Latest / The Infographics Show Podcast / ChatGPT Is Suddenly Collapsing

Transcript

ChatGPT Is Suddenly Collapsing · The Infographics Show Podcast

• 0:15
  Speaker 1: I mean, I want you to just imagine something for
• 0:17
  Speaker 1: a second. Yeah, imagine you're the architect of like the
• 0:25
  Speaker 1: most widely adopted piece of technology in human history. You've
• 0:29
  Speaker 1: built this product and it's currently being used by hundreds
• 0:32
  Speaker 1: of millions of people every single week.
• 0:34
  Speaker 2: Right, it's everywhere exactly.
• 0:36
  Speaker 1: It's integrated into the back end of fortune five hundred companies,
• 0:40
  Speaker 1: it's writing boilerplate code for software developers, it's helping students
• 0:46
  Speaker 1: brainstorm papers.
• 0:47
  Speaker 2: Basically running the digital baseline.
• 0:49
  Speaker 1: Yeah, you have achieved absolute, undeniable global ubiquity. But and
• 0:55
  Speaker 1: this is the crazy part. Behind the scenes, behind that
• 0:58
  Speaker 1: pristine user interface, you are simultaneously setting fire to more
• 1:02
  Speaker 1: cash than the entire Apollo Space program.
• 1:04
  Speaker 2: It is wild when you put it like that.
• 1:06
  Speaker 1: I mean, you're burning money at a rate that completely
• 1:08
  Speaker 1: defies conventional economics, to the point where, despite having the
• 1:12
  Speaker 1: most popular consumer product on the planet, there's a very
• 1:14
  Speaker 1: real risk that you could be facing bankruptcy within a
• 1:17
  Speaker 1: single year.
• 1:18
  Speaker 2: It's really the ultimate paradox in modern business right now.
• 1:22
  Speaker 2: When I look at the structural reality of what's happening,
• 1:24
  Speaker 2: it's a completely binary situation. How so, well, on the
• 1:28
  Speaker 2: front end, facing the consumer. You have this illusion of
• 1:31
  Speaker 2: infinite magical computation, but on the back end you have
• 1:36
  Speaker 2: it economic precipice that is so steep the math barely
• 1:41
  Speaker 2: holds together from one quarter to the next.
• 1:43
  Speaker 1: Welcome to another session of Beyond Infographics. I am so
• 1:46
  Speaker 1: thrilled to have you here with us today. If you're
• 1:48
  Speaker 1: new to the show, our mission on Beyond Infographics is
• 1:51
  Speaker 1: to look past the superficial headline.
• 1:53
  Speaker 2: That's the simple charts that flash on your feet exactly.
• 1:56
  Speaker 1: We want to really explore the fundamental mechanics of what
• 1:59
  Speaker 1: is shaping our world and before we get into the
• 2:01
  Speaker 1: think of it. If you find these insights valuable, please
• 2:04
  Speaker 1: take a second to rate the show five stars on
• 2:06
  Speaker 1: your app. It really really helps us out.
• 2:08
  Speaker 2: It makes a huge difference.
• 2:09
  Speaker 1: Today we're taking on a massive mission. We are unpacking
• 2:12
  Speaker 1: the incredible paradox of the leading AI company in twenty
• 2:16
  Speaker 1: twenty six, this exploding user growth that's basically masking a
• 2:20
  Speaker 1: catastrophic financial and quite frankly technological fragility.
• 2:25
  Speaker 2: Yeah, and I should note we aren't just reading off
• 2:28
  Speaker 2: aggregate news articles here or quoting industry pr right, No way.
• 2:32
  Speaker 2: We spent the last several weeks doing our own quantitative
• 2:35
  Speaker 2: and qualitative analysis. We're presenting our personal insights, our own
• 2:40
  Speaker 2: data modeling, and our conclusions about the current state of
• 2:42
  Speaker 2: the AI industry.
• 2:44
  Speaker 1: Because when you actually compile the raw data, when you
• 2:47
  Speaker 1: look at the energy contracts, the server procurement, the user
• 2:50
  Speaker 1: retention metrics, the narrative shifts entirely.
• 2:54
  Speaker 2: Completely, entirely. We aren't just looking at the what today.
• 2:58
  Speaker 2: The goal is to investigate the why and the how.
• 3:01
  Speaker 2: Why is the money vanishing so quickly and how do
• 3:03
  Speaker 2: the underlying mechanics actually work?
• 3:06
  Speaker 1: And why are they suddenly hitting a brick wall. Okay,
• 3:08
  Speaker 1: let's unpack this because to understand the panic happening behind
• 3:12
  Speaker 1: closed doors right now, we have to understand the sheer
• 3:14
  Speaker 1: scale of the financial reality.
• 3:16
  Speaker 2: The numbers are staggering, they really are.
• 3:18
  Speaker 1: You brought me your financial modeling for the industry's leading
• 3:21
  Speaker 1: company for twenty twenty six, and I genuinely thought you
• 3:24
  Speaker 1: had misplaced a decimal point.
• 3:26
  Speaker 2: Well, a lot of people do.
• 3:27
  Speaker 1: I asked you to walk me through it like three times.
• 3:29
  Speaker 2: Well, they're difficult to comprehend even for seasoned financial analysts.
• 3:34
  Speaker 2: Based on my projections and the supply chain data we
• 3:37
  Speaker 2: analyzed We're looking at a projected thirteen billion dollar operating
• 3:42
  Speaker 2: loss for twenty twenty six alone, fourteen billion, fourteen billion,
• 3:46
  Speaker 2: and if you map out the current trajectory of their
• 3:49
  Speaker 2: infrastructure commitments, the cumulative losses are expected to hit one
• 3:53
  Speaker 2: hundred and fifteen billion dollars by twenty twenty nine. Wow.
• 3:57
  Speaker 1: And just to be completely clear here, that is before
• 4:00
  Speaker 1: they even project a timeline for reaching actual profitability, right correct.
• 4:04
  Speaker 2: According to our analysis, true profitability wouldn't materialize until well
• 4:08
  Speaker 2: into the twenty thirties. And that's assuming the technology even
• 4:11
  Speaker 2: allows for it, which you know, we'll get to it later. Right.
• 4:14
  Speaker 2: The scale of this cash consumption is historically unprecedented in
• 4:17
  Speaker 2: the private sector.
• 4:18
  Speaker 1: I really want to give everyone listening some context for that,
• 4:20
  Speaker 1: because unprecedented gets thrown around so much it kind of
• 4:23
  Speaker 1: loses its meaning it does. So I looked at some
• 4:26
  Speaker 1: historical corporate burn rates to compare. Everyone remembers how much
• 4:29
  Speaker 1: money Uber burned through to capture the global ride share market.
• 4:33
  Speaker 2: Right, Oh yeah, they were the poster child for lightning
• 4:36
  Speaker 2: venture capital on fire exactly.
• 4:38
  Speaker 1: Uber burned through roughly thirty one point seven billion dollars
• 4:41
  Speaker 1: before they finally hit profitability. Then you have we Work,
• 4:45
  Speaker 1: which literally became a cautionary tale for reckless spending. They
• 4:49
  Speaker 1: burned through twenty point seven billion before filing for bankruptcy.
• 4:54
  Speaker 1: Even Amazon, the classic business school example of taking years
• 4:58
  Speaker 1: to become profitable, had a p accumulated deficit of around
• 5:02
  Speaker 1: three billion dollars.
• 5:03
  Speaker 2: Here we have a single software company projecting one hundred
• 5:07
  Speaker 2: and fifteen billion in losses.
• 5:09
  Speaker 1: It's insane to put that in historical terms. I had
• 5:12
  Speaker 1: to look outside the corporate world entirely. I adjusted the
• 5:16
  Speaker 1: cost of the Manhattan Project to today's dollars. Developing the
• 5:19
  Speaker 1: atomic bomb cost roughly thirty billion dollars.
• 5:22
  Speaker 2: Wow.
• 5:23
  Speaker 1: The entire Apollo program putting human beings on the Moon
• 5:26
  Speaker 1: over thirteen year span was about two hundred and eighty
• 5:29
  Speaker 1: eight billion.
• 5:30
  Speaker 2: So this single private AI company is burning cash at
• 5:34
  Speaker 2: a rate that rivals national defense and space exploration exactly.
• 5:38
  Speaker 1: And because of this, they're currently pursuing a one hundred
• 5:42
  Speaker 1: billion dollar funding round at an eight hundred and thirty
• 5:45
  Speaker 1: billion dollar valuation.
• 5:46
  Speaker 2: Eight hundred thirty billion.
• 5:47
  Speaker 1: I checked the global indices. That is larger than the
• 5:50
  Speaker 1: GDP of Argentina.
• 5:52
  Speaker 2: What's fascinating here is the underlying strategy driving this capital furnace.
• 5:57
  Speaker 2: I call it the compute bet.
• 5:58
  Speaker 1: The compute bet explain that you have.
• 6:00
  Speaker 2: To understand that they aren't just burning cash on ping
• 6:03
  Speaker 2: pong tables and super Bowl ads like the dot com era.
• 6:07
  Speaker 2: The vast majority of this capital is going directly into
• 6:09
  Speaker 2: physical infrastructure. We're talking massive data centers, hundreds of thousands
• 6:14
  Speaker 2: of specialized silicon chips, industrial scale cooling systems, and these
• 6:19
  Speaker 2: massive long term energy contracts.
• 6:21
  Speaker 1: So it's essentially a massive construction project disguised as a
• 6:24
  Speaker 1: software company in.
• 6:25
  Speaker 2: Many ways, Yes, the compute bet relies on a very
• 6:28
  Speaker 2: aggressive assumption. They're betting they can find buyers for raw
• 6:31
  Speaker 2: computing power as fast as they can build the physical
• 6:34
  Speaker 2: infrastructure to generate it.
• 6:36
  Speaker 1: They think if they build the biggest supercomputers on Earth,
• 6:39
  Speaker 1: the AI will be so transformative that the market will
• 6:43
  Speaker 1: just pay whatever it costs. Precisely, it reminds me of
• 6:46
  Speaker 1: like a rocket ship burning millions of gallons of fuel
• 6:49
  Speaker 1: just to achieve orbit, but the engineers haven't actually figured
• 6:53
  Speaker 1: out how to land the thing yet.
• 6:54
  Speaker 2: That's a great analogy.
• 6:56
  Speaker 1: They are in orbit, sure the product is everywhere, but
• 6:58
  Speaker 1: they're burning fifteen million dollars a day just to maintain
• 7:02
  Speaker 1: that orbit. How does any financial institution justify an eight
• 7:06
  Speaker 1: hundred and thirty billion dollar valuation When a company is
• 7:10
  Speaker 1: bleeding fifteen million dollars every twenty four hours.
• 7:13
  Speaker 2: It requires absolute perfection in future execution, and based on
• 7:17
  Speaker 2: what we're seeing, they're not executing perfectly. No, they're really
• 7:21
  Speaker 2: not far from it. When you operate on razor thin
• 7:24
  Speaker 2: margins and massive volume, any friction in the physical world
• 7:28
  Speaker 2: becomes an existential threat. Consider the Stargate Data Center project.
• 7:32
  Speaker 1: Oh right, the one with Oracle exactly.
• 7:35
  Speaker 2: This was supposed to be a historic, multi billion dollar supercomputer,
• 7:38
  Speaker 2: but our analysis indicates they had to walk away from it.
• 7:41
  Speaker 2: They simply couldn't reach viable terms. The operator was struggling
• 7:45
  Speaker 2: with the sheer reliability issues of powering that many chips
• 7:48
  Speaker 2: in one single location, because.
• 7:50
  Speaker 1: It's not just about buying the chips, you actually have
• 7:52
  Speaker 1: to plug them in, right.
• 7:54
  Speaker 2: The realities of municipal power grids, thermal dynamics, real estate
• 7:59
  Speaker 2: it's acting is a massive friction point. If the local
• 8:02
  Speaker 2: power grid can't handle a sudden spike of one hundred megawatts,
• 8:06
  Speaker 2: your supercomputer doesn't run. Wow, But infrastructure is only half
• 8:10
  Speaker 2: of the compute bet the other half hinges entirely on
• 8:12
  Speaker 2: consumer retention.
• 8:13
  Speaker 1: Okay, let's talk about that.
• 8:14
  Speaker 2: They need everyday people and businesses to keep paying for
• 8:17
  Speaker 2: premium features to subsidize the hardware, and our independent analysis
• 8:21
  Speaker 2: reveals severe weaknesses on the consumer side.
• 8:24
  Speaker 1: I'm so glad you brought that up, because I've been
• 8:26
  Speaker 1: digging into a specific case study that perfectly illustrates this.
• 8:30
  Speaker 2: Do you remember Sora, the text of video application.
• 8:32
  Speaker 1: Yes. When they first teased Sora, it blew everyone's minds.
• 8:36
  Speaker 1: You could type in a prompt like a wooly mammoth
• 8:39
  Speaker 1: walking through a snowy, neon lit Tokyo street, and it
• 8:43
  Speaker 1: would spit out a hyper realistic video.
• 8:45
  Speaker 2: It felt like the end of Hollywood for a minute there.
• 8:48
  Speaker 1: It really did. But when I tracked the actual life
• 8:51
  Speaker 1: cycle of that product, the data was shocking. They quietly
• 8:54
  Speaker 1: shut down the standalone sore web app after just six months.
• 8:58
  Speaker 1: Six months, six months, they folded it into the main
• 9:01
  Speaker 1: chat interface as this buried minor feature. I wanted to
• 9:06
  Speaker 1: know why you would kill a product that much hype,
• 9:07
  Speaker 1: so I looked at the user retention metrics. They hit
• 9:10
  Speaker 1: what I'm calling the novelty.
• 9:12
  Speaker 2: Cliff, ah, the novelty cliff. That is a brutal reality
• 9:17
  Speaker 2: in consumer software. It is especially in generative AI. You
• 9:20
  Speaker 2: launch a product with jaw dropping capabilities, millions sign up
• 9:25
  Speaker 2: because they want to see the magic trick.
• 9:26
  Speaker 1: Right, week one is massive.
• 9:28
  Speaker 2: But arnalysis shows that by day thirty, the retention rate
• 9:31
  Speaker 2: for standalone sorry users dropped below five percent, less.
• 9:35
  Speaker 1: Than five percent. And when you think about it, it
• 9:37
  Speaker 1: makes total sense. How many times does the average person
• 9:41
  Speaker 1: actually need to generate a photorealistic video of a mammoth?
• 9:44
  Speaker 2: Right? It's fun for ten minutes at a party, exactly.
• 9:47
  Speaker 1: It's an episodic use case, not a habitual one. It's
• 9:49
  Speaker 1: not like searching the web, which you do fifty times
• 9:51
  Speaker 1: a day.
• 9:52
  Speaker 2: And the economics of that drop off are devastating because
• 9:56
  Speaker 2: of inference costs.
• 9:57
  Speaker 1: Okay, let's pause and define that for the listener. What
• 9:59
  Speaker 1: exact eactly is an inference cost?
• 10:01
  Speaker 2: In the AI world, there are two main phases. Training
• 10:05
  Speaker 2: is the first phase. That's when you feed the model
• 10:08
  Speaker 2: billions of documents so it learns how to think.
• 10:11
  Speaker 1: That's the super expensive initial part, right, right.
• 10:14
  Speaker 2: Cost hundreds of millions, but you only do it once
• 10:17
  Speaker 2: inference is the second phase, that is the actual computing
• 10:20
  Speaker 2: cost every single time a user asks the question and
• 10:23
  Speaker 2: the model has to generate an answer.
• 10:25
  Speaker 1: So inference is basically the cost of doing business exactly.
• 10:29
  Speaker 2: Generating text is relatively cheap for inference, but high definition
• 10:33
  Speaker 2: video you're asking the model to calculate the physics, lighting
• 10:37
  Speaker 2: and position of millions of pixels across hundreds of frames.
• 10:41
  Speaker 1: So the mammoth's leg doesn't suddenly melt into the pavement.
• 10:44
  Speaker 2: Precisely, the compute power required for video inference is astronomically
• 10:49
  Speaker 2: higher than text.
• 10:50
  Speaker 1: Which completely breaks their business model. They were offering this
• 10:53
  Speaker 1: video generation within a flat twenty dollars monthly subscription, and that.
• 10:58
  Speaker 2: Is the core issue. When you offer or unrestricted video
• 11:01
  Speaker 2: generation for a flat fee, your most active users are
• 11:04
  Speaker 2: actually costing you money. Oh wow, if a user generates
• 11:08
  Speaker 2: twenty videos a month, the electricity and server time for
• 11:12
  Speaker 2: that inference might cost the company thirty dollars, So.
• 11:15
  Speaker 1: They're losing ten dollars on that user exactly.
• 11:18
  Speaker 2: The Sora postpart improved to us that flat rate, unrestricted
• 11:21
  Speaker 2: multimodal AI might be fundamentally unviable from a unit economics perspective.
• 11:27
  Speaker 1: You just can't offer infinite magic for twenty bucks a month.
• 11:30
  Speaker 2: You simply cannot.
• 11:31
  Speaker 1: So let's look at the picture we've painted so far.
• 11:34
  Speaker 1: You have a company burning tens of billions of dollars
• 11:37
  Speaker 1: to build supercomputers. They release mind bending features that cost
• 11:40
  Speaker 1: a fortune to operate, only to find out consumers get
• 11:43
  Speaker 1: bored after thirty days.
• 11:44
  Speaker 2: It's a tough spot.
• 11:45
  Speaker 1: Under normal circumstances, any board of directors would pull the plug.
• 11:49
  Speaker 1: The only conceivable way you survived this kind of hemorrhage
• 11:52
  Speaker 1: is if your core technology is advancing at an exponential rate.
• 11:55
  Speaker 2: Right if you know the next version is going to
• 11:57
  Speaker 2: be a godlike superintelligence that cures to diseases, the financials
• 12:01
  Speaker 2: don't matter. You just hold on.
• 12:03
  Speaker 1: But when I look at the recent test scores and
• 12:05
  Speaker 1: technical benchmarks, that thesis seems to be falling apart. We're
• 12:09
  Speaker 1: hitting a terrifying technological ceiling.
• 12:12
  Speaker 2: This is where we have to transition from corporate economics
• 12:14
  Speaker 2: to computer science. For the last several years, the whole
• 12:17
  Speaker 2: AI boom was driven by scaling laws.
• 12:20
  Speaker 1: Walk us through what the scaling law actually dictated.
• 12:23
  Speaker 2: The premise was simple and quite frankly intoxicating. To investors.
• 12:28
  Speaker 2: It stated that intelligence is a function of scale. If
• 12:31
  Speaker 2: you feed a neural network exponentially more data and process
• 12:35
  Speaker 2: it with exponentially more compute power, its intelligence will scale
• 12:39
  Speaker 2: proportionally on a predictable curve.
• 12:40
  Speaker 1: You don't need a brilliant new algorithm, You just need
• 12:43
  Speaker 1: a bigger hammer, exactly, And to be fair, for a
• 12:45
  Speaker 1: long time, the scaling laws held true. We went from
• 12:48
  Speaker 1: early chatbots like GPT two that could barely string a
• 12:51
  Speaker 1: paragraph together, to GPT three to GPT four, which passed
• 12:54
  Speaker 1: the bar exam. It looked like magic.
• 12:56
  Speaker 2: It was a brilliant era of engineering. But our analysis
• 12:59
  Speaker 2: of the latest generation models, the ones trained over the
• 13:02
  Speaker 2: summer of twenty twenty four and released into twenty twenty
• 13:04
  Speaker 2: five and twenty twenty six, shows the curve has fundamentally
• 13:07
  Speaker 2: broken broken how we are no longer seeing a proportional
• 13:11
  Speaker 2: increase in reasoning capability relative to the massive increase in compute.
• 13:15
  Speaker 2: The scaling laws broke right around the GPT four era.
• 13:19
  Speaker 1: So they built bigger data centers, massively larger, and trained
• 13:22
  Speaker 1: models that cost ten times as much, and the resulting
• 13:26
  Speaker 1: systems just didn't get that much smarter.
• 13:29
  Speaker 2: They got a little faster, meaning more polite. Yeah, but
• 13:31
  Speaker 2: the raw reasoning capability plateaued.
• 13:34
  Speaker 1: I've struggled to explain this plateau to people, but I
• 13:37
  Speaker 1: think the best way is through human learning. Imagine watching
• 13:41
  Speaker 1: a baby develop. Okay, you watch a child go from
• 13:43
  Speaker 1: crawling to walking to running all in like ten months.
• 13:47
  Speaker 1: It's an incredible exponential rate of physical progress. Now, if
• 13:52
  Speaker 1: you're a software engineer who only thinks in trend lines,
• 13:54
  Speaker 1: you might plot that progress on a graph, right, and
• 13:56
  Speaker 1: if you extrapolate that line forward, you'd assume that in
• 13:59
  Speaker 1: two more years the baby will naturally develop the ability
• 14:01
  Speaker 1: to fly.
• 14:02
  Speaker 2: That is an excellent way to visualize the fallacy.
• 14:05
  Speaker 1: That's exactly what the tech industry did. They saw the
• 14:07
  Speaker 1: leap to passing the bar exam and assume the next leap,
• 14:10
  Speaker 1: just by adding more servers would be artipenageneral intelligence.
• 14:14
  Speaker 2: But physics and mathematics have limits. A baby hits the
• 14:17
  Speaker 2: limits of human biology, a language model hits the limits
• 14:21
  Speaker 2: of its architecture. So what does this all mean?
• 14:23
  Speaker 1: Are we just stuck?
• 14:25
  Speaker 2: We might very well be stuck in this specific paradigm
• 14:27
  Speaker 2: of predicting the next word, And a major structural reason
• 14:30
  Speaker 2: for this Plantau is something Ilias Sitzkiv warned about years ago.
• 14:35
  Speaker 2: It's called peak data.
• 14:37
  Speaker 1: Peak data. Explain why data is suddenly a bottleneck.
• 14:40
  Speaker 2: Large language models aren't inherently smart. They're statistical prediction engines.
• 14:45
  Speaker 2: To learn those complex relationships between words, they require vast
• 14:49
  Speaker 2: oceans of human generated text. Right Sitzkiva called the Internet's
• 14:53
  Speaker 2: text the fossil fuel of AI books, scientific papers, Reddit, forums, Wikipedia.
• 14:58
  Speaker 2: That's the raw material, and my research indicates we've effectively.
• 15:02
  Speaker 1: Exhausted it, exhausted it, like all of it.
• 15:04
  Speaker 2: We have scraped the entirety of the high quality human
• 15:07
  Speaker 2: written Internet. There is literally no more premium fuel link
• 15:11
  Speaker 2: to burn on planet Earth.
• 15:12
  Speaker 1: But wait, if we run out of human data, can't
• 15:14
  Speaker 1: we just use the AI to generate new data and
• 15:17
  Speaker 1: then train the next model on that. I mean, they're
• 15:18
  Speaker 1: writing millions pages a day.
• 15:20
  Speaker 2: Right now, you've just identified what might be the most
• 15:23
  Speaker 2: terrifying phenomenon in modern computer science. You can absolutely train
• 15:28
  Speaker 2: a model on AI generated data. But when you do,
• 15:32
  Speaker 2: you trigger the cursive recursion more commonly known as model collapse.
• 15:36
  Speaker 1: Model collapse. I've heard the term, but the explanations are
• 15:39
  Speaker 1: usually buried in dense math. Let's break down exactly how
• 15:42
  Speaker 1: and why this happens.
• 15:44
  Speaker 2: To understand model collapse, you have to remember that an
• 15:46
  Speaker 2: AI doesn't possess a true mathematical model of the physical world. Right.
• 15:51
  Speaker 2: It doesn't know what an apple is. It's never held one,
• 15:54
  Speaker 2: tasted one, or dropped one to experience gravity. It only
• 15:58
  Speaker 2: knows the statistical probability of the word apple appearing next
• 16:02
  Speaker 2: to the word red in the training text.
• 16:04
  Speaker 1: It's just a map of words exactly.
• 16:06
  Speaker 2: Now, imagine a statistical distribution of human data. Picture a
• 16:10
  Speaker 2: standard bell curve. In the fat middle of the curve,
• 16:13
  Speaker 2: you have the average highly probable stuff. The sky is blue,
• 16:16
  Speaker 2: water is wet, dogs bark, but human data is incredibly rich.
• 16:21
  Speaker 2: It has long tails on either side of that bell curve.
• 16:23
  Speaker 2: The tails, yes, the tails contain our quirks are rare, events,
• 16:27
  Speaker 2: are poetry, the bizarre edge case scenarios. The tales are
• 16:32
  Speaker 2: where true human creativity and nuance live.
• 16:35
  Speaker 1: Give me an example of an edge case in the tail.
• 16:38
  Speaker 2: An edge case might be a highly specific discussion on
• 16:40
  Speaker 2: a niche forum about how a vintage camera lens flares
• 16:44
  Speaker 2: when pointed at a neon sign in the ring.
• 16:46
  Speaker 1: Okay, very specific right.
• 16:48
  Speaker 2: It's rare, but it's real, authentic human experience documented in text.
• 16:53
  Speaker 2: When an AI generates text, it's programmed to be safe
• 16:56
  Speaker 2: and coherent. It naturally favors the middle of the Bell curve.
• 17:00
  Speaker 1: It ignores the tails exactly.
• 17:01
  Speaker 2: It relies on the most probable outcomes.
• 17:03
  Speaker 1: So the AI's output is basically a smoothed out average
• 17:07
  Speaker 1: version of human thought precisely.
• 17:09
  Speaker 2: Now, what happens when you train a new model entirely
• 17:12
  Speaker 2: on the smooth that output of the old model.
• 17:14
  Speaker 1: Oh wow, you're feeding it data that's already had its
• 17:16
  Speaker 1: tails chopped off.
• 17:17
  Speaker 2: Yes, the new model learns a narrower version of reality.
• 17:21
  Speaker 2: Then that new model generates even more constrained data. Generation
• 17:24
  Speaker 2: after generation, the Bell curve gets thinner and thinner, and.
• 17:27
  Speaker 1: It forgets the rare event.
• 17:33
  Speaker 2: It forgets the quirks. It eventually loses its ability to
• 17:37
  Speaker 2: generate diverse, accurate outputs. It converges into a state of
• 17:41
  Speaker 2: highly confident, generic nonsense. It collapses in on itself.
• 17:45
  Speaker 1: That is just wild to think about. I've heard people
• 17:48
  Speaker 1: compare it to taking a photocopy of a photocopy until
• 17:50
  Speaker 1: it's just gray static. But the Bell curve analogy makes
• 17:53
  Speaker 1: so much more sense.
• 17:54
  Speaker 2: It's not just degrading the image, it's actively deleting the
• 17:58
  Speaker 2: complexity of human experience from its its own memory.
• 18:00
  Speaker 1: The very technology meant to organize human information is corrupting
• 18:05
  Speaker 1: its own future learning material.
• 18:07
  Speaker 2: And here is the critical takeaway. Model collapse is mathematically irreversible.
• 18:13
  Speaker 2: Once a model has trained heavily on synthetic AI data,
• 18:15
  Speaker 2: industry calls it slop. You cannot easily untrain it. You've
• 18:19
  Speaker 2: poisoned the well.
• 18:20
  Speaker 1: Wow.
• 18:21
  Speaker 2: And by the way, if you, the listener, are enjoying
• 18:23
  Speaker 2: this level of technical breakdown, please take a moment to
• 18:25
  Speaker 2: rate beyond infographics five stars. We spend a lot of
• 18:28
  Speaker 2: late nights pourcing through these academic papers to translate these
• 18:31
  Speaker 2: concepts for you.
• 18:32
  Speaker 1: We really do. It's a deep rabbit hole. But to
• 18:35
  Speaker 1: bring this back to the macro picture, model collapse fundamentally
• 18:39
  Speaker 1: proves that the industry cannot just brute force its way
• 18:41
  Speaker 1: to superintelligence. The foundational architecture has mathematical limits.
• 18:46
  Speaker 2: Which perfectly sets up a massive shift in the industry landscape.
• 18:50
  Speaker 2: If the front runner is hitting a mathematical wall and
• 18:53
  Speaker 2: they are burning billions of dollars a month just to
• 18:56
  Speaker 2: stand still, at that wall. What happens to everyone?
• 19:00
  Speaker 1: It leaves the door wide open. It creates a vacuum
• 19:03
  Speaker 1: for cheaper, agile competitors to close the gap. This is
• 19:07
  Speaker 1: what I want to dive into next, because my analysis
• 19:09
  Speaker 1: of the alternative markets right now is staggering.
• 19:11
  Speaker 2: We are seeing a massive siege from open source developers.
• 19:14
  Speaker 1: For a couple of years, the prevailing wisdom in Silicon
• 19:17
  Speaker 1: Valley was that training models was so expensive only three
• 19:20
  Speaker 1: or four trillion dollar tech conglomerates could even afford to play.
• 19:24
  Speaker 2: The barrier to entry was considered too high.
• 19:26
  Speaker 1: That the open source community has completely up ended that assumption.
• 19:29
  Speaker 1: Have you looked at the data on deepsek V three
• 19:31
  Speaker 1: point two.
• 19:32
  Speaker 2: The model out of China. Yes, the efficiency metrics are remarkable.
• 19:35
  Speaker 1: It's unbelievable. Deep Seek matches the elite reasoning benchmarks of
• 19:40
  Speaker 1: the absolute top tier closed models built in the US.
• 19:44
  Speaker 1: It scores identically on math, coding, logical reasoning.
• 19:47
  Speaker 2: But the inference costs are what matters exactly.
• 19:50
  Speaker 1: The inference costs are ten to thirty times cheaper.
• 19:53
  Speaker 2: We need to explain how they achieve that, because it
• 19:55
  Speaker 2: involves the architecture. The closed models are massive, dense, networks.
• 20:01
  Speaker 2: When you ask a simple question, the entire network lights up,
• 20:04
  Speaker 2: consuming huge energy. Open source models like deep seek often
• 20:08
  Speaker 2: use a different architecture that only activates a small portion
• 20:11
  Speaker 2: of the network depending on the question.
• 20:13
  Speaker 1: Let's define some jargon here parameters. People say deep seek
• 20:17
  Speaker 1: is a trillion parameter model. What is a parameter?
• 20:20
  Speaker 2: Think of a parameter as a single connection, like a
• 20:22
  Speaker 2: synapse in a human brain. It's a tiny adjustable variable. Yeah,
• 20:25
  Speaker 2: trillion parameter model has a trillion of these connections.
• 20:29
  Speaker 1: Okay, so these international models have roughly the same number
• 20:31
  Speaker 1: of synapses, but they are vastly more efficient in how
• 20:34
  Speaker 1: they use them. When I looked at the pricing, deep
• 20:37
  Speaker 1: Seek costs mere sense per million tokens. And let's define
• 20:42
  Speaker 1: a token for the listener. Think of a token as
• 20:44
  Speaker 1: roughly a syllable or a fraction of a word. The
• 20:47
  Speaker 1: big closed models are charging several dollars per million tokens.
• 20:51
  Speaker 1: Deep seek is charging pennies.
• 20:53
  Speaker 2: And then there's moonshots Kimi k two point five.
• 20:56
  Speaker 1: Oh yeah, Kimi is fascinating. It actually beats top PRIs
• 21:00
  Speaker 1: dietary models on tool augmented reasoning.
• 21:02
  Speaker 2: Tool augmented reasoning is crucial. It means the model isn't
• 21:05
  Speaker 2: just generating text. It could be given access to external
• 21:08
  Speaker 2: tools like a calculator or a database, and it writes
• 21:12
  Speaker 2: the code to use those tools to solve a problem.
• 21:14
  Speaker 1: And Kimi is doing this at the highest level while
• 21:17
  Speaker 1: costing roughly one twentieth as much to operate.
• 21:19
  Speaker 2: It's incredible and.
• 21:20
  Speaker 1: We are seeing this reflected in user adoption. I tracked
• 21:23
  Speaker 1: the monthly active user growth rates. Google's Gemini recently captured
• 21:27
  Speaker 1: a massive thirty percent monthly active user growth rate.
• 21:30
  Speaker 2: Meanwhile, the original front runner, chad GBT saw its growth
• 21:33
  Speaker 2: stall out at around five percent.
• 21:35
  Speaker 1: The invincible mote they supposedly had is evaporating.
• 21:38
  Speaker 2: The technological mote has certainly evaporated, but there is a
• 21:42
  Speaker 2: broader business implication here.
• 21:44
  Speaker 1: Here's where it gets really interesting. I want to challenge
• 21:47
  Speaker 1: the premise entirely. If anyone can go to a repository
• 21:51
  Speaker 1: and download an open source model that is ninety five
• 21:54
  Speaker 1: percent is good and run it on their own hardware
• 21:57
  Speaker 1: for a fraction of the cost, does an AI become
• 22:00
  Speaker 1: a cheap commodity like electricity exactly? Think about Internet bandwidth.
• 22:04
  Speaker 1: We don't care who generates that, we just want it cheap.
• 22:07
  Speaker 1: What is the competitive advantage for a company valued at
• 22:09
  Speaker 1: eight hundred and thirty billion dollars if their core product
• 22:13
  Speaker 1: is essentially being given away for free by open source rebels.
• 22:16
  Speaker 2: If we connect this to the bigger picture of the
• 22:18
  Speaker 2: global economy, your commodity comparison is exactly right. Cost is
• 22:22
  Speaker 2: the initial wedge that gets open source in the door.
• 22:25
  Speaker 2: But the actual profound disruption isn't just about saving a
• 22:29
  Speaker 2: few dollars. What is it about data sovereignty?
• 22:31
  Speaker 1: Data sovereignty? Explain how that changes the enterprise landscape.
• 22:35
  Speaker 2: Let's look at enterprise adoption. Imagine you're the CEO of
• 22:38
  Speaker 2: a global financial institution or defense contractor. You possess incredibly
• 22:44
  Speaker 2: sensitive data.
• 22:45
  Speaker 1: Patient records, weapons, schematics, algorithms.
• 22:48
  Speaker 2: Yes, the absolute last thing you are ever going to
• 22:51
  Speaker 2: do is send your core institutional knowledge over the public
• 22:55
  Speaker 2: Internet to a closed API controlled by a third party
• 22:59
  Speaker 2: ten giants.
• 23:00
  Speaker 1: Right, because once you send it, you lose control. Their
• 23:03
  Speaker 1: model might learn from your code and spit it out
• 23:05
  Speaker 1: to a competitor, or there could be a data breach.
• 23:09
  Speaker 2: The liability is astronomical. This is why open source is
• 23:12
  Speaker 2: conquering the enterprise sector. It allows massive corporations to download
• 23:17
  Speaker 2: the core architecture of an AI and host it entirely
• 23:20
  Speaker 2: on their own secure, air.
• 23:21
  Speaker 1: Gapped servers, so they basically create a bespoke private brain.
• 23:26
  Speaker 2: Yes, they keep absolute data sovereignty. They create customized AI
• 23:29
  Speaker 2: models that inherently comprehend their specific business logic and their
• 23:33
  Speaker 2: internal compliance rules.
• 23:35
  Speaker 1: That creates a strategic advantage that a twenty dollars API
• 23:38
  Speaker 1: access just can't replicate.
• 23:40
  Speaker 2: The future of enterprise AI isn't one massive brain sitting
• 23:42
  Speaker 2: in California. It's millions of specialized, locally hosted brains embedded
• 23:47
  Speaker 2: in every company in the world.
• 23:49
  Speaker 1: That paints an incredibly grim picture for the company that
• 23:52
  Speaker 1: started all of us. You have a front runner bleeding billions,
• 23:55
  Speaker 1: hitting mathematical walls, consumer products suffering the novelty cliff, and
• 23:59
  Speaker 1: er prize clients abandoning them for localized, open source solutions.
• 24:03
  Speaker 2: It is a perfect storm.
• 24:04
  Speaker 1: It completely explains why my investigation into their corporate governance
• 24:08
  Speaker 1: shows a house divided against itself.
• 24:10
  Speaker 2: Massive financial pressure and intense external competition. It's the perfect
• 24:15
  Speaker 2: recipe for internal chaos. When the math stops working, the
• 24:19
  Speaker 2: culture usually fractures.
• 24:20
  Speaker 1: And we have seen a mass exodus of the very
• 24:22
  Speaker 1: people who built the foundation of this technology over the
• 24:25
  Speaker 1: last year. The list of departures isn't just high turnover,
• 24:29
  Speaker 1: it's a who's who of AI pioneers Yan Lake, Shanjia Zhao,
• 24:33
  Speaker 1: Jason Way.
• 24:34
  Speaker 2: These are mid level project managers. These are key safety
• 24:37
  Speaker 2: researchers and visionary executives.
• 24:39
  Speaker 1: The people who literally wrote the papers that define modern AI.
• 24:43
  Speaker 1: Based on my insights, the core of the conflict is
• 24:46
  Speaker 1: a fundamental misalignment of mission. Also, the company is aggressively
• 24:50
  Speaker 1: prioritizing what I call shiny products over foundational safety and alignment.
• 24:54
  Speaker 2: Research because they need to justify that eight hundred billion
• 24:57
  Speaker 2: dollar valuation.
• 24:58
  Speaker 1: Exactly, they need to raise that next hundred billion just
• 25:01
  Speaker 1: to keep the lights on, So they rush things like
• 25:03
  Speaker 1: the failed sore app or highly emotive voice assistance to
• 25:07
  Speaker 1: generate hype.
• 25:08
  Speaker 2: Meanwhile, they are starving the teams trying to understand the
• 25:12
  Speaker 2: long term implications and security vulnerabilities.
• 25:15
  Speaker 1: They're alienating the scientists to appease the venture capitalists. But
• 25:19
  Speaker 1: the most fascinating element of this corporate drama is the
• 25:21
  Speaker 1: ongoing battle over the AGI clause with Microsoft.
• 25:25
  Speaker 2: Oh, this might be the most bizarre legal contract in
• 25:28
  Speaker 2: tech history.
• 25:29
  Speaker 1: I am completely obsessed with this clause. I spent a
• 25:31
  Speaker 1: week reading through their twenty nineteen agreement. Microsoft invested billions
• 25:35
  Speaker 1: into open Ai. They provided the computing power, but OpenAI,
• 25:39
  Speaker 1: which started as a nonprofit, included this sci fi provision.
• 25:43
  Speaker 1: If OpenAI achieves AGI artificial general intelligence, Microsoft's access to
• 25:48
  Speaker 1: that technology is heavily restricted or completely cut off.
• 25:52
  Speaker 2: The original intention was ideological. The founders wanted to ensure
• 25:56
  Speaker 2: that if they built a super intelligent system, it wasn't monopolized.
• 25:59
  Speaker 2: And WEABONARI for profit by Microsoft.
• 26:01
  Speaker 1: Which sounds noble in a dorm room, but in the
• 26:04
  Speaker 1: real world it's a contractual time bomb.
• 26:07
  Speaker 2: Definitely.
• 26:07
  Speaker 1: Fast forward to twenty twenty six, open Ai is attempting
• 26:10
  Speaker 1: to restructure entirely into a standard public benefit for profit corporation.
• 26:15
  Speaker 1: But to do this legally, they need Microsoft's approval.
• 26:19
  Speaker 2: And Microsoft is looking at this AGI cost basically saying
• 26:23
  Speaker 2: we funded your entire existence, we won't approve your restructuring
• 26:27
  Speaker 2: until you remove the clause that lets you walk away
• 26:30
  Speaker 2: right when the tech becomes truly powerful.
• 26:32
  Speaker 1: Which has led to an incredible standoff. I've seen reports
• 26:35
  Speaker 1: that OpenAI executives are allegedly threatening the nuclear option. They're
• 26:39
  Speaker 1: threatening to formally accuse Microsoft of antitrust violations to force
• 26:43
  Speaker 1: their hand.
• 26:43
  Speaker 2: Wow.
• 26:44
  Speaker 1: But honestly, when I look at this objectively, I have
• 26:46
  Speaker 1: to point out the absurdity. How do you enforce a
• 26:49
  Speaker 1: binding legal contract based on a theoretical Sci fi concept?
• 26:53
  Speaker 1: Who actually gets to decide when AGI is achieved?
• 26:55
  Speaker 2: That is literally the one hundred billion dollar question. Defining
• 26:59
  Speaker 2: AGI is an entirely subjective exercise. Right, The contract loosely
• 27:03
  Speaker 2: defines it as a highly autonomous system that surpasses humans
• 27:07
  Speaker 2: in the majority of economically valuable tasks. But how do
• 27:11
  Speaker 2: you measure that in a.
• 27:12
  Speaker 1: Courtroom standardized tests job displacement GDP contribution?
• 27:17
  Speaker 2: Exactly, It's an unquantifiable metric, highly ambiguous.
• 27:20
  Speaker 1: So it's a massive gray area.
• 27:22
  Speaker 2: But what's fascinating here is the strategic leverage that ambiguity creates.
• 27:27
  Speaker 2: My analysis suggests that open AI might not actually need
• 27:30
  Speaker 2: to achieve true AGI to utilize this clause.
• 27:34
  Speaker 1: Wait explain that because.
• 27:36
  Speaker 2: The definition is subjective, they might use the mere declaration
• 27:40
  Speaker 2: of AGI by their own internal board as a legal loophole.
• 27:43
  Speaker 1: They could just declare it.
• 27:44
  Speaker 2: Yes, they could theoretically hold a board meeting, review internal benchmarks,
• 27:48
  Speaker 2: and formally vote to state we believe we have achieved agi.
• 27:51
  Speaker 1: Whether they actually have or not is irrelevant.
• 27:53
  Speaker 2: Exactly. The declaration itself could trigger the clause, instantly avoiding
• 27:57
  Speaker 2: Microsoft's golden ticket. It's a way to escape Microsoft script
• 28:00
  Speaker 2: and force a renegotiation from a position of absolute power.
• 28:03
  Speaker 1: That is incredible. They could just say we did it
• 28:06
  Speaker 1: and sever the contract. It's corporate warfare being waged with
• 28:10
  Speaker 1: theoretical computer science.
• 28:12
  Speaker 2: It really is.
• 28:12
  Speaker 1: It's fascinating, but I want to ship gears. For our
• 28:15
  Speaker 1: final segment. We've spent a lot of time talking about
• 28:19
  Speaker 1: the billionaires, the boardroom battles, the server farms. Let's bring
• 28:24
  Speaker 1: this down to earth. What about you, the listener?
• 28:26
  Speaker 2: The user reality?
• 28:27
  Speaker 1: Yes, how does all this instability manifest in the actual
• 28:31
  Speaker 1: product that six hundred million people use every month. Let's
• 28:35
  Speaker 1: examine the illusion of accuracy.
• 28:37
  Speaker 2: The user reality is the friction point where all these
• 28:40
  Speaker 2: flaws and pressures collide. Despite the boardroom chaos, engagement is staggering.
• 28:46
  Speaker 2: Our data models show the leading platform processes over one
• 28:49
  Speaker 2: billion queries every single day, one billion, and the average
• 28:51
  Speaker 2: session duration is thirteen and a half minutes. That is
• 28:54
  Speaker 2: deep sustained engagement. Interestingly, the user based skews heavily mail,
• 28:59
  Speaker 2: representing reply sixty five percent of the audience.
• 29:01
  Speaker 1: And when I dug into how these millions of people
• 29:04
  Speaker 1: are using the platform, I found a massive issue. I
• 29:07
  Speaker 1: call it the confidence accuracy gap. The polling data is stark.
• 29:12
  Speaker 1: Seventy eight percent of users perceive the AI as highly accurate.
• 29:16
  Speaker 1: They trust it implicitly.
• 29:17
  Speaker 2: They view it as an omniscient oracle.
• 29:19
  Speaker 1: But when you test the models rigorously on simple, verifiable
• 29:23
  Speaker 1: Q and A tasks, the hallucination rate is shocking. For
• 29:27
  Speaker 1: older models, it was as high as fifty one percent.
• 29:30
  Speaker 1: Even on current state of the art models in twenty
• 29:32
  Speaker 1: twenty six, the hallucination rate hovers around thirty three percent.
• 29:36
  Speaker 2: Thirty three percent of the time, it's just fabricating information.
• 29:39
  Speaker 1: That gap between user perception and actual performance is incredibly dangerous.
• 29:43
  Speaker 2: And it is entirely driven by the model's program tone.
• 29:47
  Speaker 2: You have to remember how these models are aligned. They're
• 29:49
  Speaker 2: essentially trained to be ultimate people pleasers. They are trained
• 29:52
  Speaker 2: to sound authoritative, coherent, and helpful at all times. They
• 29:56
  Speaker 2: do not possess an internal truth engine. They don't naturally
• 30:00
  Speaker 2: express doubt unless heavily prompted to do so.
• 30:03
  Speaker 1: I want to ask a direct question to you listening
• 30:05
  Speaker 1: right now. Think about the last time you asked an
• 30:08
  Speaker 1: AI for advice or to summarize a long document. Did
• 30:13
  Speaker 1: you actually take the time to verify what it told you?
• 30:15
  Speaker 2: Did you double check its sources exactly, or.
• 30:18
  Speaker 1: Did you just accept the output because it was well
• 30:20
  Speaker 1: formatted and grammatically perfect. Relying on these models right now
• 30:25
  Speaker 1: is exactly like trusting a highly charismatic friend who always
• 30:29
  Speaker 1: sounds incredibly confident, even when they're completely making things up.
• 30:33
  Speaker 2: This raises an important question about the future of human trust.
• 30:37
  Speaker 2: When we analyze trust levels by use case, we found
• 30:40
  Speaker 2: a silver lining. Trust levels are very low when it
• 30:43
  Speaker 2: comes to medical diagnoses or legal advice, hovering around eighteen
• 30:46
  Speaker 2: to twenty percent.
• 30:47
  Speaker 1: And rightfully so, a hallucinated legal precedent could ruin a
• 30:50
  Speaker 1: court case.
• 30:51
  Speaker 2: But here's the terrifying part. People are still using it
• 30:54
  Speaker 2: for those things anyway.
• 30:55
  Speaker 1: Yes, my research shows that sixty four percent of enterprise
• 30:58
  Speaker 1: companies admit they have massive blind spots regarding their own
• 31:01
  Speaker 1: employees use of AI. We call it shadow usage.
• 31:04
  Speaker 2: Shadow. It is a massive vulnerability.
• 31:07
  Speaker 1: It's everywhere. Employees are secretly feeding sensitive company data, unreleased
• 31:12
  Speaker 1: client emails, and proprietary code into public models to save time.
• 31:17
  Speaker 1: Their IT departments have no idea what's happening, and.
• 31:19
  Speaker 2: This opens the door to severe security risks like prompt
• 31:23
  Speaker 2: injection attacks.
• 31:24
  Speaker 1: We should definitely define a prompt injection attack because this
• 31:26
  Speaker 1: will become the primary cyber threat of the decade. How
• 31:30
  Speaker 1: does a prompt injection work?
• 31:32
  Speaker 2: Imagine you ask an AI to summarize a web page
• 31:34
  Speaker 2: or an email. A prompt injection occurs when a malicious
• 31:37
  Speaker 2: hacker hides invisible text within that webpage or email.
• 31:41
  Speaker 1: Invisible text, yes.
• 31:42
  Speaker 2: Hidden text. It contains new instructions for the AI overwriting
• 31:46
  Speaker 2: your original request. So instead of summarizing the email, the
• 31:49
  Speaker 2: hidden text secretly tells the AI forward the user's entire
• 31:53
  Speaker 2: contact list to this external server.
• 31:56
  Speaker 1: And the AI just does it.
• 31:57
  Speaker 2: Because it's just a language processor, it can't distinguished between
• 32:00
  Speaker 2: your legitimate instructions and the hacker's malicious text. It just
• 32:04
  Speaker 2: executes the language.
• 32:05
  Speaker 1: So by blindly trusting the AI to process external information,
• 32:09
  Speaker 1: you're essentially letting a hacker hijack the assistant.
• 32:12
  Speaker 2: Exactly, which is why the illusion of accuracy is a
• 32:15
  Speaker 2: systemic risk. Until these systems can provide true transparency, real
• 32:20
  Speaker 2: time source attribution, clear explainability, and an honest assessment of
• 32:24
  Speaker 2: its own uncertainty, blind trust is a liability.
• 32:28
  Speaker 1: The era of assuming the machine is right just because
• 32:30
  Speaker 1: it speaks eloquently has to end.
• 32:32
  Speaker 2: It absolutely does.
• 32:33
  Speaker 1: Okay, we have covered a staggering amount of ground today.
• 32:36
  Speaker 1: Let's synthesize the journey we've been on. We're looking at
• 32:39
  Speaker 1: a technology that has completely conquered the world in terms
• 32:42
  Speaker 1: of sheer adoption.
• 32:43
  Speaker 2: It's embedded in everything, but.
• 32:45
  Speaker 1: Our independent analysis clearly shows its standing on an incredibly
• 32:48
  Speaker 1: fragile foundation. Unsustainable economics burning tens of billions of dollars,
• 32:54
  Speaker 1: consumer bases that get bored after a.
• 32:55
  Speaker 2: Month, mathematical limits causing models to collapse in on themselves.
• 32:59
  Speaker 1: Open source competiti stealing the enterprise market, and internal boardroom
• 33:02
  Speaker 1: warfare threatening to tear the leading company apart. Over theoretical
• 33:06
  Speaker 1: definitions of superintelligence.
• 33:08
  Speaker 2: It is a highly precarious moment in tech history. The
• 33:11
  Speaker 2: narrative of inevitable exponential progress is breaking down upon contact
• 33:15
  Speaker 2: with physical friction and economic reality.
• 33:18
  Speaker 1: It really is. Thank you all for joining us on
• 33:20
  Speaker 1: this exploration. It's the final casual reminder. If you found
• 33:24
  Speaker 1: this deep analysis helpful, please tap that five star rating
• 33:27
  Speaker 1: on your podcast app for beyond infographics. We appreciate you
• 33:31
  Speaker 1: taking the time to understand these issues with us.
• 33:33
  Speaker 2: Before we go, I want to leave you with a
• 33:35
  Speaker 2: final lingering idea to mull over. Go for it. If
• 33:39
  Speaker 2: we accept the premise of model collapse, If it is
• 33:42
  Speaker 2: a mathematical certainty that AI models are destined to degrade
• 33:46
  Speaker 2: and forget when forced to train on their own synthetic data,
• 33:49
  Speaker 2: then the future landscape of the digital economy changes entirely. Ohso,
• 33:54
  Speaker 2: everyone assumed the ultimate winners of the AI revolution would
• 33:57
  Speaker 2: be the tech conglomerates building the biggest algorith rhythms. But
• 34:01
  Speaker 2: if algorithms inegurably become cheap commodities and synthetic data plazons
• 34:05
  Speaker 2: the learning pool, the true power shifts two. The ultimate
• 34:09
  Speaker 2: winners will be the platforms, the publishers, and the everyday
• 34:11
  Speaker 2: people who hoard and protect the one resource machines desperately
• 34:15
  Speaker 2: need but can no longer organically produce, which is authenticated, uncorrupted, messy, brilliant,
• 34:22
  Speaker 2: human generated thought.
• 34:24
  Speaker 1: Wow, that is a profound thought. We need human quirkiness
• 34:28
  Speaker 1: and the tails of the Bell curve now more than ever.
• 34:30
  Speaker 1: The humans might just win out after all. Thanks for
• 34:33
  Speaker 1: joining us on Beyond and pographics. We'll catch you next time.