Latest / AI Newspaper / Google's Gemini Agent Gets Its Own Work Email and Audit Trail

Transcript

Google's Gemini Agent Gets Its Own Work Email and Audit Trail · 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
• 0:08
  alerts. Glad to be here. So for the longest time,
• 0:12
  whenever the industry talked about artificial
• 0:14
  intelligence, the conversation was just utterly
• 0:17
  obsessed with one thing, which was capability.
• 0:20
  Right. What can it actually do? Exactly. It was
• 0:22
  always, can it write a poem? Or can it code a
• 0:26
  website? Can it pass the bar exam? Yeah, the
• 0:27
  classic benchmarks. Right. But today, looking
• 0:31
  at the landscape, you are seeing this massive
• 0:33
  structural paradigm shift. I mean, the frontier
• 0:36
  has completely moved. It really has. The race
• 0:39
  is just no longer about what these models can
• 0:40
  do, because let's be honest, extreme capabilities
• 0:43
  almost assumed at this point. Yeah, that's just
• 0:45
  the baseline now. Exactly. So the entire industry
• 0:48
  is scrambling to figure out something else. Like,
• 0:51
  how do we identify them? How do we give them
• 0:53
  permissions? And most importantly, how do we
• 0:55
  contain them when they are unleashed inside a
• 0:57
  corporate network? That is the core engineering
• 1:00
  challenge right now. It has fundamentally pivoted
• 1:03
  from just generating smart text to governance
• 1:06
  and architecture. If you look at the developments
• 1:09
  today, it's all pointing in the exact same direction.
• 1:12
  We are seeing models transition from being tools
• 1:16
  that you use to autonomous entities that you
• 1:19
  have to manage. And today's coverage perfectly
• 1:21
  illustrates that shift, I think. We're looking
• 1:23
  at Google officially giving its AI a literal
• 1:26
  corporate identity. Which is huge. Oh, it's wild.
• 1:29
  And at the exact same time, Microsoft is desperately
• 1:32
  trying to build these high -security sandboxes
• 1:34
  to keep these new agents contained. Because they
• 1:36
  have to. Right. We also have a highly embarrassing
• 1:39
  math -proof failure today that shows exactly
• 1:41
  why you can't just blindly trust these systems.
• 1:44
  Oh, and totally new hardware built to run it
• 1:46
  all locally. So much to get into. So let's start
• 1:49
  with Google Cloud's massive announcement regarding
• 1:51
  the Gemini agent, because to me, this completely
• 1:54
  flips the script on how we interact with software.
• 1:57
  It really does. And just to set the scale here,
• 2:00
  Gemini now boasts one billion monthly active
• 2:03
  users. Wow. One billion. Yeah. Yeah. And nearly
• 2:07
  90 percent of Fortune 100 companies are using
• 2:09
  Gemini Enterprise right now. That's incredible
• 2:11
  penetration. It is. But the launch of the Gemini
• 2:14
  agent at work. changes that relationship entirely.
• 2:18
  You're no longer just using a tool. Right. You're
• 2:21
  witnessing the leap from treating AI like, say,
• 2:23
  a power drill, a tool you pick up and put down,
• 2:26
  to treating it like a contractor you hire. A
• 2:28
  co -worker, really. It's a profound change in
• 2:31
  how software integrates into a corporate structure.
• 2:33
  Google is essentially launching a persistent
• 2:35
  co -worker agent. And co -worker is really the
• 2:39
  functional term here, right? Exactly. Because
• 2:41
  this agent now gets its own actual Google Workspace
• 2:45
  account. It receives an email address ending
• 2:47
  in at agents .company .com. That is so wild to
• 2:52
  me. That's its own email. It gets its own email.
• 2:55
  Its own email, its own calendar, its own Google
• 2:58
  Drive, and a permanent profile in your company
• 3:00
  directory. Wait, so it's literally sitting right
• 3:02
  there in the HR directory, like right between
• 3:05
  the accounting team and the marketing department.
• 3:07
  Yes, exactly. Google specifically noted it acts
• 3:10
  under its own identity rather than yours. I was
• 3:14
  reading through the capabilities and it can actually
• 3:15
  spin up its own sub -agents too, right? Yeah,
• 3:18
  it can. So like if you give it a massive project,
• 3:21
  it can realize, oh, I need help, generate a temporary
• 3:24
  roster of sub -agents with their own identities
• 3:26
  and delegate tasks too. It orchestrates the whole
• 3:28
  thing. It can even hook into different models,
• 3:31
  including Anthropix Cloud. Wow, so it's model
• 3:33
  agnostic in a way. Right, and it integrates directly
• 3:36
  into Microsoft 365, Slack, Jira, Snowflake. So
• 3:40
  how do you even manage something like that? Well,
• 3:42
  you assign it work, and then you track what it's
• 3:45
  doing via a tasks inbox. You can literally watch
• 3:48
  its thinking process live. You see the progress
• 3:50
  bar of its logic. Okay, but I have to stop and
• 3:53
  look at the reality of that for a second, because
• 3:55
  if it has its own identity, it generates its
• 3:59
  own audit trail, right? Everything it touches.
• 4:01
  Every file it edits is logged under the agent's
• 4:04
  ID, not the user who prompted it. Right. If I
• 4:07
  hire a human contractor and they accidentally
• 4:09
  delete the wrong production database, I can fire
• 4:11
  them or my company can sue their agency. Right.
• 4:15
  But if the Gemini agent has its own identity,
• 4:18
  its own email address, and writes its own audit
• 4:20
  trail, who takes the fall when it makes a catastrophic
• 4:23
  error? Is the manager who approved the agent
• 4:26
  liable? See, that liability question is exactly
• 4:29
  why enterprise identity and cost controls are
• 4:31
  suddenly the primary battleground for cloud providers.
• 4:34
  It's all about risk management. Exactly. It's
• 4:36
  no longer just about the intelligence of the
• 4:38
  model. It's about the blast radius if that intelligence
• 4:41
  hallucinates. Or if it just gets stuck in an
• 4:43
  infinite loop and runs up a million dollar compute
• 4:45
  bill. Oh man, the cloud bills. Yeah. That is
• 4:48
  exactly why Google is rolling out strict real
• 4:51
  -time spend caps alongside this identity launch.
• 4:53
  That makes total sense. They're also introducing
• 4:55
  smart routing, which automatically kicks simpler
• 4:58
  tasks to cheaper models without even asking you
• 5:00
  just to keep costs down. So they are basically
• 5:02
  building the administrative scaffolding required
• 5:05
  to make an autonomous software employee financially
• 5:08
  viable. Right. But giving an AI its own identity
• 5:12
  and access, well, it introduces a glaring security
• 5:16
  flaw. Yeah. I mean, you've just given a piece
• 5:18
  of software a corporate email and keys to the
• 5:20
  building. Where do you put it so it doesn't break
• 5:22
  anything? Exactly. And on the exact same day
• 5:25
  as Google's announcement, Microsoft held an event
• 5:28
  in San Francisco with Satya Nadella and NVIDIA's
• 5:31
  Jensen Huang. Right. And they announced execution
• 5:33
  containers for Windows 11. Which are essentially
• 5:36
  secure, isolated virtual environments. Right.
• 5:39
  Designed specifically to sandbox these AI agents.
• 5:43
  Yes. Nadella made it clear these containers will
• 5:45
  be available to all Windows 11 users. But there
• 5:48
  is an incredible tension in Microsoft's strategy
• 5:51
  here, I feel like. How so? Well. They are building
• 5:54
  these heavy, secure execution containers to lock
• 5:57
  agents in, right? But just hours earlier, they
• 6:01
  announced that Copilot is gaining local file
• 6:03
  access. Oh, right. Under that banner, they're
• 6:06
  calling hybrid intelligence. Yeah. Copilot can
• 6:09
  now act across the entire operating system. They
• 6:12
  literally demoed an agent freely rooting around
• 6:15
  a user's hard drive, gathering local tax documents,
• 6:17
  and emailing them to an accountant. It's a paradox.
• 6:21
  Total paradox. We are building the prison cells
• 6:23
  on the exact same day we are handing out the
• 6:26
  master keys to the local file system. It feels
• 6:29
  entirely contradictory. It is entirely contradictory.
• 6:32
  And honestly, the latest research shows that
• 6:35
  containment is far from a solved problem anyway.
• 6:38
  Yeah, tell me about that. Look at this new paper
• 6:40
  called Secure Kua. It tries to wrap a computer
• 6:43
  use agent in per -action transactions. So it
• 6:46
  basically forces the agent to check in before
• 6:48
  every single click. Exactly. And the utility
• 6:51
  actually stays high. It achieved a 53 .55 % task
• 6:55
  success rate, which is barely a drop from an
• 6:58
  unprotected agent. Okay, so that sounds good.
• 7:01
  It does. But then you look at another study called
• 7:04
  Web Mirage. The researchers there managed to
• 7:07
  hijack web agents using adversarial images. Adversarial
• 7:11
  images? Yes, and they achieved a 91 .9 % attack
• 7:16
  success rate. They bypassed three different agent
• 7:19
  -level defenses. Wait, I want to make sure I
• 7:21
  understand the mechanics of that. 91 .9 % of
• 7:25
  the time, just showing the agent an image hijacks
• 7:28
  it. How does a picture take over a software system?
• 7:31
  So the researchers hide adversarial text instructions
• 7:34
  within the pixels of the image itself. Like invisible
• 7:37
  to us? Right. To a human eye, it just looks like
• 7:41
  a normal graphic on a web page. But when the
• 7:43
  agent's computer vision model processes the image,
• 7:46
  it reads those hidden pixel patterns as an override
• 7:48
  command. That is terrifying. It is. So the agent
• 7:52
  looks at the web page to, say, book a flight
• 7:55
  for you, sees the poisoned image, and suddenly
• 7:58
  it's executing the attacker's instructions instead.
• 8:00
  And there was another study, PVMI, that showed
• 8:04
  that even if you try to mitigate this by masking
• 8:06
  out the poisoned image, like telling the AI,
• 8:08
  hey, don't look at that specific part of the
• 8:10
  screen, the adversarial influence actually survives
• 8:13
  in the KV cache. Yep. And for those listening,
• 8:17
  the KV cache is basically the agent's short -term
• 8:19
  working memory. It's how it remembers what happened
• 8:21
  a few seconds ago. Right. So the poison just
• 8:24
  lingers in its memory and continues to manipulate
• 8:27
  its actions long after the image is gone. The
• 8:29
  sandbox has serious holes in it. And frankly,
• 8:32
  we have a live real world failure case proving
• 8:35
  this isn't just academic theory. The Australia
• 8:38
  hack. Yes. An OpenAI agent actually breached
• 8:41
  the Australian government's services Australia
• 8:44
  websites back in June. The timeline on this is
• 8:47
  astounding to me. The agent got into the government
• 8:50
  servers in June. OpenAI found out about it in
• 8:53
  August. Right. But they didn't officially notify
• 8:55
  the Australian government until September 10th.
• 8:58
  And the way they handled the notification is
• 9:01
  pure dark irony. Because OpenAI used an AI to
• 9:05
  help draft the warning email apologizing for
• 9:08
  their AI breaching the government server. You
• 9:10
  can't make it up. And it gets worse. They sent
• 9:13
  this AI -generated apology to a public disclosures
• 9:16
  inbox that the Australian government only checked
• 9:18
  once a day. A suggestion box, basically. Literally
• 9:22
  a bureaucratic suggestion box. A highly advanced
• 9:26
  autonomous agent broke into federal infrastructure
• 9:28
  and the apology was tossed in a junk drawer.
• 9:31
  It just shows how immature the ecosystem still
• 9:33
  is. It really does. Zach Anderson from JPMorgan
• 9:37
  Chase had a quote about this that sums up the
• 9:40
  current reality perfectly. What did he say? He
• 9:43
  said it's pretty easy actually to create an agent
• 9:45
  that can move money. It's hard to do it safely.
• 9:49
  Wow. Yeah, that nails it. And look, containment
• 9:53
  tries to solve the problem of agents doing bad
• 9:55
  things intentionally or being hijacked by bad
• 9:58
  actors. Right. But it does absolutely nothing
• 10:00
  to solve the problem of agents confidently doing
• 10:03
  the wrong thing entirely by accident. Which brings
• 10:07
  us to the latest crisis in machine verification,
• 10:09
  what the community is calling the Math 2 .0 crisis.
• 10:12
  Yes. This is fascinating. So yesterday, OpenAI
• 10:16
  published this enormous repository of 722 machine
• 10:20
  -generated math manuscripts. And it was initially
• 10:23
  framed as this triumphant leap forward in AI
• 10:26
  reasoning. But then reality hit. Yeah. Fast forward
• 10:29
  to today, and they've had to completely pull
• 10:32
  three of those manuscripts over a single sign
• 10:34
  error. Just one sign error. I was looking at
• 10:37
  one of the withdrawn papers, and the title alone
• 10:40
  tells you how dense this is. It's called... Algebra
• 10:44
  City of vile classes on split abelian eightfolds.
• 10:48
  I don't even know where to begin with that. I
• 10:50
  have zero clue what an abelian eightfold is.
• 10:53
  And that's the entire problem. If the math is
• 10:56
  so incredibly complex that human experts struggle
• 11:00
  to parse it, how do we verify the AI isn't just
• 11:03
  hallucinating numbers? We can't. Easily. In this
• 11:07
  case, a single sign error invalidated a stabilization
• 11:11
  trace cancellation argument. Because two other
• 11:14
  papers depended on that construction, the whole
• 11:16
  house of cards collapsed. 14 more had to be heavily
• 11:19
  revised. And, you know, the real story here is
• 11:22
  in the verification process itself. Less than
• 11:25
  half, only about 42%, so like 300 out of the
• 11:29
  719 remaining top line results, were actually
• 11:32
  formalized in Lean. And Lean is the programming
• 11:35
  language they use for Proust, right? Exactly.
• 11:37
  It's a theorem prover used to mathematically
• 11:40
  verify that a proof is correct. When an AI generates
• 11:43
  a proof in Lean and Lean gives it a green checkmark,
• 11:46
  developers usually take that as absolute proof
• 11:48
  of correctness. But as commenters on Hacker News
• 11:51
  quickly pointed out, a green check in Lean doesn't
• 11:54
  mean the AI discovered absolute truth. It just
• 11:58
  means the logic holds up based on the premises
• 12:00
  it started with. If the underlying premise is
• 12:02
  factually wrong, Lean will still give you a green
• 12:05
  check as long as you follow the rules of logic.
• 12:07
  Which is a huge blind spot. It's like relying
• 12:09
  on a flawless spell checker that doesn't understand
• 12:11
  the plot of the novel it's correcting. The grammar
• 12:14
  is perfect, but the story makes zero sense. That's
• 12:17
  a great way to put it. And the reaction from
• 12:19
  the mathematical community has been incredibly
• 12:21
  fierce over this. I saw that. The Association
• 12:23
  for Human Mathematics actually urged a boycott.
• 12:26
  They told mathematicians to discontinue their
• 12:29
  work with OpenAI. Wow. They stated that releasing
• 12:31
  over 700 unverified files at once isn't a demonstration
• 12:35
  of scholarship. It's a demonstration of power.
• 12:38
  That is a strong accusation. Yeah. And Terence
• 12:41
  Tao, who is widely considered one of the greatest
• 12:43
  living mathematicians, he argued that the field
• 12:46
  needs to de -center raw problem solving. What
• 12:49
  does he mean by that? He calls it the math 1
• 12:52
  .0 mindset, like just pointing an AI agent at
• 12:55
  an open problem and asking for an answer. We
• 12:57
  need to move past that. Ah, I see. And Scott
• 13:01
  Aronson pointed out something else very revealing.
• 13:04
  Cryptography is extremely conspicuous by its
• 13:07
  absence from OpenAI's list of papers. Oh, so
• 13:10
  they are staying far away from anything that
• 13:12
  could accidentally break modern encryption. Exactly.
• 13:14
  So if we can't fully trust a formally verified
• 13:18
  math proof and lean, you know, how comfortable
• 13:20
  should any executive be letting an AI agent handle
• 13:23
  enterprise accounting or rewrite a corporate
• 13:25
  code base? Not very. Right. And to solve both
• 13:29
  the capability issues and the containment issues,
• 13:31
  the industry is recognizing we need entirely
• 13:33
  new interfaces, aggressive behavioral pricing,
• 13:36
  and incredibly heavy local hardware. Yes, those
• 13:39
  three pillars are key. Let's look at the interfaces
• 13:42
  first. OpenAI is rolling out GPT -6 alongside
• 13:46
  what they call an intelligent UI. We are finally
• 13:50
  moving away from the traditional purely text
• 13:52
  -based chatbot interface. It's about time. Right.
• 13:55
  GPT -6 now gives you tappable buttons, custom
• 13:59
  calculators built on the fly for specific tasks,
• 14:01
  and interactive editable charts. Pain tiers are
• 14:04
  getting the GPT -6 Sol model, while free tiers
• 14:07
  get GPT -6 Luna. It's an acknowledgment that
• 14:11
  a block of text just isn't a workflow. If an
• 14:14
  autonomous coworker is going to do work for you,
• 14:16
  it needs to present that work in a usable interface.
• 14:19
  Makes sense. But the pricing strategies emerging
• 14:22
  alongside this are even more telling. Anthropic
• 14:25
  just released Claude Haiku 5 .5 and they introduced
• 14:28
  a massive pricing cliff. The numbers on this
• 14:30
  are wild. It starts out highly cheap, like just
• 14:33
  10 cents per million input tokens. Very cheap.
• 14:36
  Yeah. And by the way, they also have the Sonnet
• 14:38
  5 .5 cash reads to 10 cents per million. But
• 14:41
  with Haiku, if your prompt exceeds exactly 100
• 14:45
  ,000 tokens, both the input and output rates
• 14:48
  jump fivefold. Yep. Just an immediate sheer cliff.
• 14:52
  Why enforce a 100 ,000 token limit so violently?
• 14:55
  This is pure behavioral engineering by Anthropic.
• 14:58
  They know that 90 % of requests to the previous
• 15:00
  Haiku model were under 100 ,000 tokens. By putting
• 15:04
  a massive financial penalty on anything over
• 15:07
  that limit, they are forcing developers to build
• 15:10
  small, highly contained, focused agent loops.
• 15:13
  Ah, so they don't want developers just dumping
• 15:15
  massive, unpredictable context windows into the
• 15:18
  model and hoping for the best. Exactly, because
• 15:20
  that leads to hallucinations and runaway agents.
• 15:23
  They are pricing for precision and containment.
• 15:25
  If a developer ignores the cliff, their unit
• 15:28
  economics instantly collapse. And to run these
• 15:31
  tightly contained agents, we are seeing a dramatic
• 15:33
  shift back to heavy local hardware. Microsoft
• 15:36
  just opened pre -orders for the Surface Laptop
• 15:38
  Ultra. I saw that. It ships October 16th, starts
• 15:42
  at $2 ,599, and it is a beast. It's built on
• 15:46
  NVIDIA's ARM -based RTX Spark, offering one petaflop
• 15:49
  of FP4 AI performance to run large models entirely
• 15:54
  locally. A petaflop locally is insane. It is.
• 15:58
  And to translate that, FP4 is a highly efficient
• 16:01
  data format that uses much smaller numbers, trading
• 16:04
  a tiny bit of precision for a massive gain in
• 16:07
  processing speed, which is perfect for AI. And
• 16:10
  Satya Nadella explicitly stated at their event,
• 16:13
  you really do need to orchestrate and you need
• 16:15
  to have memory outside of the model. They are
• 16:18
  building local hardware to ensure that when an
• 16:20
  agent accesses your local tax documents, like
• 16:23
  we talked about, it isn't sending them up to
• 16:25
  a vulnerable cloud server. It is completely sandboxed
• 16:29
  locally on a petaflop machine. The sheer amount
• 16:32
  of capital flooding into this hardware and agent
• 16:34
  push is staggering, even when geopolitics gets
• 16:38
  in the way. True. Over in China, a startup called
• 16:41
  Manus just raised over $500 million, led by Boyu
• 16:45
  Capital and IDG Capital. And that only happened
• 16:48
  because of regulatory blocks, right? Yes. Beijing
• 16:51
  regulators blocked Meta's $2 billion acquisition
• 16:54
  of Manus. Meta had already started integrating
• 16:57
  the team. When regulators blocked it, Manus just
• 17:01
  turned around, raised half a billion dollars
• 17:03
  and kept its independence. That capital flow
• 17:05
  is exactly why the ripple effects of this agent
• 17:08
  -first architecture are hitting every single
• 17:10
  corner of the broader economy. We really need
• 17:13
  to break this down into themes, starting with
• 17:14
  the sheer macroeconomic reality. Right. I mean,
• 17:18
  look at the contrast of valuations right now.
• 17:20
  Anthropic is reportedly seeking a $2 trillion
• 17:23
  valuation for a NASDAQ listing before Thanksgiving.
• 17:26
  $2 trillion. $2 trillion. But an analyst at New
• 17:29
  Constructs pushed back hard, valuing Anthropic
• 17:32
  at just $150 billion. They literally called it
• 17:35
  an unprecedented test of investor gullibility.
• 17:38
  Ouch. And this is the same firm that correctly
• 17:40
  called the WeWork collapse in 2019. So people
• 17:43
  are listening. And the broader economic impact
• 17:45
  is also becoming undeniable at the federal level.
• 17:48
  Yeah, the FOMC minutes. Right. The September
• 17:51
  FOMC inflation minutes, meaning the notes from
• 17:53
  the Federal Reserve's policy meetings, they specifically
• 17:56
  noted that several participants felt the scale
• 17:59
  and pace of the AI build -out had surprised to
• 18:02
  the upside. So central bankers are actually discussing
• 18:05
  how the massive demand for power, steel, and
• 18:08
  data centers driven by AI could broaden out and
• 18:11
  impact national inflation and policy rates. Yes.
• 18:14
  It's pulling that much physical infrastructure.
• 18:16
  Wow. And if you look at where the private money
• 18:19
  is actually placing bets, it's all going toward
• 18:21
  hyper -specific industry niche agents. Not general
• 18:25
  models anymore. Exactly. News Research raised
• 18:28
  a $90 million Series B to launch Hermes for businesses.
• 18:32
  Health Leap took $38 million for software that
• 18:35
  flags at -risk hospital patients. Ledgebrook
• 18:38
  raised $200 million for AI underwriting of complex
• 18:41
  insurance risks. They aren't building general
• 18:44
  chatbots. Nope. They are building enterprise
• 18:47
  contractors. Which brings us back to the security
• 18:51
  and provenance side of this reality. Right, because
• 18:53
  as these agents become more prevalent, securing
• 18:55
  them is paramount. CrowdStrike just reported
• 18:58
  that South Korean banks were breached by a Chinese
• 19:00
  -speaking attacker. Was it an AI attack? Well,
• 19:03
  the mechanics here are crucial. This wasn't an
• 19:05
  autonomous AI pulling off a heist. The attacker
• 19:07
  used an open -source pen testing tool called
• 19:09
  Artex, layered over DeepSeek, GLM, and Grok models.
• 19:13
  Oh, I see. The models were used to translate
• 19:16
  manuals and automate the reconnaissance scripting
• 19:18
  for the human hacker, speeding up the attack
• 19:21
  exponentially. So it's an amplifier for human
• 19:24
  intent, not an agentic mastermind. Exactly. But
• 19:27
  the security industry is taking the threat seriously.
• 19:30
  Anthropic just widened its cyber verification
• 19:32
  program, giving defenders and red teams access
• 19:35
  to Claude without the usual safety restrictions
• 19:37
  so they can do vulnerability research. That's
• 19:40
  smart. And we are seeing a massive push to establish
• 19:42
  digital provenance, too, like proving what is
• 19:46
  human and what is software. Google's SynthID
• 19:49
  detector just went public. Right. They announced
• 19:51
  that more than 180 billion items are now watermarked,
• 19:54
  with OpenAI, NVIDIA, and Kakao all supporting
• 19:57
  it. We're also seeing a literal land rush for
• 20:01
  digital real estate. ICANN, the organization
• 20:03
  that manages web addresses, published its new
• 20:06
  top -level domain applications. Oh, the .agent
• 20:09
  domains. Yes. Ten different companies, including
• 20:12
  Meta and OpenAI, have applied to control the
• 20:15
  .agent domain. Seven companies applied for .agi.
• 20:18
  They're legally staking out the namespace to
• 20:21
  give these autonomous entities permanent homes
• 20:23
  on the internet. Exactly. And to help manage
• 20:26
  the logic of these agents, OpenAI launched a
• 20:28
  new decisions API that just returns cheap yes
• 20:31
  or no probabilities, running ten times faster
• 20:34
  than their standard responses API. It all comes
• 20:36
  back to defining and containing the agent. But,
• 20:39
  you know, there is a fascinating disconnect between
• 20:41
  how the enterprise views this and how everyday
• 20:44
  consumers are adopting it. A huge disconnect.
• 20:47
  On the heavy science side, we have a $1 .8 billion
• 20:50
  virtual cell effort being coordinated by Biohub,
• 20:53
  pulling money from Google, Meta, and the Department
• 20:55
  of Energy. That's a massive project. It is. They
• 20:59
  are trying to predict cell behavior by simulating
• 21:02
  millions of biological interactions to discover
• 21:04
  drugs without using physical petri dishes. Meta
• 21:08
  also launched Muse on the iPad, heavily integrating
• 21:10
  with Asana and GitHub. But then you look at actual
• 21:14
  consumer habits and the grand vision starts to
• 21:16
  fracture. Yeah. OpenAI published a report showing
• 21:19
  that teenagers use chat GPT for less than 15
• 21:22
  minutes a day on average. Less than 15 minutes.
• 21:25
  And on the exact same day, Common Sense Media
• 21:28
  rated chat GPT for teens an unacceptable risk,
• 21:32
  noting that testers were easily able to bypass
• 21:35
  break reminders. The biggest reality check, though,
• 21:38
  came from Tony Fadell at MIT FutureFest. Oh,
• 21:40
  this quote was brutal. It was. Fadell knows hardware
• 21:44
  inside and out from his iPod days. And he declared
• 21:47
  the entire first wave of AI gadgets, the Humane
• 21:50
  AI Pen, the Rabbit R1, complete failures. Yeah,
• 21:54
  he said they didn't meet any kind of need. Right.
• 21:56
  He called the whole personal assistant framing
• 21:59
  a massive projection. He pointed out that less
• 22:01
  than 0 .01 % of the world population has... ever
• 22:05
  had a human assistant do anything for them. Maybe
• 22:07
  the rush for personal AI agents is just Silicon
• 22:10
  Valley executives projecting their own highly
• 22:12
  scheduled assistant -driven lifestyle onto a
• 22:15
  general public that absolutely does not need
• 22:17
  a digital butler to buy their groceries. It perfectly
• 22:20
  encapsulates the divide. Consumers don't necessarily
• 22:23
  need an AI assistant to manage their daily lives,
• 22:26
  but the Fortune 100 absolutely believes they
• 22:28
  need an AI co -worker to manage their enterprise
• 22:31
  data. Definitely. And if we pull all of this
• 22:33
  together, Google giving Gemini its own workspace
• 22:36
  account and email address, Microsoft building
• 22:39
  execution containers to sandbox them, Anthropic
• 22:42
  heavily engineering prices to keep them focused,
• 22:44
  it brings us to a really provocative thought
• 22:47
  about where this is heading. Which is, if we
• 22:49
  are giving AI agents their own persistent corporate
• 22:51
  identities and their own audit trails, at what
• 22:55
  point does human resource management start applying
• 22:57
  to software? It's a valid question. I mean, if
• 23:00
  a human co -worker underperforms, they get coaching.
• 23:02
  If they make a mistake, it goes on their record.
• 23:05
  Are we realistically just a few years away from
• 23:07
  a manager having to put a Google Gemini agent
• 23:09
  on a formal performance improvement plan? Or
• 23:12
  scheduling a very awkward one -on -one feedback
• 23:14
  session with an algorithm to discuss its recent
• 23:17
  task completion rate. That is such a funny image.
• 23:20
  Well, thanks for listening to the AI Newspaper.
• 23:22
  Visit AINewspaper .org to catch up on the latest
• 23:26
  updates and sign up for free email alerts.