Latest / AI Fire Daily / 🎙 EP 384: Grok's Odyssey Episode 1 & Apollo Economist Warns S&P 500 Sees Zero AI Payoff

Transcript

🎙 EP 384: Grok's Odyssey Episode 1 & Apollo Economist Warns S&P 500 Sees Zero AI Payoff · AI Fire Daily

• 0:00
  Billions are being poured into AI every single
• 0:02
  month, but outside of big tech, corporate profits
• 0:05
  linked to AI are exactly zero. I mean, how did
• 0:08
  we get here? It is a massive disconnect, honestly.
• 0:11
  The tech is just accelerating so rapidly, but
• 0:13
  the financial reality is completely stalling
• 0:16
  out. Yeah, and today, for you listening, we're
• 0:18
  going on a deep dive into this exact phenomenon.
• 0:20
  We're exploring this massive, you know, growing
• 0:23
  gap between AI's flashy technical milestones
• 0:26
  and its actual human payoff. I really think it's
• 0:29
  the most important story in tech right now. We're
• 0:31
  going to trace this disconnect from the bleeding
• 0:33
  edge of Hollywood straight down to the balance
• 0:35
  sheets of corporate America. Okay, let's unpack
• 0:38
  this. We have to start with the visual frontier,
• 0:40
  which is generative AI in filmmaking. Right,
• 0:43
  because we just saw a massive test case for this.
• 0:45
  Yeah, Grok Imagine just released a seven -minute
• 0:47
  video, and they're calling it episode one of
• 0:50
  an AI -generated odyssey. And this arrives about
• 0:53
  three months after a very big promise from Elon
• 0:56
  Musk. He promised a full length, historically
• 0:58
  accurate film. Exactly. So they collaborated
• 1:01
  with Wonder Studios and they brought in director
• 1:03
  Dev Walker. And visually, I have to admit, it
• 1:07
  is absolutely stunning. Oh, the pixel generation
• 1:10
  is incredible. It really is. The ships look completely
• 1:13
  real and the simulation of the sea is flawless.
• 1:16
  Even the human faces, you know, they're surprisingly
• 1:18
  consistent and highly detailed. Well, the models
• 1:21
  are doing exactly what they were trained to do,
• 1:22
  right? They're predicting the next logical pixel
• 1:25
  in a sequence. But the soul is just completely
• 1:28
  missing. It really is. It falls totally flat.
• 1:31
  Yeah, reviewers absolutely panned the release.
• 1:34
  I mean, they criticized the incredibly stiff
• 1:36
  dialogue. Right, and the pacing was just painfully
• 1:39
  slow. Exactly. Yeah. The whole thing just lacked
• 1:42
  any real emotional weight. It's a beautiful rendering.
• 1:46
  I'll give it that. But it feels entirely empty.
• 1:48
  And Musk originally positioned this project as
• 1:51
  a direct challenge to Christopher Nolan. But,
• 1:53
  you know, Nolan's work relies so heavily on massive
• 1:56
  practical filmmaking. He uses real humans on
• 1:59
  real physical sets. Yeah. And right now, Grok
• 2:02
  hasn't even confirmed when the full movie will
• 2:04
  arrive or honestly, if it will ever arrive. I
• 2:08
  have to be honest here. I still wrestle with
• 2:10
  prompt drift myself. Getting an AI to just maintain
• 2:13
  a coherent narrative across like a few paragraphs
• 2:17
  is incredibly frustrating. Oh, absolutely. That
• 2:19
  is entirely normal. It's a fundamental architectural
• 2:23
  problem with how these models actually work.
• 2:25
  It kind of feels like the AI is great at painting
• 2:27
  a beautiful ship, but it's just terrible at actually
• 2:30
  staring it. That is a perfect way to frame it.
• 2:33
  This really represents the core AI dilemma right
• 2:36
  now. Because creating individual, highly convincing
• 2:39
  shots is essentially a solved problem. Right.
• 2:42
  The underlying math works flawlessly for that.
• 2:45
  Exactly. The hard problem is turning those isolated,
• 2:49
  beautiful shots into a cohesive movie. You know,
• 2:52
  you have to build a narrative that a human actually
• 2:54
  wants to watch. But wait, we see AI analyzing
• 2:57
  sentiment all the time, right? It can write a
• 2:59
  sad poem. Why is maintaining emotional tension
• 3:02
  in a video so uniquely difficult? Well, because
• 3:05
  cinematic storytelling requires deep emotional
• 3:08
  memory. It requires subtle continuity. And AI
• 3:13
  does not actually feel the stakes of a scene.
• 3:16
  It's just doing math. Right. It's just taking
• 3:18
  static noise and organizing it into a picture
• 3:21
  based on probability. It predicts the next statistically
• 3:25
  likely frame. Or the next statistically likely
• 3:28
  word. Exactly. But human emotion is rarely statistically
• 3:31
  likely. It's incredibly messy. It relies entirely
• 3:35
  on subtext. I mean, a character might say one
• 3:37
  thing but mean the exact opposite. And generative
• 3:41
  models really struggle with subtext. They absolutely
• 3:43
  do. They give you exactly what you prompt them
• 3:46
  to give you. They just can't read between the
• 3:49
  lines of their own generated scripts. So can
• 3:52
  raw compute power eventually solve the problem
• 3:55
  of bad writing? I really don't think so. Storytelling
• 3:59
  fundamentally requires lived... human experience.
• 4:02
  I mean, more GPUs just give you faster rendering
• 4:04
  times. They don't give the machine a broken heart.
• 4:07
  They don't give it the human condition. So better
• 4:10
  GPUs won't magically fix a boring script. Makes
• 4:12
  sense. It's a profound limitation for sure. So
• 4:15
  if general AI is stumbling so hard with art and
• 4:18
  narrative, where is this technology actually
• 4:20
  succeeding? I mean, we have to look at environments
• 4:22
  that are less about messy human emotion. Yeah,
• 4:25
  we're seeing a massive pivot away from Hollywood.
• 4:28
  The real momentum is moving into hyper -specialized
• 4:30
  vertical markets. Which brings us to healthcare.
• 4:33
  Microsoft has a system called MAI -DXO, and it
• 4:36
  is making huge waves right now. Oh, the clinical
• 4:39
  data on this is wild. It more than doubled doctors'
• 4:42
  accuracy on complex medical cases. Doubled it.
• 4:44
  I mean, that is a staggering metric in modern
• 4:47
  medicine. It is. And it also significantly cut
• 4:50
  testing costs across the board. So now it's heading
• 4:52
  to real clinical trials at hospitals in Boston.
• 4:55
  What's fascinating here is the sheer contrast.
• 4:57
  The AI completely fails at writing a compelling
• 5:00
  sea captain, right? But it absolutely excels
• 5:03
  at diagnosing a rare, complex disease. So why
• 5:07
  the massive difference? What's the actual mechanism
• 5:09
  making it so much better at medicine? Well, it's
• 5:12
  because medicine operates on a highly structured
• 5:13
  knowledge base. It's essentially pattern recognition
• 5:16
  on a massive planetary scale. Ah, so it's not
• 5:20
  trying to invent subtext. No, not at all. It's
• 5:23
  matching symptoms against a massive, static database
• 5:26
  of human anatomy and pathology. AI completely
• 5:28
  thrives on strict rules and highly categorized
• 5:32
  data sets. And we are seeing this exact same
• 5:34
  success outside of health care, too. Like, look
• 5:37
  at the legal and finance sectors. Yeah, the adoption
• 5:39
  curves there are nearly vertical right now. There's
• 5:42
  a specific legal tech company called Legora,
• 5:44
  and their legal AI currently serves over 140
• 5:47
  ,000 professionals. Across more than 80 different
• 5:50
  global markets, which is huge. And their annual
• 5:52
  recurring revenue just exploded. It jumped from
• 5:54
  $1 million to over $200 million, and they did
• 5:57
  that in under two years. That is pure exponential
• 6:00
  scale. And they just raised $550 million in new
• 6:04
  funding, which puts them at an $8 .8 billion
• 6:06
  valuation. Whoa, I mean... Imagine scaling revenue
• 6:10
  to $200 million that fast. It is almost unheard
• 6:13
  of. But, you know, it happens because legal work
• 6:16
  fundamentally involves reviewing thousands of
• 6:18
  dense documents. It's entirely text -heavy. Exactly.
• 6:21
  Contracts are rule -bound text. So the AI uses
• 6:24
  semantic search to extract specific clauses.
• 6:27
  It doesn't just look for keywords. It actually
• 6:29
  understands the structural logic of the contract.
• 6:32
  So the return on investment for a big law firm
• 6:34
  is immediate. Right. It's obvious on day one.
• 6:36
  It drastically cuts down their billable hours
• 6:39
  on tedious discovery work. And the models driving
• 6:42
  these highly specific tools are getting insanely
• 6:44
  powerful, too. Like a model called Kimi K3 .1
• 6:49
  Pro just launched. It boasts 2 .8 trillion total
• 6:53
  parameters. But the key metric there is its 104
• 6:56
  billion active parameters. Wait, let me break
• 6:59
  that down. Active parameters are the specific
• 7:01
  parts of the AI actually thinking right now.
• 7:04
  Right. And that makes the model incredibly efficient.
• 7:07
  Because it doesn't have to fire up its entire
• 7:09
  brain for every single prompt. Exactly. It routes
• 7:13
  your query to the specific expert neural network
• 7:15
  that's best suited to answer it, and that drastically
• 7:18
  lowers the compute cost. We are also seeing this
• 7:20
  highly targeted efficiency in video tools, like
• 7:23
  ByteDance just released CDance 2 .5 4K. Oh, this
• 7:27
  completely changes the workflow for creators.
• 7:30
  Yeah, you experiment and build your video in
• 7:32
  a very low -cost draft mode, and then, once it's
• 7:35
  perfect, you upscale the final result to full
• 7:38
  4K resolution. Which drastically lowers the barrier
• 7:41
  to entry. You know, you don't need a massive
• 7:43
  server farm to generate a high quality ad anymore.
• 7:46
  And the open source models are aggressively keeping
• 7:48
  pace too. Nand Beige 4 .3 -33B just landed on
• 7:53
  the scene and it is consistently beating much
• 7:55
  larger competitors on standard industry tests.
• 7:58
  Yeah, open source is moving incredibly fast right
• 8:00
  now. We're even seeing totally new edge tools
• 8:03
  popping up. There's an app called MOLI, which
• 8:05
  is basically an AI browser proxy. That one is
• 8:08
  genuinely fascinating. It essentially automates
• 8:10
  websites. It parses the actual structure of the
• 8:13
  web page and extracts data for you autonomously.
• 8:15
  And then there's Badgy. Oh, the iPhone one, right?
• 8:18
  Yeah. It literally turns your iPhone screen into
• 8:21
  a smart name badge that uses the camera to actually
• 8:24
  face people as you move around a room. I mean,
• 8:27
  these are highly specific workflow integrated
• 8:29
  tools. And they work beautifully because their
• 8:32
  scope is so incredibly narrow. But this specialized
• 8:34
  power brings massive new security concerns, and
• 8:37
  the U .S. Treasury is actually stepping in right
• 8:39
  now. Yeah, they're proposing some very tight
• 8:41
  regulatory limits. Right. On American investments
• 8:44
  over $50 ,000, they are specifically targeting
• 8:47
  investments in AI, supercomputing, and biotech.
• 8:51
  Primarily tied to nations like China, Russia,
• 8:53
  Iran, North Korea, and Cuba. So they clearly
• 8:56
  view this specialized AI as critical national
• 8:58
  infrastructure. They absolutely do. Right. And
• 9:00
  the stakes are massive. I mean, we're also seeing
• 9:02
  a huge rise in what cybersecurity experts are
• 9:05
  calling Black October AI scams. The threats are
• 9:08
  definitely not slowing down. Well, governments
• 9:11
  are quickly recognizing that narrow AI is highly
• 9:13
  weaponizable. A general AI writing a slightly
• 9:16
  bad movie script is totally harmless. It's just
• 9:19
  embarrassing for the studio. Exactly. But a hyper
• 9:22
  -specialized AI discovering novel, highly toxic
• 9:24
  chemical compounds? Now that is a very serious
• 9:28
  national security issue. So are we officially
• 9:30
  moving away from one -size -fits -all models?
• 9:33
  Absolutely. The era of relying on one monolithic
• 9:36
  chatbot to do everything is rapidly evolving.
• 9:38
  The real value, and honestly the real danger,
• 9:41
  lies entirely in niche supremacy. Niche AI dominates
• 9:45
  where general models stumble. That is a crucial
• 9:47
  takeaway. 100%. All right. Before we dive into
• 9:51
  the macro picture, a quick shout out to the sponsors
• 9:53
  who make this deep dive possible. We'll be right
• 9:55
  back. Stick around. Okay. We are back. So we've
• 10:01
  seen how AI can make absolute millions for specific
• 10:04
  startups. Companies like Legora are absolutely
• 10:06
  thriving right now. But when we zoom out to the
• 10:09
  macro level, you know, the broader economy, the
• 10:11
  picture changes entirely. It really does. The
• 10:14
  broader economy is telling a very different,
• 10:16
  very sobering story. A shocking one, honestly.
• 10:20
  Torsten Slocke is the chief economist at Apollo
• 10:22
  Global Management, and he recently stated there
• 10:25
  is zero financial payoff for corporate America
• 10:28
  outside of big tech. Zero. I mean, that specific
• 10:32
  word should terrify every single institutional
• 10:34
  investor. Let's look closely at his evidence.
• 10:38
  He focuses heavily on the S &P 493. And just
• 10:43
  to clarify for everyone, that is the S &P 500
• 10:45
  index, but excluding the Magnificent Seven tech
• 10:47
  stocks. Right. So you exclude Alphabet, Amazon,
• 10:51
  Apple, Meta, Microsoft, NVIDIA and Tesla. And
• 10:54
  when you look at the remaining 493 companies,
• 10:57
  revenue per worker is completely flat. And their
• 11:00
  profit margins are entirely flat, too. Meanwhile,
• 11:02
  the Mag 7 is steadily rising. Profit margins
• 11:05
  are climbing rapidly for the company selling
• 11:07
  the AI infrastructure, but they are staying totally
• 11:10
  flat everywhere else. It's the classic gold rush
• 11:13
  dynamic, right? The people selling the pickaxes
• 11:15
  are getting incredibly rich. But the people actually
• 11:18
  digging for the gold are mostly just getting
• 11:20
  tired. Yeah, Apollo says there is absolutely
• 11:23
  no economy -wide productivity acceleration. There's
• 11:26
  nothing positively linked to AI on a macro level
• 11:28
  yet. And in some cases, it's actually hurting
• 11:31
  the bottom line. Yes, token spending is becoming
• 11:34
  a massive problem. Many legacy companies are
• 11:37
  starting to view AI entirely as a cost center.
• 11:40
  They do not see it as a growth engine at all.
• 11:42
  Because they're paying constantly for API calls.
• 11:45
  Explain how that drains the budget so quickly.
• 11:47
  Well, every single time an employee queries a
• 11:50
  language model, the company pays a tiny fraction
• 11:53
  of a cent. And sure, it seems cheap at first.
• 11:56
  But spread that across 50 ,000 employees asking
• 11:59
  questions all day long? The costs skyrocket.
• 12:03
  Exactly. They are bleeding millions of dollars
• 12:05
  in compute costs, but they aren't seeing any
• 12:08
  tangible return on that investment. And Apollo
• 12:10
  analyzed multiple traditional sectors. They looked
• 12:14
  at... consumer staples, industrials, real estate,
• 12:17
  energy, and materials, and they just found weak
• 12:20
  or completely flat margin trends across the board.
• 12:24
  And honestly, the adoption numbers make this
• 12:26
  macro failure even more fascinating. Yeah, a
• 12:29
  2026 survey just looked at 9 ,000 different international
• 12:34
  executives and 78 % of these firms already use
• 12:37
  AI in some capacity. That is massive, unprecedented
• 12:41
  market penetration. Yet more than 80 % of them
• 12:44
  previously reported no measurable impact whatsoever.
• 12:48
  No positive impact on productivity, no impact
• 12:51
  on employment numbers. I mean, they bought all
• 12:54
  the expensive tools, but nothing actually changed
• 12:56
  in their output. It's kind of like stacking Lego
• 12:59
  blocks of data, but having absolutely no idea
• 13:02
  what you're trying to build. You're just buying
• 13:04
  expensive plastic. If we connect this to the
• 13:06
  bigger picture, we see a massive integration
• 13:08
  lag happening in real time. There is a huge fundamental
• 13:11
  difference between adoption and integration.
• 13:14
  But wait, isn't adoption just handing my team
• 13:18
  a login for chat GPT? I mean, what does integration
• 13:21
  actually require? Are we talking about firing
• 13:23
  people or entirely changing how we run basic
• 13:27
  meetings? It means fundamentally redesigning
• 13:29
  the human workflow from the ground up. The AI
• 13:32
  simply does not work if your old legacy processes
• 13:34
  remain exactly the same. So bolting a smart...
• 13:38
  chatbot onto a dumb process just gives you a
• 13:41
  slightly faster dumb process. Precisely. Adoption
• 13:44
  is buying the enterprise subscription. Integration
• 13:47
  is tearing down the old system so the AI can
• 13:49
  actually do the heavy lifting autonomously. And
• 13:52
  there is a very interesting counterpoint in that
• 13:54
  2026 research. Yeah, we do actually see real
• 13:58
  productivity gains in a few specific areas. primarily
• 14:01
  in finance, high -skill services, and corporate
• 14:03
  accounting. But there is a massive catch. The
• 14:06
  gains only happen when companies entirely redesign
• 14:09
  their workflows around the AI. Right. You can't
• 14:12
  just tell an accountant to use an AI to draft
• 14:14
  a faster email. That saves, what, three minutes?
• 14:17
  It doesn't move the needle. Exactly. But if you
• 14:19
  connect the AI directly to the raw data pipeline
• 14:21
  and let it autonomously reconcile the entire
• 14:24
  global ledger overnight, now that is structural
• 14:28
  change. But that requires incredible human ingenuity.
• 14:31
  It requires painful, slow structural change.
• 14:35
  And most large corporations are terribly, notoriously
• 14:38
  slow at executing structural change. Which is
• 14:41
  exactly why the S &P 493 remains flat. They're
• 14:44
  just sitting around waiting for a magic bullet.
• 14:46
  So is the AI corporate bubble bursting or just
• 14:49
  resting? Oh, it is definitely not a bust. It's
• 14:52
  purely an integration lag. The core technology...
• 14:55
  Absolutely works. But the corporate immune system
• 14:58
  is currently rejecting the necessary workflow
• 15:00
  changes. Not a burst just waiting for humans
• 15:03
  to actually adapt their daily workflows. Precisely.
• 15:05
  The burden of execution is entirely on us now.
• 15:09
  So what does this all mean? We have covered a
• 15:11
  vast amount of ground today. We really have.
• 15:14
  We went all the way from Elon Musk's Hollywood
• 15:16
  ambitions straight down to corporate profit margins.
• 15:19
  We started with a beautifully rendered but...
• 15:21
  totally stiff AI script in Grok's Odyssey. And
• 15:24
  then we ended with completely flat margins across
• 15:27
  the S &P 500. But the core lesson here seems
• 15:30
  identical across the board. Yeah, the raw foundational
• 15:33
  technology has completely arrived. It has. But
• 15:37
  the human execution is lagging far, far behind.
• 15:40
  I mean, whether we are trying to tell a deeply
• 15:42
  compelling emotional story or trying to totally
• 15:44
  redesign a legacy corporate workflow, the human
• 15:47
  element is always the bottleneck. And it's a
• 15:50
  highly critical moment for the industry. AI is
• 15:53
  rapidly becoming a standard baseline utility.
• 15:55
  It's becoming just like electricity or the internet.
• 15:58
  Right. Everyone eventually has access to the
• 16:00
  exact same foundational models. Exactly. So the
• 16:03
  most lucrative, highly sought after skill of
• 16:05
  the next decade might really surprise people.
• 16:08
  It isn't necessarily going to be coding. It's
• 16:11
  going to be workflow design. The unique ability
• 16:14
  to critically tear down and rebuild traditional
• 16:16
  human systems. Yes. To actually extract the massive
• 16:19
  financial value we have been promised for years,
• 16:22
  we have to actively bridge that frustrating gap
• 16:25
  between the raw compute power and the actual
• 16:28
  human application. I really want you to think
• 16:31
  about this in your own life this week. Look closely
• 16:34
  at your own daily tasks at work. Are you just
• 16:37
  buying AI tokens as a brand new cost center?
• 16:39
  Or are you actively redesigning how you actually
• 16:42
  work? The billions being poured into AI models
• 16:45
  won't generate a magical return on their own.
• 16:47
  The real profit, the actual breakthrough you're
• 16:50
  looking for, is just waiting on you to integrate
• 16:53
  it properly.