Latest / AI Fire Daily / 🎙️ EP 267: Google’s New "Googlebook" & Mira Murati’s TML Breakthrough
Transcript
- 0:00Think about the last time you spoke to an AI.
- 0:02Yeah. It is usually a pretty jarring feeling.
- 0:05You try to interrupt the machine. An awkward
- 0:07silence stretches out. The system clunks through
- 0:10a messy reset. It just feels completely unnatural.
- 0:14It really does. It breaks the illusion immediately.
- 0:16Now, change that picture entirely. What if your
- 0:19AI could, you know, see you frown? What if it
- 0:21stopped talking mid -sentence and asked, what
- 0:23is wrong? That completely changes the paradigm.
- 0:25It does. Welcome to this custom deep dive. We
- 0:28are very glad you are here with us. Today, we
- 0:30are exploring a massive technological shift.
- 0:33We are tracking the move from clunky AI tools
- 0:36to fluid proactive partners. It is a big one.
- 0:40First, we will unpack Google's brand new Android
- 0:43updates. Then we are going to examine the staggering
- 0:46capital engines operating behind the scenes.
- 0:48We are talking about OpenAI, SoftBank, and Anthropic.
- 0:52The scale of the money is just wild. It truly
- 0:54is. Finally, we will look at a major breakthrough
- 0:57from Miramarati's Thinking Machines Lab. It is
- 1:00a fundamental shift that promises to end the
- 1:02awkward AI silence forever. It is a profound
- 1:05transition. We are moving past the novelty phase
- 1:09of AI. Yeah. The technology is weaving itself
- 1:11into our daily routines. It is basically becoming
- 1:14an invisible layer of reality. That is exactly
- 1:17what struck me about Google's latest announcements.
- 1:19They just wrapped up the Android show. It was
- 1:22their special IO edition. Right. And honestly,
- 1:24the whole ecosystem is shifting. We are looking
- 1:27at a radical redesign of how we interact with
- 1:29software. We are seeing a definitive move away
- 1:32from reactive tools. Consumer tech is becoming
- 1:35a proactive assistant. It anticipates what you
- 1:38need. It doesn't just wait around for a type
- 1:40command anymore. Right. And the clearest example
- 1:43of this is the Google book. They announced this
- 1:45entirely new type of computer. It is built from
- 1:48the ground up for Gemini. The operating system
- 1:50architecture is basically inverted. How so? Well,
- 1:53traditionally, the OS manages files and apps.
- 1:55Here, the OS centers entirely on the AI model.
- 1:59The AI is the kernel. Wow. They showcased a feature
- 2:02called the Magic Pointer too. It is essentially
- 2:04an AI -powered cursor. Yeah, that was fascinating.
- 2:07What is really interesting is how it bridges
- 2:09devices. You can use all your mobile apps on
- 2:12the big screen seamlessly. That removes a massive
- 2:15layer of friction. And the cursor itself... actually
- 2:19understands context. Right. It uses computer
- 2:22vision. Exactly. It knows what you are hovering
- 2:25over. If you hover over an address, it predicts
- 2:28you want a map. It anticipates your next action
- 2:30before you click. Okay, but the vibe -coded widgets
- 2:32feature is what actually stopped me in my tracks.
- 2:35Oh, man. You just describe a widget you want,
- 2:37maybe something incredibly niche that doesn't
- 2:39exist yet, and Android just builds it for you
- 2:42on the spot. Oh, it is wild. It completely flips
- 2:45how we interact with our phone's interface. You
- 2:47don't search an app store anymore? Nope. You
- 2:49just articulate a desire. The operating system
- 2:51handles the creation instantly. It is like magic.
- 2:55Under the hood, a lightweight language model
- 2:57writes the code. It compiles the UI components
- 3:00in real time. But wait, let's unpack that. Isn't
- 3:03that a massive threat to developers? Oh, absolutely.
- 3:05If the OS dynamically generates custom widgets,
- 3:08nobody is buying third -party apps for those
- 3:11microtasks anymore. Right? And that is the hidden
- 3:13disruption here. It democratizes software creation
- 3:16for the user. but it absolutely cannibalizes
- 3:20the lower tier of the developer ecosystem. The
- 3:23middleman is eliminated. That makes sense. We
- 3:26are also seeing major upgrades to Gemini intelligence.
- 3:29It is actually taking real -world actions now.
- 3:32Yes. And this is the shift from a passive oracle
- 3:34to an active agent. Right. It moves beyond generating
- 3:38text into executing multi -step workflows. The
- 3:42example they showed was brilliant. You take a
- 3:44photo of a concert flyer on a telephone pole.
- 3:46You simply tell Gemini to book a hotel. It goes
- 3:49out, navigates the travel sites autonomously,
- 3:52and finds a room near the venue. Think about
- 3:54the underlying mechanics there. It parses the
- 3:56image using a vision model. Yeah. It extracts
- 3:59the location, the band name and the date. It
- 4:01cross references your calendar. It accesses your
- 4:04payment preferences. It is doing all of that
- 4:06in the background. Exactly. Then it executes
- 4:08a complex sequence of web navigation tasks. It
- 4:11acts exactly like a human assistant. It is staggering.
- 4:14And alongside these massive shifts, there are
- 4:16smaller quality of life improvements. Right,
- 4:18like in Gboard. Yes. Gboard now has a feature
- 4:21called Rambler. It scrubs your voice -to -text
- 4:24inputs before they ever hit the screen. It automatically
- 4:27removes your awes and mid -sentence corrections.
- 4:31It uses on -device models to filter acoustic
- 4:34garbage. It is so helpful. It makes human communication
- 4:37look much cleaner. It basically edits our natural
- 4:40verbal stumbling in real time. Then there's the
- 4:43pause point feature. This one feels deeply necessary.
- 4:47I love this one. It forces you to wait 10 seconds
- 4:49before opening distracting apps, things like
- 4:52TikTok or Instagram. It gives you a moment to
- 4:54decide if you want to read a book instead. It
- 4:57is a great intervention. I have to admit something
- 4:59here. I still wrestle with doom scrolling myself.
- 5:01We all do. Our brains are hardwired for that
- 5:04quick dopamine hit. Yeah. PausePoint acts as
- 5:06a digital speed bump. Yes. It inserts a crucial
- 5:09moment of reflection into a compulsive habit.
- 5:11It fights algorithm -induced addiction with a
- 5:14counter -algorithm. They also announced some
- 5:16practical quick share updates. You can easily
- 5:18share files with iPhones via QR codes now. Moving
- 5:21your whole digital life from iOS to Android is
- 5:24getting much smoother. They are actively breaking
- 5:27down the walled gardens. The strategy is to make
- 5:30switching ecosystems entirely painless. Exactly.
- 5:34Lowering the barrier to entry. Security got a
- 5:37fascinating AI upgrade too. There is a new theft
- 5:41detection lock. This is very clever. The AI actually
- 5:44senses when a phone has been snatched out of
- 5:46your hand. It automatically locks the device.
- 5:49It also aggressively reduces the number of PI
- 5:52attempts a thief can make. That is a brilliant
- 5:54use of hardware sensors. It uses the phone's
- 5:57accelerometers and gyroscopes. Right. The machine
- 5:59learning model was trained on the physical signature
- 6:01of a theft. It knows the exact G -force and angle
- 6:05of a phone being grabbed and someone running
- 6:07away. Wow. It reacts faster than human reflexes
- 6:10could. And on a lighter note, they fully redesigned
- 6:13all 4 ,000 3D emojis. Right, because it is the
- 6:16little things that keep users tethered to an
- 6:18ecosystem. Looking at all this together, the
- 6:20ecosystem shift is vivid. It's like turning your
- 6:23phone from a passive filing cabinet into an active
- 6:25copilot. That is the perfect analogy. The device
- 6:29is no longer just waiting for your manual input.
- 6:31It is observing, predicting, and acting. But
- 6:34this raises a profound question. Does removing
- 6:37all friction with things like Gemini intelligence
- 6:40make us lose our own cognitive maps? That is
- 6:43a very valid concern. If the AI does everything,
- 6:45do we forget how to do things ourselves? That
- 6:48is the classic dilemma of cognitive offloading.
- 6:51Every new tool fundamentally changes our baseline
- 6:54expectations. Like when we got GPS. Exactly.
- 6:57When we offloaded navigation to GPS, our spatial
- 7:01memory demonstrably weakened. But it freed up
- 7:03mental bandwidth for other things. Right. We
- 7:06are outsourcing routine digital logistics to
- 7:08AI now. We might actually forget how to manually
- 7:11navigate a travel site or build a spreadsheet.
- 7:13But the theory is we gain time and energy for
- 7:16higher level thinking. It is a permanent tradeoff.
- 7:18So we trade a little self -reliance for ultimate
- 7:20daily efficiency. Got it. We're redefining what
- 7:23we consider a worthwhile use of our human time.
- 7:25If we are outsourcing our daily logistics to
- 7:28these seamless assistants, it takes a terrifying
- 7:31amount of server power to keep that illusion
- 7:33alive without lagging. Oh, the compute power
- 7:36required is mind boggling. Which brings us to
- 7:38the actual physical engines running all this.
- 7:41The numbers we are seeing are massive. The scale
- 7:44of capital deployment right now is unprecedented.
- 7:46We are watching a historical shift in global
- 7:49resource allocation. Let us look at OpenAI. They
- 7:53just had a massive liquidity event last fall.
- 7:55It was valued at $6 .6 billion. It is hard to
- 8:00even fathom that number. Roughly 600 employees
- 8:02cashed out. Around 75 of those employees walked
- 8:05away with over $30 million each. That fundamentally
- 8:09changes the internal dynamics of a company. You
- 8:12suddenly have a large group of incredibly wealthy
- 8:14engineers. Yeah. It shifts the culture from a
- 8:17scrappy research lab to a legacy institution
- 8:19managing massive wealth. It certainly changes
- 8:21the stakes. But the infrastructure spending happening
- 8:24outside the company is even bigger. Way bigger.
- 8:27SoftBank is making aggressive moves. Their CEO
- 8:30is discussing a $100 billion AI investment in
- 8:33France. That is an astonishing figure. It is
- 8:36literally larger than the G. It includes building
- 8:40massive new AI data centers. They want to radically
- 8:43expand the country's computing power infrastructure.
- 8:47This is deeply tied to national AI sovereignty.
- 8:50Countries are realizing that compute power is
- 8:53the new oil. If you don't control your own compute,
- 8:56you are at the mercy of foreign tech giants.
- 8:59Ain't true. SoftBank is positioning itself as
- 9:02the primary financier of this new global infrastructure.
- 9:05They are pouring the concrete. The corporate
- 9:07drama behind the scenes is just as intense as
- 9:10the spending. Sam Altman recently testified under
- 9:13oath. That testimony was wild. He shared some
- 9:16fascinating and honestly bizarre details about
- 9:19his relationship with Elon Musk. The history
- 9:21between those two founders is incredibly complicated.
- 9:23It is the defining origin story of modern AI.
- 9:27Altman testified that Musk once suggested open
- 9:29AI could pass to his children if he died. I still
- 9:32cannot believe he said that. Altman called it
- 9:33one of the hair -raising moments of their relationship.
- 9:36It highlights the philosophical clashes they
- 9:38had early on. They were just arguing over a software
- 9:41product. They were arguing over who should legally
- 9:43control artificial general intelligence. It sounds
- 9:46like science fiction, but to them, the stakes
- 9:49were very real. Meanwhile, Anthropic is throwing
- 9:51its own weight around. They are hiring a literal
- 9:54Claude evangelist. The job pays up to $315 ,000
- 9:59a year. They want someone specifically to help
- 10:03founders and VCs adopt their AI products. It
- 10:06shows how critical developer adoption has become.
- 10:09The best model doesn't win automatically. You
- 10:12have to fight a ground war for mindshare among
- 10:14the people building the apps. Anthropic is also
- 10:16flexing its capital in acquisitions. reportedly
- 10:19in late -stage talks to buy a startup called
- 10:22Stainless. This is a huge deal. It builds crucial
- 10:25developer tools. OpenAI and Google actually already
- 10:28use Stainless. And Throdnick might buy it for
- 10:31over $300 million. That is a brilliant, aggressive
- 10:34chess move. If Anthropic owns the underlying
- 10:37tools that their biggest rivals rely on, they
- 10:40gain a massive structural advantage. It is about
- 10:43owning the plumbing of the entire ecosystem.
- 10:46Speaking of advantages, OpenAI is fighting back
- 10:48with new technology. They just released something
- 10:50called Daybreak. This is their new security play.
- 10:53It is their direct answer to anthropic -sclawed
- 10:55mythos. Security is rapidly becoming the next
- 10:58major battleground for these models. Daybreak
- 11:01combines two distinct systems. It uses GPT -5
- 11:04.5 cyber and codex security, an AI guard dog
- 11:08catching software bugs before hackers do. The
- 11:11AI writes the code. But now it also audits the
- 11:15code autonomously. Which is crazy to think about.
- 11:18It runs adversarial attacks against itself. It
- 11:21hardens critical systems against zero -day exploits
- 11:23at machine speed. There are also some incredibly
- 11:26fun details emerging from all this coding. The
- 11:29team behind Cloud Code shared a strange tip.
- 11:32Oh, the markdown thing. Yes. They said developers
- 11:34should avoid using markdown files for instructions.
- 11:37Funny enough, they called the tip unreasonably
- 11:39effective. Sometimes the behavior of these massive
- 11:41models defines logical explanation. It really
- 11:44does. It is likely a quirky artifact hidden deep
- 11:46in the training data. The model just pays better
- 11:49attention to raw text. And everyday people are
- 11:53using these massive models for entirely different,
- 11:55hilarious things. Like what? A fan used AI video
- 11:59tools to insert himself into the movie Titanic.
- 12:02He somehow fixed every single fan complaint.
- 12:06in a single video. He saved Jack. Right, and
- 12:08he basically became the hero the movie needed.
- 12:11Exactly. It shows the incredible, chaotic, creative
- 12:14potential of these tools once they hit the mainstream.
- 12:17But looking at the big picture, a massive question
- 12:19remains. With SoftBank dropping $100 billion
- 12:22and OpenAI minting millionaires overnight, are
- 12:25we building foundational plumbing or is this
- 12:28just an unsustainable arms race? It certainly
- 12:30looks like an arms race from the outside. The
- 12:32valuations are undeniably astronomical. Right.
- 12:34But we have to look at what they're actually
- 12:35building with that cash. The data centers are
- 12:38real. The power grids are real. They are physical
- 12:40assets. Exactly. They are pouring actual concrete
- 12:43and laying thousands of miles of fiber optics.
- 12:46Even if an economic bubble bursts, that physical
- 12:49infrastructure remains. It will power the next
- 12:52generation of software, regardless of which specific
- 12:56company wins the AI war. Massive capital is literally
- 13:00laying down the physical plumbing for the future.
- 13:03Makes sense. Yes. It is very similar to the telecom
- 13:06boom of the late 90s. How so? The companies went
- 13:08bust, but the fiber they laid gave us the modern
- 13:11Internet. Wait, I have to push back there. Telecom
- 13:14laid neutral fiber that anyone could use. These
- 13:17AI data centers are highly proprietary walled
- 13:20gardens. SoftBank isn't building a public park.
- 13:23Isn't that fundamentally different? That is a
- 13:25very sharp distinction. You are right. Yeah.
- 13:27The physical hardware exists, but the access
- 13:30is strippedly gated by the corporate giants.
- 13:33It is infrastructure, but it is privatized infrastructure.
- 13:37Sponsor. The billions of dollars we just discussed
- 13:40are all chasing one holy grail. They all want
- 13:43to make computers feel completely, seamlessly
- 13:45human. That is the ultimate goal. This leads
- 13:48perfectly into Mira Marotti's new venture. It's
- 13:51perhaps the most exciting technical development
- 13:52we are covering today. It changes the core interaction
- 13:55paradigm completely. Maradi is the former CTO
- 13:58of OpenAI. She recently founded Thinking Machines
- 14:01Lab. We will call it TML. Right. They just dropped
- 14:04a research preview that is stunning. It focuses
- 14:06on the one thing we actually hate about talking
- 14:08to AI. The awkward silence. Yeah. The fundamental
- 14:12lack of human conversational rhythm. Exactly.
- 14:14They are trying to end the chat bubble interface
- 14:17forever. TML built a system that collaborates
- 14:21in a live, continuous streaming loop. More waiting.
- 14:24It processes your voice, video, and text simultaneously.
- 14:28This is a radical departure from turn -based
- 14:30interactions. Standard AI operates like a walkie
- 14:33-talkie. Yeah. You speak, you wait. It speaks.
- 14:36With TML, you no longer wait for the model to
- 14:38finish generating a response. The model perceives
- 14:41and processes data in tiny rapid bursts. It operates
- 14:44in chunks of just 200 milliseconds at a time.
- 14:47That is incredibly fast. It is. This allows it
- 14:49to make natural human noises. It says, mm -hmm,
- 14:52yeah, and got it while you are actively speaking.
- 14:54Human conversation is incredibly complex sociologically.
- 14:57We rely heavily on those subtle back -channel
- 15:00signals. They are so important. They let us know
- 15:02the other person is tracking our thoughts. Standard
- 15:05AI lacks this entirely, which is why it feels
- 15:07dead. It even reacts to your facial expressions
- 15:10before you finish your sentence. Wow. If you
- 15:12look confused mid -thought, the AI stops and
- 15:15clarifies its point. That requires incredibly
- 15:17low system latency. 200 milliseconds is the magic
- 15:20number. Why that specific number? It matches
- 15:23human conversational gap times perfectly. If
- 15:26it takes longer than that, our brains register
- 15:28it as an awkward pause. To stay that impossibly
- 15:32fast... TML uses a unique two layer architecture.
- 15:36This is a brilliant engineering solution to a
- 15:38very hard physics problem. Because you cannot
- 15:40run a massive intelligence model in 200 milliseconds.
- 15:44Exactly. You just can't. So they split it up.
- 15:46They use a live model to handle immediate superficial
- 15:49interaction. Then a background model does the
- 15:52heavy intellectual lifting. One brain for quick
- 15:54reactions, another brain for deep thinking. Exactly.
- 15:57It is like the human amygdala versus the prefrontal
- 16:00cortex. That makes sense. The live model manages
- 16:03the social dynamics and the hmms. The background
- 16:06model pulls the factual data and structures the
- 16:09complex argument. They work in tandem seamlessly.
- 16:13Because it processes video in real time, the
- 16:15real -world capabilities are wild. The model
- 16:19can literally watch you exercise through your
- 16:22phone camera. Oh, yeah. It will count your workout
- 16:24reps out loud. It acts as an interactive physical
- 16:26coach. It sees your posture, analyzes your form,
- 16:30and corrects it instantly. It can also translate
- 16:32live speech directly from a TV screen. It can
- 16:36speak up proactively if it sees something change
- 16:38in your environment. Right, like telling you
- 16:39your coffee has got to stop. It has genuine situational
- 16:42awareness. That is a massive leap forward from
- 16:45a text box. Think about standard AI voice modes
- 16:48today. If you have ever tried to interrupt one,
- 16:51you know how clunky it feels. Oh, it is terrible.
- 16:53It refuses to stop or it glitches. It breaks
- 16:56the illusion of intelligence immediately. TML
- 16:58is placing a massive contrarian bet here. They
- 17:01believe the future of AI isn't just about raw,
- 17:04escalating intelligence. No, it is deeply about
- 17:07latency, empathy, and conversational flow. Whoa.
- 17:11Beat it. Imagine scaling those 200 millisecond
- 17:15reactions to a billion queries. It requires an
- 17:18entirely different approach to server architecture.
- 17:20That is why the two -layer system is so crucial.
- 17:23Yeah. They are basically rewriting how data moves.
- 17:27But I have to ask a strategic question. If TML
- 17:30is betting everything on latency and flow rather
- 17:32than just raw smarts, can they actually survive
- 17:35as a standalone company? That is the million
- 17:37dollar question. Or is this just a preview of
- 17:39a feature that Google or OpenAI will eventually
- 17:42absorb? That is the ultimate Silicon Valley dilemma
- 17:44right now. Interaction models do create a powerful
- 17:47user mode. True. If users fall in love with that
- 17:50fluidity, they will not want to go back to clunky
- 17:53walkie -talkie AI. However, tech giants have
- 17:56immense distribution power. Right. Google already
- 17:59has the phones. Exactly. Google can push a software
- 18:01update to billions of Android phones overnight.
- 18:04TML has to build their user base from absolute
- 18:07scratch. That is a tough climb. They are pioneering
- 18:09the interface, but the giants are watching closely,
- 18:12and they have infinite capital to copy it. They
- 18:14either redefine the whole industry or get swallowed
- 18:17by a giant. fascinating it will be a thrilling
- 18:20narrative to watch unfold over the next year
- 18:22let us step back and look at the big picture
- 18:24how do all these disparate pieces fit together
- 18:26today we can trace A very clear through line
- 18:30across all these stories. AI is fundamentally
- 18:32moving from reactive text boxes to omnipresent
- 18:36real time partners. We see it in Google Androids,
- 18:40proactive widgets building themselves. We see
- 18:42it in the massive physical infrastructure being
- 18:45built globally to support these complex agents.
- 18:47Right. And we see it most clearly in TML's 200
- 18:50millisecond conversational bursts. The new metric
- 18:53for AI isn't just how smart it is on a test.
- 18:56Exactly. The metric is. how seamlessly it flows
- 18:59with human latency. It is about emotional intelligence
- 19:02and frictionless interaction. It is a profound
- 19:05shift in our relationship with technology. I
- 19:07want to leave you with a final thought to mull
- 19:09over. Two sec silence. We talked heavily about
- 19:11ending the awkward silence today. We talked about
- 19:13fluid, proactive interactions. If an AI can read
- 19:16your micro expressions perfectly. and interrupt
- 19:19you with the precise level of empathy in just
- 19:21200 milliseconds, at what point does talking
- 19:24to a machine become more comforting than talking
- 19:26to a human? Wow. Thank you for joining us on
- 19:29this deep dive. We will see you next time. U2RO
- 19:32Music.