Latest / AI Fire Daily / 🎙️ EP 268: The Demis Hassabis Fixation & Anthropic Surpasses OpenAI
Transcript
- 0:00A single recruitment decision beat. It ended
- 0:03a legendary billionaire friendship. It sparked
- 0:07a 10 -year silence. Think about that for a second.
- 0:10We assume the future is built on pure science.
- 0:13On logic. Exactly. We assume the trajectory of
- 0:16artificial intelligence is driven by cold, hard
- 0:18math and objective algorithms. But sometimes
- 0:21the entire timeline of human innovation hinges
- 0:24on, well, on one deeply personal betrayal. It
- 0:29is a really messy reality. I mean, you look at
- 0:31the neural networks and the staggering compute
- 0:32clusters, and it all feels so systematic. But
- 0:35behind the code, you just have people. You have
- 0:37incredibly powerful, deeply flawed individuals
- 0:40making decisions based on fear and pride. Welcome.
- 0:45You're joining us for a deep dive into the raw
- 0:47reality of AI today. No fluff. Not at all. We
- 0:50are mapping out a fascinating journey through
- 0:52a stack of newly unsealed documents, market data,
- 0:55and engineering specs. We are moving from the
- 0:58bruised egos and boardroom paranoia of the AI
- 1:00elite straight into the physical footprint of
- 1:03these models. Which is huge. It is. And that
- 1:05footprint is expanding from gas -guzzling data
- 1:08centers in Mississippi all the way into literal
- 1:12orbit. And we're wrapping up with a massive kind
- 1:16of quiet shift in the market. The loudest voices
- 1:19dominating the headlines are not actually the
- 1:21runs winning the enterprise war. Okay, let's
- 1:23unpack this. We have to start with the newly
- 1:25revealed legal filings from the ongoing lawsuit
- 1:28between Elon Musk and OpenAI. Legal discovery
- 1:31is always illuminating because it strips away
- 1:34all the PR messaging. Right. We are finally getting
- 1:37a raw window into their internal communications.
- 1:40from around 2018. I was reading through those
- 1:43documents and a very specific theme emerges immediately.
- 1:46There is this deep -seated, almost obsessive
- 1:49preoccupation with Demis Hassabis. The CEO of
- 1:52Google DeepMind. Yeah. At the time, DeepMind
- 1:54had just conquered the game of Go. Yeah. They
- 1:56were achieving technological leaps that most
- 1:58experts thought were decades away. And internal
- 2:01communications show that Musk and his inner circle
- 2:03viewed Hassabis as a literal threat. Like they
- 2:06didn't just see a business rival. No. They actively
- 2:08believed he endangered That level of existential
- 2:17dread fundamentally alters how a company operates.
- 2:21I mean, it shifted their entire strategy from
- 2:23a methodical research pace into an absolute sprint.
- 2:28That fear destroyed major personal relationships.
- 2:31Yeah, Musk actually testified about the fallout.
- 2:33He recognized that to counter DeepMind, he needed
- 2:37the absolute best talent on the planet. So he
- 2:39recruited a brilliant researcher named Ilya Sutskiver.
- 2:43Right. He poached him directly from Google. And
- 2:45that specific poaching incident was the breaking
- 2:47point. It completely ended Musk's friendship
- 2:50with Larry Page. Page tried desperately to keep
- 2:53Sutskiver. Sergey Brin tried. Hassabis tried.
- 2:56They threw everything at him. They really did.
- 2:58They threw everything at him to stop him from
- 3:00leaving. But Sud Skiver chose the vision at OpenAI,
- 3:03and the fallout was immediate. Larry Page felt
- 3:06completely betrayed. He stopped speaking to Musk
- 3:08entirely, and that silence has lasted for over
- 3:11a decade. Wow. Two titans of the tech industry,
- 3:14completely estranged over a single hire. It's
- 3:17like stacking Lego blocks of data, but the architects
- 3:20are burning down the house fighting over the
- 3:22instruction manual. Two sec silence. That's a
- 3:25brilliant way to picture it. The math is perfect,
- 3:27but the architects are deeply unstable. And what
- 3:30becomes apparent in these documents is the stark
- 3:33contrast in what was driving these architects.
- 3:36Right. Musk was entirely consumed by the threat
- 3:39of a Google monopoly on intelligence. He wanted
- 3:42to shatter that monopoly. But when you look at
- 3:44Sam Altman's private messages from the same period,
- 3:47you see a completely different psychological
- 3:49drive. Altman was analyzing corporate moats.
- 3:52His messages reveal a meticulous pursuit of industry
- 3:55dominance. Like he was focused on the actual
- 3:57structure of power. Securing compute. Yes, securing
- 4:00compute and positioning open AI as the gatekeeper
- 4:03of the next computing paradigm. And that boardroom
- 4:05paranoia is still dictating the company's structure
- 4:08today. Even in late 2023, open AI executives
- 4:12were sending frantic messages. Mira Marotti was
- 4:15emailing Microsoft leadership, actively panicking
- 4:18about losing their top researchers back to Hassabis.
- 4:20Right, and pleading for more computing resources
- 4:23just to keep their talent happy. The ghost of
- 4:25DeepMind never actually left their operations.
- 4:28B, I have to ask you something about this dynamic.
- 4:31Is the race toward artificial general intelligence
- 4:34actually being driven by science or just the
- 4:36fragile egos of billionaires? That raises a critical
- 4:40question about the foundation of this technology.
- 4:42The algorithms themselves are undeniably rigorous.
- 4:46But the timeline, the aggressive funding rounds,
- 4:50and the ruthless poaching, that is entirely driven
- 4:53by human flaws. It is propelled by a paranoid
- 4:57need to be first, regardless of the scientific
- 4:59readiness. So billionaire bruised egos are literally
- 5:02steering the future of human technology. They
- 5:05absolutely are. And those bruised egos demand
- 5:07more than just market share. They require a staggering
- 5:10amount of physical infrastructure. They need
- 5:12compute power today, which leads to drastic,
- 5:14often dirty physical measures. Because these
- 5:16billionaires cannot wait five years to build
- 5:18sustainable infrastructure, which brings us to
- 5:21the actual physical reality of this technological
- 5:23arms race. Right. The energy requirements for
- 5:26training frontier models are hitting a hard physical
- 5:29wall. You cannot just plug 100 ,000 GPUs into
- 5:32a standard city grid. Look at what Elon Musk's
- 5:35XAI is doing right now. They are running a massive,
- 5:38colossus data center down in Memphis and Mississippi.
- 5:42Yeah. To power it quickly, they utilized a regulatory
- 5:45loophole. They bypassed the years -long wait
- 5:48for grid interconnects by bringing in nearly
- 5:5050 massive gas turbines. 50 gas turbines burning
- 5:53fossil fuels around the clock just to push electricity
- 5:56into a neural network. The local pushback has
- 5:58been severe. Community groups are protesting
- 6:01because the sheer volume of emissions is drastically
- 6:04worsening the local air quality. Understandably.
- 6:06Yeah. It is a brute force engineering solution
- 6:08to a grid that simply cannot support the AI timeline.
- 6:11It exposes a massive structural bottleneck. The
- 6:14planet's transmission lines take a decade to
- 6:17upgrade, but these models are doubling in size
- 6:19every few months. And that bottleneck is directly
- 6:22fueling an explosion in specialized hardware.
- 6:25Companies realize that brute forcing energy isn't
- 6:28sustainable long term. There is a London -based
- 6:30startup called Fractile that just closed a $220
- 6:34million funding round. That is massive capital
- 6:37for a hardware startup trying to compete with
- 6:40NVIDIA. I mean, NVIDIA still dominates this entire
- 6:43sector with a $5 .49 trillion market cap. Fractile
- 6:48is taking a different approach. entirely focused
- 6:50on building faster AI inference hardware. They're
- 6:54trying to make the end user experience drastically
- 6:56cheaper and faster. And for you listening to
- 6:59this right now, when we talk about inference
- 7:00hardware, meaning the specific computer chips
- 7:02that run in AI after it's trained, that's really
- 7:04where the everyday battle is fought. Training
- 7:06a model is like sending a kid to college. Inference
- 7:08is that kid going to work every single day. Beautifully
- 7:11put. The demand for that daily computation is
- 7:14skyrocketing, which leads us to the ultimate,
- 7:16almost science fiction escalation in this arms
- 7:18race. I was reading the engineering proposals
- 7:21for this, and it is staggering. Google and SpaceX
- 7:24are seriously considering moving AI servers into
- 7:27actual orbit. It sounds absurd until you look
- 7:30at the thermodynamics. Musk argues that orbit
- 7:33is the fastest way to scale compute without fighting
- 7:35terrestrial power companies. Google is already
- 7:39planning prototype data satellites by 2027. Data
- 7:42centers on Earth spend roughly half their energy
- 7:45budget just on cooling systems. Right. In deep
- 7:47space. you have access to the endless ambient
- 7:50cold of the vacuum for heat dissipation. Furthermore,
- 7:54you have unmetered constant solar energy. You
- 7:57completely bypass the planetary power grid. Whoa!
- 8:00Imagine scaling to a billion queries from literal
- 8:03space. It completely shatters our traditional
- 8:05understanding of digital infrastructure. You
- 8:08beam the processing requests up, the orbital
- 8:10cluster performs the inference, and it beams
- 8:12the result back down. But looking at the Mississippi
- 8:15gas turbines and the plans for orbital servers,
- 8:17are we fundamentally running out of resources
- 8:19on Earth to sustain this arms race? If we connect
- 8:22this to the bigger picture, the answer is yes.
- 8:25Our physical reality grows linearly. We can only
- 8:29pour concrete and string copper wire so fast.
- 8:33But algorithmic demand is compounding exponentially.
- 8:36Space solves the thermodynamic and energy constraints
- 8:39of Earth, assuming the launch economics hold
- 8:42up. Got it. We're outgrowing the planet's power
- 8:45grid, forcing this wild hardware race. Precisely.
- 8:48And all of this massive infrastructure, from
- 8:50gas turbines to space servers, exists to power
- 8:53a new generation of tools. And these tools require
- 8:56an incredible amount of behavioral data to function.
- 8:59Because we are shifting away from simple chatbots
- 9:01toward autonomous agents, agents that need to
- 9:04know exactly how you navigate your daily workflows.
- 9:07Which lands this technology squarely in the regulatory
- 9:09crosshairs. Yeah. The data collection required
- 9:11to build these agents is becoming aggressively
- 9:13invasive. Let's examine the internal culture
- 9:15war happening at Meta right now. employees recently
- 9:17organized a widespread internal protest over
- 9:20new data collection policies. They discovered
- 9:23the company had deployed internal mouse tracking
- 9:25tools. They were logging the micro -movements
- 9:27of employee mouse cursors to train future AI
- 9:31agents. To build a large action model, the AI
- 9:34doesn't just need text. It needs to know how
- 9:37a human physically navigates a user interface.
- 9:40The employees were deeply unsettled. I mean,
- 9:42it represents a profound shift in corporate surveillance.
- 9:45An algorithm is quietly mapping your hesitation,
- 9:48your scrolling speed, your exact workflow patterns.
- 9:51It feels incredibly dystopian. You're just trying
- 9:53to clear your inbox. And an invisible architecture
- 9:56is learning how to replicate your physical inputs.
- 9:58Meta is facing intense friction everywhere right
- 10:02now. Over in Europe, they are navigating a regulatory
- 10:04minefield with the Digital Markets Act. They
- 10:07are desperately trying to dodge a potential multi
- 10:10-billion dollar fine from the EU regarding their
- 10:13closed ecosystems. Their strategy is actually
- 10:15quite clever. To avoid being labeled an illegal
- 10:18gatekeeper, Meta is temporarily opening up WhatsApp.
- 10:21Right. They're allowing free direct access to
- 10:23rival AI chatbots like ChatGPT and Claude right
- 10:27inside the WhatsApp interface. They are essentially
- 10:29turning their most guarded messaging app into
- 10:32an open port. It is a wild concession, but it
- 10:36highlights how crowded and desperate this ecosystem
- 10:39is getting. Everyone is fighting for space on
- 10:41your devices. Notion recently announced a massive
- 10:44platform pivot. They're no longer just a workspace.
- 10:48No. They're a full hub for AI agents that interact
- 10:51with live company databases. Then you have the
- 10:53explosion of ambient hardware. There is a new
- 10:56device called the Momokit Gem. It is a tiny all
- 10:59-day AI wearable. It passively captures your
- 11:01meetings, your phone calls, and your casual conversations.
- 11:04It runs localized inference to record and analyze
- 11:07your daily decisions. The developer tools are
- 11:10getting equally invasive. A platform called Latitude
- 11:12now traces every single session inside Claude
- 11:15Code. It maps exactly where developers are burning
- 11:17tokens so companies can audit the thought process
- 11:20of the AI. We are also seeing the rise of self
- 11:23-building tools like CraftBot. It allows a general
- 11:25AI agent to write, test, and evolve custom internal
- 11:28applications entirely on its own. Even traditional
- 11:31hardware is adapting to the surveillance layer.
- 11:34The new Google book launching this fall is built
- 11:37entirely around native Gemini intelligence, constantly
- 11:40reading screen context. The integration into
- 11:43our lives is becoming completely frictionless.
- 11:45But the reality of that friction disappearing
- 11:48is what worries me. I see the utility in a system
- 11:51that knows my preferences perfectly. But is the
- 11:53sheer convenience of these pervasive AI tools
- 11:56blinding us to the reality of constant workplace
- 11:59surveillance? This raises one of the most important
- 12:02questions of this decade. The tradeoff is absolute.
- 12:05We all want bespoke AI assistance. We want an
- 12:09agent that preemptively drafts the perfect email
- 12:11because it knows exactly how we communicate.
- 12:14Right. But to achieve that level of personalization,
- 12:17the AI requires total behavioral transparency.
- 12:19It has to watch you work. We are voluntarily
- 12:22trading our behavioral autonomy just to save
- 12:24a few keystrokes a day. Right. We're normalizing
- 12:27intense surveillance just to build slightly smarter
- 12:29digital assistants. And corporate adoption is
- 12:31the primary engine forcing this normalization.
- 12:34Businesses are eager to deploy these agents.
- 12:36Which brings us to the actual market reality.
- 12:39We're going to dive into exactly which businesses
- 12:41are adopting what right after this quick break.
- 12:44Mid -roll sponsor, Reed Placeholder. And we are
- 12:48back. We have spent the first half of this deep
- 12:51dive looking at the noisy, headline -grabbing
- 12:53chaos. We explored Elon Musk's paranoia, Meta's
- 12:57internal tracking protests, and OpenAI's boardroom
- 13:01drama. It is very easy to get distracted by the
- 13:03billionaire drama. It is loud by design. But
- 13:07when you strip away the noise and look at the
- 13:08actual enterprise deployment data... A completely
- 13:11different narrative emerges. Someone else is
- 13:13quietly winning the business war. The telemetry
- 13:16data tells a story of hyper -focused execution.
- 13:18We just received the fresh May 2026 enterprise
- 13:22software data from Ramp, and it reveals a massive
- 13:25fundamental shift in the market hierarchy. Anthropic
- 13:28has officially surpassed OpenAI in verified business
- 13:31customers. That is a staggering milestone. The
- 13:34data shows that 34 .4 % of businesses are now
- 13:37paying for Anthropic's enterprise tier. OpenAI
- 13:40has steadily dropped down to 32 .3%. OpenAI certainly
- 13:44maintains a volume lead in broader non -technical
- 13:46consumer applications. Sure. But Anthropic has
- 13:49captured the high -value, high -adoption enterprise
- 13:52sectors. They have completely locked down finance,
- 13:56technology, and professional services. And if
- 13:58you look at developer platforms like the open
- 14:00router Leaderboard, Anthropic has consistently
- 14:03ranked above OpenAI since December of 2025. Ramp
- 14:08economists attribute this victory to a very deliberate
- 14:11technical first strategy. While OpenAI was building
- 14:14consumer -friendly voice modes and image generators,
- 14:17Anthropic ignored the broad consumer market.
- 14:20They focused entirely on developers and specialized
- 14:23technical professionals. And you can see the
- 14:25results in their feature rollouts. Look at Claude
- 14:28Code. They just upgraded it with a native multi
- 14:31-agent view. Which is a massive paradigm shift
- 14:33for software engineering. Yeah. Instead of a
- 14:35developer prompting one AI to write code, the
- 14:38multi -agent view allows an entire fleet of specialized
- 14:40agents to check each other's work, run tests,
- 14:43and debug in a continuous loop. Professional
- 14:45services are deploying tools like Cloud Cowork
- 14:48for incredibly dense, complex document analysis.
- 14:51And in the finance sector, the demand for precision
- 14:54is absolute. Finance has a massive demand for
- 14:57a large context window size. It means exactly
- 15:00how much text the AI can remember at once. If
- 15:03a financial analyst feeds a 500 -page corporate
- 15:06prospectus into a model, the AI cannot drop the
- 15:09thread. No. It cannot hallucinate a single digit
- 15:12on page 400 because it forgot the parameters
- 15:15set on page one. And I have to admit, I still
- 15:17wrestle with prompt drift myself. I'll start
- 15:20a long, complex research conversation with the
- 15:22chatbot, and an hour later, it completely forgets
- 15:25my initial formatting instructions. It starts
- 15:27hallucinating broad summaries. It is the most
- 15:30frustrating bottleneck in consumer AI right now.
- 15:32The models degrade over long sessions. It completely
- 15:36ruins professional workflows. So it makes total
- 15:38sense why enterprise users are abandoning general
- 15:41chatbots. They are migrating to Anthropic because
- 15:43the structural integrity of the long context
- 15:45memory is simply superior. Corporate clients
- 15:47do not care about a model. Does this mean the
- 15:57era of the general purpose chatbot is already
- 15:59dead in the corporate world? I believe that phase
- 16:02of the market has permanently closed. Businesses
- 16:05demand precision execution. They have no use
- 16:08for conversational parlor tricks in a boardroom.
- 16:10If a model cannot reliably parse a massive legal
- 16:13contract or autonomously debug thousands of lines
- 16:17of legacy code without supervision, it offers
- 16:19zero enterprise value. They need focused utility,
- 16:23not a generic talking companion. Basically, businesses
- 16:26want a specialized execution tool, not a jack
- 16:30-of -all -trades chatbot. Exactly. The enterprise
- 16:32market has finally matured past the novelty phase.
- 16:35This has been a wild journey through the stack
- 16:37today. We covered a tremendous amount of ground,
- 16:39from orbital cooling systems to pixel -level
- 16:42mouse tracking. If we connect this to the bigger
- 16:44picture, a profound irony emerges from all this
- 16:46data. The loudest voices in this industry, the
- 16:49pioneers like Musk and Altman. are fighting bitterly
- 16:52over theoretical power. They're burning bridges,
- 16:54building 50 -turbine power plants, and planning
- 16:57servers in space to win a perceived monopoly.
- 17:00But while they wage this loud, chaotic war fueled
- 17:04by ego, they're being quietly outflanked. They're
- 17:07losing the actual, highly lucrative business
- 17:10sector to Anthropic's quiet, hyper -focused,
- 17:13drama -free execution. The enterprise market
- 17:16ultimately rewards utility, not paranoia. So
- 17:20what does this all mean, Beat? Think about this
- 17:23as you go about your week. If the foundational
- 17:25pioneers of this technology are so deeply driven
- 17:28by paranoia, ego, and the pursuit of power that
- 17:31they will stop speaking to their best friends
- 17:33for a decade, what inherent human biases are
- 17:36being permanently baked into the foundations
- 17:37of the objective AI agents we will soon rely
- 17:40on for every decision we make? There's a deeply
- 17:42unsettling thought to leave on. Something to
- 17:44mull over as these agents integrate into your
- 17:46daily life. Thank you for joining us on this
- 17:47deep dive. We will see you next time. Outro music.