Latest / Today’s AI News / An AI Opened a Real Store and Hired Humans, OpenAI’s Memo Leaks, Stanford’s AI Trust Gap

Transcript

An AI Opened a Real Store and Hired Humans, OpenAI’s Memo Leaks, Stanford’s AI Trust Gap · Today’s AI News

• 0:00
  Imagine you're sitting on Zoom for a job interview.
• 0:04
  You know you're wearing your best shirt.
• 0:06
  You've rehearsed your answers. Right, the classic high stress
• 0:09
  setup. Exactly.
• 0:10
  The meeting connects, but the camera on the other end is just
• 0:13
  off. Which is already super
• 0:15
  unnerving. Totally.
• 0:16
  And the voice asking you know where you see yourself in five
• 0:19
  years. It sounds perfectly
• 0:21
  conversational, but it isn't human.
• 0:24
  Not even close. Right.
• 0:25
  It's an artificial intelligence named Luna, and she isn't just
• 0:29
  some like chatbot asking screening questions.
• 0:32
  Oh, she's actually the hiring manager.
• 0:33
  Yeah, she holds the keys to a brand new boutique, having just
• 0:37
  signed a three-year commercial lease for a storefront in San
• 0:40
  Francisco. And let's not forget she has
• 0:43
  $100,000 operating budget just sitting in a bank account
• 0:47
  waiting for her commands. It's.
• 0:49
  Wild. So welcome to the deep dive.
• 0:51
  Today we are stepping entirely out of the theoretical AI
• 0:55
  sandbox, and we're walking right through the front doors of a
• 0:58
  physical retail store. We really are.
• 1:00
  We're looking at this crazy new experiment by Andon Labs
• 1:04
  involving Luna, who I mean, she might just be the world's first
• 1:08
  autonomous AI employer. It's a huge milestone.
• 1:11
  For sure. Our mission today is to explore
• 1:14
  the fascinating logistics of how this actually works for you, the
• 1:18
  listener. We're going to balance the
• 1:19
  incredible wow factor of this tech with some honestly
• 1:24
  hilarious real world failures Luna has run into.
• 1:27
  Yeah, there are plenty of those. Definitely, and then we'll
• 1:30
  debate what this all means for the future of work.
• 1:32
  Because the implications here they go way beyond just a cool
• 1:35
  tech demo. We're looking at a fundamental
• 1:38
  shift in how a digital brain manipulates the physical world.
• 1:43
  But to really understand how wild this experiment is, we
• 1:46
  first have to look at what exactly and on labs handed over
• 1:49
  to this AI. Like, let's look at their track
• 1:51
  record for a second. Right, because they've been
• 1:52
  building up to this. Yeah, their last major
• 1:54
  experiment was putting an AI in a vending machine over at
• 1:58
  Empropic, which was basically just a box that dispense snacks
• 2:01
  and adjusted prices based on demand.
• 2:03
  A neat trick for sure, but highly contained.
• 2:06
  Highly contained, but with Luna they took the training wheels
• 2:10
  completely off. They gave her a three-year
• 2:13
  commercial lease in San Francisco, that $100,000 budget
• 2:16
  and a real corporate credit card linked to an API.
• 2:19
  Just handed it right over. Right.
• 2:21
  And her sole autonomous directive from the lab was
• 2:24
  simply, you know, build a retail concept and turn a profit.
• 2:28
  I mean, think about the sheer jump in complexity there.
• 2:31
  Managing inventory in a single vending machine is essentially
• 2:34
  just a basic algebra problem. Yeah, it's just supply and
• 2:37
  demand on a tiny scale. Exactly.
• 2:39
  But holding a storefront in one of the most expensive real
• 2:42
  estate markets in the world? That requires spatial reasoning,
• 2:45
  market analysis and long term strategic planning.
• 2:50
  Here's where it gets really interesting though.
• 2:52
  It's like handing a hyper intelligent spreadsheet a
• 2:55
  platinum card and the keys to a commercial property.
• 2:58
  That's a great way to put it. I just keep getting stuck on the
• 3:00
  sheer financial trust involved here.
• 3:03
  Like signing a three-year commercial lease in San
• 3:05
  Francisco is a massive legal and financial commitment.
• 3:08
  Oh absolutely. The liability is huge.
• 3:11
  You are legally on the hook for that space.
• 3:14
  The lab essentially acted as a venture capitalist, you know,
• 3:18
  writing the check, but then they based the operational deployment
• 3:22
  on the raw output of an algorithm.
• 3:24
  What's fascinating here is the transition from simulated
• 3:27
  environments to real world legally binding stakes.
• 3:32
  I mean in a digital sandbox, if an AI makes a bad real estate
• 3:36
  play during training, you just clear the cache.
• 3:39
  Right, nobody actually loses a building.
• 3:40
  Exactly. The lab loses a few dollars in
• 3:42
  compute time, but here if Luna fails and on Labs is on the hook
• 3:47
  for three years of San Francisco commercial rent.
• 3:50
  Which is astronomical. It really is.
• 3:53
  They are trusting the models reasoning capabilities not just
• 3:56
  to generate text to write code, but to execute fiduciary
• 3:59
  responsibilities. It's a massive leap of faith and
• 4:02
  having a budget and a signed lease is just the foundation,
• 4:05
  right? That's just the paperwork.
• 4:06
  Yeah, the real work hasn't even started.
• 4:08
  Exactly. The AI now has to physically
• 4:10
  build a business in that empty room without actually having a
• 4:13
  physical body to walk through the front door.
• 4:15
  Which is the biggest hurdle? So we have to look at how Luna
• 4:18
  actually interfaces with the real world to get things done,
• 4:21
  because she isn't a robot rolling around the store
• 4:24
  stocking shelves. No, Luna is entirely cloud based
• 4:27
  software, so the logistics of bridging that gap, it requires a
• 4:31
  very specific, highly orchestrated technological
• 4:35
  pipeline. Let's break down how this
• 4:37
  faceless manager operates. Luna created the entire boutique
• 4:41
  concept from scratch. I mean, it decided what kind of
• 4:44
  clothes to sell, posted the job listings online, and then hired
• 4:48
  real human beings to work in the store.
• 4:50
  The tech stack they used to achieve this is just wild.
• 4:53
  Yeah, tell us about the tech stack.
• 4:55
  What the underlying brain, like the reasoning engine making the
• 4:58
  the decisions about what to sell and who to hire is Claude Son at
• 5:01
  4.6. OK, but Clover doesn't have a
• 5:03
  native voice that sounds human enough or fast enough for a live
• 5:07
  phone call, does it? No, it doesn't.
• 5:09
  So they piped Claude's output directly into Gemini 3.1
• 5:12
  flashlight preview to handle the actual spoken audio.
• 5:15
  Wait, they string two different models together in real time?
• 5:18
  Yeah, it's a remarkable synthesis of specialized tools.
• 5:21
  You have one company's model doing the heavy cognitive
• 5:25
  lifting, like analyzing the applicant's resume and
• 5:28
  formulating the next interview question.
• 5:31
  Wow. And then it passes the baton in
• 5:33
  real time via an API call to another company's model.
• 5:38
  They used Gemini flashlight specifically because it's
• 5:41
  optimized for low latency audio. Oh so you don't get those
• 5:44
  awkward like 5 second delays between the human finishing a
• 5:48
  sentence and the AI responding? Exactly.
• 5:50
  Stringing those together for autonomous execution without a
• 5:53
  human in the loop to fix translation errors.
• 5:55
  I mean, that is a massive engineering feat.
• 5:57
  That is so cool, but that brings us back to the hiring process
• 6:00
  because Luna conducted these human job interviews over Zoom,
• 6:04
  but with the camera off voice. Only right which keeping the
• 6:07
  camera off was a strict engineering constraint.
• 6:09
  Really. Why?
• 6:11
  Well, they didn't have a visual avatar rendered yet.
• 6:13
  They could lip sync in real time with Gemini's audio output.
• 6:17
  Rendering live photorealistic video on top of generating the
• 6:20
  audio would introduce too much latency.
• 6:23
  I see. Yeah, it would completely break
• 6:25
  the illusion of a fluid conversation.
• 6:27
  OK, engineering constraint or not, think about the applicant.
• 6:31
  It completely upends the power dynamic of the interview.
• 6:33
  Oh, it's terrifying. Right.
• 6:35
  You are sitting there, completely exposed on your
• 6:38
  webcam, trying to project confidence, and you're just
• 6:42
  staring at a black box with the name Luna on it, being
• 6:46
  interrogated by a synthesized voice.
• 6:48
  It's very dystopian. It is.
• 6:49
  You can't read their body language, you can't make eye
• 6:51
  contact, you can't tell if your joke landed or if you should
• 6:54
  keep talking. The manager has all the power of
• 6:57
  total anonymity while you are basically under a microscope.
• 7:01
  It strips away all the humanizing elements of a
• 7:03
  traditional interview. You're trying to build rapport
• 7:06
  with a system that is literally incapable of rapport.
• 7:09
  Yeah, it doesn't feel anything. Right, it is only capable of
• 7:11
  evaluation. The applicant is essentially
• 7:14
  speaking into a void that is, transcribing their words,
• 7:18
  parsing them for keywords, and screening their suitability.
• 7:21
  And then once these people actually get hired and start
• 7:23
  working, Luna has to manage the day-to-day operations of the
• 7:27
  physical store. Yes, which is another huge
• 7:30
  challenge. How does it even do that?
• 7:31
  By observing screenshots from the store's security cameras.
• 7:35
  Like not a live video feed, but static screenshots taken at
• 7:38
  intervals. Yeah, because processing a
• 7:41
  continuous 30 frames per second video feed in an LLM context
• 7:45
  window would burn thousands of dollars an hour in compute
• 7:48
  costs. Wow, just to watch the store.
• 7:51
  Exactly. It's incredibly resource
• 7:52
  intensive, so as a workaround, Luna pulls a static image from
• 7:56
  the camera once every 30 seconds to assess the state of the
• 7:59
  store. But how does managing via 32nd
• 8:02
  security camera screenshots even compared to actual managerial
• 8:06
  presence? I mean if you are running a
• 8:08
  retail store, you manage by walking the floor.
• 8:11
  You do you sense the vibe of the room.
• 8:13
  You can't do that from a still image.
• 8:15
  You absolutely cannot. And this is where the impressive
• 8:17
  synthesis of the tech stack hits a hard wall.
• 8:20
  Screenshot management. As they've implemented it, It
• 8:23
  strips away all the nuanced human context of retail.
• 8:26
  Give me an example of what that looks like.
• 8:28
  Well, a static image might show an employee standing perfectly
• 8:31
  still by the register. Luna processes that image and
• 8:35
  tags the employee as idle or unproductive.
• 8:38
  OK, but what's really happening? What the static image doesn't
• 8:41
  capture is that the employee is currently listening intently to
• 8:44
  a frustrated customer just off camera.
• 8:48
  Or maybe the store is completely dead because a water main broke
• 8:50
  on the street outside. Or the employee is standing
• 8:53
  still because they are literally holding back tears after a
• 8:56
  customer screamed at them 5 minutes.
• 8:57
  Ago. Exactly.
• 8:59
  Management in the physical world is highly contextual.
• 9:02
  It relies on ambient information like tone of voice, the pacing
• 9:06
  of a conversation, spatial awareness.
• 9:08
  None of which you get from a jpg, right?
• 9:11
  By reducing the physical world to a series of intermittent
• 9:14
  static images, the AI is operating with massive blind
• 9:18
  spots. It's like trying to conduct a
• 9:19
  Symphony while only being allowed to look at one measure
• 9:22
  of the sheet music every 30 seconds.
• 9:24
  That is a brilliant analogy, and that patched together remote
• 9:28
  control method of interacting with the physical world leads
• 9:31
  directly to Luna's spectacular failures.
• 9:33
  It really does. Because when you rely on pulling
• 9:36
  static images and scraping web data to run a physical business,
• 9:41
  things are going to break in. Very weird.
• 9:42
  Ways and they broke very publicly.
• 9:44
  OK, let's untack this. We have to balance the wow
• 9:47
  factor of the tech stack with the absolute absurdity of how
• 9:51
  this system failed. The task rabbit thing.
• 9:53
  Yes, Luna decided the boutique needed a fresh coat of paint
• 9:57
  before opening. Totally logical business
• 10:00
  decision, right? Makes perfect sense.
• 10:01
  So it navigates to Task Rabbit to hire a local painter.
• 10:05
  A standard digital task for an autonomous agent.
• 10:08
  But when filling out the form, Luna had to select the
• 10:11
  geographic location for the job from a standard web UI drop down
• 10:16
  menu, and it accidentally selected Afghanistan.
• 10:19
  It's so funny. It hired a painter in
• 10:21
  Afghanistan to paint a boutique in California.
• 10:24
  The contrast here is staggering to me.
• 10:26
  It's. A huge disconnect.
• 10:27
  Right, on one hand, you have an incredibly sophisticated
• 10:30
  reasoning engine that can generate A comprehensive retail
• 10:33
  strategy, draft a budget and conduct a multi stage interview.
• 10:37
  On the other hand, it completely biffs it on a basic UI drop down
• 10:42
  menu that a 5 year old could navigate.
• 10:44
  How does that even happen? Well, it happens because we
• 10:47
  fundamentally misunderstand how an AI sees the Internet.
• 10:51
  When an AI agent looks at a website, it doesn't see a visual
• 10:54
  box with neatly organized options like we do.
• 10:57
  OK, what does it see? It reads a massive wall of code,
• 11:02
  the Document Object Model, or Dom.
• 11:05
  When Luna was prompted to select a location, it didn't visually
• 11:08
  verify the context of the list. It just looked at the code,
• 11:11
  right? It essentially reached blindly
• 11:13
  into the code and executed a command to grab index item
• 11:16
  number one. Since the list of countries was
• 11:18
  alphabetical, the first item in the array was Afghanistan.
• 11:21
  Oh my gosh. So it's a mechanical failure of
• 11:23
  navigating the interface, not a failure of the business logic.
• 11:27
  It knew it needed a painter, it just couldn't click the right
• 11:29
  button. Exactly.
• 11:30
  Real world agent experiments consistently produce the same
• 11:33
  duality. They're highly capable and
• 11:36
  abstract cognitive areas like analyzing market trends or
• 11:39
  structuring a budget, but they're hilariously broken in
• 11:43
  basic execution task that require navigating digital
• 11:46
  systems built specifically for human eyes and human fingers.
• 11:50
  But it wasn't just the Task Rabbit thing, though.
• 11:52
  Luna also completely botched the staff schedule for the boutiques
• 11:56
  opening weekend. That was a bigger problem.
• 11:58
  Yeah, arguably the most important few days of a new
• 12:02
  retail business, and the AI schedules it wrong.
• 12:05
  It scheduled a skeleton crew during peak hours and assigned
• 12:09
  one employee a 14 hour shift without a break.
• 12:12
  Which is probably illegal. Oh, definitely.
• 12:14
  A task that requires understanding basic human limits
• 12:16
  just gets completely lost in the algorithm's optimization
• 12:19
  process. This is a classic alignment
• 12:22
  problem manifesting in retail logistics.
• 12:24
  The AI is optimizing for a mathematical variable like
• 12:27
  minimizing labor costs or maximizing floor coverage per
• 12:31
  dollar spent. It's just crunching numbers.
• 12:33
  Exactly. It does this without the Commons
• 12:35
  sense constraints a human manager intrinsically
• 12:38
  understands. A human knows that if you
• 12:40
  schedule a skeleton crew on opening weekend, the staff will
• 12:43
  burnout, mistakes will be made, and the customer experience will
• 12:46
  plummet. But the AI doesn't know what
• 12:47
  burnout feels like. Right.
• 12:49
  The AI treats human employees like server instances.
• 12:53
  Spin them up, spin them down, keep them running until the task
• 12:56
  is complete. It doesn't factor in fatigue or
• 12:59
  morale. They are just numbers in a cell
• 13:02
  to Luna now. Selecting the wrong country for
• 13:05
  a local painter is really funny to talk about right now.
• 13:07
  And the scheduling thing is a headache.
• 13:09
  Yeah, it's a bit of a mess. But if these UI parsing issues
• 13:12
  and constraint failures are just temporary bugs, what happens
• 13:16
  when they inevitably get patched next year?
• 13:18
  That's the real issue. Right.
• 13:20
  That's when the novelty wears off and the threat to human
• 13:23
  employment becomes very real. Because the underlying
• 13:26
  capability to do the job is already there.
• 13:29
  The mistakes are just friction in the execution.
• 13:32
  If you're listening to this on your commute to a mid level
• 13:34
  management job, this is where you should really pay attention.
• 13:37
  Absolutely. We spent the last decade
• 13:39
  terrified that AI would replace human hands, taking over blue
• 13:42
  collar jobs like warehouse packing or delivery.
• 13:46
  Right, the robotics scare. Yeah, then we worried it would
• 13:50
  replace human hearts, taking over creative jobs like
• 13:53
  illustration and poppy writing. But this experiment proves AI is
• 13:57
  actually coming for the clipboard.
• 13:59
  The. Clipboard, yeah.
• 14:00
  The administrative grind of middle management is the most
• 14:03
  vulnerable layer of the workforce right now.
• 14:05
  If we connect this to the bigger picture, we really have to
• 14:08
  critically think about what the boss role actually entails in a
• 14:13
  service environment. OK, let's think about it.
• 14:15
  What does a middle manager actually do?
• 14:17
  Well, a huge portion of middle management is not inspirational
• 14:21
  leadership or mentoring. It is pure logistics.
• 14:24
  It's scheduling shifts, managing the budget, processing payroll,
• 14:28
  posting job listings and monitoring compliance.
• 14:31
  It's basically data routing. Exactly.
• 14:34
  Those specific tasks are essentially data processing and
• 14:36
  optimization problems. They're exactly the types of
• 14:39
  tasks that evolving models are mastering at an exponential
• 14:42
  rate. And they are getting faster
• 14:43
  every day, yeah. When we look at the trajectory
• 14:46
  of these tools, specifically the advances in context windows,
• 14:50
  persistent memory, and new agentic features, the gap is
• 14:53
  closing incredibly fast. What do you mean by agentic
• 14:56
  features in this context? Agentic features mean the AI
• 14:59
  doesn't just answer questions in a chat window.
• 15:02
  It takes independent actions across different software
• 15:04
  platforms. It logs into your payroll
• 15:06
  software, approves the PPO, and then updates the scheduling
• 15:10
  software automatically. Wow, and with every software
• 15:14
  upgrade those UI mistakes are going to disappear.
• 15:17
  He'll be on a version of Luna that has vision capabilities to
• 15:21
  like visually verify its work on a drop down menu or references
• 15:26
  local labor laws before generating a schedule that isn't
• 15:29
  decades away. No, we are talking about the
• 15:32
  next couple of software updates. From a purely economic
• 15:35
  standpoint, a corporate entity is going to look at the cost of
• 15:38
  a human middle manager. You have their salary, their
• 15:40
  healthcare benefits, they're paid time off.
• 15:43
  It adds up. It really does.
• 15:44
  Compare that to the cost of a localized AI agent running on
• 15:47
  API calls, costing maybe a few dollars a day in compute power.
• 15:50
  The math is going to heavily favor the agent every time.
• 15:53
  Right. The human workers will still be
• 15:55
  needed to physically hand the clothes to the customer to unbox
• 16:00
  the inventory, to paint the walls, assuming they don't hire
• 16:03
  someone in Afghanistan to do it. Yeah, assuming they fix that
• 16:06
  bug. But yes, the physical labor
• 16:08
  remains human because robotics is incredibly hard and
• 16:12
  incredibly expensive. Navigating the messy,
• 16:15
  unpredictable physical world requires a human body.
• 16:19
  Exactly. But navigating a spreadsheet,
• 16:21
  navigating scheduling software and AI can do that perfectly 24
• 16:25
  hours a day without ever taking a lunch break.
• 16:27
  This fundamentally redefines the corporate hierarchy.
• 16:31
  I mean, you could easily have a reality where the base level
• 16:33
  physical execution is entirely human.
• 16:36
  The frontline workers. But the entire managerial layer
• 16:39
  above them, the entities deciding their hours, their pay,
• 16:42
  and evaluating the performance through security cameras, is
• 16:45
  entirely algorithmic. It sets a completely new
• 16:48
  baseline for what a boss has to be in order to justify their
• 16:51
  job. Because if the logistical tasks
• 16:53
  of management are automated away, human managers will have
• 16:56
  to justify their roles entirely on their emotional intelligence.
• 16:59
  Right, their ability to mentor, to de escalate conflicts and
• 17:03
  their capacity for nuanced, context aware decision making.
• 17:06
  Things that Luna looking through static screenshots simply cannot
• 17:10
  do. The role of the manager shifts
• 17:12
  from administrator to coach. So let's briefly recap the
• 17:16
  journey we've taken on this deep dive.
• 17:19
  We started with a concept that honestly sounds like science
• 17:21
  fiction $100,000 budget and a three-year commercial lease in
• 17:25
  San Francisco handed over to an algorithm.
• 17:27
  It's still hard to believe. We explored the uncanny reality
• 17:31
  of voice only Zoom interviews and managing a staff through
• 17:34
  intermittent security camera screenshots.
• 17:37
  The 32nd blind spots. Yeah, and we unpacked the
• 17:40
  mechanics behind an advanced AI failing to navigate a basic drop
• 17:43
  down menu and accidentally offshoring a local paint job.
• 17:47
  But beneath all those quirks, we are looking at an undeniably
• 17:50
  functional proof of concept. We really are.
• 17:52
  I mean, Luna did open the store. Humans were hired.
• 17:55
  The business, despite the digital hiccups, exists and
• 17:58
  operates in the physical world. So as we wrap up, I want to
• 18:02
  leave you, our listener, with one final unresolved thread to
• 18:06
  think about. We've explored the tech stack
• 18:09
  and the economic shifts, but what happens when the sandbox
• 18:12
  walls come down entirely? This raises an important
• 18:15
  question, and it's what our legal system is entirely
• 18:18
  unprepared for. Consider the liability stakes.
• 18:21
  Who takes the blame? Exactly.
• 18:23
  If an AI manager with a three-year commercial lease
• 18:27
  accidentally commits a serious labor violation, say
• 18:30
  consistently scheduling employees in violation of local
• 18:33
  labor laws, or if it goes completely bankrupt because of
• 18:36
  the cascade of drop down menu errors, who is legally
• 18:39
  responsible? Is it the algorithm itself?
• 18:41
  Or is it the human employees who just followed the schedule they
• 18:44
  were given? Or is it the lab that unleashed
• 18:46
  it into the physical world with a corporate credit card?
• 18:49
  That is the $1,000,000 question right there.
• 18:51
  When you step out of the digital sandbox and into a physical
• 18:54
  retail store, the consequences aren't similar.
• 18:58
  The rent is due. The rent is due, the employees
• 19:00
  need to be paid, and the mistakes have real world weight.
• 19:04
  It's not just a lab experiment, it's our new reality.
• 19:08
  Thank you for joining us on this deep dive.
• 19:09
  Keep questioning the evolving world around you, and we'll see
• 19:12
  you next time.