Latest / Deep Values: Value Investing Deep Dives / What's Wrong, Google?
Transcript
- 0:00Imagine for a second that your company decides to spend $2.7
- 0:05billion. Right, which is just an
- 0:07astronomical amount of money. It's insane.
- 0:09And they don't spend it to acquire some hot new startup.
- 0:13And you know they aren't building a massive new data
- 0:15center. They spend $2.7 billion just to
- 0:18rehire one single employee who left a few years ago.
- 0:22Just one guy, yeah. Now imagine that same employee
- 0:25quitting again almost immediately.
- 0:27I mean it sounds like a joke, but that is the actual reality
- 0:29inside Google right now. Exactly.
- 0:31And it's not an isolated incident.
- 0:33So for everyone listening today, our mission for this deep dive
- 0:38is to unpack this a massive, industry shaking earthquake that
- 0:42is currently tearing through the absolute highest levels of
- 0:45Google's AI division. Yeah.
- 0:46It's a huge shift. Because we're winding the clock
- 0:48just a little bit, like to late 2025, Google's Gemini 3 was
- 0:52essentially on top of the world, right?
- 0:53Oh. Absolutely.
- 0:54It had overtaken Open AI. It was crushing Anthropic just
- 0:56topping all the benchmark leaderboards, right?
- 0:59But let's look at where we are today.
- 1:01In late 2026, Google's Gemini 3.5 Pro has missed what, 3
- 1:08separate release deadlines? Yeah, it slipped from June to
- 1:11July, and then it kind of kept falling off the pounder
- 1:13entirely. And in the midst of that delay,
- 1:162 massive pieces of news dropped simultaneously.
- 1:20Demas Savas stepped down from running Deepmind's daily
- 1:22operations, moving into a Chairman and Chief Scientist
- 1:25role. And on the exact same day Jeff
- 1:28Dean, who was a 27 year Google veteran, an absolute legend in
- 1:31the field, he announces he's leaving to start a completely
- 1:34new company. It is just a stunning reversal
- 1:36of fortune, honestly. It really is.
- 1:38Because when you look at the sheer concentration of
- 1:40intellectual capital walking out the door, I mean, it forces you
- 1:43to completely reassess the foundational stability of the AI
- 1:46landscape as we know it. We're going to dive deep into
- 1:49these secret internal strike teams and the catastrophic
- 1:52product blind spots in the second, but first, quick note on
- 1:55how you can track these exact market movements for yourself.
- 1:58To access more advanced researchers, follow Guru's
- 2:00latest movement lane and leverage our latest Deep Values
- 2:03agents. Welcome to download Deep Values
- 2:06APP from App Store to track stock market temperature.
- 2:10Welcome to download iOS app Buffett Indicator from Apple App
- 2:13Store or Bubble Watch from Google Play.
- 2:16Links in the description. Disclaimer All analysis is based
- 2:19on publicly available sources and may be subject to revisions
- 2:22or differing interpretations. All contents from this channel
- 2:25is for research and entertainment purposes only and
- 2:27does not constitute any investment or financial advice
- 2:30or service. Conduct your own due diligence
- 2:32before making investment financial decisions.
- 2:35OK, let's unpack this. So when the market sees this
- 2:37news, Google stock drops 4% in a single day.
- 2:42Beyond just the shock value, walk us through the mechanics of
- 2:44this executive shake up. Just how devastating is this
- 2:47talent drain? Well, to understand the market
- 2:49panic, we have to look at the specific capabilities of the
- 2:52people they're losing, you know, right?
- 2:54This isn't middle management turnover.
- 2:56This is the foundation of modern artificial intelligence packing
- 3:00up their desks. You mentioned Jeff Dean.
- 3:02Yeah. He isn't just leaving to retire,
- 3:04right? He took 3 core Google
- 3:07engineering leaders with him to found a new startup called
- 3:10Discovery Loop. But the exodus goes much deeper
- 3:14than that. How so?
- 3:15Well, David Silver, who was Demisosabus's early partner and
- 3:19essentially the pioneer of reinforcement learning.
- 3:22The technology that allowed AI to beat human champions at the
- 3:25game of Go, right? Exactly that tech he left to
- 3:28start a company called Ineffectable Intelligence.
- 3:31Wow. And I mean, we can't ignore the
- 3:33$2.7 billion man I mentioned earlier.
- 3:36Oh, right, Noam Shazir. I mean, he is one of the
- 3:38original authors of the famous Transformer paper, which is, you
- 3:41know, the underlying architecture for virtually every
- 3:44large language model in existence.
- 3:46Yeah, Google paid that astronomical sum to bring him
- 3:49back specifically to lead a specialized AI strike team.
- 3:52And he's gone again. He's.
- 3:54Gone, yeah, But perhaps the most conceptually fascinating
- 3:57departure is John Jumper. Right, the Nobel winner.
- 4:00Yeah, he won the 2024 Nobel Prize in Chemistry for his work
- 4:05on Alphafold, which completely revolutionized our understanding
- 4:09of protein structures in biology.
- 4:11And he just left to join Anthropic specifically to pursue
- 4:14something called agentic science.
- 4:16Wait, let me stop you right there, because that phrase gets
- 4:18thrown around a lot, but it rarely gets explained.
- 4:21What exactly is the mechanism behind agentic science?
- 4:25Like how does it actually differ from what Jumper was already
- 4:28doing with Alphafold at Google? What's fascinating here is the
- 4:31shift from passive tools to active researchers.
- 4:34So in the old paradigm, a human scientist would use an AI like
- 4:38alpha fold as a highly advanced calculator.
- 4:41OK, the human comes up with a hypothesis.
- 4:44The human feeds the chemical data into the AI, the AI
- 4:47predicts how the protein will fold, and then the human
- 4:50analyzes the result. So the human is doing all the
- 4:52heavy cognitive lifting? Exactly.
- 4:54Agentic science removes the human bottleneck entirely.
- 4:58You are building autonomous AI agents that can actually design
- 5:01the experiment, run the virtual simulations, evaluate the
- 5:05chemical binding results, and then they.
- 5:07Do all of that themselves. All of it, and then they
- 5:10automatically adjust their own hypothesis to run 1000 more
- 5:13experiments overnight. Jumper basically believes the
- 5:16future of biology or requires AI that acts like an independent
- 5:20scientist, not just a prediction software.
- 5:22So he's leaving to build an AI that can basically run a digital
- 5:26laboratory entirely on its own? Yes, that is wild.
- 5:30And capping all this off is Demisa Savas's new role, calling
- 5:34it a promotion to chairman and chief scientist.
- 5:36Sounds great on a press release, but in Silicon Valley terms,
- 5:39he's been kicked upstairs, hasn't he?
- 5:40Oh, 100%. Like they gave him a pristine
- 5:42title just to sideline him from daily operations.
- 5:45It's like a legendary rock band breaking up right before a
- 5:48massive stadium tour. That's a great analogy actually.
- 5:51But here's what I really don't understand, and I think a lot of
- 5:54people listening are probably wondering this too.
- 5:57Google has a near infinite war chest.
- 6:00They have oceans of data, sprawling server farms, custom
- 6:04chips. Infinite money, basically.
- 6:06Right. So how do you lose your top
- 6:08three lead guitarists in a matter of months when you can
- 6:12offer them literally any resource they could possibly ask
- 6:15for? Because it fundamentally isn't
- 6:17about the money or the compute access anymore.
- 6:20It is about a deep philosophical disagreement on the mechanical
- 6:24trajectory of artificial intelligence.
- 6:26Meaning what exactly? Well, for the last several
- 6:28years, the dominant paradigm across the industry has been
- 6:31scaling the belief that if you just throw more text data and
- 6:35more computing power at a large language model, it will linearly
- 6:38get smarter. Just bigger is better.
- 6:40Right. But these starting scientists,
- 6:42they are looking at the math and realizing that brute force
- 6:45scaling is hitting a wall. We have effectively scraped the
- 6:48entire Internet. So we are running out of high
- 6:50quality human data to train these models on.
- 6:54So they were jumping ship because they realized the
- 6:57massive cargo vessel they built is steering directly into a dead
- 7:01end and they want to build speed boats instead.
- 7:03Yeah, that's it, perfectly. They want to move away from just
- 7:06predicting the next word in a sentence and shift toward
- 7:09reinforcement learning and autonomous agents that can
- 7:12figure things out without needing a human textbook.
- 7:14Wow. Which brings up the actual
- 7:16mechanism that catalyzed this entire corporate shake up in the
- 7:201st place. You know Google as massive,
- 7:22historically catastrophic blind spot regarding coding agents.
- 7:26Right. We know that while Google's
- 7:29leadership was debating the philosophical path to artificial
- 7:32general intelligence, anthropic and open AI unlocked a
- 7:36multibillion dollar market almost overnight with coding
- 7:39agents. Yeah, they just scooped the
- 7:40market. But I want to make sure we don't
- 7:42understate this for you listening.
- 7:43We are just talking about an AI that helps a programmer write a
- 7:46Python script faster, right? We're talking about code as the
- 7:50ultimate API for the real world. That is the crucial distinction.
- 7:54Coding agents are the foundational gateway to
- 7:56automating enterprise knowledge work.
- 7:58Because it's not just making an app, right?
- 8:01If an AI model can reliably generate, debug, and execute
- 8:04complex code, it can be adapted to navigate complex corporate
- 8:08workflows. It can pull data from a legal
- 8:11database formatted into a financial model and e-mail a
- 8:15summary to ACEO. Code is just the language of
- 8:19modern business infrastructure. Exactly.
- 8:21So whoever controls the best coding agent essentially holds
- 8:25the skeleton key to enterprise automation.
- 8:28But Demisysabis historically hated the entire concept of
- 8:32large language models. He famously dismissed them as
- 8:35ungrounded language games. Yeah, He looked at ChatGPT and
- 8:38essentially said, you know, this is just a sophisticated parrot.
- 8:41It doesn't actually understand reality.
- 8:43Right. Because of that bias, Google
- 8:46heavily prioritized multimodal world models instead.
- 8:49Which is a huge difference structurally.
- 8:51Yeah, explain that for a second. Well, a world model isn't just
- 8:54trying to predict text, it is trying to process video, audio,
- 8:58and physics simultaneously to build a spatial understanding of
- 9:01reality. OK.
- 9:01That makes sense. Hasabis believed that True AGI
- 9:04had to interact with environments, much like how his
- 9:07early Project Alpha Star learned intelligence by actually playing
- 9:11the video game StarCraft. I set a logic there.
- 9:13Honestly, if you want an AI to understand gravity, having it
- 9:18watch a video of an apple falling is fundamentally
- 9:22different than having it just read the word gravity in a text
- 9:24file. Absolutely.
- 9:25It's a much deeper understanding.
- 9:27But by obsessing over that physics based multimodal future,
- 9:31they completely missed the text based billion dollar opportunity
- 9:35of coding agents sitting right in front of them.
- 9:38And that strategic miss triggered an absolute panic at
- 9:40the highest levels of alphabet. The sources outline a scenario
- 9:44in April 2020 Sixth that is almost hard to believe.
- 9:47Oh yeah, Sergey Brin thing. Yes, Google Co founder Sergey
- 9:50Brin literally comes out of retirement.
- 9:53He leaves his yacht, starts showing up at the Mountain View
- 9:55office three days a week, and forms an internal strike team
- 9:58specifically designed to build a world class coding agent.
- 10:02And he doesn't ask Kasavas to run it.
- 10:03No, he doesn't. He taps Corey Kavukulglu, who is
- 10:07Deepmind's CTO, Hasavas's own deputy.
- 10:09Yeah, it's brutal. I want you listening to put
- 10:12yourself in Hasavas's shoes for a second.
- 10:14You founded DeepMind. You practically willed the
- 10:19modern AI movement into existence.
- 10:21You employ Nobel Prize winners, right?
- 10:24But suddenly your boss's boss comes out of retirement, taps
- 10:27your deputy on the shoulder and says, you work for me now, and
- 10:31we're building the exact product you told us was a waste of time.
- 10:35I mean, the internal politics of that must be incredibly toxic.
- 10:38It really is a textbook manifestation of the innovator's
- 10:41dilemma. The underlying mechanism here is
- 10:43just corporate paralysis. Me neither, too big to move
- 10:46quickly. Exactly.
- 10:47Google has consistently struggled to pivot its
- 10:49architecture quickly because they are terrified of
- 10:53cannibalizing their own cash cow, which is the traditional
- 10:56search engine. Right, because if autonomous
- 10:58agents just do the task for you, nobody is clicking on Google
- 11:01sponsored ad links. Exactly, and furthermore they
- 11:04are terrified of reputational damage.
- 11:06Remember the BRDRN launch years ago?
- 11:08That. Was a disaster.
- 11:09Yeah, the AI hallucinated a single fact during a public demo
- 11:13and the stock tanked 7 to 8% instantly.
- 11:16That trauma is just baked into their organizational DNA now.
- 11:19So they are structurally too large to take risks, but in the
- 11:23AI sector, if you aren't risking failure, you are literally
- 11:27dying. Right?
- 11:28Here's where it gets really interesting, though.
- 11:30When you dig into the sources, you find out this cultural
- 11:33friction isn't actually a new phenomenon.
- 11:35No, not at. All DeepMind actually tried to
- 11:37engineer an escape route from Google's corporate structure
- 11:40years ago, didn't they? Yeah, it's a wildly fascinating
- 11:43piece of secret history called Project Mario.
- 11:46When Google was restructuring itself into the Alphabet holding
- 11:48company, DeepMind saw a massive window of opportunity.
- 11:53OK Hassabis and his Co founder Mustafa Suleiman secretly
- 11:56negotiated a deal with Sergey Brin to split off entirely.
- 12:00The blueprint was for DeepMind to take $15 billion billion with
- 12:04AB and operate as a completely independent pure research entity
- 12:08outside of alphabets commercial demands.
- 12:10Wow, just imagine having the leverage to say, hey, give us
- 12:14$15 billion and leave us entirely alone so we can do our
- 12:17science experiments. Right.
- 12:18And Brynn was actually on board with the structural split.
- 12:21Really. Yeah.
- 12:22But at the final hour, CEO Sundar Pichai vetoed the entire
- 12:26deal. He looked at the landscape,
- 12:28realized AI was the existential future of the company, and knew
- 12:32he just couldn't let his crown jewel walk out the door.
- 12:35That must have been a huge blow to the deep mind guys.
- 12:38Oh, the resulting fallout was incredibly messy.
- 12:42Suleiman was later quietly pushed out after some HR
- 12:44complaints surfaced, which, you know, we're very conveniently
- 12:48leaked to the financial press. DeepMind was essentially chained
- 12:51to the Google Machine. And when you connect that
- 12:54historical desperation for independence to Jeff Dean's
- 12:56departure today, the pattern just becomes undeniable.
- 13:00Exactly. Dean is leaving to start
- 13:02Discovery Loop because he believes the era of massive
- 13:05thousands of people engineering teams is structurally obsolete.
- 13:09His core thesis is that the future of AI for scientific
- 13:12discovery requires incredibly small, hyper agile teams.
- 13:17We're talking fewer than 10 people. 10 people.
- 13:20That's tiny. Yeah, he wants to utilize a
- 13:22highly specific machine learning method called rubric rewarding
- 13:26combined with autonomous agents. Let's breakdown the mechanics of
- 13:30that because I know we have reinforcement learning from
- 13:32human feedback where humans grade AI outputs.
- 13:36How does rubric rewarding actually work without needing
- 13:391000 engineers in the loop? Well, think of it like giving an
- 13:43AIA highly specific mathematical grading rubric, setting it loose
- 13:47in a simulated environment, and letting it run millions of
- 13:50experiments autonomously. OK, so you don't need a massive
- 13:53team to write step by step code. Right, you just need a few
- 13:56brilliant scientists to set the objective function.
- 13:59For example, maximize the thermal efficiency of this new
- 14:02material alloy. Got it.
- 14:04And then the AI generates thousands of chemical
- 14:06structures, test them against the physics engine, and grades
- 14:09its own success based strictly on the rubric you provided.
- 14:13The AI does all the heavy lifting, correcting its own
- 14:16neural pathways based on those automated rewards.
- 14:18So a team of five scientists using rubric rewarding can
- 14:21potentially out innovate a team of 500 traditional software
- 14:24engineers. Easily.
- 14:26But I have to challenge this for a second.
- 14:27Why couldn't Jeff Dean just build a ten person team inside
- 14:33Google to do exactly that? I mean they have divisions for
- 14:36everything. Because of the hardware and
- 14:37infrastructure mismatch, Google has spent the last decade of
- 14:41building TPU tensile processing units.
- 14:44These are custom, highly specialized chips designed
- 14:47fundamentally to run Google's ads, search algorithms and
- 14:50massive text models efficiently. OK, so they're optimized for
- 14:53text and search? Exactly, but for pure scientific
- 14:56discovery running complex physics simulations and rubric
- 14:59rewarding environments, standard GPU's like the ones NVIDIA
- 15:03architects are actually far superior.
- 15:05Wow, so he was literally on the wrong hardware.
- 15:07Yeah, Dean recognized he was trapped in a company that is
- 15:11structurally, culturally and physically hardware acquired for
- 15:14commercial software, while he wants to cure diseases and
- 15:17discover new laws of physics. The irony there is just
- 15:20incredibly thick. Google spent billions acquiring
- 15:23the greatest scientific minds of a generation to secure its long
- 15:26term future. But by forcing them into the
- 15:29trenches of a commercial chatbot war, they alienated the exact
- 15:34generational talent they paid all that money to acquire.
- 15:37It's tragic in a corporate sense.
- 15:40But Google isn't just losing talent to local Silicon Valley
- 15:42rivals. This internal paralysis is
- 15:45happening at the exact moment the global AI landscape is
- 15:48completely fracturing. Oh.
- 15:50Absolutely. And before we get into this next
- 15:51part, we should probably just explicitly state to everyone
- 15:54listening, right? We're just impartially reporting
- 15:57the geopolitical tech race analysis that's provided in our
- 16:00sources, right? Yeah.
- 16:01We're not endorsing any political viewpoints here.
- 16:03Exactly. We're just conveying the
- 16:04materials observations on market dynamics.
- 16:07Just following the data. Right.
- 16:08And the data from Q2 and Q3 of 2026 shows a dramatic, almost
- 16:13unthinkable shift. Chinese open source models,
- 16:16specifically from companies like DeepSeek, Kimi and Zipu, have
- 16:20reached the trillion parameter scale.
- 16:21It's a massive milestone. Yeah, the performance gap, which
- 16:25used to look like a multi year insurmountable Canyon has closed
- 16:29to just a matter of months. And this rapid convergence has
- 16:32triggered a massive change in how AI applications are actually
- 16:35being engineered at the foundational level.
- 16:38U.S. Companies including highly
- 16:40valued startups like the coding assistant Cursor and the Legal
- 16:43AI Cool Harvey are now heavily utilizing what's called a
- 16:47routing strategy in their API gateways.
- 16:50This mechanism is brilliant when you really look at it it it's
- 16:54exactly like building a house. You don't hire a world renowned
- 16:57master architect and pay them $1000 an hour to come and hammer
- 17:00in every single nail on the floorboards, you know?
- 17:03That would bankrupt you. Exactly.
- 17:05You hire affordable, highly capable contractors for the
- 17:08baseline structural work and you only bring in the master
- 17:11architect for the complex bespoke design challenge.
- 17:14That is the perfect analogy for API routing.
- 17:17These U.S. companies have built dynamic systems that evaluate
- 17:20the complexity of a user's prompt the millisecond it is
- 17:23typed. OK, So what happens then?
- 17:24Well, they're secretly routing the easy baseline queries, like
- 17:28summarizing a short document or fixing a basic code syntax error
- 17:33to these highly capable, incredibly cheap Chinese open
- 17:36source models. Which saves the startups massive
- 17:39amounts of money on compute costs.
- 17:40Exactly. Then when a user asks a highly
- 17:44complex reasoning question that requires deep logic, the router
- 17:48redirects that specific query to an expensive top tier US model
- 17:53like Anthropics Claude. It is wild to think about the
- 17:56infrastructure required for that.
- 17:58For you listening right now, the AI applications you use on your
- 18:01phone are likely ping ponging your request back and forth
- 18:04across the globe, balancing between US and Chinese models in
- 18:07real time, all in fractions of a second.
- 18:10All just so the app developer can save half a cent on compute
- 18:12costs per query. It's incredible.
- 18:15And if we connect this to the bigger picture, the sources
- 18:17indicate we've effectively reached a global tie, but it is
- 18:20a tie with highly asymmetrical strength.
- 18:23Meaning they're good at different things.
- 18:24Right, the US still holds the definitive lead in raw data
- 18:28center computing power and the absolute peak performance
- 18:31capabilities of Frontier models. But the data shows China is
- 18:35leveraging A vastly cheaper energy grids, rapidly iterating
- 18:39on open source architectures, and deploying these AI models
- 18:42across a much more diverse range of physical industries like
- 18:45advanced manufacturing and robotics.
- 18:48And objectively speaking, this dual Track World might actually
- 18:51be a net positive for the acceleration of the technology
- 18:54itself. Really.
- 18:55How So? What?
- 18:56When you have two radically different ecosystems operating
- 18:59with different structural advantages, it creates a global
- 19:03system of checks and balances. It prevents a single corporate
- 19:06monopoly from dictating the architectural future of AI that.
- 19:09Makes a lot of sense. Historically, this kind of
- 19:12fierce, distributed, and somewhat fractured competition
- 19:15acts as the ultimate accelerant for global innovation.
- 19:18Wow. We have covered incredible
- 19:20ground today. We really have.
- 19:21We started by looking at the staggering $2.7 billion
- 19:25executive shake up at Google, with Demas Esabas getting kicked
- 19:29upstairs and Jeff Dean leaving his 27 year tenure to build
- 19:32speedboats. We unpacked the internal tug of
- 19:35war, diving into the historical secrets of Project Mario and the
- 19:39technical mechanisms behind the coding agent blind spot.
- 19:43And finally, we zoomed out to see how this internal corporate
- 19:46drama is playing out against the backdrop of a shifting global
- 19:49balance of power, where open source API routing is
- 19:53fundamentally changing the economics of how startups
- 19:55operate. It really highlights how fragile
- 19:58even the most well funded tech empires can be.
- 20:01You know you can have all the money and server racks in the
- 20:04world, but when the underlying scientific paradigm shifts
- 20:08beneath your feet. If you can't adapt structurally,
- 20:11your talent will simply leave to build the future somewhere else.
- 20:14And as we wrap up, there is one final almost haunting thought
- 20:18for the sources that I want to leave you with.
- 20:20It goes back to David Silver, reinforcement learning pioneer
- 20:23who just left DeepMind. He recently co-authored A
- 20:26fascinating paper with his old professor titled Welcome to the
- 20:30Era of Experience. Yes, this raises an incredibly
- 20:33important question about the limitations of human knowledge.
- 20:35What's the core argument? Their premise is that AI has
- 20:39essentially hit the data wall. It has exhausted all the useful
- 20:44human generated text on the Internet.
- 20:46So if an AI model only learns from human data, it is
- 20:50mathematically constrained by human limitations.
- 20:52It will only ever be as smart as the smartest person.
- 20:55The next phase, this era of experience, requires AI to
- 20:58generate its own synthetic data by interacting with
- 21:01environments, running its own simulations, and learning from
- 21:04its own unique experiences without human feedback.
- 21:08O what does that mean for us? If scientists like Jeff Dean,
- 21:12John Jumper, and David Ilver are leaving traditional tech giants
- 21:15to build autonomous agents that learn from their own synthetic
- 21:18rubric rewarded experiences, are we about to enter a reality
- 21:22where AI isn't just regurgitating human knowledge
- 21:25anymore? That's the implication.
- 21:27Yeah. Are we on the verge of AI
- 21:28actually discovering new chemical compounds, new laws of
- 21:31physics, and entirely new conceptual ideas that human
- 21:34beings have never even conceived of?
- 21:35We started this deep dive talking about a massive
- 21:38corporate machine losing its lead engineers.
- 21:40But maybe the real story here is that those engineers didn't just
- 21:45leave to build a new product, they left to birth a completely
- 21:49alien kind of intelligence. It is a profound structural
- 21:52shift in the relationship between humans and machines.
- 21:56We are actively moving from an era where AI acts as an
- 21:59archivist of human history to an era where AI is an active,
- 22:03independent participant in scientific discovery.
- 22:06That is definitely something for you to Mull over the next time
- 22:08you use a chat bot just to write a polite e-mail for you.
- 22:11Thank you so much for joining us on this deep dive.
- 22:13Keep questioning the world around you and we'll see you
- 22:15next time.