Latest / Elon Musk Podcast / Anthropic Overtakes OpenAI as Microsoft Bans Claude

Transcript

Anthropic Overtakes OpenAI as Microsoft Bans Claude · Elon Musk Podcast

• 0:00
  Anthropic has surpassed Open AI in enterprise adoption, while
• 0:03
  Microsoft is actively cancelling its own employees access to
• 0:08
  Anthropic's clawed code to push them onto Microsoft's GitHub
• 0:11
  Copilot. Yeah, you have a major
• 0:13
  corporation partnering intimately with an AI firm, yet
• 0:16
  they are simultaneously banning that partner's primary rival
• 0:20
  internally, even as the market clearly prefers the rival,
• 0:24
  right? We are far past the initial
• 0:26
  novelty phase of who has the most articulate chatbot.
• 0:29
  The competition is strictly over who controls the daily
• 0:31
  operations, the software workflows and the active
• 0:34
  operational memory of the corporate world.
• 0:37
  Today we are working through a stack of internal memos,
• 0:39
  industry data, and developer benchmarks.
• 0:42
  Our mission is to figure out how autonomous systems are quietly
• 0:45
  rewriting the standard corporate playbook from the inside out.
• 0:48
  We have to consider a specific scenario here.
• 0:51
  OK, if alternative AI tools are effective enough that even
• 0:54
  Microsoft's own engineers prefer them over the products their
• 0:56
  employer builds, what happens to the software industry when these
• 1:00
  tools start relacing entire oerational layers rather than
• 1:03
  just assisting the people who currently staff them?
• 1:05
  Well, the recent data tracking business adoption shows a very
• 1:09
  clear preference forming for a long while.
• 1:12
  I mean, consumer buzz dictated the narrative.
• 1:15
  People focused on app downloads or website traffic.
• 1:18
  But the enterprise metric tells a completely different story.
• 1:21
  The data shows Anthropic pulling ahead of Open AI when it comes
• 1:26
  to actual integration into workplace environments.
• 1:29
  And the adoption numbers really reflect that, that Anthropic is
• 1:34
  reaching a higher percentage of corporate penetration.
• 1:38
  Open AI is obviously reacting to this.
• 1:39
  Yeah, they have. To They're launching A
• 1:41
  multibillion dollar enterrise division entirely focused on
• 1:45
  embedding their own engineers directly into client
• 1:47
  cororations. So they're stepping away from
• 1:49
  the simple software subscription model.
• 1:51
  Because they found the ceiling. The bottleneck for enterprise
• 1:54
  deployment has. It has nothing to do with the
• 1:57
  intelligence of the model itself.
• 1:58
  Right, the models are highly capable.
• 2:00
  Exactly, the friction comes entirely from integration.
• 2:03
  You have companies running endless pilot programs that just
• 2:05
  die on the mine. They never reach production
• 2:08
  because integrating a non deterministic system into legacy
• 2:11
  architecture. It triggers endless compliance
• 2:14
  reviews, security audits and data governance panics.
• 2:17
  Yeah, open AI by embedding human engineers is basically an
• 2:21
  admission that the only way past that friction is to physically
• 2:26
  hold the client's hand through the architectural overhaul.
• 2:29
  They are attempting to become an infrastructure partner.
• 2:31
  Which means we are talking about the shift from chat bots to
• 2:35
  agentic infrastructure. Because I mean, a chat bot
• 2:38
  requires constant human initiation.
• 2:40
  You type a prompt, it returns a text block and you decide what
• 2:42
  to do with it. It's reactive.
• 2:44
  Right. And the ejectic system executes
• 2:46
  multi step work with minimal supervision.
• 2:48
  You assign it an objective and it navigates the software
• 2:51
  environment, executes commands, reads the errors, adjusts its
• 2:54
  approach, and then reports back upon completion.
• 2:57
  See, a chatbot makes a single worker faster at composing a
• 3:00
  document or summarizing a long thread.
• 3:03
  An autonomous agent doing multi step tasks challenges the
• 3:06
  baseline assumption of whether you need that human worker in
• 3:09
  the loop at all. And major corporations are
• 3:11
  already restructuring their operating models around these
• 3:14
  exact efficiency gains. Amazon, Meta, Coinbase, Dell,
• 3:20
  GitLab, they're all actively changing their org charts.
• 3:23
  I was looking at the GitLab stuff.
• 3:25
  Yeah, the CEO of GitLab was entirely direct about the
• 3:27
  strategy, stating plainly that soccer will be built by machines
• 3:31
  and directed by people. They are flattening management
• 3:34
  layers. They're reorganizing into much
• 3:36
  smaller teams. The corporate communications
• 3:38
  around this, you know, they always lean on the word
• 3:40
  augmentation. They spend considerable
• 3:42
  resources assuring everyone that technology will just augment
• 3:45
  existing employees, but the labor market contradicts that
• 3:48
  narrative entirely. How so?
• 3:50
  Traditional software companies are facing severe market
• 3:52
  pressure right now because agentic systems are absorbing
• 3:55
  the functions that used to be handled by layers of support
• 3:57
  staff and standard software as a service products.
• 4:01
  Right, because if a business operates on the assumption that
• 4:04
  fewer people are required for routine knowledge work, the
• 4:07
  physical footprint of the company shrinks.
• 4:10
  Entry level coders, junior data analysts, internal support,
• 4:13
  ticketing teams. The company requires fewer
• 4:16
  humans while the output theoretically remains static or,
• 4:19
  you know, increases. This creates a direct
• 4:22
  vulnerability for the traditional consulting industry
• 4:24
  to look at firms like Accenture, Deloitte, PwC or McKenzie.
• 4:29
  Yeah, historically, if a multinational corporation
• 4:32
  required a total operational overhaul, they hired thousands
• 4:36
  of external consultants to map the processes, recommend
• 4:39
  software, train the staff. Now, AI vendors are attempting
• 4:43
  to own that transformation directly.
• 4:45
  By embedding their own deployment teams.
• 4:46
  Exactly. They bypass the traditional
• 4:48
  consulting firms entirely. They want to hold the
• 4:51
  relationship from the foundational model all the way
• 4:53
  up to the employee workflow. And that transition from human
• 4:57
  execution to autonomous systems is driven entirely by the tools
• 5:01
  developers are using right now. The interface where humans write
• 5:04
  code is the exact location of this conflict.
• 5:08
  It really comes down to a fundamental difference between 2
• 5:10
  dominant tools, GitHub Copilot CLI and clawed code.
• 5:15
  The architectural approach each tool takes explains a lot about
• 5:18
  where software development is heading.
• 5:20
  Copilot is an IDE first platform.
• 5:23
  It lives inside the integrated development environment, which
• 5:26
  is the specific software application where developers
• 5:28
  actually type their code. It focuses on inline
• 5:31
  autocomplete. It suggests the next line of
• 5:34
  code or the next function based on what you are currently
• 5:36
  typing. So it's very immediate, highly.
• 5:38
  Immediate It is engineered for speed, specifically to reduce
• 5:42
  the micro frictions in the moment to moment process of
• 5:44
  writing. Whereas Claude Code takes a
• 5:47
  completely different route. It is a terminal first
• 5:50
  autonomous agent. So instead of sitting in your
• 5:52
  text editor suggesting lines, it lives in the command line
• 5:55
  interface. It reads entire repositories and
• 5:58
  plans multi step implementations to explain the difference.
• 6:02
  You know Copilot is like having an incredibly attentive sous
• 6:05
  chef standing right next to your cutting board.
• 6:07
  You reach for a knife, They hand you the onions.
• 6:09
  You start whisking. They pour in the oil.
• 6:11
  They're just matching your rhythm.
• 6:12
  Exactly. They anticipate your immediate
• 6:15
  physical needs to keep your momentum going.
• 6:17
  But Claude Code is like telling a catering company you need a
• 6:20
  wedding menu for 200 people and they disappear into a commercial
• 6:24
  kitchen and emerge 3 hours later with a fully plated tasting
• 6:28
  menu. You just have to taste it and
• 6:30
  approve it. And the technical specifications
• 6:32
  back up that division of Labor. Quad Code operates with a
• 6:36
  1,000,000 token context window. We should also look at its
• 6:39
  performance on the SWE bench software engineering evaluation.
• 6:43
  Yeah, explain SW Bench for a second.
• 6:45
  Right. For context, SW Bench is an
• 6:47
  evaluation framework that takes highly complex real world bugs
• 6:51
  from actual open source projects.
• 6:54
  We're talking bugs that required human engineers days to resolve.
• 6:58
  The evaluation hands the raw issue description to the AI to
• 7:02
  see if it can navigate the entire code base and resolve it
• 7:04
  without human intervention. And Claude code did well.
• 7:07
  It achieved an exceptionally high score on this benchmark.
• 7:11
  That large context window means it holds the architecture of a
• 7:14
  massive project in its active memory.
• 7:17
  It understands how modifying 1 variable in a core file cascades
• 7:21
  through 40 other dependent files across the system.
• 7:25
  The pricing models really reflect these different use
• 7:27
  cases. Though Copilot is accessible, it
• 7:30
  operates on a flat, low monthly rate.
• 7:33
  A company can easily justify deploying it across an entire
• 7:36
  engineering organization. Yeah, it's basically a rounding
• 7:38
  error for most budgets. Right, but Claude Code has tiers
• 7:42
  aimed at professional engineering workflows that cost
• 7:44
  significantly more, reaching hundreds of dollars a month
• 7:47
  depending on the usage tier and the underlying model processing
• 7:50
  the requests. You really have to look at the
• 7:52
  payment structure there. Claude code often relies on AP
• 7:55
  AI billing, meaning you pay per token processed.
• 7:58
  Was the every piece of text. Yeah, every piece of text the
• 8:00
  agent reads or writes costs a fraction of a cent, and there is
• 8:04
  a steep learning curve to managing an autonomous agent.
• 8:07
  If a developer initiates a complex debugging session, the
• 8:10
  agent might enter a loop. Oh, like a failure.
• 8:12
  Loop exactly. It tries to fix a failing test,
• 8:15
  fails, reads massive amounts files again, writes a new fix,
• 8:18
  fails again. You can burn through a huge
• 8:20
  amount of tokens very quickly in that loop.
• 8:23
  The financial unpredictability could erase the efficiency gains
• 8:27
  if a single unattended debugging session costs a company $50.
• 8:31
  Which points to a cultural shift among developers that people are
• 8:34
  calling Vibe coding. Developers use natural language
• 8:38
  to describe the intent of an application, and the AI handles
• 8:42
  the actual syntax and structure. The human transitions from being
• 8:46
  a syntax writer to a supervisor verifying the output.
• 8:49
  Describe what you want, let the AI build the foundation and
• 8:52
  iterate through conversation. It is a clear trade off between
• 8:54
  speed and depth. Copilot reduces friction for
• 8:57
  daily granular coding tasks. Claude Code acts as the
• 9:01
  autonomous engineer that takes a task, reads the relevant code
• 9:04
  across the entire repository, and produces pull request ready
• 9:07
  edits that span dozens of files. And the tension between these
• 9:10
  two approaches is currently playing out inside the company
• 9:12
  that owns Copilot. The internal conflict at
• 9:16
  Microsoft is revealing. The Experiences and Devices
• 9:19
  division, which is responsible for core products like Windows,
• 9:22
  Microsoft 365 and Teams, gave thousands of its own employees
• 9:27
  access to clawed code. Which is kind of wild when you
• 9:30
  think about it. Yeah, the tool became incredibly
• 9:32
  popular internally, and it wasn't just traditional software
• 9:35
  engineers using it. Non-technical staff, designers,
• 9:38
  product managers, they all started using clawed code to
• 9:42
  prototype features and automate their personal workflows.
• 9:45
  And what happened? Internal usage of Microsoft S
• 9:48
  own copilot CLI began to drop as a direct result.
• 9:51
  So Microsoft. Decided to mandate a migration
• 9:53
  back to Copilot CLI for these employees.
• 9:56
  They tied the cutoff for clawed code access to the end of their
• 9:59
  fiscal year. They framed the decision
• 10:02
  internally as a necessary step to reduce operating expenses and
• 10:05
  consolidate overlapping tooling. I mean, the financial rationale
• 10:09
  works on a spreadsheet, but we are discussing Microsoft here.
• 10:11
  They can comfortably afford anthropic licenses for their
• 10:14
  developers. Obviously the underlying issue
• 10:16
  is ecosystem sovereignty. They cannot normalize a rivals
• 10:20
  tool internally while aggressively asking their
• 10:23
  enterprise customers to standardize on Copilot
• 10:25
  externally. They have to protect their own
• 10:28
  ecosystem. Which touches on the concept of
• 10:30
  operational memory. When an AI agent works
• 10:33
  continuously in the terminal, it learns the specific environment.
• 10:37
  It learns which tests routinely fail for no reason, how the
• 10:40
  specific architecture of an internal project connects, and
• 10:43
  where the fragile undocumented points in the code base exist.
• 10:47
  Right, all the quirks. Exactly.
• 10:49
  If Microsoft's employees rely on a third party tool from
• 10:53
  Anthropic, all of that institutional intelligence is
• 10:56
  processed and temporarily held in a competitor system.
• 10:59
  And the friction this policy causes for the individual
• 11:01
  developers is severe. AI coding tools are highly
• 11:04
  personal. An agent becomes integrated into
• 11:06
  a developer's reasoning process. Switching an agent that has
• 11:09
  learned your repository conventions and your prompt
• 11:11
  style is disruptive. You aren't just swapping out a
• 11:14
  text editor, you are severing a connection with an entity that
• 11:17
  helps you think through architectural problems.
• 11:20
  It introduces the risk of dark debt.
• 11:23
  Dark debt occurs when an AI generates the vast majority of
• 11:26
  the code for a complex system, leaving the human developer with
• 11:30
  only a surface level understanding of how that system
• 11:33
  functions, right? If those developers are suddenly
• 11:36
  forced to switch to a different AI environment that lacks the
• 11:39
  accumulated context or the same reasoning style, they might
• 11:42
  completely lose grasp on how their own internal systems
• 11:45
  operate. The code works, but the human
• 11:48
  understanding of it is hollow. The realization that AI tools
• 11:52
  hold the operational memory of a company brings up a really tough
• 11:56
  mechanical challenge. How do these systems actually
• 11:59
  manage and retain all that information over time?
• 12:01
  That's the big question. Because the terminal agent knows
• 12:04
  the repository today, but it needs a reliable way to store
• 12:08
  and retrieve that knowledge effectively without starting
• 12:11
  from scratch every time you open a new session tomorrow.
• 12:13
  Anthropic addressed this retention problem with a new
• 12:16
  feature called Dreaming. It is a scheduled process that
• 12:19
  runs in the background between active user sessions.
• 12:23
  The agent reviews its own past actions, extracts patterns from
• 12:26
  those sessions, and writes new memory entries.
• 12:28
  OK, so the next time it encounters a similar task, it
• 12:32
  doesn't repeat the exact same mistakes.
• 12:34
  It operates much like human memory consolidation during
• 12:37
  sleep. The system reviews the events of
• 12:39
  the day, evaluates which interactions yielded positive
• 12:42
  results, and basically decides what information is relevant
• 12:45
  enough to keep in its long term context and what can be safely
• 12:48
  discarded. There is a legal AI startup
• 12:50
  called Harvey that implemented this Dreaming feature.
• 12:54
  They saw their task completion rates jump sixfold after turning
• 12:57
  it on sixfold. Yeah, before Dreaming, the
• 13:00
  agents continually hit the same small errors.
• 13:03
  They would repeatedly forget a specific workaround required for
• 13:06
  a certain legacy file type, which caused the same drafting
• 13:09
  jobs to fail over and over. With Dreaming, the system
• 13:12
  internalized the workarounds. That solves the mundane reasons
• 13:16
  why AI fails. In real production environments,
• 13:19
  we spend a lot of time focused on the theoretical reasoning
• 13:22
  limits of these models. But in a functional enterprise
• 13:25
  setting, an agent usually fails because it forgets a boring
• 13:28
  formatting rule or misreads a comma in a database query.
• 13:32
  Persistent memory directly attacks that specific failure
• 13:34
  mode. Entropic also introduced
• 13:37
  Outcomes, which functions as a self grading loop.
• 13:40
  A separate evaluator model scores the primary agents output
• 13:43
  against a written rubric and tells the agent what to fix
• 13:46
  before the human user ever sees the result, right alongside
• 13:50
  outcomes. They rolled out multi agent
• 13:52
  orchestration, often called agent teams.
• 13:55
  A lead agent divides a complex job into discrete chunks and
• 13:58
  hands them off to specialist sub agents that run in parallel.
• 14:02
  Netflix's platform team uses this multi agent orchestration
• 14:05
  right now to process build logs from hundreds of source
• 14:08
  repositories simultaneously. I retain a degree of skepticism
• 14:12
  regarding some of these performance metrics though.
• 14:14
  I mean, A6 fold increase in completion rates at one specific
• 14:18
  legal startup is an impressive data point, but it relies
• 14:21
  heavily on internal benchmarks and curated customer
• 14:24
  testimonials. You think it won't scale?
• 14:27
  You just have to question whether a perfectly tuned
• 14:29
  workflow at a nimble startup translates smoothly to the
• 14:32
  messy, contradictory workflows of a massive legacy enterprise.
• 14:37
  Well, running parallel sub agents, evaluating outcomes on a
• 14:40
  rubric, and processing massive amounts of vector memory
• 14:44
  simultaneously requires unprecedented physical computing
• 14:47
  power. The software capabilities we
• 14:49
  were discussing are entirely dependent on the physical
• 14:52
  hardware layer underneath them. Which shifts the focus to
• 14:55
  companies challenging the standard hardware paradigms.
• 14:58
  Cerebras is a primary example with their Wafer Scale engine.
• 15:02
  The traditional method of manufacturing chips involves
• 15:04
  creating a large circular silicone wafer and slicing it
• 15:07
  into dozens of small individual chips.
• 15:10
  Like little squares. Right.
• 15:12
  But Cerebras chose a different path.
• 15:13
  They built a single continuous chip the size of a dinner plate.
• 15:17
  The physical engineering feat is staggering because of the yield
• 15:20
  challenges. In traditional chip
• 15:22
  manufacturing, if a microscopic defect appears on the silicon
• 15:25
  wafer, you just throw away that one small square chip and keep
• 15:29
  the rest. With Cerebras, because the
• 15:31
  entire wafer is 1 giant chip, a defect anywhere could
• 15:35
  potentially ruin the whole. Thing they had.
• 15:37
  Engineer routing systems that can physically bypass flaws on
• 15:40
  the silicon. Despite those challenges, they
• 15:43
  have gone public and secured a massive multibillion dollar deal
• 15:47
  with Open AI to provide hundreds of megawatts of computing power.
• 15:51
  They've also turned profitable and secured investment from AMD,
• 15:55
  Which happens to be a major competitor to NVIDIA.
• 15:59
  The architectural advantage of the Cerebris chip is rooted in
• 16:02
  physics. Standard GPU's have to connect
• 16:04
  many small chips together using complex networking fabric.
• 16:08
  Moving data across those physical connections between
• 16:11
  chips creates latency and requires immense power.
• 16:14
  Because it's travelling physically further.
• 16:16
  Yeah, exactly. Cerebris eliminates that by
• 16:18
  keeping all the processing cores in memory on one massive piece
• 16:22
  of silicon. Data doesn't have to travel
• 16:25
  across a motherboard. That drastically lowers power
• 16:28
  consumption and increases speed. Specifically for AI inference
• 16:31
  tasks where you need an answer instantly it.
• 16:34
  Gives them a structural advantage for those inference
• 16:36
  workloads where response speed dictates the user experience.
• 16:40
  However, they carry clear business vulnerabilities.
• 16:43
  They have heavy customer concentration.
• 16:45
  The vast majority of their revenue previously came from
• 16:48
  specific entities located in the UAE.
• 16:50
  Which presents obvious geopolitical and financial risks
• 16:52
  if those relationships change or face regulatory scrutiny.
• 16:56
  The larger vulnerability is the software ecosystem gap though.
• 16:59
  Cerebras has undeniable hardware performance, but NVIDIA has a
• 17:02
  decade long head start with its CUDA software platform.
• 17:06
  The Compute unified device architecture.
• 17:09
  Right. That's the software platform
• 17:10
  that lets developers talk directly to the GPU.
• 17:14
  Millions of developers are trained specifically on Nvidia's
• 17:17
  ecosystem. Every university course uses it.
• 17:20
  The tools are deeply entrenched in the history of computing.
• 17:24
  Mathematically superior hardware often loses to entrenched
• 17:28
  software ecosystems because the switching costs for developers
• 17:31
  to learn a new system are simply too high.
• 17:34
  Despite the hardware battles and the gaps in software ecosystems,
• 17:37
  these agentic tools are actively being deployed right now.
• 17:41
  PwC expanded their alliance with Anthropic, launching a
• 17:44
  completely clawed native finance group.
• 17:47
  They're moving past isolated experimentation and are
• 17:49
  achieving specific, measurable business outcomes.
• 17:52
  The reported outcomes are very tangible.
• 17:55
  Insurance underwriting cycles have been reduced from weeks to
• 17:57
  days. Mainframe modernization projects
• 17:59
  dealing with massive legacy COBOL code bases are tracking
• 18:03
  under budget and on time. We should clarify that COBOL was
• 18:06
  a programming language from decades ago.
• 18:08
  Yeah, the people who originally wrote those systems are
• 18:11
  retiring. You have banks running critical
• 18:13
  infrastructure on code that nobody in the building actually
• 18:16
  knows how to read. Agentic systems are reading that
• 18:19
  code, mapping the logic and translating it into modern
• 18:21
  languages. That's huge.
• 18:23
  Furthermore, cybersecurity incident response has
• 18:25
  accelerated from hours to minutes because agent operations
• 18:29
  can review anomalous code and automate containment protocols
• 18:32
  without waiting for a human analyst to wake up.
• 18:35
  And Advocate Health is deploying these systems across a workforce
• 18:39
  of over 100,000 people. This is not theoretical.
• 18:42
  Future technology is running in production today at a massive
• 18:46
  enterprise scale. A critical technical piece
• 18:48
  making this deployment possible is the Model Context Protocol,
• 18:52
  or MCP. This protocol allows AI agents
• 18:55
  to connect directly to enterprise databases, internal
• 18:58
  documentation, and custom tooling in a standardized way.
• 19:01
  So it's not isolated. Right.
• 19:03
  The agent isn't just reading text in a vacuum.
• 19:06
  It is aware of the broader engineering stack.
• 19:09
  It can pull ticket context from project management tools, query
• 19:13
  live data to validate a code change, and push updates
• 19:16
  directly to internal servers. When you look at these specific
• 19:19
  deployments across healthcare, life sciences and private
• 19:22
  equity, you have to consider how your own industry's economic
• 19:26
  equations are changing. Absolutely.
• 19:28
  If a core administrative process that used to take weeks and
• 19:31
  required dozens of specialized people suddenly takes a fraction
• 19:34
  of the time and resources, the competitive baseline of your
• 19:37
  industry is permanently altered. The true disruption happening
• 19:41
  here is not just about writing code faster or summarizing
• 19:44
  emails. It's about redesigning the
• 19:46
  fundamental operating structures of a company around the solid
• 19:49
  assumption that fewer humans will be required to execute
• 19:52
  routine knowledge work. If intent and imagination become
• 19:56
  the primary interfaces for building software and operating
• 19:59
  a business. How do we redefine the value of
• 20:01
  technical expertise in a world where the machine handles all
• 20:04
  the execution? If you're not subscribed yet,
• 20:06
  take a second and hit follow on whatever a you're using.
• 20:09
  It helps us keep making this. We appreciate you being here.
• 20:13
  Also, check out our YouTube channel for more business and
• 20:15
  tech updates. There's a link in the
• 20:17
  description.