Latest / Tech Talks With Kinsoft / Last Week in Tech – Google's Agentic Gemini, the Coding-Agent Gold Rush, and CISA's Own Key Leak

Transcript

Last Week in Tech – Google's Agentic Gemini, the Coding-Agent Gold Rush, and CISA's Own Key Leak · Tech Talks With Kinsoft

• 0:00
  You know, usually when you hire an assistant,
• 0:01
  there's this expectation of supervision. Right.
• 0:06
  Yeah. Like you have to manage them. Exactly.
• 0:08
  It's sort of like being a driving instructor.
• 0:10
  You're sitting right there in the passenger seat.
• 0:12
  You've got your own set of brakes. And, you know,
• 0:15
  when the student driver veers a little too close
• 0:17
  to the curb. You just reach over and correct
• 0:19
  the wheel. Right. You gently correct the steering
• 0:22
  wheel. Because ultimately, you're the one in
• 0:24
  control. You are fundamentally in control of
• 0:26
  the vehicle. Yeah. The machine. Or the assistant
• 0:29
  assists, but the human decides. It's a very comforting
• 0:34
  safety net. Because the ultimate responsibility
• 0:37
  rests in your hands. But if you look at the tech
• 0:41
  landscape right now, it's like suddenly that
• 0:43
  passenger side brake pedal is just gone. Completely
• 0:46
  gone. We're looking at a technological shift
• 0:48
  where you aren't just, you know, letting go of
• 0:50
  the wheel. You're basically standing on the sidewalk
• 0:52
  watching the car drive itself across the country.
• 0:56
  Right. And it's negotiating for its own gas,
• 0:58
  booking its own hotel rooms along the way. Yes.
• 1:01
  And that's exactly what we're getting into today.
• 1:03
  Welcome to today's deep dive. Our entire mission
• 1:06
  today is unpacking this massive paradigm shift.
• 1:10
  We are moving entirely out of the assistant era
• 1:12
  and diving headfirst into the agentic era. It
• 1:15
  is a massive transition. It really is. And we're
• 1:18
  pulling all this from a really fantastic source
• 1:21
  today. It's the Kinsoft Tech Talks update from
• 1:24
  May 25, 2026. So just a concise, high -level
• 1:28
  tech industry recap that cuts through all the
• 1:30
  noise. And the reason you, the listener, really
• 1:32
  need to care about this specific update is because
• 1:35
  this transition to the agentic era, it fundamentally
• 1:38
  alters how businesses have to think about automation.
• 1:40
  Oh, absolutely. It brings these multibillion
• 1:43
  dollar financial ripple effects and honestly,
• 1:45
  some security implications that are just terrifying.
• 1:48
  Yeah, we're going to get to those later. And
• 1:49
  they are wild. They really are. And whether you're
• 1:52
  just a casual tech observer or you're running
• 1:54
  a business, you need to understand this. This
• 1:56
  is the new reality. OK, let's unpack this. We
• 1:59
  have this specific week. in May 2026 that serves
• 2:02
  as like a tipping point. And it starts with Google
• 2:05
  I .O. Right. Their big developer conference.
• 2:07
  Yeah. They stood on stage and essentially went
• 2:10
  all in on this agentic era. They announced Gemini
• 2:14
  3 .5, which is geared specifically for agents
• 2:16
  and coding. But the one that really grabbed me
• 2:19
  was Gemini Omni. Oh, Gemini Omni is a completely
• 2:22
  different beast. That model represents a massive
• 2:25
  architectural leap. How so? Well, the distinction
• 2:28
  between older systems and Omni is really critical
• 2:31
  to understand. Like earlier AI models were essentially
• 2:35
  these separate brains that were just clumsily
• 2:37
  taped together. Right, like different departments
• 2:39
  in a company. Exactly. A text model processed
• 2:41
  the text, an image model processed pixels, and
• 2:44
  an audio model handled the sound waves. So when
• 2:47
  you asked a complex question that involved...
• 2:50
  looking at a photo. They had to hand the information
• 2:52
  back and forth to each other. Right. It's kind
• 2:54
  of like a committee trying to describe a movie
• 2:55
  scene to each other over the phone, like it's
• 2:58
  lost in translation. A highly inefficient committee
• 3:01
  and with incredible latency too. But Gemini Omni
• 3:05
  changes the underlying architecture by being
• 3:08
  natively multimodal. Meaning it's all one system.
• 3:12
  Yes. It doesn't translate images into text to
• 3:15
  understand them. It ingests the video, the audio,
• 3:18
  text and images all through the exact same neural
• 3:21
  network simultaneously. Oh, wow. So it's processing
• 3:24
  reality a lot closer to how we do, right? Like
• 3:26
  human sensory integration. Exactly. Which allows
• 3:29
  it to instantly understand deep context. And
• 3:32
  then it can generate high quality video or audio
• 3:35
  right out the other side without all that lag.
• 3:37
  That's incredible. But, you know, Google wasn't
• 3:40
  acting in a vacuum here. The pace described in
• 3:42
  this Kinsoft update is just staggering. It really
• 3:45
  was a crazy week. I mean, within a span of three
• 3:47
  days, three entirely separate frontier -grade
• 3:51
  coding agents dropped. Three days. Yeah. Cursor
• 3:54
  released Composer 2 .5. Anthropic dropped code
• 3:57
  with Claude. And Alibaba launched Quinn 3 .7
• 4:01
  Max. Three massive releases in basically a weekend.
• 4:05
  And you really have to notice the terminology
• 4:06
  shift there. We're moving from assistants to
• 4:08
  agents. That isn't just marketing spin, I promise
• 4:10
  you. No, it means something very different. It
• 4:12
  represents a fundamental change in how the software
• 4:15
  actually operates. I love the intern versus senior
• 4:18
  manager analogy for this. The assistant era we're
• 4:22
  leaving behind. It was like having a very eager
• 4:24
  but very fast intern. Right. They mean well,
• 4:27
  but you have to check everything. Exactly. You
• 4:30
  could ask for a block of code, but you had to
• 4:32
  read every single line, spot the hallucinated
• 4:34
  variables, fix the syntax errors. You were the
• 4:37
  manager constantly correcting the intern's work.
• 4:39
  While an agent, on the other hand, operates on
• 4:41
  a continuous autonomous reasoning loop. Autonomous
• 4:44
  being the key word there. Exactly. You give the
• 4:46
  agent a high -level goal. Say you tell it, uh...
• 4:50
  build a secure login page that connects to my
• 4:52
  customer database. Okay, a standard task. Right.
• 4:55
  The agent plans the architecture, writes the
• 4:58
  code, but then it actually spins up a virtual
• 5:00
  environment and runs the code itself. Wait, it
• 5:03
  tests its own work? Yes. When the code inevitably
• 5:06
  fails, the agent reads the terminal error message,
• 5:09
  understands why the database connection dropped,
• 5:12
  rewrites the problematic function, and then tests
• 5:14
  it again. So it's doing the debugging in the
• 5:16
  dark, basically, without bother you at all. It
• 5:19
  iterates completely on its own until the goal
• 5:21
  is achieved. And what's fascinating here is how
• 5:24
  this completely upends the economics of software
• 5:26
  creation. Oh, totally. Because the cost of frontier
• 5:29
  -grade coding help is just falling off a cliff
• 5:32
  right now. Because you have Google, Anthropic,
• 5:36
  Cursor and Alibaba caught in this vicious price
• 5:40
  war. They are trying to outcompete each other
• 5:43
  week by week. And it's a price war that fundamentally
• 5:45
  alters the value proposition of the entire tech
• 5:48
  sector. I mean, think about a mid -level SaaS
• 5:51
  startup software as a service. Let's say they
• 5:54
  charge $50 a month for a specialized inventory
• 5:56
  management tool. Historically, businesses paid
• 6:00
  that subscription because building their own
• 6:02
  custom tool would require, you know, hiring a
• 6:05
  development team for six months. Right. And spending
• 6:08
  $100 ,000 easily. Exactly. But if a business
• 6:10
  owner can just open their laptop, tell an AI
• 6:13
  agent, build me an inventory tracker tailored
• 6:16
  to my exact warehouse layout and workflow. And
• 6:18
  the agent just executes that entire reasoning
• 6:20
  loop for, what, $3 in compute costs? Yeah, exactly.
• 6:23
  Why would they ever pay for a generic SaaS subscription
• 6:25
  again? They wouldn't. The barrier to creating
• 6:28
  complex, bespoke applications is dropping to
• 6:31
  almost zero. So the dynamic between the people
• 6:34
  who build software and those who buy it is just
• 6:36
  completely compressing. The intrinsic value is
• 6:39
  no longer in writing the code itself. The value
• 6:41
  shifts entirely to the massive infrastructure
• 6:44
  running these agents and, of course, the proprietary
• 6:47
  data you feed them. So what does this all mean
• 6:49
  for the companies actually footing the bill?
• 6:52
  Because, you know, having an AI senior manager
• 6:55
  constantly running these autonomous coding loops,
• 6:58
  testing and retesting, that requires a ridiculous
• 7:01
  amount of compute power. An unbelievable amount.
• 7:04
  You can't just run these frontier models on a
• 7:06
  standard laptop, which really explains the terrifying
• 7:09
  financial numbers we are seeing behind the scenes.
• 7:11
  Yeah, we got some rare financial sunlight on
• 7:13
  that front recently, thanks to a SpaceX S1 filing,
• 7:16
  oddly enough. Right, because of the Elon Musk
• 7:18
  connection. Exactly. Because he intertwines his
• 7:21
  company's regulatory filings for one often exposed
• 7:24
  data for another. And this one accidentally contained
• 7:27
  the financials for XAI for the year 2025. And
• 7:30
  the numbers are just wild. I mean, XAI brought
• 7:33
  in about $3 .2 billion in revenue. Which is huge.
• 7:37
  Earning $3 .2 billion in a single year for a
• 7:40
  relatively new company is phenomenal. Against
• 7:43
  that revenue, they posted a staggering $6 .4
• 7:47
  billion operating loss. Yeah, a massive burn
• 7:50
  rate. Wait, stop. I have to push back here. Because
• 7:53
  in any normal industry, burning twice what you
• 7:56
  make isn't a business strategy. It's a bankruptcy
• 7:58
  hearing. It really is. How do you lose over $3
• 8:01
  billion and convince investors you're doing a
• 8:03
  great job? Is this actually a sustainable business
• 8:05
  model? Well, it looks like financial ruin if
• 8:08
  you view it through the lens of a traditional
• 8:10
  software company. Right, where your expenses
• 8:12
  are just servers and payroll. Exactly. But AI
• 8:15
  infrastructure requires a completely different
• 8:17
  mental model. That $6 .4 billion isn't going
• 8:20
  to, like, marketing campaigns or bloated executive
• 8:23
  salaries. Either. Where is it going? It is going
• 8:26
  toward a physical manufacturing race. We are
• 8:30
  talking about purchasing hundreds of thousands
• 8:32
  of specialized liquid -cooled microchips. Oh
• 8:36
  right, the hardware. Yeah, they're constructing
• 8:38
  massive data centers from concrete and steel
• 8:40
  and trying to secure the energy contracts to
• 8:43
  power facilities that literally consume as much
• 8:46
  electricity as a small city. This is a capital
• 8:49
  expenditure. They're treating AI compute like
• 8:52
  they're... I don't know, laying transatlantic
• 8:55
  fiber optic cables or building a national electrical
• 8:58
  grid. That's a perfect way to look at it. They
• 9:00
  are funding it ahead of profits because they
• 9:02
  believe the winner takes all dynamic in AI is
• 9:04
  absolute. If you aren't first, you're last. Whoever
• 9:07
  builds the most capable physical infrastructure
• 9:09
  today will essentially own the cognitive autonomy
• 9:12
  of tomorrow. Wow. If a company pauses its capital
• 9:15
  expenditure to become profitable right now, their
• 9:17
  models just fall behind the competition and they
• 9:20
  become entirely irrelevant in six months. That
• 9:22
  is a brutally high stakes game of poker. Extremely
• 9:26
  high stakes. But what really jumped out at me
• 9:28
  from the Kinsoft update is the duality happening
• 9:30
  simultaneously. Like right next to that hypercapitalist
• 9:34
  reality, we see Anthropic and the Gates Foundation
• 9:36
  launching a $200 million four -year partnership.
• 9:40
  Yes, a highly targeted effort to build AI tools
• 9:43
  specifically for health, education, and agriculture
• 9:47
  in underserved regions. It's just such a fascinating
• 9:49
  contrast. You have these multi -billion dollar
• 9:51
  bloodbaths between tech giants trying to dominate
• 9:54
  the future economy. Juxtaposed with this heavily
• 9:57
  -sunded philanthropic initiative to deploy that
• 10:00
  exact same AI to help, like... A farmer in a
• 10:03
  developing nation optimized their crop yield
• 10:05
  today. It really captures the central tension
• 10:07
  of the AI boom beautifully. The foundational
• 10:09
  models are so incredibly resource intensive that
• 10:12
  they demand unprecedented corporate capital just
• 10:14
  to create. Right, the $6 .4 billion losses. Exactly.
• 10:18
  Yet their potential to solve fundamental human
• 10:20
  problems is so immediate that major philanthropic
• 10:23
  organizations are just bypassing traditional
• 10:25
  aid models entirely. They want to deploy AI directly
• 10:28
  into the hands of the people who need it most.
• 10:30
  But, and here's the catch. As powerful as these
• 10:33
  tools are, whether they are writing custom software
• 10:36
  autonomously or optimizing global agriculture,
• 10:38
  they operate on networks secured by humans. Yes.
• 10:42
  And humans, unfortunately, make mistakes. We
• 10:44
  do. And this raises a really important question.
• 10:47
  What happens when you hand the keys to an incredibly
• 10:50
  powerful autonomous AI agent, but the lock on
• 10:53
  the door was installed by a human who just forgot
• 10:55
  to tighten the screws? Which brings us perfectly
• 10:58
  to the security breakdown in the Kinsoft update.
• 11:01
  And here's where it gets really interesting because
• 11:03
  the cautionary tale they share is just dripping
• 11:06
  with irony. We are talking about CISA. America's
• 11:09
  cyber defense agency. The people whose literal
• 11:11
  mandate is to secure the nation's digital infrastructure.
• 11:14
  The absolute... ATECs of security professionals.
• 11:17
  You don't get higher than CISA. And yet, a contractor
• 11:20
  for CISA left a public GitHub repository exposing
• 11:24
  AWS GovCloud access keys. Just sitting there
• 11:28
  on the open internet over a weekend. It's almost
• 11:30
  hard to believe. To visualize this for you listening,
• 11:33
  a GitHub repo isn't just a digital bulletin board.
• 11:35
  It is the architectural blueprint of an application
• 11:38
  where developers collaborate and store their
• 11:41
  code. Right. So leaving GovCloud keys in a public
• 11:44
  repo is like publishing the blueprints to the
• 11:47
  Federal Reserve online and accidentally leaving
• 11:50
  the master vault code scribbled in the margins
• 11:53
  for literally anyone to read. It is a devastating
• 11:56
  lapse in operational security, and the source
• 11:59
  material uses this to make a highly sobering
• 12:01
  point for anyone running a business right now.
• 12:04
  Yeah, if it can happen to them. Exactly. If a
• 12:06
  contractor working for the nation's premier cyber
• 12:09
  agency can leak high -level cloud keys into a
• 12:11
  public environment, everyday developers working
• 12:14
  at a mid -sized company are perfectly capable
• 12:16
  of making the exact same error. Especially if
• 12:19
  those everyday developers are suddenly using
• 12:21
  the AI coding agents we just talked about to
• 12:24
  generate... software 10 times faster. If we connect
• 12:27
  this to the bigger picture, this is exactly why
• 12:30
  the rise of AI agents makes stringent automated
• 12:33
  security protocols entirely non -optional now.
• 12:36
  You can't just rely on the honor system anymore.
• 12:38
  No, because the velocity of development has fundamentally
• 12:40
  changed. If a human is generating and deploying
• 12:43
  code at lightning speed because an agent is doing
• 12:46
  the heavy lifting, they are also capable of deploying
• 12:49
  catastrophic security vulnerabilities at lightning
• 12:52
  speed. Right. The mistakes scale with the productivity.
• 12:54
  Exactly. When a developer is managing a massive
• 12:57
  volume of AI -generated code, human oversight
• 13:01
  just gets stretched too thin. You can't manually
• 13:03
  catch a stray string of text containing a hard
• 13:05
  -coded password. Exactly why Kinsoft lays out
• 13:09
  three mandatory security rules for this new era.
• 13:12
  And the first one is secret scanning on all code.
• 13:15
  And we really need to look at the mechanics of
• 13:17
  how that works in practice. Secret scanning is
• 13:20
  an automated pipeline step. Meaning it happens
• 13:22
  in the background? Yes. Before a developer's
• 13:24
  code is allowed to be published or merged into
• 13:27
  a shared repository, an automated tool actively
• 13:30
  reads every single line. It's mathematically
• 13:32
  searching for patterns that look like passwords,
• 13:35
  API tokens, or access keys like those AWS GovCloud
• 13:39
  keys that were leaked. So it intercepts the blueprint
• 13:41
  and scrubs the margins before it ever reaches
• 13:43
  the public domain. Exactly. It takes the human
• 13:46
  error out of the equation. Okay. And the second
• 13:49
  rule they emphasize is short -lived credentials.
• 13:51
  Think of short -lived credentials like a digital
• 13:54
  hotel key card rather than a metal house key.
• 13:57
  Oh, that's a great way to put it. Right, because
• 13:59
  instead of issuing a developer or an AI agent
• 14:02
  a permanent password to access a sensitive database,
• 14:05
  the system generates a temporary mathematical
• 14:07
  token, and that token expires automatically after
• 14:11
  a set period, say, four hours. So even if that
• 14:15
  token... Does leak onto the public internet because
• 14:18
  someone bypassed the secret scanner somehow?
• 14:21
  The window of opportunity for an attacker is
• 14:23
  incredibly narrow. The key simply turns to dust.
• 14:26
  I love that. And the final rule is least privilege
• 14:29
  access. This is the principle of containment.
• 14:32
  A developer or an AI agent acting on their behalf
• 14:35
  should only have the exact system permissions
• 14:38
  required to do their specific task and absolutely
• 14:41
  nothing more. You don't give them the keys to
• 14:43
  the kingdom. Exactly. If an intern is hired to
• 14:45
  restock the lobby fridge, you don't give them
• 14:48
  the master keys to the CEO's office. By enforcing
• 14:51
  least privilege access digitally, if an account
• 14:54
  is compromised, the attacker is trapped in that
• 14:57
  specific room. The rest of the network remains
• 14:59
  secure. Precisely. Which is so critical right
• 15:02
  now because the threat landscape Kintoff describes
• 15:05
  is relentlessly hostile. I mean, they highlight
• 15:08
  a 23 -year -old in Canada who was just arrested
• 15:10
  for running a massive Internet of Things botnet.
• 15:13
  Yeah, that was wild. And it was responsible for
• 15:15
  huge denial of service attacks. And Internet
• 15:18
  of Things, or IoT. Botnet relies on hijacking
• 15:22
  poorly secured smart devices. We're talking about
• 15:25
  internet -connected security cameras, smart refrigerators,
• 15:28
  or even office thermostats. Stuff you don't even
• 15:30
  think about having a password. Right. The attacker
• 15:33
  exploits basic vulnerabilities to compromise
• 15:35
  hundreds of thousands of these low -security
• 15:37
  devices, essentially turning them into a zombie
• 15:40
  army. And then what do they do with them? They
• 15:41
  command all of those devices to send junk data
• 15:44
  to a single target server simultaneously. It
• 15:47
  just overwhelms its capacity and knocks it completely
• 15:49
  offline. It's terrifying. And in that exact same
• 15:53
  week, a Vietnamese government network was breached.
• 15:56
  I mean, the attacks are constant. They're highly
• 15:58
  sophisticated. And as we saw with CISA. All it
• 16:01
  takes is one exhausted contractor making a human
• 16:04
  error to let them in. The agentic era accelerates
• 16:07
  both sides of this equation, unfortunately. Right.
• 16:10
  We are gaining unprecedented capabilities with
• 16:12
  multimodal models like Gemini Omni that synthesize
• 16:15
  reality and autonomous coding agents that build
• 16:18
  complex software loops from scratch. But trusting
• 16:22
  that automation requires an ironclad... automated
• 16:25
  foundation of security. We can't have one without
• 16:27
  the other. Exactly. We are rapidly reaching a
• 16:29
  point where the capability of the technology
• 16:31
  is no longer the bottleneck. Our ability to safely
• 16:34
  contain and manage it is the true limit. So let's
• 16:37
  just step back and look at the journey we've
• 16:38
  taken through the source material today. I mean,
• 16:40
  we started by exploring this massive paradigm
• 16:43
  shift. Moving from the assistants to the agents.
• 16:45
  Right. Moving from AI assistants you have to
• 16:47
  constantly supervise to... AI agents that autonomously
• 16:51
  execute complex reasoning loops. And that's all
• 16:54
  driven by breakthroughs in multimodality and
• 16:56
  fierce industry competition. And then we examined
• 16:59
  the staggering financial reality of funding that
• 17:02
  shift. We looked at the billions being burned
• 17:05
  on physical compute infrastructure to win this
• 17:07
  AI arms race, while simultaneously seeing hundreds
• 17:10
  of millions deployed to solve immediate global
• 17:13
  challenges in health and agriculture. And finally,
• 17:15
  we grounded all of that soaring potential in
• 17:18
  the unforgiving reality of cybersecurity. recognizing
• 17:21
  that as we let AI accelerate our development
• 17:23
  cycles, we have to implement automated fail -safes
• 17:26
  like secret scanning and temporary credentials.
• 17:29
  Because human error scales just as fast as human
• 17:32
  productivity. Exactly. For you listening to this
• 17:35
  deep dive, the application is direct. Look at
• 17:38
  your own workflows or your own business. The
• 17:40
  tools to automate your daily friction are getting
• 17:42
  cheaper and more powerful by the hour. But you
• 17:45
  really must ask yourself, are your security practices
• 17:47
  and oversight mechanisms evolving at that exact
• 17:49
  same speed? Because if they aren't, you are building
• 17:53
  a skyscraper on a foundation of sand. And we
• 17:55
  really are. And that leaves us with one final...
• 17:58
  lingering question to ponder as we wrap up today.
• 18:01
  As these coding agents become cheaper and flawlessly
• 18:04
  capable, it is totally inevitable that we will
• 18:07
  eventually task AI agents with writing the security
• 18:09
  protocols to lock down our cloud systems. But
• 18:12
  in a world where AI is actively defending the
• 18:15
  digital vault against other AI attackers, will
• 18:18
  we finally trust the machines completely with
• 18:21
  our safety? Or will good old -fashioned human
• 18:23
  error, a tired developer, a misconfigured setting,
• 18:26
  or a forgotten GitHub repo remain the ultimate
• 18:29
  unpatchable vulnerability that leaves the front
• 18:32
  door wide open? Something to chew on. Stay curious,
• 18:35
  keep digging, and we will catch you next time.