Latest / Connecting the Dots / Specialized AI Hardware, Contextual Workspace, and Enterprise Agents

Transcript

Specialized AI Hardware, Contextual Workspace, and Enterprise Agents · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarfle Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Imagine an organizational chart where
• 0:18
  Like 80 % of your workforce isn't actually human
• 0:21
  right which sounds crazy. It really does We're
• 0:24
  talking about a company where software programs
• 0:27
  are independently monitoring inventory spinning
• 0:30
  up marketing campaigns and You know actively
• 0:33
  hiring other software programs to help them finish
• 0:36
  their assigned tasks. It sounds totally like
• 0:37
  science fiction Yeah, exactly. But today we are
• 0:41
  looking at the exact physical hardware and the
• 0:44
  new corporate platforms making that a reality.
• 0:49
  Welcome to Connecting the Dots, the show where
• 0:51
  we break down the day's top tech news and figure
• 0:53
  out exactly how it impacts your daily life. I'm
• 0:56
  Alex. And I'm Morgan. And it is Wednesday, April
• 0:59
  22nd, 2026. We are broadcasting live from Snarfle
• 1:02
  Solutions Group right here in Sacramento, California.
• 1:05
  Beautiful Sacramento. Always. Before we unpack
• 1:08
  the morning's news, let's take a quick look out
• 1:10
  the window and at the numbers. If you are heading
• 1:12
  out in Sacramento today, you might want to keep
• 1:14
  an umbrella handy in your car. Yeah, definitely.
• 1:16
  We've got a chance of showers and thunderstorms
• 1:18
  early on, followed by a slight chance later in
• 1:20
  the afternoon. Overall, it's going to be a mild
• 1:23
  high of 64 degrees Fahrenheit with some very
• 1:26
  light winds blowing around six miles per hour.
• 1:28
  Honestly, it sounds like a perfect day to just
• 1:30
  stay inside and watch the markets, which gave
• 1:33
  us some serious turbulence yesterday. For sure.
• 1:36
  Yeah. Looking at Tuesday's closes, it was a bit
• 1:38
  of a rough ride on the floor. The Dow dropped
• 1:40
  538 .99 points, closing at 49 ,149 .38. And the
• 1:47
  S &P 500 also took a dip, shedding 58 .63 points
• 1:51
  to close out at 7 ,064 .01. Context is everything,
• 1:56
  though. I mean, both of those indices are still
• 1:58
  up for the week. So no one needs to hit the panic
• 2:01
  button just yet. And if you happen to be holding
• 2:03
  crypto, you were having a completely different
• 2:06
  Wednesday morning. Oh, totally different. Yeah,
• 2:07
  we saw a massive juxtaposition today with the
• 2:09
  Bitcoin surge. It jumped over 2 ,133 points to
• 2:13
  close at 78 ,474 .84. So depending on what is
• 2:17
  in your portfolio, you're either taking a deep
• 2:19
  breath or popping champagne. Very true. At least
• 2:22
  our local baseball teams brought some solid stability
• 2:25
  to the table on Tuesday. Thank goodness for that.
• 2:27
  Right. The SF Giants managed to beat the Dodgers
• 2:30
  3 -1, and down the bay, the Oakland A's took
• 2:33
  down the Mariners 5 -2. It is always a good morning
• 2:37
  when both local teams pull through with a win.
• 2:39
  Definitely. Well, it is time to unpack the Morningstar
• 2:43
  Tech Report. We've got a stack of stories today
• 2:45
  that really show how the fabric of our working
• 2:48
  lives is being rewoven by artificial intelligence.
• 2:51
  It's a huge shift. It is. But before we look
• 2:53
  closer at the physical engines driving all of
• 2:55
  this, let's hear from our sponsor. Today's episode
• 2:57
  is brought to you by Strong Coffee, the premium
• 3:00
  instant coffee that tastes like a traditional
• 3:02
  brew but is packed with 15 grams of protein,
• 3:05
  M4Cs, and adaptogens like ashwagandha. Get clean,
• 3:08
  sustained energy without the jitters. Use promo
• 3:11
  code SNARFL for 20 % off at strongcoffeecompany
• 3:14
  .com. Moving from the energy boost of our sponsor,
• 3:17
  we need to talk about the massive computational
• 3:19
  energy driving the next generation of tech. The
• 3:22
  raw power. Exactly. To understand where our daily
• 3:25
  workplace tools are heading, we first have to
• 3:28
  step into the engine room. And right now, the
• 3:30
  biggest engine room in the world belongs to Google.
• 3:33
  They just unveiled their new 8th generation custom
• 3:35
  tensor processor units. or TPUs. Yeah, and this
• 3:39
  represents a massive architectural shift in how
• 3:42
  artificial intelligence is physically built.
• 3:44
  For years, the entire industry relied on a unified
• 3:48
  approach. Like a one -size -fits -all kind of
• 3:50
  thing. Right. You had a chip, and that chip had
• 3:52
  to do two distinct jobs. First, it had to train
• 3:56
  the AI model, which means crunching decades of
• 3:58
  human knowledge for months at a time. Which takes
• 4:01
  incredible power. Exactly. Then, once the model
• 4:04
  was trained, that same type of chip had to run
• 4:06
  the model, answering user questions in real time.
• 4:09
  But with the rise of the agentic era, where AI
• 4:12
  programs aren't just giving you a text response,
• 4:13
  but actually executing multi -step workflows
• 4:16
  in continuous loop. Doing actual tasks for you.
• 4:18
  Yeah. That unified hardware was just hitting
• 4:20
  a wall. So Google is doing something completely
• 4:23
  unprecedented here. They are physically splitting
• 4:26
  their AI silicon into two separate purpose -built
• 4:30
  architectures. Which makes so much sense when
• 4:32
  you think about it. It really does. We have the
• 4:34
  TPU -8T, which is built exclusively for training
• 4:36
  and the TPU -8I built exclusively for inference.
• 4:41
  Let's unpack the training side first because
• 4:43
  the scale is it's almost impossible to wrap your
• 4:46
  head around. It really is. The TPU -8T is the
• 4:49
  workhorse. Google says it offers 2 .8 times better
• 4:53
  price to performance than their previous ironwood
• 4:55
  chip. A single super pod can network 9 ,600 liquid
• 4:59
  -cooled chips together, delivering 121 exaflops
• 5:03
  of compute power. Broadcom reportedly helped
• 5:05
  design this silicon. 121 exaflops? Yeah, let's
• 5:08
  pause on that word, exaflops. For the listener
• 5:10
  trying to visualize this, an exaflop isn't just
• 5:11
  a big number. It is a quintillion calculations
• 5:14
  per second. A billion billion. Wow. If every
• 5:17
  human on Earth did a math problem every second
• 5:19
  for four years, that is what this hardware does
• 5:22
  in one single second. It is just staggering raw
• 5:25
  power. And the reason Broadcom's involvement
• 5:28
  is so crucial comes down to traffic management.
• 5:31
  When you are training a frontier AI model, you
• 5:34
  are feeding it petabytes of data over several
• 5:36
  months. Right. The bottleneck isn't usually the
• 5:39
  chip itself, right? The bottleneck is how fast
• 5:42
  the chips can pass that data to each other. The
• 5:45
  TPU -8T solves this by using what's called a
• 5:48
  3D Taurus network topology. Which sounds like
• 5:51
  something out of a geometry textbook. It does,
• 5:53
  but functionally it's quite simple. Instead of
• 5:56
  connecting chips in a straight line or like a
• 5:58
  flat grid, imagine a multi -dimensional donut
• 6:01
  shape. Okay, a donut. Yeah. The chips are connected
• 6:04
  in a continuous loop in multiple directions,
• 6:06
  so there are no dead ends. Data can flow around
• 6:08
  the network without ever hitting a traffic jam,
• 6:10
  which is absolutely vital when you are pushing
• 6:12
  that much information around a super pod. Okay,
• 6:15
  so the TPU -AT is doing all the heavy lifting
• 6:17
  to build the model. But once the model is built,
• 6:20
  it gets handed over to the TPU8i, which is designed
• 6:23
  purely for inference. Right, the actual server.
• 6:25
  Right. Inference is the actual running of the
• 6:28
  model. When you ask a question, the inference
• 6:30
  chip generates the answer. Google claims the
• 6:32
  8i offers an 80 % performance per dollar improvement.
• 6:35
  That's a huge jump. Massive. It uses a new network
• 6:38
  called Boardfly and pairs 288 gigabytes of high
• 6:42
  bandwidth memory with 384 megabytes of on -chip
• 6:46
  SRAM. to prevent lag. MediaTek reportedly designed
• 6:50
  this one, but I want to focus on that on -chip
• 6:53
  memory. Why is having the memory physically on
• 6:56
  the chip such a game changer here? Well, it all
• 6:58
  comes down to latency. You know, the time it
• 7:00
  takes for a system to react. In standard computing,
• 7:03
  a processor usually has to ask the main memory
• 7:05
  for data, wait for that data to travel across
• 7:07
  the motherboard, and then use it. Like a commute.
• 7:10
  Exactly. It's a tiny delay, but it adds up. SRAM
• 7:13
  is memory built directly onto the brain of the
• 7:15
  processor itself. There is zero travel time.
• 7:17
  By keeping the active working set of an AI model
• 7:20
  entirely on the chip, you just eliminate the
• 7:22
  fetch time. In a world where an AI agent is reasoning
• 7:25
  through a problem and talking to three other
• 7:28
  software programs in real time to execute a task
• 7:30
  for you, latency is the ultimate enemy. you need
• 7:34
  lightning fast reflexes. So the on -chip memory
• 7:37
  provides that. It does. It's the difference between
• 7:39
  a chef having all their ingredients chopped and
• 7:42
  resting directly on their cutting board versus
• 7:44
  having to walk to the walk -in fridge every time
• 7:47
  they need a pinch of salt. Oh, that is a perfect
• 7:49
  analogy. Yeah. And it brings me perfectly to
• 7:51
  the analogy I was picturing for this whole TPU
• 7:54
  split. OK, let's hear it. For the last few years,
• 7:57
  the AI hardware industry has basically been operating
• 8:00
  out of one mega restaurant. The chefs are in
• 8:03
  the back, sweating, trying to invent brand new,
• 8:06
  incredibly complex recipes from scratch, while
• 8:09
  simultaneously trying to plate and serve thousands
• 8:12
  of hungry customers in the dining room at the
• 8:14
  exact same time. It's chaotic, it's expensive,
• 8:18
  and it causes delays. And splitting the architecture
• 8:21
  fundamentally separates the building. Exactly.
• 8:24
  The TPU -AT is the massive prep kitchen in the
• 8:27
  back. That's where the recipes, the models, are
• 8:30
  perfected over months of grueling work with massive
• 8:32
  ingredients. Then, once the recipe is absolutely
• 8:35
  perfect, they hand it over to the TPU ADI. The
• 8:38
  front -of -house. Yes, the ADI is the lightning
• 8:40
  -fast front -of -house staff. They aren't inventing
• 8:43
  the food. Their only job is serving the food
• 8:45
  to millions of customers simultaneously, as cheaply
• 8:48
  and quickly as possible. The prep kitchen focuses
• 8:51
  on sheer volume and throughput, while the front
• 8:53
  of house focuses purely on speed and cost per
• 8:56
  plate, or, you know, in the cloud business, cost
• 8:58
  per query. But here is where I have to push back
• 9:00
  a little on Google's strategy. Oh, OK. If Google's
• 9:03
  new front of house and back of house chips are
• 9:05
  so incredibly powerful and cost effective, why
• 9:08
  did they just sign a multi -billion dollar deal
• 9:10
  with Miramarati's Thinking Machines Lab to use
• 9:13
  Nvidia's GB300 chips? Ah, right. the Nvidia deal.
• 9:16
  Yeah, like Google is spending fortunes building
• 9:19
  custom silicon to take on Nvidia, but they're
• 9:22
  also spending billions buying Nvidia hardware
• 9:25
  to put inside their own data centers. If the
• 9:28
  TPU8 architecture is the future, this seems totally
• 9:32
  contradictory. I mean, it looks contradictory
• 9:34
  on the surface, but it really gets to the heart
• 9:36
  of the modern tech economy. You have to remember
• 9:39
  that Google is a cloud provider first. Right.
• 9:42
  NVIDIA still offers an incredibly powerful unified
• 9:45
  chip. And more importantly, millions of developers
• 9:48
  are deeply entrenched in NVIDIA's software ecosystem.
• 9:51
  They're used to it. Exactly. If a massive, lucrative
• 9:54
  customer like Thinking Machines Lab explicitly
• 9:56
  wants to train on NVIDIA chips... Google is going
• 9:59
  to provide them, because otherwise they risk
• 10:01
  losing that multi -billion dollar customer to
• 10:04
  Microsoft or Amazon. So they can't turn away
• 10:06
  a paying customer, even if the customer wants
• 10:08
  to use the competitor's oven, so to speak. Precisely.
• 10:11
  But the deeper strategy is about the long game
• 10:13
  of inference. As these AI models graduate from
• 10:16
  the laboratory into the real world where they
• 10:18
  are answering billions of queries a day for normal
• 10:20
  people, the unit economics completely flip. It
• 10:23
  becomes all about the serving cost. Yes. At a
• 10:26
  massive global scale, the cost per query matters
• 10:29
  more than almost anything else. By splitting
• 10:31
  their chips and optimizing the TPU 8i purely
• 10:34
  for low latency, low cost serving, Google is
• 10:37
  positioning itself to be the cheapest, fastest
• 10:39
  place to actually run the models, regardless
• 10:42
  of where they were built. Oh, I see. Even if
• 10:44
  a company trained their model on NVIDIA hardware,
• 10:47
  Google wants them to serve it to their customers
• 10:49
  using the TPU 8i. They're playing a very specific
• 10:52
  game of margins here. Okay, that makes a lot
• 10:54
  of sense. So Google is spending billions to optimize
• 10:58
  these 8i chips for instant cheap reflexes. What
• 11:01
  does that actually look like when you or I sit
• 11:03
  down at our laptops? That's the real question.
• 11:05
  How does that massive raw computational power
• 11:08
  change the 9 to 5 workflow of a busy professional
• 11:11
  or a parent trying to get out of the office on
• 11:13
  time? Let's transition to Google's newly announced
• 11:16
  Workspace Intelligence. Yeah, this is where the
• 11:19
  abstract world of exaflops and SRAM memory translates
• 11:22
  into tangible relief at your desk. Workspace
• 11:25
  Intelligence is a new background context layer
• 11:27
  running across Gmail, Docs, Slides, and Sheets.
• 11:30
  It's all connected now. Exactly. It is powered
• 11:33
  by advanced Gemini reasoning, making use of that
• 11:36
  fast inference hardware we just talked about.
• 11:37
  And the most fascinating part of this is how
• 11:39
  it breaks down what Google refers to as context
• 11:42
  walls. Historically, your software has been deeply
• 11:45
  fragmented. Incredibly fragmented. Your email
• 11:48
  was totally siloed from your spreadsheet, which
• 11:50
  had no idea what was in your fly deck. But Workspace
• 11:54
  Intelligence acts as a foundational layer of
• 11:56
  situational awareness. It understands the complex,
• 12:00
  semantic relationships between your active projects,
• 12:03
  your collaborators, and your broader company
• 12:05
  knowledge. Right. It actively knows what projects
• 12:08
  you are assigned to. It knows who your direct
• 12:10
  reports are. It knows the formatting templates
• 12:13
  your company prefers for quarterly reviews. It
• 12:15
  just knows. Yeah. And it surfaces all of that
• 12:18
  awareness through features like Ask Gemini in
• 12:21
  Google Chat. which essentially becomes a unified
• 12:23
  command line for your entire professional life.
• 12:26
  You can just type into chat, generate a slide
• 12:28
  deck for the Q3 marketing review, and instead
• 12:31
  of giving you a generic outline, it will pull
• 12:34
  the latest conversational data from your recent
• 12:36
  emails, grab the exact revenue numbers from your
• 12:39
  sheets, apply your company's specific visual
• 12:41
  style and logo, and drop the finished, highly
• 12:44
  specific presentation right into the chat. It's
• 12:47
  amazing. It can instantly triage your inbox to
• 12:49
  find the three emails that actually matter, or
• 12:52
  negotiate meeting times that work for five different
• 12:54
  people across three time zones. And it goes even
• 12:57
  further by learning your unique work style, your
• 12:59
  voice, and your specific formatting preferences,
• 13:02
  ensuring that any drafted outputs sound authentically
• 13:05
  like you. And this is exactly where I have to
• 13:08
  throw a flag on the play. Uh -oh. What's the
• 13:10
  flag? This concept has a major uncanny valley
• 13:13
  aspect to it. The idea of an AI silently studying
• 13:16
  years of my email so it can mimic my exact voice.
• 13:19
  I think a lot of people will find that deeply
• 13:21
  unsettling. It is a little eerie. Is it going
• 13:23
  to sound like a stiff robot pretending to be
• 13:26
  a human? How does a piece of software actually
• 13:30
  know my specific quirks? And frankly, do I even
• 13:33
  want it to? Having a machine impersonate my personality
• 13:37
  to my coworkers feels like a boundary violation.
• 13:40
  That is a completely understandable reaction.
• 13:43
  The idea of digital mimicry makes a lot of us
• 13:45
  uncomfortable because it feels, well, invasive.
• 13:49
  But let's look at the real world application
• 13:51
  here and frame it a slightly different way. Yeah,
• 13:53
  I'm listening. Think of it less as a robot stealing
• 13:56
  your identity and more like having a brilliant,
• 13:59
  highly observant executive assistant who has
• 14:02
  worked 10 feet away from you for five years.
• 14:04
  An assistant who knows my habits. Exactly. A
• 14:07
  great executive assistant knows that you always
• 14:09
  use a specific blue font for your headers. They
• 14:11
  know you despise long paragraphs and prefer bullet
• 14:14
  points. Guilty as charged. Right. They know that
• 14:16
  when you write a weekly update to your boss,
• 14:18
  you keep the tone strictly analytical and formal,
• 14:21
  but when you email the creative design team,
• 14:23
  a tone is much more casual and collaborative.
• 14:26
  Workspace intelligence isn't trying to capture
• 14:28
  your soul. Well, that's a relief. It is simply
• 14:31
  trying to save you the cognitive load of formatting
• 14:34
  and context switching. it's the ultimate time
• 14:37
  saver. Imagine being a busy professional who
• 14:40
  just wants to leave work at 5 p .m. to see their
• 14:42
  family instead of spending 45 agonizing minutes
• 14:46
  reformatting a spreadsheet and drafting six generic
• 14:50
  update emails. So it's less invasion of the body
• 14:53
  snatchers and more like autocomplete on steroids.
• 14:56
  Yes. perfect way to put it. It captures the repetitive
• 14:58
  mechanics of how I work so I don't have to manually
• 15:01
  execute those mundane steps over and over again.
• 15:04
  If I always bold the top line of a memo it just
• 15:07
  does it. It removes the friction from your digital
• 15:09
  workflow. It handles the structural busy work
• 15:12
  so you can reserve your mental energy for actual
• 15:14
  decision making and creative problem solving.
• 15:16
  That does sound significantly more appealing
• 15:18
  when you put it like that. So we've got the physical
• 15:20
  silicon providing the raw power. We've got the
• 15:23
  personal assistant layer helping you as an individual.
• 15:26
  But what happens when an entire enterprise decides
• 15:29
  to scale this up? The fleet management side.
• 15:32
  Exactly. If everyone has a personal assistant
• 15:35
  and the company decides to deploy thousands of
• 15:38
  its own autonomous AI agents, that's a lot of
• 15:41
  invisible hands in the cookie jar. We move from
• 15:44
  personal productivity to massive corporate fleet
• 15:47
  management. And that brings us to the launch
• 15:49
  of the Gemini Enterprise Agent platform. This
• 15:51
  represents a major evolution from Google's previous
• 15:54
  Vertex AI tools. It is a completely revamped
• 15:57
  developer platform designed to manage the full
• 15:59
  lifecycle of AI agent fleets. And fleets is the
• 16:03
  critical word here. We're not just talking about
• 16:05
  one chatbot. Not at all. We aren't talking about
• 16:07
  deploying a single customer service chatbot.
• 16:10
  We are talking about an interconnected workforce
• 16:12
  of software programs. The platform allows developers
• 16:15
  to build, tune, and manage over 200 different
• 16:18
  models. You can use Google's own Gemini 3 .1
• 16:21
  Pro, NanoBanana 2, open models like Gemma, or
• 16:24
  even direct competitor models like Anthropix
• 16:26
  Opus 4 .7. Which is great for flexibility. Yeah,
• 16:29
  but the truly wild part is how developers can
• 16:32
  structure these agents into subnetworks. where
• 16:34
  agents autonomously pass tasks to one another
• 16:36
  without human intervention. The demo they showed
• 16:39
  at Cloud Next was a perfect example of this.
• 16:41
  They simulated a fictional furniture company
• 16:44
  where a team of software agents was working together.
• 16:47
  An inventory agent was continuously monitoring
• 16:50
  warehouse levels. Right. It noticed a massive
• 16:53
  surplus of a specific type of office chair. Without
• 16:57
  being prompted by a human, it autonomously delegated
• 17:00
  a task to a separate marketing agent, instructing
• 17:03
  it to spin up a targeted email promotional campaign
• 17:07
  to move that specific chair inventory. Wow. They
• 17:11
  collaborated exactly the way a human logistics
• 17:13
  team and marketing team would. Which is incredible
• 17:16
  for efficiency, but it also introduces a massive
• 17:18
  chaos factor. Think about the risk here. You
• 17:21
  have autonomous agents delegating tasks to other
• 17:24
  autonomous agents, pulling sensitive data from
• 17:26
  company spreadsheets, drafting external marketing
• 17:28
  emails. That's a lot of moving parts. Without
• 17:30
  human bottlenecks to double check the work, are
• 17:33
  we just one glitch away from a catastrophic security
• 17:35
  breach? What happens if a rogue AI accidentally
• 17:38
  attaches a highly confidential financial report
• 17:41
  to a client email because it misunderstood a
• 17:43
  prompt? How does an IT department sleep at night?
• 17:46
  Honestly, security and governance are arguably
• 17:49
  the biggest hurdles to enterprise AI adoption
• 17:52
  right now, because the exact scenario you just
• 17:55
  described is every chief information officer's
• 17:58
  worst nightmare. Right. Google knows this, which
• 18:00
  is why they built something called agent identity
• 18:02
  into the very core of this new platform. Cryptographic
• 18:06
  IDs sound like a highly technical way of saying
• 18:09
  a password, but at an enterprise scale, I imagine
• 18:12
  it has to be more robust than that to prevent.
• 18:14
  a rogue scenario. It is much more robust. The
• 18:17
  best way to understand agent identity is to think
• 18:19
  about how a secure office building operates.
• 18:22
  It is exactly like giving every single digital
• 18:25
  employee a corporate key card. When a human employee
• 18:28
  swipes their key card in a corporate headquarters,
• 18:30
  the IT department and security team know exactly
• 18:33
  who entered the building, which elevator they
• 18:35
  took, which floor they went to, and whether or
• 18:37
  not they tried to access a restricted server
• 18:39
  room. Agent identity works the exact same way
• 18:42
  for the AI fleet. so every action the AI takes
• 18:45
  leaves a digital footprint tied to its specific
• 18:47
  ID. every single action. It standardizes governance
• 18:50
  through a single control plane. The IT department
• 18:54
  has full real -time visibility. They can see
• 18:57
  exactly which databases an agent is touching,
• 19:00
  which other agents it is communicating with,
• 19:02
  and what external actions it is attempting to
• 19:04
  take. And if an agent suddenly tries to access
• 19:06
  the payroll database when it's only supposed
• 19:09
  to be writing marketing copy, the system flags
• 19:11
  it. I imagine IT can revoke that access instantly.
• 19:15
  Instantly. It provides the same level of strict
• 19:18
  line -by -line auditability that you would expect
• 19:20
  in a payroll system or quarterly financial reporting
• 19:23
  software. It eliminates the blind spots. That's
• 19:26
  reassuring. Furthermore, to prevent the rogue
• 19:29
  scenarios from ever happening in the first place,
• 19:31
  they introduced a tool called Agent Simulation.
• 19:34
  So developers can put these agents in a sandbox
• 19:36
  before giving them real power. Exactly. Before
• 19:39
  you unleash a new marketing agent into your live
• 19:41
  systems, you stress test it against complex,
• 19:44
  real -world scenarios in a heavily monitored
• 19:46
  simulation. You watch how the inventory agent
• 19:49
  and the marketing agent interact. You try to
• 19:50
  break them. You find the flaws first. Right.
• 19:53
  You ensure they follow the rules before you ever
• 19:55
  give them the actual keys to the company email
• 19:57
  server. It ensures that the transition to a hybrid
• 20:00
  human and AI workforce is secure, heavily monitored,
• 20:04
  and fully auditable. That key card analogy really
• 20:07
  clarifies the mechanism. It isn't a digital free
• 20:10
  -for -all. It's a highly regulated assembly line
• 20:13
  of software. Well, we have covered a massive
• 20:15
  amount of ground today. Tracing the path from
• 20:18
  the physical hardware all the way up to corporate
• 20:20
  governance. But before we wrap up and connect
• 20:23
  all these dots one last time, let's hear from
• 20:25
  our final sponsor. Before we wrap up, a quick
• 20:27
  thanks to Old Glory, an iconic music and sports
• 20:30
  fan merch store. Perfect for the techie in your
• 20:33
  life. Use promo code SNARFL for 15 % off at oldglory
• 20:37
  .com. Let's take a step back and look at the
• 20:39
  entire journey we just took. We started all the
• 20:41
  way down in the engine room with Google's physical
• 20:43
  silicon, the new TPU -8T and 8i chips. the foundation
• 20:48
  of it all. Exactly. The architecture of AI is
• 20:51
  splitting into two purpose -built lanes. You
• 20:53
  have one massive prep kitchen dedicated to training
• 20:56
  the models with zero traffic jams, and one ultra
• 20:59
  -fast front -of -house staff dedicated to serving
• 21:02
  the models to keep inference costs as low as
• 21:05
  possible. And that raw, low -latency compute
• 21:08
  power is the foundation that makes tools like
• 21:10
  workspace intelligence possible. It provides
• 21:13
  the sheer speed necessary for an AI to act as
• 21:16
  a seamless executive assistant right on your
• 21:18
  laptop. Understanding the semantic context of
• 21:21
  your daily work. Right, remembering your formatting
• 21:23
  quirks and saving you valuable hours of repetitive
• 21:26
  busy work every single week. But once you scale
• 21:28
  that personal assistant concept up to a massive
• 21:30
  corporate level, you require serious oversight.
• 21:34
  That is where the Gemini Enterprise agent platform
• 21:36
  comes in. The management layer. Right. It acts
• 21:38
  as the IT control plane, using cryptographic
• 21:40
  key cards to ensure that massive fleets of autonomous
• 21:43
  software agents can collaborate securely without
• 21:46
  causing digital chaos or compromising sensitive
• 21:48
  data. It really is a complete top -to -bottom
• 21:51
  stack, from the physical silicon, to the desktop
• 21:54
  software, to the enterprise security. It is a
• 21:56
  profound paradigm shift in how human labor and
• 21:59
  computational power intersect. It is. And it
• 22:02
  leaves me with one final thought to mull over.
• 22:05
  As these AI agents become increasingly sophisticated,
• 22:09
  as they become capable of monitoring global inventory,
• 22:13
  designing complex marketing campaigns, and autonomously
• 22:16
  delegating tasks to other AI agents within these
• 22:19
  sub -networks. It gets wild. It does. At what
• 22:22
  point does the organizational chart of a Fortune
• 22:24
  500 company feature more software programs than
• 22:27
  human beings? And when you are managing a team
• 22:30
  that is 80 % cryptographic IDs and 20 % human
• 22:34
  staff, How does that fundamentally change the
• 22:37
  very meaning of being a manager? That is an incredibly
• 22:40
  complex question, and it is a reality we will
• 22:42
  all be navigating very, very soon. Something
• 22:44
  to think about as you head into the office today
• 22:46
  and swipe your own key card. We're here to help
• 22:49
  it snarfle, so reach out on our website if you
• 22:51
  have any more questions. That's a wrap on today's
• 22:57
  top stories. Remember, tech doesn't have to be
• 22:59
  complicated, and we're here to help connect the
• 23:02
  dots. If you've got questions or want to learn
• 23:04
  more, visit us at snarful .com. Thanks for listening,
• 23:08
  and we'll catch you next time.