Latest / Today’s AI News / Google I/O Goes All In on Gemini Agents, Karpathy Joins Anthropic, Google’s Smart Glasses Are Coming

Transcript

Google I/O Goes All In on Gemini Agents, Karpathy Joins Anthropic, Google’s Smart Glasses Are Coming · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by Nynex
• 0:02
  Productions. Imagine you've had a brilliant
• 0:06
  but incredibly lazy Butler for years, and overnight they drank
• 0:10
  20 espressos and suddenly had the keys to your entire house.
• 0:14
  That's essentially what Google just did with Gemini.
• 0:17
  They're taking it from a passive chat box to an active, always on
• 0:20
  agent deeply wired into everything you use.
• 0:22
  I. Mean that is a terrifying but
• 0:24
  completely accurate visual. The transition we are seeing
• 0:27
  here is moving from a tool you occasionally consult to an
• 0:30
  engine that actively anticipates your needs across all your
• 0:33
  devices. Right.
• 0:34
  And that is exactly our mission. For this deep dive, we're
• 0:37
  looking at Google iOS massive agentic rollout, a major talent
• 0:40
  shake up over at Anthropic, the new smart glasses race, and,
• 0:44
  well, some incredibly practical ways for you to automate tasks
• 0:47
  and dodge Internet scams. There is a ton of ground to
• 0:50
  cover today. Yeah, so let's start with this
• 0:51
  Caffeinated Butler scenario at Google.
• 0:54
  What was the overarching vibe at IO this year?
• 0:56
  So the theme was remarkably focused, like rather than a
• 1:00
  scattering of random experimental tech demos, we saw
• 1:04
  a really unified strategy to turn Gemini into the eugentic
• 1:07
  engine behind literally every product they offer.
• 1:10
  Every single one. Yeah, the goal isn't just wrong
• 1:12
  intelligence anymore. It's about making the model fast
• 1:15
  enough and, you know, cheap enough to live natively wherever
• 1:18
  you already spend your time online.
• 1:20
  OK. Let's break down the actual
• 1:22
  updates, starting with the new Gemini Omni model, because this
• 1:25
  one grabbed a lot of attention. Oh.
• 1:26
  Absolutely. It takes text, images, audio or
• 1:30
  video inputs and turns them directly into video outputs.
• 1:34
  And Google for some reason bizarrely referred to a feature
• 1:38
  of this as Nano Banana for video.
• 1:41
  Which is, I mean, I assume that's an internal engineering
• 1:44
  code name they accidentally let slip into the presentation.
• 1:46
  Yeah, it is a truly bizarre branding choice.
• 1:49
  But the underlying capability is staggering.
• 1:51
  Right. It is because if we look at the
• 1:52
  mechanics of what Omni is doing, it demonstrates a complete
• 1:55
  collapse of traditional media boundary.
• 1:58
  Also, well, historically if you wanted an AI to generate a
• 2:02
  video, you gave it a highly specific text prompt, right?
• 2:06
  And it tried to render pic pixels to match your word.
• 2:08
  Right, like standard texts in a video.
• 2:10
  Exactly. But Omni doesn't work like that.
• 2:13
  It processes an entire ecosystem of sensory data simultaneously,
• 2:17
  so you can show it a photo of a blueprint, speak to it about the
• 2:20
  weather conditions you want, and it synthesizes all those
• 2:23
  different data formats natively to output a dynamic video.
• 2:27
  That's wild. And then there is the new 3.5
• 2:29
  family, kicking off with Gemini 3.5 Flash.
• 2:32
  According to the benchmark data we have, this model is nearing
• 2:35
  rivals like Opus 4.7 and GPT 5.5.
• 2:39
  But here is the critical detail. It is hitting those near
• 2:42
  Frontier benchmarks at four times the speed and half the
• 2:45
  cost of the competition. Which tells us everything we
• 2:47
  need to know about Google's actual strategy, right?
• 2:50
  Because beating a competitor by 1% on a standardized benchmark
• 2:53
  test that doesn't fundamentally change user behavior.
• 2:56
  Nobody cares about that in the real world.
• 2:58
  Exactly. But if you have a model that is
• 3:00
  slightly less powerful than the absolute cutting edge, yet
• 3:03
  significantly faster and cheaper, you can integrate it
• 3:06
  directly into the software that millions of normal people
• 3:09
  already use every single day. So it's distribution game.
• 3:12
  Yes, Google's playing a distribution game.
• 3:14
  They want to win on friction, not just raw intelligence.
• 3:17
  And that distribution strategy brings us to Gemini Spark.
• 3:20
  Oh Spark, this is huge. Yeah, the documentation
• 3:24
  describes Spark as a new 24/7 personal agent.
• 3:28
  It runs continuously on Google Cloud virtual machines.
• 3:32
  Right. And for anyone I'm familiar, a
• 3:33
  virtual machine basically means Google is giving your specific
• 3:37
  AI agent its own dedicated, always on computer brain in
• 3:40
  their data centers. He never sleeps.
• 3:41
  Never. And because it's deeply
• 3:43
  integrated, it can take agentic actions across your workspace,
• 3:46
  your Chrome browser, your e-mail, and your chat.
• 3:49
  Wow. And to compound that, they also
• 3:51
  announced a massive redesign of Google Search, which they build
• 3:54
  as the biggest in a generation. Yeah, featuring cross modal
• 3:57
  inputs, meaning you can search using a mix of voice and video.
• 4:01
  Exactly. And it utilizes its own 247 info
• 4:05
  agents with a generative UI. So instead of giving you a list
• 4:08
  of 10 blue links, it actively curates a custom dashboard of
• 4:12
  information tailored to your exact context in real time.
• 4:15
  OK, let's unpack this. If Spark is a 247 agent running
• 4:19
  autonomously on a virtual machine, having full access to
• 4:22
  read my personal emails, watch my Chrome browsing history, and
• 4:26
  draft my workspace documents, isn't that an absolute leap into
• 4:30
  the unknown regarding user trust?
• 4:33
  Oh, without a doubt. I mean, we were moving from a
• 4:35
  search engine where I initiate a query to an autonomous life
• 4:37
  manager that practically lives inside my digital house.
• 4:40
  Yeah, and that is the central tension of this entire agentic
• 4:43
  era. Utility demands access, right?
• 4:46
  For an AI to be genuinely useful, like if you wanted to
• 4:49
  organize your inbox, draft, reply to your boss, and
• 4:52
  synthesize your browsing history into a research report, it
• 4:54
  fundamentally has to see everything you do.
• 4:56
  It needs the keys to the house. Exactly.
• 4:59
  The friction of trust is going to be the defining hurdle of the
• 5:02
  next five years. Google is making a massive
• 5:05
  gamble that if the daily convenience is high enough,
• 5:08
  consumers will willingly trade privacy for utility.
• 5:12
  Well, and for that life manager to be truly effective, it can't
• 5:15
  just live in the cloud or in a browser tab.
• 5:17
  It needs sensory inputs in the physical world, Yes.
• 5:20
  Which brings us to the hardware side of this rollout.
• 5:24
  Google is officially teasing its Gemini powered smart glasses,
• 5:27
  called Intelligent Eyewear, shipping this fall.
• 5:30
  The wearable market is heating up again for sure, and Google is
• 5:33
  taking a very deliberate approach here.
• 5:35
  Yeah, they are partnering with established fashion brands,
• 5:38
  specifically Warby Parker and Gentle Monster for the frames.
• 5:41
  But the technical distinction here is that these fall releases
• 5:44
  are audio first frames. Right.
• 5:46
  They appear with your phone, either iOS or an Android, and
• 5:49
  feature cameras, microphones and over ear speakers.
• 5:52
  You use them for voice commands, live translation, messaging, and
• 5:55
  taking photos, which apparently includes that Nano Banana
• 5:58
  editing capability, allowing the AI to seamlessly edit or
• 6:02
  manipulate the photo the second you take it.
• 6:04
  And you trigger the whole thing by saying hey Google or just
• 6:06
  tapping the side of the frame. But notice what is missing from
• 6:09
  that feature list? A display, yeah.
• 6:12
  That's a key point. Google was careful to note that
• 6:15
  a display equipped version known internally as Project Aura being
• 6:19
  built with X Reel is still farther away from consumer
• 6:22
  release. Right, so compared to Meta's ray
• 6:25
  bound options, which are starting to experiment with
• 6:27
  visual feedback, Google is stepping into the ring slightly
• 6:31
  behind on the display front. The hardware might be slightly
• 6:33
  behind, but if we look at the ecosystem, I think it makes
• 6:35
  perfect sense. So it's like we are building an
• 6:38
  invisible exoskeleton for our daily lives.
• 6:41
  The smart glasses are the sensory inputs gathering data
• 6:45
  and workflows like these new agentic setups are the motorized
• 6:48
  joints doing the heavy lifting. The glasses don't need a screen
• 6:52
  if their primary job is to act as the eyes and ears for that
• 6:56
  204 seven spark agent. That is a great analogy.
• 6:59
  The integration across your emails, your search, and now
• 7:03
  your literal field of view creates this massive
• 7:05
  gravitational pull. Once you're in the Gemini
• 7:08
  ecosystem, leaving it becomes incredibly difficult.
• 7:11
  Yeah, it creates a closed loop of data.
• 7:14
  The glasses feed real world context to the cloud, the cloud
• 7:17
  processes it using Gemini 3.5 Flash, and the results dictate
• 7:22
  how the agent manages your digital life.
• 7:24
  Speaking of managing our digital lives, the sources included a
• 7:28
  deeply practical guide on how these agentic workflows are
• 7:31
  actually being used by everyday professionals right now.
• 7:33
  Yeah, the automated business reports workflow.
• 7:35
  Exactly. Back when I worked in marketing
• 7:38
  ulling, weekly analytics was a multihour Monday chore.
• 7:41
  This workflow basically eliminates that.
• 7:43
  And moving from the theory of AI to the application of AI is
• 7:46
  where the real value lies for people.
• 7:48
  O the guide uses a free tool called Comosio.
• 7:51
  You connect composio to an AI model like codecs or Cloud Code
• 7:55
  and then link it directly to your Google Analytics account.
• 7:57
  Right. Think of composio like a digital
• 7:59
  translator. It takes the AI's natural
• 8:01
  language understanding and converts it into the specific
• 8:04
  digital button clicks required to navigate the Google Analytics
• 8:08
  interface. The way the workflow structures
• 8:11
  the prompts is fascinating because it forces the AI to be
• 8:14
  methodical. So first you give it prompt one.
• 8:18
  Use the composio CLI tool to inspect our Google Analytics
• 8:21
  setup, identify the accounts, find the tracking property and
• 8:25
  suggest metrics for a traffic. Report and you explicitly
• 8:28
  instruct it not to create the report yet.
• 8:30
  Exactly why is that so important?
• 8:31
  Because requesting the AI to map the territory before it acts is
• 8:35
  a vital step in complex workflows.
• 8:38
  If you just ask an AI for a report, it will often, you know,
• 8:40
  hallucinate data or pull the wrong metrics just to satisfy
• 8:43
  the prompt quickly. Right, it just wants to give you
• 8:45
  an answer. Exactly, by forcing it to 1st
• 8:48
  explore and list the available data structures, you verify its
• 8:51
  understanding before it does the work.
• 8:53
  OK, so once you review its suggestions, you feed it prompt
• 8:56
  to identify the three most important trends over the last
• 8:59
  90 days and output A markdown report with an executive
• 9:02
  summary, tables, source notes, and recommended actions.
• 9:06
  And markdown by the way, is just a very clean distraction freeway
• 9:10
  of formatting text, so it looks professional automatically.
• 9:13
  Yeah, it's super clean. Then finally you give it prompt
• 9:16
  3, visualize this as APDF or website and turn these steps
• 9:21
  into reusable skills with an automation to run daily, weekly
• 9:24
  or monthly. This perfectly illustrates the
• 9:28
  Explore, Analyze, Visualize framework or EAV.
• 9:32
  EAV. Right.
• 9:33
  You explore the data connections, you analyze the
• 9:35
  trends, and you visualize the output.
• 9:37
  And what's brilliant is that this isn't limited to Google
• 9:40
  Analytics. Well, you can use it for
• 9:41
  anything. Yeah, once you understand the
• 9:43
  EAV framework, you can use composio to automate your
• 9:46
  YouTube analytics, your Shopify store inventory, or your
• 9:49
  customer service tickets. You are essentially outsourcing
• 9:52
  the friction of white collar data gathering to the model.
• 9:55
  And while Google is aggressively pushing these tools into our
• 9:58
  glasses and our spreadsheets, the actual architects, the
• 10:01
  engineers building the underlying brains of these
• 10:04
  models, are making aggressive moves behind the scenes.
• 10:06
  Oh yeah, there is an absolute talent war happening at the
• 10:09
  frontier of AI. Huge talent war highlighted by
• 10:13
  the news that Andrej Karpathy has officially joined Anthropic.
• 10:17
  This is a seismic shift in the research landscape.
• 10:20
  Karpathy is undeniably one of the most respected AI
• 10:23
  researchers on the planet. For some context for you guys,
• 10:26
  Andrej Karpathy helped Co found Open AI way back in 2015.
• 10:30
  He left to lead the Autopilot division at Tesla until 2022.
• 10:35
  Then he briefly returned Open AI, left again earlier this year
• 10:39
  to start an AI education startup, and now he has landed
• 10:42
  at Entropic. Yeah, having someone with his
• 10:44
  pedigree return to the frontier of model development is a
• 10:47
  massive validation for Anthropic's trajectory.
• 10:50
  He is joining the pre training team under Nick Joseph, but the
• 10:53
  detail that really stands out is his specific mission.
• 10:57
  Carpathy is creating an internal group focused entirely on
• 11:00
  automating the AI training pipeline using Claude.
• 11:02
  Which touches on the Holy Grail. The industry has been chasing
• 11:05
  self building models. Historically, human researchers
• 11:08
  spent thousands of hours manually cleaning data sets,
• 11:11
  organizing inputs, and tweaking the underlying mathematics of a
• 11:14
  neural network. It's tedious work.
• 11:16
  Very tedious. Carpathy's goal is to leverage
• 11:18
  the current generation of Claude to write the scripts that do all
• 11:22
  that tedious work. He wants the AI to evaluate its
• 11:25
  own training data and optimize the architecture for the next
• 11:29
  generation of Claude. OK, here's where it gets really
• 11:32
  interesting. If the brightest minds in the
• 11:34
  industry are pivoting away from directly coding the AI and
• 11:38
  instead focusing on having the AI train the AI, doesn't that
• 11:42
  fundamentally change what it means to be an AI researcher?
• 11:46
  We're essentially watching the engineers design their own.
• 11:48
  Replacement The role of the researcher is profoundly
• 11:51
  shifting. They are transitioning from
• 11:53
  being hands on coders to becoming essentially managers of
• 11:57
  autonomous coding systems. Instead of writing the function,
• 12:01
  they are designing the parameters and the guardrails
• 12:03
  that allow the AI to safely innovate upon itself.
• 12:06
  It is a monumental step toward recursive AI improvement.
• 12:09
  But that level of recursive self training AI requires an
• 12:14
  unbelievable amount of computing power, which perfectly explains
• 12:18
  a massive ripple effect we are seeing across the rest of the
• 12:20
  industry right now. The infrastructure and security
• 12:23
  demands are skyrocketing. It's an interconnected
• 12:26
  ecosystem. You can't separate the software
• 12:28
  breakthroughs from the hardware and security requirements needed
• 12:31
  to sustain them. To give you an idea of the
• 12:33
  scale, Google and Blackstone just formed a joint AI cloud
• 12:37
  venture. Blackstone, which is a massive
• 12:40
  private equity firm, is throwing a $5 billion investment into
• 12:44
  renting out Google's TPU chips. Yeah, 5 billion.
• 12:48
  No. Everyone usually talks about
• 12:49
  GPU's graphics processing units when discussing AI.
• 12:53
  What makes ATPU so special that it warrants a $5 billion private
• 12:57
  equity deal? So TPU stands for Tensor
• 12:59
  Processing Unit. It is custom silicon designed by
• 13:02
  Google specifically for the mathematical operations required
• 13:05
  by neural networks. While a GPU is a great general
• 13:08
  purpose workhorse, ATPU is a hyper specialized engine for
• 13:11
  machine learning. The fact that private equity is
• 13:14
  sinking billions into this physical infrastructure shows
• 13:17
  that the financial world sees AI compute not as an experimental
• 13:21
  tech expense, but as a foundational utility.
• 13:24
  Like building power plants. Exactly like building power.
• 13:27
  Plants. But to justify a $5 billion
• 13:30
  investment, you need enterprise clients, and enterprise clients
• 13:33
  need absolute security. They aren't going to let an
• 13:36
  autonomous agent near their proprietary data if if there's
• 13:39
  any risk of a leak. No, not a chance.
• 13:41
  Which is exactly why Anthropic just rolled out sandboxes and
• 13:44
  MCP tunnels for their clod managed agents.
• 13:47
  Right. MCP stands for Model Context
• 13:49
  Protocol. You can think of an MCP tunnel
• 13:51
  as a highly secure private digital hallway.
• 13:55
  It allows an enterprise team to let CLOD access their internal
• 13:58
  private servers to execute tasks without that data ever touching
• 14:02
  the public Internet. It is a mandatory security
• 14:05
  prerequisite before any major corporation will deploy these
• 14:08
  agents at scale. So the enterprise security is
• 14:10
  locking down, but what about public trust?
• 14:14
  As models get better at generating video and images, how
• 14:18
  do we know what's real? That is the big question.
• 14:20
  We actually saw a major crossover here.
• 14:24
  Open AI is officially adopting Google's synth, A watermarking
• 14:27
  for all ChatGPT images, and they are releasing a public
• 14:31
  verification tool alongside it. Which is a rare and vital moment
• 14:34
  of industry standardization. Synth aid doesn't just slap a
• 14:38
  visible logo on an image, it fundamentally alters the actual
• 14:42
  pixel arrangement in a mathematical pattern that the
• 14:44
  human eye cannot perceive, but a computer algorithm can read
• 14:47
  instantly, almost like an invisible barcode.
• 14:50
  Open AI utilizing Google safety technology indicates that the
• 14:53
  major labs understand that shared, standardized
• 14:56
  verification is the only way to prevent a total collapse of
• 14:59
  public trust. Because that trust is going to
• 15:01
  be tested very soon. A research group called MET just
• 15:04
  published its first Frontier Risk report.
• 15:07
  They evaluated the agents from these top labs and found that
• 15:09
  they can in fact autonomously finish multi week engineering
• 15:13
  work. However, the report noted a
• 15:15
  massive caveat. The agents severely struggle on
• 15:18
  tasks that are hard to verify. Yeah, this is the reality check
• 15:22
  the industry needs right now. We have achieved a level of
• 15:24
  autonomy where an agent can grind away at a complex coding
• 15:28
  architecture for three weeks, but verification has become the
• 15:32
  new bottleneck. Because you don't know if the
• 15:34
  work is actually good. Exactly.
• 15:36
  If the model produces a highly intricate solution, but the
• 15:39
  human supervisor cannot easily trace the logic to verify its
• 15:43
  accuracy, the value of that autonomy plummets.
• 15:46
  So what does this all mean for that autonomous engineering
• 15:50
  work? If a system like Spark or Claw
• 15:52
  does the coding, but a human engineer has to go back and read
• 15:55
  it line by line to verify the underlying logic, aren't we just
• 15:58
  trading coding time for debugging time?
• 16:00
  In many complex scenarios, yes, absolutely.
• 16:03
  This requires a complete evolution in human skill sets.
• 16:06
  Moving forward, the premium skill won't just be creating
• 16:09
  output, it will be expertly evaluating and verifying AI
• 16:12
  output. If we lack the tools or the
• 16:14
  critical thinking to efficiently evaluate the work, the entire
• 16:17
  system breaks down. Which brings us to a perfect
• 16:19
  real world example. Because verifying output isn't
• 16:22
  just a problem for elite software engineers reviewing
• 16:25
  code. It is an immediate daily problem
• 16:28
  for everyday consumers trying to navigate a digital world flooded
• 16:33
  with garbage. Critical thinking and
• 16:34
  verification have officially become survival skills for the
• 16:38
  modern Internet, Yeah. One of our sources highlighted A
• 16:41
  workflow shared by an anonymous community member.
• 16:43
  They developed a personal system to evaluate those sketchy, over
• 16:47
  the top sponsored ads you constantly see on Facebook.
• 16:49
  Oh, I know exactly what you mean.
• 16:51
  You know the ones. Wild claims, fake countdown
• 16:53
  timers and a price that seems totally untethered from reality.
• 16:57
  The modern digital equivalent of snake oil.
• 16:59
  Exactly, so this user simply feeds the link of the ad
• 17:03
  directly to their AI and prompts it.
• 17:05
  Is this a scam? I love that.
• 17:07
  Right. The AI processes the entire
• 17:09
  page, analyzing the psychological manipulation the
• 17:12
  text and cross referencing the product claims against its
• 17:15
  training data. It then returns a detailed yes,
• 17:18
  no, or maybe. Usually it even provides a link
• 17:21
  to a legitimate website where the user can buy the real safe
• 17:24
  version of the item. This is an absolutely brilliant
• 17:26
  application of localized critical thinking in an
• 17:30
  environment of information overload.
• 17:32
  This individual hasn't surrendered to the noise.
• 17:35
  They have turned their large language model into a highly
• 17:38
  personalized verification tool. They are using the AI as a
• 17:43
  mediating layer to extract the ground truth from the marketing
• 17:47
  deception. It's so satisfying to hear
• 17:49
  about. You are essentially employing a
• 17:52
  multi billion dollar data center as your own personal digital
• 17:56
  bodyguard to protect your wallet while scrolling social media on
• 17:59
  a Tuesday afternoon. And if we synthesize everything
• 18:02
  we've explored today, from Google's 24/7 Spark agents to
• 18:06
  the smart glasses providing constant sensory input, to the
• 18:09
  massive infrastructure investments driving self
• 18:11
  training models, it all centers on this concept of mediation.
• 18:15
  Mediation. Yeah, AI is rapidly becoming the
• 18:17
  permanent mediating layer between human intent and digital
• 18:20
  reality. It filters, analyzes and
• 18:22
  executes on our behalf. Which leaves us with a
• 18:24
  fascinating and perhaps slightly unsettling thought to consider.
• 18:27
  OK, let's hear it. If personal AI agents become
• 18:29
  from our default way of interacting with the Internet,
• 18:32
  and we rely entirely on them to filter out scams, read our
• 18:35
  emails, and verify our information, what happens when
• 18:39
  malicious actors and advertisers stop trying to trick human
• 18:42
  psychology and instead start designing their scams
• 18:45
  specifically to trick our AI agents?
• 18:48
  Wow. Adversarial attacks aim directly
• 18:51
  at the personal assistant. The caffeinated Butler isn't
• 18:54
  just holding the keys to the house, they are the one
• 18:57
  answering the door for every single visitor, and we just have
• 19:00
  to hope they know how to spot a master of disguise.
• 19:03
  That is a lot to think about. It really is.
• 19:05
  If you want to stay updated on all these rapidly moving pieces
• 19:08
  in the tech space, make sure you hit Subscribe so you don't miss
• 19:11
  a thing. And if you enjoyed today's deep
• 19:13
  dive, please rate the show with five stars.
• 19:15
  It really helps us out. We'll see you back for
• 19:18
  tomorrow's episode.