Latest / Elon Musk Podcast / AI UPDATE: Anthropic's Pentagon Ultimatum and OpenAI Ads

Transcript

AI UPDATE: Anthropic's Pentagon Ultimatum and OpenAI Ads · Elon Musk Podcast

• 0:00
  GPT 5.2 didn't just process data this week, it actually
• 0:04
  conjectured a brand new formula for single minus gluon tree
• 0:09
  amplitudes. Which, if you aren't a
• 0:11
  theoretical physicist, is a problem that human
• 0:14
  mathematicians generally describe as incredibly long and
• 0:17
  painful. Yeah, we are talking about what
• 0:19
  is typically 1/4 page of really dense messy algebra.
• 0:24
  And the AI identified a hidden pattern that simplified all of
• 0:28
  that into a single, elegant product structure.
• 0:31
  And that is massively significant because this goes
• 0:33
  way beyond a calculator doing arithmetic faster than a human
• 0:37
  right. We are looking at a system
• 0:39
  analyzing a mess of complexity and spotting A symmetry that
• 0:42
  human experts had missed for decades, right?
• 0:45
  It is the difference between simply computing a result and
• 0:47
  actually understanding the architecture of the physics
• 0:49
  problem itself. It proves we are building
• 0:51
  genuine collaborators now. But while the science side is
• 0:54
  having this massive Eureka moment, the way these AI
• 0:57
  companies operate is shifting just as drastically.
• 0:59
  We are moving entirely away from the era of chat bots where you
• 1:02
  type a question and get an answer.
• 1:04
  We are entering the era of agents and Co workers that
• 1:06
  actually execute tasks. That distinction really matters.
• 1:10
  A chat bot talks, an agent acts, and right now the industry has
• 1:15
  split into two very distinct battles to make that happen.
• 1:18
  There is one battle for raw speed and hardware
• 1:21
  infrastructure, and another for deep autonomous execution.
• 1:24
  Exactly. Which brings us to the core
• 1:27
  question we are going to be exploring today.
• 1:29
  In a week where AI is proving theorems and learning to drive,
• 1:32
  our computers, are the safety guardrails crumbling under the
• 1:35
  pressure to win massive government and enterprise
• 1:38
  contracts? It is the defining tension of
• 1:40
  everything happening right now. And the autonomous side of that
• 1:43
  battle just got a huge injection of talent.
• 1:46
  As of this morning, February 25th, 2026, Anthropic has
• 1:49
  officially acquired Vercept. This is a major signal flare.
• 1:53
  Vercept was a Seattle based startup and they were founded by
• 1:56
  some real heavy hitters from the Allen Institute for AI.
• 1:59
  They had built this desktop agent called V.
• 2:02
  I've actually seen demos of V It is a creepy and impressive and
• 2:06
  equal measure. It doesn't just read the
• 2:08
  underlying text code, it physically sees the screen
• 2:11
  elements. Right, the visual component is
• 2:13
  the main hurdle everyone has been trying to clear.
• 2:15
  Yeah, for a long time. If you wanted an AI system to
• 2:18
  use a computer, you had to hook it up to an API.
• 2:21
  That is essentially a special backdoor that lets the software
• 2:24
  talk directly to other software. But the real world is incredibly
• 2:28
  messy. Most software out there doesn't
• 2:30
  have clean API's. Exactly.
• 2:32
  So VAI interacts directly with the graphical user interface.
• 2:35
  The actual pixels on the monitor.
• 2:37
  It looks at the screen just like you do.
• 2:40
  It identifies that a specific Gray rectangle is a submit
• 2:44
  button and a specific white box is a text field.
• 2:47
  Which sounds so simple to us. It does, but for a machine it is
• 2:51
  incredibly difficult. It knows the grounding problem.
• 2:54
  You have to perfectly map pixel coordinates to semantic actions.
• 2:57
  And this acquisition perfectly explains the sudden jump in
• 3:00
  Anthropics performance numbers. We saw the new benchmarks for
• 3:03
  Claude Sonnet 4.6 this week, specifically looking at AUS
• 3:07
  World. Yes.
• 3:08
  For context, the Alice World benchmark is the standard test
• 3:11
  for measuring how well an AI can navigate an operating system.
• 3:15
  The test asks things like can it open a spreadsheet, copy a
• 3:18
  specific cell, open a web browser, paste that date into a
• 3:22
  form and hit enter. Real computer use.
• 3:25
  Yes, and in late 2024, the best models in the world were scoring
• 3:28
  under 15% on that benchmark. Which is functionally useless
• 3:32
  for a user. You would spend more time
• 3:34
  correcting its mistakes than the time it theoretically saves you.
• 3:37
  A 15% success rate is basically a toy, but Sonnet 4.6, which is
• 3:42
  clearly integrating this new Recept tech, is now hitting
• 3:45
  72.5%. That crosses a massive threshold
• 3:49
  at over 72%. You can actually walk away from
• 3:51
  your desk and let the agent run. You can trust it with a multi
• 3:54
  step workflow. That is entirely why Antropic
• 3:57
  bought them. They aren't interested in
• 3:58
  keeping the Vibe brand alive. They are integrating the team of
• 4:01
  about 20 engineers to make Claude fully capable of direct
• 4:04
  computer use. They want an AI that writes the
• 4:06
  e-mail for you, opens your client, attaches the PDF and
• 4:09
  click send. So Anthropic is heavily betting
• 4:12
  on the smart autonomous agent that navigates a messy desktop
• 4:16
  environment. Open AI, however, seems to be
• 4:19
  betting on something else entirely.
• 4:21
  This week they released GPT 5.3 codecs Spark.
• 4:26
  Spark is the operative word there.
• 4:28
  This new model is entirely obsessed with latency.
• 4:31
  And they are achieving this incredible speed through
• 4:34
  specialized hardware, right? This goes beyond standard
• 4:37
  software optimization. It is absolutely a hardware
• 4:39
  play. Spark is running on Cerebra's
• 4:41
  Wafer Scale Engine 3. I really need you to visualize
• 4:44
  this for a second because I saw a photograph of this chip
• 4:47
  recently. Standard computer chips are
• 4:48
  small. They are roughly the size of a
• 4:50
  postage stamp. This Cerebra's chip is the size
• 4:53
  of a dinner plate. Yeah, it's the size of a giant
• 4:55
  pancake. It is an entire uncut silicon
• 4:58
  wafer, usually in standard chip manufacturing where you take a
• 5:01
  silicon wafer and cut it into hundreds of tiny individual
• 5:04
  chips. So reverse just keeps the whole
• 5:06
  thing intact as one massive processor.
• 5:08
  Why does keeping it intact matter so much for speed?
• 5:11
  Because in a traditional setup, you have your memory sitting on
• 5:14
  one physical stick and the processor sitting on another.
• 5:18
  Data has to physically travel back and forth through wires to
• 5:21
  compute anything. Even if we get the speed of
• 5:24
  light, that travel takes time and consumes a lot of energy.
• 5:27
  Right on the wafer scale engine, the memory and the compute cores
• 5:31
  are right next to each other on the exact same piece of silicon.
• 5:36
  The data transfer delay is effectively eliminated.
• 5:39
  It is nearly instant. And the practical result of that
• 5:41
  architecture is over 1000 tokens per second.
• 5:44
  Well over 1000. Wait, hold on, let's back up a
• 5:47
  second. To put that in perspective, a
• 5:49
  human being reads roughly 5 words a second, generating 1000
• 5:53
  tokens. A second means a massive wall of
• 5:55
  text appears instantly on your screen.
• 5:57
  You can't even begin to read it as it generates.
• 5:59
  You absolutely cannot. But for writing code, which is
• 6:02
  exactly what Spark is designed for, it fundamentally changes
• 6:05
  the texture of the work. It is essentially 15 times
• 6:08
  faster at coding than their standard model.
• 6:10
  They are marketing this as conversational coding.
• 6:13
  Think about how you use a standard chat bot right now.
• 6:16
  You type a prompt. You wait maybe 10 seconds.
• 6:19
  The code appears block by block. You read it.
• 6:22
  It feels very much like sending a letter and waiting for a reply
• 6:25
  in the mail. A turn based interaction.
• 6:27
  Right, with Spark the code generates so fast you can
• 6:30
  interrupt at mid thought. You can see it going down the
• 6:32
  wrong logical path in line three of a function and just stop it
• 6:36
  instantly. You correct it on the fly.
• 6:37
  So it feels more like jamming with a musician in a studio.
• 6:41
  That is the perfect analogy. It creates a tight real time
• 6:44
  feedback loop. But there is a serious trade off
• 6:47
  here. You do not get that kind of
• 6:49
  speed for free. The model is less rigorous.
• 6:51
  Exactly. Spark is completely latency
• 6:54
  first. If you look at the SWE Bench Pro
• 6:57
  scores, which is the premier software engineering benchmark,
• 7:00
  Spark scores significantly lower than the full GPT 5.3 codecs
• 7:03
  model. And perhaps more worryingly,
• 7:06
  Open AI explicitly states it is not rated for high capability
• 7:10
  cybersecurity work. So we have a purposeful division
• 7:12
  now. If you want the AI to discover
• 7:15
  the gluon tree amplitude formula, you wait patiently for
• 7:19
  the slow deep thinking model. If you want to hack together a
• 7:22
  website infrastructure in 10 minutes, you use Spark for fast
• 7:26
  execution. It is a deliberate split between
• 7:29
  deep reasoning and velocity. Speaking of deep reasoning and
• 7:32
  that physics discovery we mentioned at the start, we sort
• 7:34
  of glossed over where that actually took place.
• 7:37
  That discovery did not happen in a standard chat window on a
• 7:40
  browser. It happened inside Prism.
• 7:42
  Open AI Prism. Yeah, this is their brand new
• 7:45
  workspace designed specifically for scientists, and it is
• 7:48
  fascinating because it attacks the actual workflow of doing
• 7:51
  science. It is a fully Latex native
• 7:54
  writing environment. Latex is the complex typesetting
• 7:56
  system that pretty much every physicist and mathematician uses
• 8:00
  to write their formal papers. Prism integrates the AI directly
• 8:03
  into that source mode. Plenty of standard text editors
• 8:06
  have AI plugins these days, but Prism is completely different
• 8:10
  because of the context window and the deep integration.
• 8:13
  Prism reads the entire project structure simultaneously.
• 8:16
  It sees your equations, your citations, your raw empirical
• 8:20
  data files, and all of your figures.
• 8:22
  So if I go in and tweak a variable in my core equation on
• 8:26
  page 2, Prism automatically knows to update the resulting
• 8:30
  graph in figure 3 on page 10. Yes, it validates if your
• 8:34
  empirical results actually match your theoretical model.
• 8:37
  It can check all your citations against the actual text of the
• 8:40
  reference papers to ensure you are quoting them correctly.
• 8:43
  It is doing the tedious grunt work of consistency that usually
• 8:45
  drives researchers crazy. And I asked about the business
• 8:48
  model earlier because I noticed they are offering this entirely
• 8:51
  for free for personal accounts. That is the classic Silicon
• 8:54
  Valley play. They want to become the default
• 8:56
  infrastructure for scientific discovery.
• 8:59
  Right now, a scientist might use Overly for collaborative
• 9:02
  writing, Zotero to manage their citations, and Python for data
• 9:06
  analysis. Prism is designed to replace all
• 9:08
  of those fragmented tools. They want the next Nobel Prize
• 9:11
  winning discovery to happen natively inside an open AI
• 9:14
  interface. Discovery is an incredibly
• 9:16
  valuable commodity. If you own the tool where the
• 9:19
  science happens, you get to see where human knowledge is going
• 9:22
  before anyone else does. Which brings us directly to the
• 9:25
  money, because whether it is via navigating a messy desktop or
• 9:29
  Prism writing a theoretical physics paper, this technology
• 9:33
  has to be monetized and simply selling an API key to developers
• 9:37
  isn't cutting it for these massive valuations anymore.
• 9:39
  We are seeing the rise of the true AI Co worker and the
• 9:43
  massive consulting army is required to install them.
• 9:45
  Open AI refers to these as frontier alliances.
• 9:49
  They have realized a harsh truth.
• 9:51
  You cannot just hand a Fortune 500 company access to a super
• 9:55
  intelligent model and expect them to magically become more
• 9:57
  productive. The companies literally do not
• 9:59
  know how to use it. They have no idea how to wire it
• 10:01
  into their legacy systems, so Open AI has officially partnered
• 10:05
  with McKenzie, BCG, Accenture, and Cap Gemini.
• 10:08
  These are the massive consulting firms you traditionally hire
• 10:11
  when you want to fire half your staff and completely restructure
• 10:14
  the remaining half. Or, phrase more charitably,
• 10:17
  they're the people you hire when you need to redesign your
• 10:20
  organization's central nerve, the system.
• 10:22
  These consulting firms are wiring the frontier platform
• 10:25
  directly into massive corporate data warehouses and customer
• 10:28
  relationship management systems. They are actively redesigning
• 10:32
  organizational workflows to accommodate autonomous agents.
• 10:35
  It is essentially organizational surgery.
• 10:37
  And it is very expensive surgery.
• 10:39
  Entropic is playing this enterprise game just as hard
• 10:41
  right now. They are currently hitting a $14
• 10:44
  billion revenue run rate, and to fuel that massive infrastructure
• 10:48
  and enterprise expansion, they just closed their Series G
• 10:51
  funding round. The number on that round was
• 10:53
  staggering, $30 billion. $30 billion in cash.
• 10:57
  That completely values Anthropic at $380 billion.
• 11:01
  That is an immense, almost incomprehensible amount of
• 11:04
  capital. But here is where the tension we
• 11:07
  talked about earlier really surfaces.
• 11:09
  We have $30 billion funding rounds, we have heavy military
• 11:13
  interest, and we have these massive consulting armies
• 11:15
  deploying agents. What happens to the original
• 11:18
  safety mission? Enthropic was founded
• 11:20
  specifically by people leaving Open AI to be the definitive
• 11:23
  safety company. That core mission is severely
• 11:26
  colliding with reality right now.
• 11:27
  Just yesterday, on February 24th, Anthropic released version
• 11:31
  3 Point O of their Responsible Scaling Policy, or RSPI.
• 11:35
  Read through that document last night.
• 11:36
  There is a very specific, very controversial change in the
• 11:39
  language. In the previous versions of the
• 11:41
  RSP, Anthropic had a hard written commitment.
• 11:44
  They stated they would unilaterally pause all
• 11:46
  development if they couldn't meet certain strict safety
• 11:49
  measures. If a model was deemed too
• 11:51
  dangerous or autonomous to contain, they would stop
• 11:54
  training. Period.
• 11:55
  And in version 3 point O that is gone.
• 11:57
  That unilateral commitment is entirely gone.
• 11:59
  They've replaced it with a clear bifurcation.
• 12:01
  They now differentiate between industry recommendations and
• 12:04
  company plans. So they're publicly recommending
• 12:07
  that the entire AI industry should pause if things get
• 12:10
  dangerous, but they are no longer promising that they will
• 12:13
  pause if their competitors keep going.
• 12:15
  They are framing it as a classic collective action problem.
• 12:19
  Their argument, which is highly rational from a pure business
• 12:22
  perspective, is that if Anthropic pauses to build
• 12:26
  perfect safety guards, but open AI or a massive state backed lab
• 12:30
  in China keeps scaling and throffic just loses market
• 12:34
  share. They lose their influence over
• 12:36
  the industry. Exactly.
• 12:37
  They effectively cede the future to actors who might be far less
• 12:41
  concerned with safety than they are.
• 12:43
  It is the ultimate prisoner's dilemma.
• 12:45
  If I play nice and you decide to play rough, I lose everything,
• 12:48
  so I am forced to keep playing rough.
• 12:50
  But there's another massive pressure point here that we
• 12:52
  absolutely have to discuss the Pentagon.
• 12:55
  The US Department of Defense has been getting incredibly loud
• 12:58
  about AI integration over the last year.
• 13:00
  They have. The Pentagon explicitly
• 13:03
  communicated that a strict refusal by an AI company to work
• 13:06
  on national security tasks was viewed as a supply chain risk.
• 13:09
  Supply chain risk, that is highly specific bureaucrat speak
• 13:14
  for we are not going to buy anything from you.
• 13:16
  Exactly. It is all about reliability.
• 13:19
  If the United States government is going to heavily integrate
• 13:22
  your AI agents into their defense system, they need
• 13:25
  absolute certainty that you aren't going to suddenly turn
• 13:28
  the servers off because of an internal moral qualm about how
• 13:31
  the tech is being used. So this policy shifted.
• 13:33
  Anthropic isn't just about preserving enterprise market
• 13:36
  share, it is about making themselves a viable long term
• 13:39
  government contractor. This new policy is a calculated
• 13:42
  pivot to survive in a world where the US government is
• 13:46
  suddenly the biggest, most important customer in the room.
• 13:49
  The entire concept of safety is being redefined in real time.
• 13:53
  It is no longer about slowing down to ensure alignment, it is
• 13:56
  about staying in the absolute lead so you have the power to
• 13:58
  set the rules. And while they are fiercely
• 14:00
  fighting for those lucrative government contracts, they're
• 14:03
  also simultaneously writing massive checks to make their
• 14:06
  foundational legal problems disappear.
• 14:08
  We really have to talk about the data that powers all of this.
• 14:12
  The cost of content, we finally have a concrete price tag on it.
• 14:15
  The Barts V Anthropic Settlement. $1.5 billion.
• 14:19
  That is a historic number for a copyright lawsuit.
• 14:22
  It completely sets the precedent for the entire industry.
• 14:25
  For a very long time, the standard defense from these AI
• 14:28
  labs was fair use. The argument was always that an
• 14:32
  AI learns exactly like a human learns.
• 14:35
  It reads publicly available information and internalizes the
• 14:38
  concepts. But the settlement strongly
• 14:41
  suggests that the actual cost of doing business moving forward
• 14:44
  involves paying out billion dollar class action settlements
• 14:48
  to the authors and creators. And the details of this specific
• 14:51
  case are crucial. This wasn't just a broad
• 14:54
  philosophical debate about whether an AI reading a
• 14:56
  purchased book is fair use. It was specifically focused on
• 15:00
  the book's three data set. Right, the shadow libraries.
• 15:02
  This was a massive data set that was largely composed of directly
• 15:05
  pirated books, so the legal battle shifted away from a high
• 15:09
  level debate about copyright theory to a very specific,
• 15:12
  undeniable accusation. You downloaded this material
• 15:15
  from a known pirate site and your engineers knew exactly what
• 15:18
  they were doing. A $1.5 billion penalty
• 15:21
  absolutely stings, but for a company that just raised $30
• 15:25
  billion in cash a few days ago, it is highly affordable.
• 15:28
  It is literally just a line item on a spreadsheet for them.
• 15:31
  And that is the dark irony of this entire settlement.
• 15:34
  A $1.5 billion penalty instantly destroys any new startup.
• 15:39
  It completely bankrupts the university research lab trying
• 15:41
  to build an open source model. But for Anthropic or Open AI or
• 15:46
  Google, it is just the toll they have to pay to get on the
• 15:49
  highway. It effectively entrenches the
• 15:50
  giants. They're the only entities on
• 15:52
  Earth who can actually afford to retroactively pay for the data
• 15:55
  they already scraped from the Internet.
• 15:57
  It completely clears the competitive field.
• 15:59
  So tying all of this together, we have AI models right now that
• 16:02
  are capable of genuine world changing brilliance.
• 16:05
  They are discovering hidden physics formulas, you're coding
• 16:07
  at the speed of thought, and we have autonomous agents that can
• 16:10
  finally navigate the messy pixel based reality of our everyday
• 16:13
  computers. But to actually get those agents
• 16:16
  out of the lab and into the real world, these companies are
• 16:19
  making a very specific set of compromises.
• 16:22
  They're wiring themselves deep into the corporate structure
• 16:25
  through massive consulting firms.
• 16:27
  They're softening their founding safety pledges to seamlessly
• 16:31
  aligned with military interests. And they are paying billion
• 16:34
  dollar fines to retroactively legalize the aggressive data
• 16:38
  gathering that made the model smart in the first place.
• 16:41
  We spent years intensely debating whether artificial
• 16:44
  intelligence would be safe or whether it would be perfectly
• 16:47
  aligned with human values. But looking at the reality of
• 16:50
  2026, safety isn't a philosophical stance anymore.
• 16:53
  It is simply a clause in a massive enterprise Oregon
• 16:56
  government contract. And it is being defined entirely
• 16:59
  by whoever is signing the biggest check, whether that
• 17:01
  happens to be the Pentagon or the venture capitalists.
• 17:04
  If you're not subscribed yet, take a second and hit follow on
• 17:06
  whatever app you're using. It helps us keep making this.
• 17:09
  We appreciate you being here.