  own isolated, hyper -secure cloud environments
• 16:36
  to poach terrified enterprise customers who are
• 16:39
  pulling their data away from OpenAI and Anthropic.
• 16:42
  They saw a gold rush. and decided to sell the
• 16:45
  safes to store the gold in. It is a masterful
• 16:47
  pivot. So if we look past all these costly failures,
• 16:51
  the hallucinated code, the impossible warehouse
• 16:54
  routes, the billion -dollar bills, and the security
• 16:56
  scares, how is the industry moving forward? What
• 16:58
  is the new blueprint for actually making this
• 17:01
  work? The new blueprint requires a total rejection
• 17:04
  of the plug -and -play fantasy. The companies
• 17:06
  that are succeeding are shifting toward a much
• 17:08
  slower, highly intentional, human -supervised
• 17:12
  integration process. Microsoft actually published
• 17:15
  an incredibly revealing internal memo about their
• 17:17
  experience acting as customer zero for their
• 17:20
  own AI products. They openly admitted that simply
• 17:23
  licensing AI tools to 200 ,000 employees didn't
• 17:27
  actually change how the work got done. Access
• 17:32
  and usage do not equal transformation. If you
• 17:35
  just hand an employee an AI tool without fundamentally
• 17:38
  redesigning the workflow, they just use the massive
• 17:41
  computational power to do their old, inefficient
• 17:43
  tasks slightly faster. Or worse, as Microsoft
• 17:47
  noted in that memo, if you inject an AI agent
• 17:50
  into a process that is already broken, you just
• 17:53
  speed up the failure rate at step one, which
• 17:55
  creates a massive, unmanageable bottleneck at
• 17:58
  step two. You just end up waiting in line faster.
• 18:01
  I'd take it a step further. You aren't just waiting
• 18:03
  in line faster. You are actively compounding
• 18:06
  the errors in the system. The solution Microsoft
• 18:09
  found was to stop treating AI like a software
• 18:11
  update and start treating it as a catalyst for
• 18:14
  total end -to -end workflow redesign. They highlighted
• 18:18
  a massive success story with their cloud supply
• 18:20
  chain team to prove this point. They didn't just
• 18:22
  give the supply chain planners a generic chatbot.
• 18:24
  Right. They mapped it out. Yeah. The software
• 18:26
  engineers and the human supply chain experts
• 18:28
  sat down in a room together and mapped out the
• 18:30
  entire convoluted workflow from start to finish.
• 18:34
  They centralized all their fragmented data so
• 18:36
  the AI had a single accurate source of truth.
• 18:40
  And then rather than one big AI model, they deployed
• 18:43
  over 100 highly specialized narrow AI agents
• 18:46
  to handle very specific mundane steps in the
• 18:49
  chain. And the results were exactly what the
• 18:51
  industry had been hoping for all along. Cycle
• 18:53
  times for those selected workflows dropped by
• 18:56
  an astonishing 75%. That's huge. Supply chain
• 18:59
  planners who used to spend an entire week just
• 19:02
  tracing the data to figure out why a demand forecast
• 19:04
  changed can now get human validated answers in
• 19:07
  20 minutes. The critical difference is that the
• 19:11
  AI is doing the tedious data gathering, but the
• 19:14
  human expert is still the one making the strategic
• 19:16
  decision at the end. Microsoft calls this new
• 19:19
  corporate math capability add. The formula is
• 19:23
  CI plus AI equals CA. Continuous improvement
• 19:28
  plus artificial intelligence equals capability
• 19:31
  add. It is a vital distinction in philosophy.
• 19:34
  The goal is no longer just doing the same old
• 19:37
  tasks faster with fewer people to save a few
• 19:39
  dollars on payroll. Efficiency is the floor.
• 19:42
  Capability is the ceiling. The true goal is achieving
• 19:45
  things that were previously impossible for your
• 19:47
  team. For example, a human analytics team might
• 19:51
  only have the time and bandwidth to model three
• 19:53
  different supply chain scenarios before a deadline.
• 19:55
  But a human -steered AI team can model 20 highly
• 19:59
  complex scenarios in that exact same time frame,
• 20:02
  which radically improves the quality of the final
• 20:04
  strategic decision. If we look past the corporate
• 20:07
  branding of capability ad, what these massive
• 20:11
  tech companies are finally admitting is that
• 20:12
  you can't just strap a jetpack to a horse. If
• 20:16
  you actually want to travel at jet speed, you
• 20:17
  have to fundamentally redesign the vehicle from
• 20:19
  the ground up. And you desperately need a highly
• 20:22
  trained human pilot steering the whole thing.
• 20:24
  The organizations that ultimately win this race
• 20:26
  won't be the ones who just buy the most enterprise
• 20:29
  licenses or lay off the largest percentage of
• 20:31
  their staff. The winners will be the companies
• 20:33
  that build a continuous learning loop. Where
• 20:36
  the human experts train the AI and the AI in
• 20:39
  turn empowers the humans. Exactly. You absolutely
• 20:42
  need the contextual understanding and the judgment
• 20:44
  of a human expert to teach the AI what good actually
• 20:48
  looks like. And you are finally seeing this realization
• 20:51
  hit the HR departments. Oh, definitely. Meta
• 20:53
  has completely stopped tracking employee computer
• 20:55
  activity to force AI training, and they have
• 20:58
  actually removed AI adoption metrics from their
• 21:01
  employee performance reviews. They realize that
• 21:03
  forcing adoption creates resentment and bad data.
• 21:06
  And Microsoft is taking a similar human -centric
• 21:08
  approach. They are moving away from forcing hundreds
• 21:11
  of employees to sit through broad, generic AI
• 21:14
  webinars. Instead, they launched a program called
• 21:18
  Right. This program pairs junior engineers with
• 21:21
  experienced veteran human mentors, and they simply
• 21:25
  use the AI as an assistant in the learning process.
• 21:28
  It is a deeply human first approach to upskilling.
• 21:31
  Because the smartest companies realize that their
• 21:34
  internal institutional knowledge is their only
• 21:37
  real competitive moat. If you fire all your veteran
• 21:40
  experts to save a buck. Who is going to check
• 21:43
  the AI's work when it inevitably hallucinates
• 21:45
  a complex legal argument or writes a subtle,
• 21:48
  devastating security flaw into your core product?
• 21:51
  Which leads us with a truly fascinating, slightly
• 21:53
  ironic question to ponder as we wrap up. If the
• 21:57
  entire tech industry is finally waking up to
• 22:00
  the reality that unchecked AI naturally creates
• 22:03
  a massive volume of plausible -looking errors,
• 22:06
  and that it requires intense, highly skilled
• 22:08
  human oversight to prevent catastrophic failures,
• 22:12
  Will the most lucrative, secure jobs of the next
• 22:14
  decade actually be AI auditors and AI janitors?
• 22:18
  That's a great point. Like, will the future economy
• 22:21
  belong to the human experts hired specifically
• 22:24
  to clean up the messes that the machines leave
• 22:26
  behind? The data certainly points that way. I
• 22:30
  mean, as the volume of automated output explodes,
• 22:32
  the premium on human judgment and critical thinking
• 22:35
  has really never been higher. Thanks for listening
• 22:37
  to the AI newspaper. Visit AINewspaper .org to
• 22:41
  catch up on the latest updates and sign up for
• 22:43
  free email alerts.