Latest / Today’s AI News / The Ocean Is the New Data Center, Anthropic Goes Private Equity, AI Trains Its Own Successor by 2029

Transcript

The Ocean Is the New Data Center, Anthropic Goes Private Equity, AI Trains Its Own Successor by 2029 · Today’s AI News

• 0:00
  The artificial intelligence industry has it's basically
• 0:04
  become a lot like a giant hungry Kraken.
• 0:06
  It is completely outgrown the land.
• 0:08
  It's eating up all the power infrastructure and right now it
• 0:11
  is literally heading out into the open ocean to feed.
• 0:14
  It really is. I mean, the scale is just.
• 0:16
  It's hard to wrap your head around sometimes.
• 0:18
  Welcome back to today's AI News, brought to you by 9X
• 0:21
  Productions. That is our mission for this
• 0:23
  deep dive. We are looking at the massive
• 0:27
  both physical and intellectual expansion of AI that's happening
• 0:31
  right at this exact second. Yeah, we're covering a lot of
• 0:33
  ground today, or well, ocean, I guess.
• 0:35
  Right, exactly. We're going to explore the
• 0:37
  literal oceanic length companies are going to just to find space
• 0:40
  for data centers. And then from there, we'll look
• 0:43
  at the shrinking timeline for AI took, you know, actually start
• 0:46
  building itself. Plus how massive private equity
• 0:49
  firms are forcing this tech into legacy corporations.
• 0:52
  And finally, how you can just bypass all of that and run these
• 0:55
  models right from your own pocket?
• 0:57
  Packed one for sure. It is.
• 0:59
  So let's start with this ocean data center story because this
• 1:01
  is wild. There's this organ based startup
• 1:04
  called Panthelaca, and they just raised $140 million Series B,
• 1:10
  yeah, led by Peter Thiel, which values the company at nearly a
• 1:14
  billion dollars. And what they are building are
• 1:17
  these autonomous floating compute structures that are
• 1:20
  powered entirely by ocean waves. You know, to understand why a
• 1:23
  billion dollar valuation actually makes sense here, you
• 1:26
  really have to look at the absolute physical wall the
• 1:29
  industry has hit on land. The power grid.
• 1:31
  Right. Exactly.
• 1:32
  The land grab for AI data centers is facing just immense
• 1:36
  public and infrastructural resistance right now.
• 1:39
  Communities are getting really hostile toward these
• 1:41
  constructions. I mean, I get it.
• 1:42
  They're massive. They are, and they require
• 1:44
  astronomical amounts of electricity.
• 1:47
  They strain local grids to the breaking point and honestly, the
• 1:50
  cooling fans are incredibly noisy.
• 1:53
  So rather than fighting zoning boards for the next 10 years,
• 1:56
  Panthelas is basically looking at the 70% of the Earth's
• 1:59
  surface that is mostly unoccupied.
• 2:01
  And conveniently generates its own endless kinetic energy.
• 2:05
  Right, the ocean. The Spics on these nodes are
• 2:08
  just, they're crazy. We are talking about 85 meter
• 2:12
  steel structures. So that's like, what, Nearly a
• 2:16
  football field long just bobbing out in the open ocean?
• 2:19
  Yeah, they're massive monoliths. And they use that wave motion to
• 2:22
  generate the electricity to run the onboard AI chips.
• 2:25
  Plus, because they're sitting in the sea, they just use the
• 2:28
  surrounding seawater for natural cooling.
• 2:30
  Which I mean, that solves a massive energy drain right
• 2:33
  there. Oh totally.
• 2:34
  Because a huge percentage of a land based data center's power
• 2:38
  consumption goes purely to air conditioning.
• 2:41
  Like just to keep the processors from physically melting down,
• 2:44
  right? You're paying to power the chip
• 2:46
  and paying just as much to cool the chip.
• 2:47
  Exactly. So using the ocean as a natural
• 2:50
  heat sink that bypasses the problem entirely.
• 2:53
  But the mobility is the real breakthrough here.
• 2:55
  These aren't just tethered ships, you know.
• 2:57
  Like they aren't anchored. No, and they don't even have
• 2:59
  traditional engines. They are designed to steer
• 3:02
  themselves to these optimal remote waters simply by using
• 3:06
  their whole shape to navigate the ocean currents.
• 3:08
  That is insane. And once they're in position
• 3:12
  just churning through all that data, they beam the computations
• 3:15
  back to shore via Spacex's Starlink.
• 3:18
  It paints this incredible picture and makes me think of
• 3:20
  like offshore oil rigs, but instead of a crew drilling into
• 3:25
  the Earth for crude oil, you have these silent metal giants
• 3:30
  just, you know, drilling the Ocean's kinetic energy to
• 3:33
  extract compute power. That's a great way to put it.
• 3:35
  Yeah. And we're taking this funding to
• 3:36
  finish a pilot factory near Portland.
• 3:39
  They've got the first Pacific Ocean deployment and a
• 3:41
  commercial rollout plan for 2027.
• 3:44
  Feel even told the Financial Times that Panthelassa has quote
• 3:47
  open the ocean frontier. It's very sci-fi.
• 3:51
  But that's my question, let's be real for a second, is this
• 3:53
  really the future of compute? Or is this just some overly
• 3:56
  complex billionaires sci-fi dream?
• 3:58
  Like, is the zoning problem on land really so bad that we
• 4:02
  absolutely have to resort to throwing servers into the
• 4:05
  Pacific Ocean? I mean, honestly, the necessity
• 4:08
  is very real. The sheer volume of compute
• 4:11
  required to train the next generation of models.
• 4:13
  It simply cannot be supported by our current terrestrial grid.
• 4:16
  It just can't. Wow, so we are actually out of
• 4:18
  room. Pretty much, you even have
• 4:21
  companies exploring space based compute, but that's that's still
• 4:24
  decades away from being viable because of launch costs.
• 4:27
  The ocean is practically next door, and because training a
• 4:30
  model doesn't require that exact microsecond latency that, say,
• 4:34
  high frequency stock trading requires.
• 4:36
  Yeah, right. It doesn't have to be instant.
• 4:37
  Right, bouncing the data off Starlink satellites works
• 4:40
  perfectly fine for this, so it's actually a highly pragmatic
• 4:43
  engineering solution. OK, so if we are going to these
• 4:45
  absolute extremes just to find the power to train these models,
• 4:50
  what exactly are they doing with all that compute?
• 4:52
  That's the billion dollar question.
• 4:54
  Well, it turns out they're learning to train themselves.
• 4:57
  Anthropic Co founder Jack Clark just published A blog post where
• 5:01
  he puts the odds at 60% or higher that AI systems will be
• 5:05
  training their own successors before the year 2029.
• 5:08
  Yeah, Clark is basing this on benchmark data that shows a
• 5:11
  truly exponential capability curve.
• 5:14
  We're watching AI transition from basically zero independent
• 5:18
  capability to almost total autonomy across core software
• 5:22
  development tasks. And it's happening so.
• 5:25
  Fast happening in a compressed 3 year window.
• 5:27
  The data from Metr. That's the group that evaluates
• 5:30
  AI's independent work capabilities.
• 5:32
  It really puts this in perspective.
• 5:33
  So get this, in 2022 AI could handle coding tasks that took
• 5:39
  about 30 seconds. Right, like writing a quick
• 5:41
  script. Exactly.
• 5:43
  By 2026, we are looking at AI successfully completing 12 hour
• 5:48
  independent tasks and the projection is that by the end of
• 5:51
  this year we're going to see 100 hour independent runs.
• 5:55
  That's the metric that should make everyone pause.
• 5:57
  Then you look at Southwest Bench.
• 5:59
  This is a benchmark that tests AI on real world GitHub coding
• 6:03
  issues. It jumped from the clawed 2
• 6:05
  model solving just 2% of problems to the new Mythos
• 6:08
  preview model solving 93.9%. In under three years.
• 6:12
  In under three years, and now Open AI is targeting what they
• 6:15
  call an automated research intern by September 2026, and
• 6:18
  startups like Recursive Superintelligence are chasing
• 6:21
  that exact same milestone. We really need to unpack what
• 6:24
  100 hour or independent run actually entails because it
• 6:28
  represents a complete paradigm shift.
• 6:30
  Yeah. Explain that because it sounds
• 6:31
  huge. Going from a 32nd task to a 100
• 6:34
  hour task means the AI isn't just auto completing a line of
• 6:38
  code for you anymore. It requires what we call agentic
• 6:41
  looping. Agentic looping.
• 6:42
  Yeah, it means you give the AIA high level objective on say,
• 6:46
  Friday afternoon and it spends the entire weekend researching,
• 6:49
  writing code, hitting errors, reading those error logs,
• 6:53
  debugging its own mistakes, and then testing new hypotheses.
• 6:56
  And by Tuesday morning, it just delivers a finished product.
• 6:59
  OK but if an AI can reliably do that by the end of this year,
• 7:03
  what does an automated research intern look like in practice?
• 7:07
  Does this completely remove the human bottleneck in AIRND?
• 7:10
  Like are the models gets to be locked in a room inventing their
• 7:13
  own next iterations. Well, that is the exact
• 7:15
  inflection point Jack Clark is pointing towards right now.
• 7:19
  AI development speed is constrained by humans.
• 7:21
  We are the bottleneck. Humans have to format the
• 7:23
  training data. We run the evaluation
• 7:25
  experiments. We tweak the architecture, the
• 7:27
  grunt work. Exactly.
• 7:29
  But if an automated research intern can do all that tedious
• 7:32
  grunt work, Florida State human researchers just become
• 7:35
  managers. They're managing an army of
• 7:37
  digital scientists. Once an AI system can actively
• 7:40
  build, test, and refine its own successor, that feedback loop
• 7:45
  tightens to a speed we haven't seen before.
• 7:48
  The era of self improving software is well, it's arriving
• 7:51
  much faster than the broader economy realizes.
• 7:53
  I mean, if an AI is running autonomously from Friday to
• 7:56
  Tuesday, you can totally see why legacy corporations are
• 8:00
  panicking right now. They don't have the
• 8:01
  infrastructure or the talent to handle that kind of technology.
• 8:04
  Not even close. Which is exactly why the
• 8:06
  Frontier Labs are suddenly stepping in to force this
• 8:08
  integration themselves. Anthropic is forming this
• 8:11
  massive $1.5 billion Claude Services company in partnership
• 8:16
  with private equity giants. We're talking Blackstone,
• 8:19
  Hellman and Friedman and Goldman Sachs.
• 8:21
  That's. Serious money.
• 8:22
  Very serious. They're deploying applied AI
• 8:24
  engineers directly into mid sized companies just to build
• 8:27
  custom workflows for them. And Open AI is making a nearly
• 8:31
  identical move. They are reportedly raising $4
• 8:34
  billion from 19 investors. That includes CPG, Brooklyn
• 8:38
  Field, Bain and SoftBank, and this is at a $10 billion
• 8:43
  valuation for their own private equity backed deployment
• 8:46
  company. You know, to me this feels like
• 8:49
  the companies that invented the jet engine suddenly realizing
• 8:51
  the airlines have absolutely no idea how to attach it to their
• 8:54
  planes. Oh, that's exactly.
• 8:56
  It right like the engine makers are basically being forced to
• 8:58
  start aviation consulting firms just to get their product in the
• 9:01
  air. It's a great analogy because the
• 9:03
  barrier for the enterprise sector is no longer model
• 9:06
  intelligence. And it's not really cost either.
• 9:08
  The models are genius level and they're relatively cheap to
• 9:11
  access via API. So what's holding them back?
• 9:14
  The real bottleneck is the corporate plumbing.
• 9:16
  Large scale businesses are frankly they're messy.
• 9:20
  Their internal data is unstructured, their legacy
• 9:23
  systems from the 90s don't talk to each other, and they do not
• 9:26
  have the in house engineering talent required to deploy these
• 9:29
  AI agents securely. They just don't have the guys to
• 9:32
  do it. Right.
• 9:33
  So by partnering with private equity, Anthropic and Open AI
• 9:37
  aren't just raising cash, they are buying a paved high speed
• 9:41
  highway directly into the thousands of portfolio companies
• 9:44
  that are owned by firms like Blackstone and Bain.
• 9:47
  It's a brilliant way to ensure their specific models become the
• 9:50
  foundational infrastructure of the corporate world.
• 9:54
  But, you know, while these multibillion dollar firms are
• 9:58
  restructuring entire industries, individual users are starting to
• 10:02
  pull this massive compute power completely offline.
• 10:05
  Yeah, this is a fascinating shift.
• 10:06
  There is this new movement of people completely replacing Siri
• 10:10
  on their iPhones with free local AI models that keep all your
• 10:15
  private data strictly on the device.
• 10:17
  Which is kind of the necessary counter movement to those
• 10:19
  massive ocean based data centers we talked about.
• 10:22
  You don't always need to ping a server farm 1000 miles away just
• 10:25
  to get a smart answer. Right in doing this is
• 10:27
  surprisingly simple. Now there's an app on the App
• 10:29
  Store called Locally AI, so you download it and it recommends
• 10:32
  grabbing Google's open source Gemma model.
• 10:35
  That's a great starting. Point.
• 10:36
  OK, Gemma, Yeah. Then you just go into your
• 10:38
  iPhone settings, you map that app to the physical action
• 10:41
  button on the side of your phone and you select voice mode.
• 10:44
  The first time you press the button it downloads a speech to
• 10:47
  text model and from then on you just hold the button.
• 10:50
  You hear a chime and you are talking to a highly capable AI
• 10:53
  that can translate languages, explain concepts or solve math
• 10:57
  problems. And it is completely offline.
• 11:00
  That's wild that it can do that locally now.
• 11:03
  The pro tip here is to experiment by downloading larger
• 11:06
  models to see how they impact the speed, the storage space on
• 11:10
  your phone, and the quality of the answers.
• 11:12
  Yeah, the underlying mechanism making this possible is
• 11:15
  quantization. Developers are figuring out how
• 11:18
  to impress what used to require a massive cloud server down to
• 11:21
  just a few gigabytes so it can run directly on your phone's
• 11:25
  local RAM. That's incredible compression.
• 11:27
  And when you run a model locally, your audio and your
• 11:29
  data, they never leave your hand.
• 11:31
  For professionals dealing with sensitive information, or
• 11:34
  honestly anyone who just values their privacy, that shift in
• 11:37
  data architecture is monumental. OK, sure, the privacy is great,
• 11:41
  but realistically I'm not entirely convinced the average
• 11:44
  person wants to download massive files that eat up all their
• 11:47
  phone storage and dig into shortcut menus just to bypass
• 11:50
  Siri. Is the average person actually
• 11:53
  going to use this daily, or is this still just a fun playground
• 11:56
  for techies who want to tinker with open source stuff?
• 11:59
  I mean, right now it definitely leans toward the tech
• 12:02
  enthusiast, but the trend line is what matters here.
• 12:05
  These open source models are getting smaller and faster every
• 12:08
  single month. Apple and Android will
• 12:11
  eventually integrate this natively of course, but until
• 12:13
  they perfect it, tools like Locally AI show that having the
• 12:17
  reasoning capability of a supercomputer in your pocket
• 12:20
  completely untethered from the Internet is entirely possible
• 12:23
  today. That compression of power is
• 12:25
  happening everywhere, actually, which brings us to our quick
• 12:28
  hits. These are the rapid and fire
• 12:30
  updates showing how AI is permeating every other corner of
• 12:33
  tech, business and even law. So on the tool side, Grog 4.3
• 12:37
  was just released by X AI, boasting strong cost efficiency
• 12:41
  and highly specific domain performance.
• 12:43
  General Intelligence also launched Co Founder 2 which is a
• 12:46
  new agent orchestrator, and Open AI just launched Codex Pets
• 12:51
  which are these animated digital companions that live on your
• 12:54
  screen and basically react to your programming work in Codex.
• 12:57
  You know that agent orchestration space with Co
• 13:00
  Founder 2? That's really the missing link
• 13:02
  for the enterprise integration we just talked about.
• 13:04
  It allows businesses to manage multiple autonomous AI agents
• 13:08
  working in tandem. Oh, that makes sense.
• 13:09
  Over in Law and Regulation, a new filing in the Elon Musk
• 13:13
  versus Open AI case reveals that Musk actually reached out to
• 13:17
  Open AI President Greg Brockman about a potential settlement
• 13:21
  just days before the trial. Interesting timing, Very.
• 13:25
  Also, the New York Times is reporting that the White House
• 13:27
  is seeking to create a formal review and oversight process
• 13:31
  prior to companies publicly deploying AI models.
• 13:33
  Right. And having a formal review
• 13:35
  process just indicates A structural shift in how model
• 13:38
  deployment will be handled moving forward.
• 13:40
  You know, moving from a retrospective analysis to
• 13:42
  proactive engagement. Definitely on the business
• 13:45
  front, Sierra just raised $950 million at a $15 billion
• 13:50
  valuation. They're already serving over 40%
• 13:53
  of Fortune 50 companies transforming their customer
• 13:56
  experiences with AI. Huge numbers.
• 13:58
  Yeah, and Anthropic is reportedly in talks to purchase
• 14:02
  highly efficient AI chips from Fractal, which is a three-year
• 14:05
  old startup out of London. But the update that really stood
• 14:07
  out to me is in robotics. Oh, the Ruma guy.
• 14:10
  Yes, Colin Angle, the creator of the Roomba and former CEO of
• 14:13
  iRobot just introduced a product called Familiar.
• 14:17
  It is a bulldog size AI pet robot aimed specifically at
• 14:21
  retirees who have aged out of traditional pet ownership.
• 14:24
  Targeting that specific demographic addresses a really
• 14:27
  profound sociological issue, loneliness in aging populations.
• 14:30
  It is a very powerful application of embodied AI.
• 14:32
  But think about the through line there.
• 14:34
  You have digital codex pets keeping developers company on
• 14:37
  their screens, and then you have physical bulldog robots keeping
• 14:41
  retirees company in their homes. We spend so much time talking
• 14:45
  about AI as a pure utility, you know, something to parse
• 14:48
  spreadsheets or write code. But it's clearly crossing over
• 14:51
  into this deeply personal space of being a companion.
• 14:54
  It reflects our natural human tendency to anthropomorphize
• 14:58
  intelligence. When a system can speak to you
• 15:00
  naturally, and it remembers your specific contact and anticipates
• 15:04
  what you need, that line between software and companion blurs
• 15:08
  very, very quickly. It really does.
• 15:10
  And actually, let's ground all of these massive concepts, the
• 15:13
  billion dollar valuations, the ocean compute notes, the robot
• 15:16
  companions in a very real human story.
• 15:18
  This comes from our community, from a reader named Adam M in
• 15:21
  Berlin, Germany. OK, Adam's best friend Eric was
• 15:24
  in a terrible horse riding accident in a remote part of
• 15:26
  South Africa, and the accident tragically left Eric partially
• 15:30
  paralyzed from the neck down, requiring a helicopter, medevac
• 15:33
  and multiple really intense surgeries.
• 15:36
  Wow. Navigating a medical emergency
• 15:38
  of that scale is incredibly daunting anyway, but especially
• 15:42
  in a remote location far from home.
• 15:44
  Absolutely, and medical insurance only covered a portion
• 15:47
  of the massive costs, so Adam and their network used AI to
• 15:52
  help set up a GoFundMe campaign and generate social media
• 15:55
  updates to raise money. But what is truly remarkable is
• 15:59
  how Eric is using the tech himself.
• 16:02
  Because he cannot easily use his hands, Eric is relying heavily
• 16:05
  on mobile voice AI to share his journey, record personal
• 16:10
  updates, and stay connected with his support system during his
• 16:12
  recovery. That is the immediate promise of
• 16:14
  accessibility technology right there.
• 16:16
  Voice to text paired with advanced reasoning.
• 16:19
  It completely restores the level of communication and
• 16:21
  independence for someone dealing with severe motor impairment.
• 16:24
  Yeah, and it goes beyond just communicating with friends, too.
• 16:27
  They are actively using Claude to handle the back end
• 16:30
  bureaucracy of the crowdfunding platform.
• 16:32
  Oh, that's smart. GoFundMe has strict
• 16:34
  documentation requirements to prove Eric is a real person and
• 16:38
  to ensure the campaign isn't flagged as a scam.
• 16:41
  So whenever the platform sends a complex request for medical
• 16:44
  documentation, Adam and his team just paste those requirements
• 16:48
  into Claude. The AI parses the legalistic
• 16:51
  jargon and format the exact proper response needed.
• 16:55
  It speeds up the entire approval process and ensures the funds
• 16:58
  are actually released to pay for Eric's ongoing surgeries.
• 17:01
  Anyone who has dealt with medical bureaucracy knows just
• 17:04
  how much friction is involved in a crisis.
• 17:07
  The last thing you want to do is spend hours figuring out the
• 17:10
  specific formatting of compliance emails.
• 17:13
  Having an AI act as an administrative assistant removes
• 17:15
  that barrier so they can focus entirely on his recovery.
• 17:18
  Exactly, They also used AI to help estimate what their complex
• 17:22
  international funding goal should be, which they ended up
• 17:25
  setting at half €1,000,000 and Adam says they are well on their
• 17:28
  way to hitting it. That's amazing.
• 17:29
  It is, you know, We spent the first half of his deep dive
• 17:32
  exploring the philosophical implications of a 100 hour
• 17:35
  automated R&D run and weather models will self improve by
• 17:39
  2029. But right now, today, in the
• 17:43
  messy reality of a hospital room, this technology is giving
• 17:46
  a paralyzed man his voice, his agency, and a literal financial
• 17:50
  lifeline. We can so easily get lost in the
• 17:53
  macro level industry narratives, you know, the private equity
• 17:56
  deals and the data centers. But the micro level
• 17:59
  applications, the ways individuals leverage these
• 18:02
  models to solve immediate painful problems, that is where
• 18:06
  the true value of the technology lies.
• 18:08
  It empowers people exactly when they are most vulnerable.
• 18:10
  It puts everything into perspective, doesn't it?
• 18:13
  Well, that wraps up our look at the expanding universe of AI for
• 18:15
  today. If you want to see updated on
• 18:17
  all these massive shifts, please hit that subscribe button.
• 18:19
  And if you enjoyed our deep dive today, we would really love it
• 18:21
  if you could rate the show with five stars.
• 18:23
  It really helps us out. And we'll see you back for
• 18:25
  tomorrow's episode. Before you go, think about
• 18:28
  Pantholaca and those 85 meter steel compute nodes we discussed
• 18:32
  at the start. If this technology scales to
• 18:34
  meet the massive demand for AI and the open ocean is suddenly
• 18:38
  populated by thousands of these giant heat radiating, wave
• 18:41
  absorbing steel monoliths, what happens to the marine ecosystem?
• 18:45
  Will solving the public hostility of land based data
• 18:47
  centers just end up creating an entirely new environmental
• 18:50
  crisis right in the middle of the Pacific?
• 18:52
  Something to ponder.