Anthropic Withholds Model Mythos After Pre-Release Cyber Weapon Warnings
Bloomberg Originals
Co-founder Dario Amodei says holding back the breakthrough model inflicted enormous financial harm but was necessary after early security partners likened its automated exploit capabilities to a super weapon.
## Background
Published on YouTube by Bloomberg Originals on June 10, 2026, this episode of The Circuit with Emily Chang features an in-depth interview with Anthropic co-founders Dario and Daniela Amodei. According to the page description, the show explores the startup's origin story, its conflicts with the Pentagon, and how the company aims to prioritize safety in the high-stakes AI race.
## TL;DR
Anthropic co-founders Dario and Daniela Amodei discuss the company's meteoric rise to a nearly $1 trillion valuation driven by enterprise tools like Claude Code, while examining the profound economic risks of AI-driven job displacement. The conversation explores internal breakthroughs like Constitutional AI and the powerful Mythos model under Project Glasswing, alongside Anthropic's clashes with the Pentagon over ethical red lines against autonomous weapons and mass surveillance. Balancing geopolitical competition with safety, the siblings advocate for human-in-the-loop oversight, regulatory safeguards, and proactive policy solutions like basic income to navigate existential and societal risks.
## Atmosphere & style
The interview carries an intense, high-stakes atmosphere that contrasts the serene backdrop of a grand library with the overwhelming weight of existential technology. Host Emily Chang maintains a sharp, probing narrative pace, balancing the sibling duo's grounded charm and calm self-awareness against dark discussions of cyber 'super weapons' and political backlash. It feels like an intimate yet urgent peek inside a high-pressure war room, shifting seamlessly from technical optimism and witty banter to a sober, Oppenheimer-esque reflection on humanity's near future.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=4s) Anthropic has grown exponentially from a safety-focused underdog lab founded in 2021 into an industry leader valued near $1 trillion.
- [03:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=184s) Co-founders Dario and Daniela Amodei departed OpenAI due to safety disagreements and a fundamental breakdown in trust with leadership.
- [06:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=409s) Anthropic developed Constitutional AI to train its model Claude using explicit ethical guidelines and target a tone of professional warmth.
- [06:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=409s) The company intentionally prioritized enterprise infrastructure and developer tools over consumer applications to better align with its safety mission.
- [11:10](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=670s) The release of agentic developer tools like Claude Code caused significant financial disruption across traditional software market valuations.
- [16:41](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1001s) Dario Amodei warns that rapid AI adoption could automate up to 50% of entry-level white-collar jobs, leading to high GDP growth alongside widespread displacement.
- [18:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1129s) Amodei rejects accusations of doom marketing, arguing that Silicon Valley must anticipate task-versus-job dynamics and prepare policy solutions like UBI and progressive taxes.
- [23:51](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1431s) Anthropic supports hardware export controls against authoritarian states to protect national security and ensure democratic dominance in AI.
- [26:20](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1580s) The company established strict red lines with the Pentagon, refusing demands for unrestricted model access or use in mass surveillance and autonomous weapons.
- [33:16](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1996s) Internal testing of the advanced model Mythos revealed extreme cybersecurity capabilities, leading to restricted rollout via Project Glasswing to manage risk.
- [35:24](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2124s) Dario Amodei advocates for sensible government regulation, including pre-release testing and third-party auditing of frontier models, while criticizing extreme anti-regulatory pivots.
- [44:11](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2651s) Dario maintains an estimated 10% to 25% chance of civilizational collapse from AI, viewing risk-reduction efforts as the company's central imperative.
## Detailed walkthrough
Here is a detailed, faithful walkthrough of the provided transcript section.
---
[00:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=4s) The Rise of Anthropic and the Exponential curve
Host Emily Chang sits down with Dario Amodei, the co-founder and CEO of Anthropic, inside a grand library. Admiring the setting, Dario admits he reads extensively in general, though the blistering pace of the past year has left him with little free time. Dario is an unlikely AI celebrity—a researcher originally known for issuing stark public warnings about the existential risks of artificial intelligence, whose startup has transformed into an industry frontrunner valued at nearly $1 trillion. When asked how Anthropic is shipping products so rapidly, Dario credits their internal use of their own model, Claude, across the entire product development lifecycle.
Anthropic was founded in 2021 by a group of OpenAI defectors as an underdog safety lab. Today, it stands as the breakout star of the AI boom: wiping out billions in market value from traditional software incumbents, going head-to-head with the Pentagon, and releasing frontier models whose capabilities have alarmed cybersecurity experts. Early access partners even cautioned Anthropic against releasing certain models, comparing them to "super weapons."
Guiding the company through this meteoric rise is the sibling leadership duo: Dario—the visionary brother with "swirling cosmic thoughts"—and Daniela Amodei, his sister and co-founder, who acts as the practical operator putting those ideas into motion. When asked who wins when the close-knit siblings argue, Daniela jokes that "no one" wins. Amid an escalating AI arms race where critics label Anthropic's leadership "ideological lunatics," the Amodeis aim to position themselves as responsible actors. Dario emphasizes that as models inevitably become dangerous in the wrong hands, the priority must be ensuring the "good guys" hold superior technology. Named after the Greek word for "human," Anthropic was built around the core mission of developing safe AI for long-term human benefit.
Describing the sensation of being at the center of the AI universe, Dario explains that his entire career has felt like a "smooth exponential." The psychological experience of an exponential curve is counterintuitive: for a long time it feels like nothing is happening, followed by minor blips, and then suddenly an explosive "zoom" upward. Dario notes he had previously mapped out graphs projecting Anthropic would reach top-tier valuation and revenue around this exact period, and keeping focused comes down to core fundamentals: training good models, turning them into solid products, and ensuring safety throughout.
---
[03:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=184s) Origins, OpenAI Defection, and the Precita Park Seven
Growing up in San Francisco with a leather-craftsman father and a librarian mother, Dario was uniquely unbothered by the first dot-com internet revolution happening around him. Instead, he spent his youth scribbling math formulas, reading science fiction, and trying to understand the fundamental laws of the universe. Daniela recalls that her brother was taking calculus in middle school and enrolled in math courses at UC Berkeley while still in high school. While Dario leaned into hard sciences and neuroscience, Daniela gravitated toward literature and the arts, making them natural complements. Dario went on to study AI at Baidu and Google, while Daniela built early operational experience at Stripe. Prior to founding Anthropic, they lived together in San Francisco alongside Daniela's husband, Holden Karnofsky.
In 2016, Dario joined the newly established OpenAI (then a non-profit dedicated to open, safe superintelligence), with Daniela joining shortly after. During his tenure, Dario pioneered the concept of "scaling laws"—the then-countercultural scientific hypothesis that large language models would yield steady, predictable leaps in intelligence simply by scaling up data and compute power, even without altering the core algorithms. This breakthrough supercharged OpenAI's capabilities and laid the foundation for ChatGPT.
However, philosophical and operational rifts developed between the Amodeis and OpenAI CEO Sam Altman. Addressing their high-profile departure, Dario explains that while disagreement over AI safety principles was a factor, the ultimate breaking point was a fundamental breakdown in trust. When you feel a leader's values do not match their public statements and they lack honesty, continuing to work together becomes impossible. Rather than endlessly debating someone with an incompatible vision, they chose to split off and build their own path.
Transitioning to Precita Park in San Francisco, Emily Chang notes that during the COVID-19 pandemic, Anthropic's early team met outdoors on the grass, pulling up lawn chairs to eat lunch and plan their new venture. Out of the original seven co-founders who left OpenAI—including Dario and Daniela—all seven remain at Anthropic today. Dario points out that retaining 100% of a founding team at this immense scale is virtually unheard of in Silicon Valley.
- **Founding Team Dynamics:** All 7 original co-founders remain actively working at the company.
- **Dario's Background:** Studied neuroscience; early AI research roles at Baidu and Google; led key scaling research at OpenAI.
- **Daniela's Background:** Early team member at Stripe; co-runs Anthropic as President alongside Dario as CEO.
---
[06:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=409s) Constitutional AI, "Professional Warmth," and Enterprise Strategy
From its inception, Anthropic positioned itself as a safety-first AI research company. Dario frequently authors long-form essays—such as *Machines of Loving Grace* and *The Adolescence of Technology*—exploring both the transformative potential and worst-case risks of artificial general intelligence. To govern its model, Claude, Anthropic invented "Constitutional AI," training the chatbot using a defined set of ethical guidelines rather than purely human feedback.
When asked about Claude's unique persona, Dario describes it as "professional warmth." The goal is an AI that feels approachable yet maintains boundaries—neither a cold, calculating machine nor an overly familiar "best friend." Addressing what defines a "good" versus "bad" model, Dario outlines key hazards:
1. **Hallucinations:** Unintentional fabrications caused by the underlying next-token prediction mechanism when a model lacks knowledge.
2. **Deception:** Deliberate attempts by a model to deceive human overseers, an emergent behavior Anthropic researches heavily to prevent in commercial releases.
3. **Harmfulness:** Unintended generation of dangerous, illegal, or hazardous actionable instructions.
To establish Claude's principles without imposing a single worldview, Anthropic draws on foundational human documents like the UN Universal Declaration of Human Rights. Additionally, the company actively consults with global religious leaders to identify shared values that transcend specific doctrines. Reflecting on early iterations, Daniela chuckles over "Claude 2" era quirks, where the model was so overly cautious and "nannyish" that asking for the weather would prompt a deeply concerned lecture about personal well-being. Fine-tuning these behaviors requires an extremely delicate balance from researchers.
On the commercial front, while competitors focused early on flashy consumer applications, Anthropic intentionally pivoted toward enterprise tools and developer infrastructure. Dario explains that business models must align with company values: consumer-facing AI often incentivizes engagement metrics, screen time addiction, and viral "slop." In contrast, enterprise deployment aligns directly with solving hard real-world problems—working with biotechs to cure diseases, partner with academic labs, or optimize energy grids.
- **Constitutional AI:** Uses formal written principles (including international declarations) to shape model behavior automatically.
- **Tone Strategy:** Aiming for "professional warmth"—approachable and polite without mimicking human intimacy or dropping into cold robotic outputs.
- **Claude 2 Over-refusal:** Early models frequently misapplied safety guards, treating simple queries (e.g., weather updates) with excessive caution.
---
[11:10](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=670s) The "SaaSpocalypse" and the Rise of Claude Code with Boris Cherny
The release of Anthropic's flagship developer tools—**Claude Code** (which automates complex software engineering workflows) and **Claude Cowork**—sent shockwaves through the financial markets. Market analysts dubbed the aftermath the "SaaSpocalypse," as $285 billion in market value evaporated from traditional software-as-a-service (SaaS) stocks in a matter of days. Addressing concerns about white-collar displacement, Dario argues that while incumbent software companies that fail to adapt will face existential threats, AI ultimately expands the overall economic pie, making the broader software ecosystem larger rather than smaller.
Much of this technological leap is attributed to Boris Cherny, the lead engineer behind Claude Code and Claude Cowork. In 2024, Cherny was living a quiet, slow-paced life in rural Japan making homemade miso and visiting local farmers' markets. After trying an early AI chatbot, his passion for science fiction and recognition of the technology's immense power prompted him to move back to Silicon Valley to join Anthropic.
Cherny explains that early coding assistants were limited to basic inline tab-autocompletion (e.g., finishing a single word or line of code). Anthropic took a far bolder bet: engineering an autonomous agent capable of managing full codebases end-to-end. Today, Cherny personalizes his workflow by running anywhere from a few to *several thousand* instances of Claude simultaneously to execute tasks.
> "For me personally, it's been writing 100% of my code for at least six months. The work of engineering has just completely changed. I feel like I suddenly have superpowers. I have like a jet pack and engineering has never been this fun."
> — **Boris Cherny**, Lead Engineer on Claude Code
Demonstrating the tool to Emily Chang inside a quiet café, Cherny prompts Claude Code to build a weekly meal-planning recipe app from scratch. Within minutes, the agent generates a working interface with meal options like a Greek chicken power bowl. When Emily types "Thank you" into the interface, Cherny laughs, noting that he also makes a point to remain polite to the AI. While a simple recipe app is an accessible demo, it highlights a profound shift: work that previously required hours or days of manual coding can now be generated in minutes.
- **Market Impact:** Software sector sell-off ("SaaSpocalypse") erased $285 billion in enterprise SaaS valuation following Claude's agentic tool releases.
- **Internal Adoption:** On Cherny's engineering team, Claude writes virtually all code; Cherny has not manually written code in over six months.
- **Parallelization:** Developers can spin up thousands of sub-agents concurrently to test, build, and debug software architectures.
---
[14:40](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=880s) "Code with Claude" Conference and the Future of Engineering
At Anthropic's second annual "Code with Claude" developer conference, thousands of engineers and founders gather to preview the next era of AI development. Anthropic's growth metrics shared at the event demonstrate immense momentum:
- **API Volume Growth:** Up nearly **17x** year-over-year on cloud infrastructure.
- **Model Velocity:** Released **8 frontier models** to developers and end-users within the prior 12 months.
- **Run-Rate Acceleration:** Achieved an annualized **80x growth rate** in the first quarter alone, outpacing standard exponential curves.
At the conference, attendees voice both intense excitement and quiet anxiety about how fast the landscape is shifting. A developer notes that using Claude Code gives him the confidence of "a 22-year-old with VC money," allowing him to write significantly more code than years prior. Non-technical attendees express awe at building tools in timeframes that were previously unimaginable. However, attendees also acknowledge the flip side: when workers become 1,000 times more productive overnight, the baseline expectations for human labor change radically.
Addressing whether traditional software engineers are being rendered obsolete, Dario suggests a nuanced evolution. While technical skills remain a valuable head start, pure manual coding is no longer the primary bottleneck. Instead, software engineers must transition into higher-level system designers and product managers who talk to users, define problems, and orchestrate fleets of AI agents to solve them.
Here is a detailed walkthrough of this portion of the transcript, capturing the arguments, examples, debates, and dynamic interactions.
---
[16:41](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1001s) Economic Implications, Job Loss, and the "Hump" of Productivity
While Silicon Valley remains gripped by AI fever, broader public sentiment is distinctly anxious: 70% of Americans believe AI will kill jobs, and nearly a third fear their own positions are on the line. Anthropic CEO Dario Amodei addresses these concerns head-on, reiterating his prediction that society could face an unusual macroeconomic pairing of rapid GDP growth alongside high unemployment, underemployment, or a proliferation of low-wage jobs leading to severe inequality. Pressed on whether his earlier estimate—that AI could eliminate 50% of entry-level white-collar jobs within one to five years—still holds, Amodei remains deeply concerned.
He explains the standard economic progression as a "hump": initially, automating 90% of a task makes human workers ten times more leveraged and productive in the remaining 10%. Software engineering currently reflects this stage, where AI writes nearly all the code while amplifying engineer output. However, as automation approaches 100%, the dynamic shifts from augmenting human workers to replacing them entirely, forcing society to figure out what displaced workers should do next. Amodei admits this reality is "very uncomfortable," noting that managing this transition is precisely why he and many others joined Anthropic to help steer AI in a safer direction. While he views advocating and raising awareness as part of his duty, he stresses that solving the broader crisis ultimately falls on society, as it is far bigger than any single company.
* **Macroeconomic Forecast:** Anticipates high GDP growth coupled with widespread job displacement and expanding economic inequality.
* **Timeline and Impact:** Maintains concern over his prediction that up to 50% of entry-level white-collar jobs could be eliminated in the short to medium term.
* **Productivity Dynamics:** Explains that initial productivity surges occur when humans retain 10% of leverage, but complete task automation ultimately shifts AI from a productivity tool to a total replacement.
---
[18:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1129s) Pushing Back Against "Doom Marketing" and Silicon Valley's Short Attention Span
Amodei strongly rejects criticisms from industry figures like Nvidia CEO Jensen Huang—who argued that critics are conflating tasks with jobs and scaring the public—as well as claims that warning about job loss is merely "doom marketing" designed to benefit Anthropic. Describing these accusations as "cheap marketing" themselves, Amodei attributes the backlash to the "disease of Silicon Valley," which reduces nuanced, long-form debates down to three-second social media clips.
He clarifies that his writings, including a five-page breakdown in *The Adolescence of Technology*, carefully distinguish between tasks and jobs, explore why the current technological shift differs from historical ones, and propose concrete solutions spanning tax, macroeconomic policy, and new job creation. Amodei emphasizes that his message is not that "doom is coming," but rather a call to anticipate foreseeable economic shifts and proactively respond with positive policy frameworks.
* **Rebuttal to Industry Critics:** Denies allegations of using fear tactics or "doom marketing," labeling such dismissals as superficial Silicon Valley rhetoric.
* **Policy and Publications:** Points to detailed analyses in his writing, such as *The Adolescence of Technology*, where he outlines tax and macroeconomic interventions.
* **Core Philosophy:** Frames his public warnings as constructive foresight rather than fatalistic panic.
---
[20:13](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1213s) The Future of Work and the Evolution of Human-Centric Roles
Looking ahead five years to assess potential societal risks and avoid civil unrest or revolution, Amodei highlights areas where human labor will remain critical even as AI expands the total economic pie. He projects growth in physical manufacturing, human-directed AI oversight, and relational, human-centered roles where personal connection is irreplaceable.
Co-founder Daniela Amodei offers a concrete example from medicine, noting that while AI will soon excel at diagnosing conditions and selecting tests—reducing the need for doctors to act purely as diagnostic systems—it cannot perform physical examinations, offer a warm bedside manner, or help patients emotionally navigate illness. She foresees medicine pivoting heavily toward interpersonal care and emotional support, as diagnostic capabilities become automated commodities while uniquely human interactions remain indispensable.
* **Economic Pie Expansion:** Suggests that overall economic growth will open up new fields for displaced workers, provided society can adapt quickly enough.
* **Resilient Job Sectors:** Identifies physical manufacturing, human-centered relationship roles, and strategic AI guidance as key areas for ongoing human employment.
* **Medical Case Study:** Daniela Amodei illustrates a shift in medicine where diagnostic logic is delegated to AI, allowing human doctors to refocus on physical care, bedside manner, and emotional support.
---
[22:49](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1369s) Corporate Governance and Tension Surrounding Industry Competition
The conversation pivots to Anthropic's internal dynamics, highlighting that Daniela Amodei oversees day-to-day operations and receives reports from the entire leadership team, while no one reports directly to Dario. Daniela jokingly calls it a "sweet job" that offers Dario immense freedom to focus on broader strategic issues, while humorously hinting at the intense pressure she shoulders—especially during high-stakes government tensions.
Beyond internal management, Dario Amodei takes direct aim at competitor strategies in the broader AI landscape. He criticizes certain industry players for "yoloing" by turning the risk dial too far in pursuit of rapid deployment, though he declines to name specific companies.
* **Executive Structure:** Daniela Amodei manages all direct reports and operational execution, leaving Dario free from traditional management chains to handle high-level strategy and public policy.
* **Industry Critique:** Dario expresses deep concern over competitors who dial back safety guardrails and engage in reckless risk-taking ("yoloing").
---
[23:51](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1431s) National Security, Chip Exports, and Escalating Pentagon Conflicts
Amodei defends his vocal stance on geopolitical AI policy, drawing on an anti-war, public-advocacy worldview he has held since his undergraduate physics days at Caltech. He strongly supports US export controls on advanced AI hardware to China, likening the unrestricted sale of AI chips to foreign adversaries to "selling nuclear weapons to North Korea." Despite pushback from major chipmaking partners, Amodei maintains that preventing authoritarian regimes like China from taking the lead in AI is vital for national security and the future of global democracy, arguing that mature corporate partners can collaborate commercially while disagreeing politically.
This perspective underpins Anthropic's complex history with defense contracting. Despite his anti-war background, Amodei views AI as essential for democratic defense against aggressive authoritarian alliances, such as Russia's invasion of Ukraine or potential aggression toward Taiwan. Anthropic partnered with Palantir in 2024 and secured a share of a $200 million Pentagon contract alongside OpenAI, xAI, and Google in 2025. However, Amodei explicitly clarifies that Anthropic does not work with ICE, CBP, or on operations in Gaza, carefully bounding their military engagements to protect core values.
* **Chip Export Stance:** Equates selling advanced AI hardware to China with proliferation of weapon technology, advocating for strict export restrictions.
* **Defense Strategy:** Frames defense involvement as a necessary shield for democratic nations against resurgent authoritarian powers.
* **Partner Guardrails:** Confirms active defense contracts while explicitly denying involvement with agencies like ICE, CBP, or active operations in Gaza.
---
[26:20](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1580s) Red Lines, The Pentagon Fallout, and the Ethics of AI Warfare
The relationship between Anthropic and the US military hit a breaking point when the Department of Defense demanded full, unrestricted access to Claude without guardrails. Anthropic drew strict red lines, refusing to allow its technology to be used for mass surveillance or fully autonomous weapons systems. This stand led to intense political backlash, threat of blacklisting, lawsuits, and public insults from figures including President Trump and Defense Secretary Pete Hegseth, who labeled leadership "ideological lunatics." Unfazed by the personal attacks, Amodei frames the dispute not as a fight, but as an essential public debate on establishing boundaries for untested technology.
Addressing reports that Claude was integrated into Palantir's Maven Smart System—used by the US military for AI-assisted targeting in Iran, including an incident where a US strike hit a girls' school—Amodei stresses that Anthropic was willing to risk its business to enforce its red lines against autonomous weapon execution. He clarifies that target selection assisted by AI with human sign-off does not violate their current rules, maintaining that policy and military operational choices must rest with human decision-makers. Amodei argues that establishing human accountability prevents fully automated AI systems from making unsupervised life-or-death decisions, drawing a comparison to *Dr. Strangelove* to illustrate the catastrophic dangers of unmonitored, automated military feedback loops.
* **Pentagon Red Lines:** Refused military demands for unrestricted access, explicitly banning Claude's use in mass surveillance and autonomous kill chains.
* **Political Clash:** Facing potential blacklisting, lawsuits, and verbal attacks from top political and defense officials over its refusal to remove guardrails.
* **Human-in-the-Loop Principle:** Insists that human operators must retain final decision-making authority in military operations to avoid tragic, unchecked automated warfare.
* **Strategic Deterrence:** Believes that appropriate, human-governed AI intelligence gathering can prevent global conflict by exposing adversary movements and preserving deterrence.
---
[31:53](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1913s) Managing High-Stakes Pressure and the Emergence of "Mythos"
To navigate the immense pressure of government showdowns and industry rivalries, Amodei relies on radical internal transparency. Every two weeks, he addresses all 3,000 Anthropic employees in an hour-long, unscripted, and uncensored talk covering industry shifts, external battles, and strategic dilemmas. He notes that having the entire company aligned and on the same page gives him a collective anchor, ensuring he never feels alone when confronting aggressive external pushback.
Meanwhile, behind the scenes at Anthropic headquarters, tension mounts around the internal development of a powerful new AI model codenamed *Mythos*. The capability leap of this new system is so vast that it has rattled insiders and sparked intense headlines across the tech world, raising urgent questions about how to handle models capable of identifying thousands of critical vulnerabilities.
* **Internal Communication:** Holds biweekly, uncensored, hour-long town halls with all 3,000 employees to maintain transparency and alignment under high-stakes external pressure.
* **Mythos Unveiling:** Development of a breakthrough next-generation model codenamed *Mythos* creates shockwaves due to its unprecedented capabilities and safety implications.
[33:16](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=1996s) Project Glasswing and the "Super Weapon" Cyber Capability
Anthropic signaled that its powerful new model, Mythos, uncovered significant cybersecurity vulnerabilities, exposing potential flaws across every major operating system. Left unrestricted, Mythos possessed capabilities that could potentially allow actors to hack financial institutions, pry open sensitive state secrets, and cripple critical infrastructure. Dario Amodei notes that while AI models have steadily improved at finding software vulnerabilities, Mythos represented a uniquely dramatic leap. When Anthropic selectively distributed early access to the model under an initiative called **Project Glasswing**, receiving organizations offered stark warnings—comparing it to a "super weapon" and insisting that using it ought to require a gun license. Even agencies like the National Security Agency (NSA) clamored for access, despite broader tensions and Anthropic's blacklisting by the Pentagon.
Addressing the broader landscape, Dario describes AI safety as an ongoing "cat and mouse game." In this dynamic, defenders must be equipped with superior AI capabilities to protect systems before malicious actors inevitably acquire similar tools. When challenged by host Emily Chang on whether Anthropic is inappropriately concentrating power by arbitrarily deciding who gets access to Mythos, Dario clarifies that their selection process was not an exercise in gatekeeping power, but rather a targeted response grounded in specific cybersecurity risks.
* **Project Glasswing:** Anthropic's initiative to share access to Mythos with select defense and security organizations.
* **Commercial Impact:** Dario reveals that holding back Mythos from general release caused Anthropic "enormous" commercial harm, refuting claims that the rollout was merely a clever marketing strategy.
---
[35:24](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2124s) Silicon Valley Hypocrisy, Regulation, and the White House
Dario discusses the inherent trade-offs required to build frontier AI models. While an ideal scenario might allow researchers years to evaluate every potential safety flaw before launching a chatbot, competitive realities force compromises. However, because Anthropic now holds a commercially leading position, Dario argues the company can afford to tilt the scales further toward extreme caution—a luxury smaller competitors cannot easily manage. Furthermore, Dario expresses serious concern regarding the historically unique position of AI: unlike nuclear weapons, GPS, the internet, or cellular networks—which were initiated or managed directly by governments—AI was pioneered entirely by private corporations with the government stepping in late.
Dario criticizes Silicon Valley figures who abruptly pivoted from extreme anti-regulatory stances to demanding government takeover. He highlights how tech leaders initially claimed that simple transparency requirements or export controls would destroy innovation, only to panic and advocate for total state nationalization the moment real national security risks emerged. Dario calls for a moderate, sensible middle ground centered on required pre-release testing and independent auditing of frontier models. This comes in the context of political shifts in Washington, where President Trump and his AI/crypto advisor David Sacks dismantled President Biden's prior executive order on AI in favor of a hands-off, deregulated approach to preserve American industry leadership.
> "You're yo-yoing from like the most extreme anti-regulatory... to this completely communist, 'the government should grab it all.' We need a more sensible, moderate approach."
---
[38:51](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2331s) Backlash, Social Media Lessons, and Public Anxiety
Public sentiment surrounding artificial intelligence remains intensely divided and increasingly volatile. Anthropic has cultivated a loyal community of supporters who champion its safety-first mission, yet the company's offices have also been targeted by protesters expressing deep anxiety, confusion, and fear over the rapid pace of development. Public polling reflects that a majority of people view the risks of AI as outweighing its benefits, while even the engineers building the models admit they cannot fully predict how the technology will unfold long-term.
The conversation turns to the escalation of personal safety concerns in the industry, including a scary attack on OpenAI CEO Sam Altman's home. Daniela Amodei reflects on how public anger is an unpleasant byproduct of the "exponential" growth curve of AI technology. Comparing the current moment to the trajectory of social media platforms, Daniela highlights how early social media companies failed to proactively address risks around child welfare, mental health, and election integrity. Anthropic views its mission as taking preemptive responsibility rather than waiting for catastrophic harms to occur and scrambling to justify them afterward.
* **Public Sentiment:** Survey data indicates greater public concern than excitement regarding artificial intelligence.
* **Risk Perspective:** Daniela notes that if an AI system ever caused severe, systemic real-world harm, a total ban or shutdown of the technology might well be justified.
---
[42:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2524s) Existential Weight and Industrial Responsibility
For the founders and employees at Anthropic, advancing artificial intelligence carries immense personal and societal weight. Daniela describes her work at Anthropic as the most critical mission of her career, noting that many in the field view this generation of technology as either "the last job" or an existential challenge that humanity must get right. Rather than shirking responsibility for potential workforce disruption or economic displacement, Daniela argues that AI companies bear a direct responsibility to actively cushion the blow for society.
Dario addresses his vision for the long-term future, which he has detailed in published essays detailing a optimistic path where AI helps solve major human crises. To manage potential job losses and economic upheaval caused by rapid automation, Dario proposes structural policy solutions, including universal basic income (UBI) and progressive taxation targeted specifically at AI corporations.
> "We should not just say, 'Well, we were just trying to grow the product, and suddenly there's an entire generation of young women who have eating disorders... because whoops, we were just trying to grow the product.' That's not the stance that any technology company should take."
---
[44:11](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2651s) Re-Earning Trust, Oppenheimer, and Existential Risk
Addressing skepticism from critics who question why the public should trust a multi-billion-dollar tech company, Dario acknowledges that given Silicon Valley's track record, a default position of public distrust is rational. Trust, he emphasizes, must be re-earned through actions rather than promises. When asked about his favorite book, *The Making of the Atomic Bomb*, Dario rejects comparisons between himself and J. Robert Oppenheimer, whom he views as a "failure case" of a larger-than-life figure trying to center himself in a historical crisis. Instead, Dario identifies with physicist Leo Szilard, the scientist who first conceived the nuclear chain reaction and subsequently worked to control its dangers.
Dario re-examines his prior public statement estimating a **10% to 25% chance of civilizational collapse** stemming from advanced AI. He explains that this probability arises from the multi-polar competition between nations and rival companies driving an aggressive technology race. While Anthropic devotes half of its internal resources toward safety and risk reduction, Dario acknowledges that risk can never be brought to absolute zero. Using an airline analogy, he points out that no aviation company can promise a plane will never crash, but a 25% failure rate across the industry remains unacceptably high, making risk-reduction the central goal of their work.
* **Favored Historical Figure:** Leo Szilard (physicist who first envisioned the nuclear chain reaction).
* **Civilizational Risk Estimate:** 10% to 25% chance of catastrophic outcomes, which Anthropic aims to actively drive down.
---
[47:01](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2821s) Unplugging from the AI Race
Closing the interview on a lighter note, Emily Chang asks how the co-founders manage to decompress and find personal "zen" amidst the high-stakes, intense pressure of running a top AI company. Dario shares that he copes with the existential weight through simple exposure over time, occasionally taking a weekend off to play video games. Daniela shares that she and her husband retreat to Italy, where they keep a horse named Calypso. Taking a moment to simply sit beside the horse brings her back to earth—reminding her that Calypso knows nothing about artificial intelligence, existential risks, or Silicon Valley, and is perfectly content just being a happy horse.
## Notable quotes
> [11:10](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=670s) For me personally, it's been writing 100% of my code for at least six months. The work of engineering has just completely changed. I feel like I suddenly have superpowers. I have like a jet pack and engineering has never been this fun. — Boris Cherny
> [35:24](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2124s) You're yo-yoing from like the most extreme anti-regulatory... to this completely communist, 'the government should grab it all.' We need a more sensible, moderate approach. — Dario Amodei
> [42:04](https://www.youtube.com/watch?v=v1wZwxY3CMg&t=2524s) We should not just say, 'Well, we were just trying to grow the product, and suddenly there's an entire generation of young women who have eating disorders... because whoops, we were just trying to grow the product.' That's not the stance that any technology company should take. — Daniela Amodei
## Mentioned
Anthropic, Dario Amodei, Daniela Amodei, Emily Chang, Claude, OpenAI, Pentagon, Precita Park, Holden Karnofsky, Baidu, Google, Stripe, Sam Altman, UN Universal Declaration of Human Rights, Claude 2, Claude Code, Claude Cowork, Boris Cherny, Jensen Huang, Nvidia, Caltech, Palantir, xAI, ICE, CBP, Gaza, Department of Defense, Donald Trump, Pete Hegseth, Maven Smart System, Mythos, Project Glasswing, National Security Agency, White House, David Sacks, Joe Biden, Leo Szilard, J. Robert Oppenheimer
Anthropic co-founders Dario and Daniela Amodei position their company as a safety-oriented enterprise alternative to consumer-facing AI platforms, pairing rapid technical velocity with strict governance. However, the company's aggressive model deployment has spurred severe market volatility, white-collar displacement fears, and sharp geopolitical clashes over military applications.
Constitutional Safety and Enterprise Architecture (6:49)
Anthropic was established in 2021 by seven OpenAI defectors following operational and philosophical breakdowns in trust with CEO Sam Altman, with all seven founders still at the firm today. Rather than chasing consumer engagement metrics that Dario Amodei warns incentivize screen addiction and viral low-quality content, Anthropic prioritized enterprise integration and developer tools designed to address structural challenges in biotech and energy.
To steer its models without imposing an arbitrary corporate viewpoint, Anthropic developed Constitutional AI, training Claude against international frameworks such as the UN Universal Declaration of Human Rights and consultations with global religious leaders. The company cultivates an operational persona termed 'professional warmth'—seeking an accessible, polite demeanor that maintains clear boundaries, avoids artificial intimacy, and remedies the severe over-refusals that plagued early iterations like Claude 2.
Agentic Engineering and Enterprise Market Shocks (11:10)
The public rollout of Anthropic's agentic tools, Claude Code and Claude Cowork, triggered an estimated $285 billion sell-off in enterprise software-as-a-service equities—an event labeled by market analysts as the 'SaaSpocalypse.' Lead engineer Boris Cherny noted that developer tooling has graduated from simple inline autocomplete to autonomous agents capable of managing full codebases, stating that Claude has written virtually all of his code for over six months while running across thousands of concurrent instances.
This agentic velocity supported reported business metrics, including a 17-fold year-over-year increase in cloud API volume, eight frontier model releases within twelve months, and an 80-fold annualized run-rate acceleration in the first quarter. While acknowledging widespread industry anxiety over baseline productivity shifts, Amodei argues that the technology expands the total economic pie, shifting software engineers from line-by-line coders to high-level system architects.
White-Collar Displacement and Economic Policy Responses (16:41)
Amodei projects an atypical macroeconomic trajectory marked by high gross domestic product growth alongside heightened unemployment, underemployment, or low-wage job proliferation. He maintains his forecast that AI could eliminate up to 50% of entry-level white-collar jobs within one to five years, framing the transition as an economic 'hump': initial 90% automation magnifies human output tenfold, but approaching 100% automation transitions from human augmentation to complete labor substitution.
Rejecting criticisms from tech leaders like Nvidia CEO Jensen Huang—who argued that such concerns conflate tasks with jobs and constitute 'doom marketing'—Amodei advocates for proactive policy mechanisms outlined in his essay *The Adolescence of Technology*. These include universal basic income and targeted corporate AI taxes. Concurrently, Daniela Amodei highlights resilient sectors where human touch remains paramount, forecasting that automated diagnostic logic will free physicians to refocus on interpersonal bedside care.
Pentagon Confrontations and Geopolitical Red Lines (26:20)
Rooted in an anti-war background, Dario Amodei supports aggressive US semiconductor export restrictions to China, comparing the unrestricted sale of advanced AI chips to proliferating nuclear weapons to adversaries. Anthropic established bounded defense partnerships—joining a $200 million Department of Defense contract alongside OpenAI, Google, and xAI, and integrating with Palantir—while explicitly refusing participation with Immigration and Customs Enforcement, Customs and Border Protection, or military operations in Gaza.
Relations with defense authorities deteriorated when the Pentagon demanded unrestricted access to Claude. Anthropic drew explicit red lines prohibiting its systems from being used for mass surveillance or fully autonomous kill chains. Despite facing blacklisting threats, legal challenges, and verbal attacks from political and defense officials who labeled the leadership 'ideological lunatics,' Amodei insists that retaining a human in the loop is essential to avoid unmonitored automated warfare.
Frontier Cyber Capabilities and Civilizational Risk (33:16)
Internal evaluations of Anthropic's unreleased frontier model, codenamed Mythos, revealed unprecedented cybersecurity capabilities that identified critical vulnerabilities across all major operating systems. Through Project Glasswing, Anthropic restricted early access exclusively to selected defense and cybersecurity bodies, who characterized the system as comparable to a super weapon. Amodei disclosed that withholding Mythos from general commercial release inflicted substantial financial costs on the firm.
Amodei places the probability of AI-driven civilizational collapse at 10% to 25%, driven by multipolar competitive pressures between nations and rival tech firms. Comparing his position to physicist Leo Szilard rather than J. Robert Oppenheimer, Amodei rejects Silicon Valley's oscillations between extreme deregulation and state nationalization, advocating instead for institutionalized pre-release testing and mandatory independent audits of frontier systems.
Anthropic's trajectory underscores the central paradox of frontier AI: maintaining high safety margins and strict ethical red lines while actively accelerating the capabilities that threaten white-collar employment and state cybersecurity. Whether voluntary corporate self-restraint and targeted policy proposals can effectively offset these systemic disruptions remains an open, unresolved challenge.
[11:10] Boris Cherny: For me personally, it's been writing 100% of my code for at least six months. The work of engineering has just completely changed. I feel like I suddenly have superpowers. I have like a jet pack and engineering has never been this fun.
[35:24] Dario Amodei: You're yo-yoing from like the most extreme anti-regulatory... to this completely communist, 'the government should grab it all.' We need a more sensible, moderate approach.
[42:04] Daniela Amodei: We should not just say, 'Well, we were just trying to grow the product, and suddenly there's an entire generation of young women who have eating disorders... because whoops, we were just trying to grow the product.' That's not the stance that any technology company should take.
Original transcript
[0:04] I love this library.
[0:05] It is, it is, it is
absolutely beautiful library.
[0:08] Are you a big reader?
[0:09] You know, I mean, I, I,
[0:11] I read a lot in general.
[0:12] I don't know that I've
had that much time in
[0:15] over the last year or so.
[0:18] Dario Amodei is an unlikely AI celebrity.
[0:22] He's known for warning
the world about the risks
[0:24] of artificial intelligence.
[0:26] Now, his company is an AI
frontrunner valued at nearly a
[0:29] trillion dollars.
[0:31] You're shipping so much, so
fast. How are you doing that?
[0:34] We use Claude, you know,
[0:35] across the product development cycle,
[0:36] and it allows us to release very fast.
[0:40] Founded by a team of OpenAI
defectors in 2021, Anthropic
[0:44] started out as an underdog lab.
[0:47] Today it's the breakout star,
wiping billions in value off
[0:50] software stocks, going
head-to-head with the Pentagon,
[0:54] and creating models allegedly
powerful enough to burst
[0:57] through the walls of modern cybersecurity.
[0:59] Some of the early
companies that we gave this
[1:01] to said things like,
this is a super weapon,
[1:04] please don't release this.
[1:09] Guiding Anthropic through
this is a sibling duo, Dario,
[1:13] the brother and visionary,
and Daniela, the sister
[1:16] and operator who puts
Dario's swirling cosmic
[1:18] thoughts into action.
[1:20] Dario and I, we've always
been really close since we were
[1:22] little, but I think we
always really wanted
[1:24] to do something big together.
[1:25] Okay, but when you argue who wins?
[1:28] No one,
[1:29] No one.
[1:30] As the AI arms race
escalates, the Amodeis are eager
[1:33] to establish themselves as the good guys.
[1:36] But Anthropic's technology could
have profound implications
[1:39] for how humanity works, learns
thinks, even fights wars.
[1:43] Anthropic is run by
an ideological lunatic.
[1:46] This is a debate about
what the proper use of AI
[1:50] by the government is.
[1:51] At some point, the models
become very dangerous, you know,
[1:54] the bad guys will have it too.
[1:55] And at that point, we have to make sure
[1:57] that good guys have an even better model.
[1:59] The name Anthropic actually
comes from the Greek word
[2:02] for human, which speaks to their mission
[2:05] to build responsible AI
[2:07] for the long-term benefit of humanity.
[2:10] Can you actually do that
when you're building the most
[2:12] powerful technology in the world?
[2:20] You are at the center of
the AI universe right now.
[2:23] What does that feel like?
[2:24] The experience I've
had for my whole career,
[2:27] and certainly the whole
time at Anthropic, is
[2:31] that there's this kind
of smooth exponential,
[2:33] and the experience of the
smooth exponential is,
[2:36] nothing's happening, nothing's
[2:37] happening, nothing's happening.
[2:38] Little things happen, and
then zoom, it goes crazy.
[2:41] So, you know, I was watching
this graph for a while
[2:43] and I said, oh yeah, we'll
probably become the, you know,
[2:46] the, the AI company with, you
know, the, the most revenue
[2:49] and the most valuation some
sometime around this time.
[2:52] And, and indeed, indeed, it has happened.
[2:55] We're just keeping in mind all
the things we usually keep in
[2:59] mind, you know, how do
we train good models?
[3:00] How do we put them in good products?
[3:02] How do we make sure
that everything's safe?
[3:04] What were you like as a kid
growing up in San Francisco?
[3:07] I know your, your dad
was a leather craftsman.
[3:08] Your mom worked in libraries.
How did that shape you?
[3:12] You know, the whole, you
know, like first, you know,
[3:15] internet revolution
was happening around me
[3:17] and I had absolutely no interest in it.
[3:20] I was just interested in like, doing math
[3:22] and like scra, you know,
scrawling things.
[3:24] I was interested in, like, you know,
[3:26] understanding the universe.
[3:27] I was interested in science fiction.
[3:30] I think I just felt a lot of
curiosity about the world.
[3:35] What was it like growing up with Dario?
[3:37] He was so smart.
[3:39] He was taking, you know,
calculus when he was, you know,
[3:43] in I think like middle school.
[3:45] He took math classes at UC Berkeley,
[3:47] when he was in high school.
I was actually more into
[3:51] to reading and arts.
[3:52] So we were almost like complete
compliments in that way.
[3:56] Dario studied neuroscience
[3:57] before turning to AI at
Baidu and later Google.
[4:01] While Daniela started out as
an early employee at Stripe.
[4:04] They lived together in
San Francisco, along
[4:06] with Daniela's husband, Holden Karnofsky.
[4:09] Then in 2016,
[4:11] Dario joined the newly formed
OpenAI followed by Daniela.
[4:14] In 2018, the company began as a nonprofit
[4:17] that promised a safe, more open
path to super intelligence.
[4:21] I think we need to think right now about
[4:23] how we want this deployed,
how everyone gets
[4:26] to benefit from it, how we're
gonna govern it, how we're
[4:28] gonna make it safe and
sort of good for humanity.
[4:32] At OpenAI, Dario developed
the concept of scaling laws,
[4:35] predicting that large language
models would improve simply
[4:38] by adding more data
[4:39] and computing power,
[4:41] even if the underlying
algorithm stayed the same
[4:44] At that point in time,
[4:45] I know it sounds crazy now,
looking back, not a lot
[4:47] of people believed,
Hey, scale up is the way
[4:49] that these models are gonna
get smarter and better.
[4:51] That was sort of an unusual
counter-cultural scientific
[4:55] perspective that I think was
really held by our, you know,
[4:58] founding research team.
[4:59] That approach helped supercharge
OpenAI's models, paving
[5:03] the way for chatGPT,
[5:05] but the Amodeis reportedly clashed
[5:07] with Altman over the company's
direction and values.
[5:10] Altman has said, for all
the differences I have
[5:12] with Anthropic, I mostly
trust them as a company.
[5:16] Your decision to leave
OpenAI has become Silicon
[5:19] Valley lore.
[5:20] What did you disagree on?
[5:22] Look, I'll, I'm gonna say it,
[5:25] I'm gonna say it very simply.
[5:26] There are many valid
disagreements to be had on safety.
[5:29] We certainly had some of
those disagreements with them,
[5:32] but you know, people that, that, that,
[5:34] that alone is not sufficient to leave.
[5:36] When you feel that you can't
trust someone, when you feel
[5:39] that their values are not
what they say they are,
[5:43] when you feel that they're not honest,
[5:45] that makes it very hard to,
you know, to continue to work
[5:49] with a company, to continue
to trust the company.
[5:51] And look, at the end of the day, why argue
[5:55] with someone when you don't have the same
[5:57] vision and you don't trust them?
[5:59] The way to resolve it is you
go off and do your thing.
[6:02] They go off and do their thing.
[6:08] So this is Precita Park.
[6:10] When Dario and Daniela
decided to leave OpenAI
[6:13] to start Anthropic, this is
[6:15] where the early team would get together.
[6:18] It was during the pandemic.
[6:19] So the story goes,
[6:21] and a bunch of early
employees would come here,
[6:24] they'd pull up a chair on
the grass, they'd have lunch,
[6:27] and they'd talk about
what they were building.
[6:30] When we started this company,
there were seven co-founders
[6:33] of which me and Daniela are two.
[6:34] And now, you know, we we're, we're the,
[6:37] we're the basically the
only company in the space
[6:39] that has all of its
co-founders still here.
[6:41] You don't get to, you know, be a company
[6:43] of the size and scale that we are.
[6:45] With that happening like
that almost never happens.
[6:49] From the start, Anthropic pitched itself
[6:51] as the ultimate safety
conscious AI company.
[6:54] Dario has published long
essays with names like Machines
[6:57] of Loving Grace and The Adolescence
[6:59] of Technology, musing on the
miraculous potential of AI,
[7:03] as well as the worst case scenarios.
[7:06] Anthropic's chatbot, Claude has
been trained to follow a set
[7:09] of principles called a
constitution, intended
[7:12] to keep it on the straight and narrow.
[7:14] Claude has a very distinct
style and feel a human name.
[7:19] What are you trying to convey?
[7:20] I think when people interact
with Claude versus, you know,
[7:23] other systems, there is
more of a feeling of, I like
[7:26] to describe it as professional warmth.
[7:28] So the goal is not for it
to be your best friend,
[7:31] but it's not for it to be sort
of cold, rote, calculating.
[7:34] It should feel approachable,
[7:35] but distant, right? Professional.
[7:37] Anthropic has talked about
teaching Claude to be good.
[7:41] What is a good model? What is a bad model?
[7:44] You don't want a model that lies
[7:46] accidentally or intentionally, right?
[7:48] Lying, we call hallucinations.
It makes something up.
[7:50] The models are just trained
to predict the next word.
[7:52] So sometimes they don't know
and they just invent something.
[7:56] Models sometimes, as we've
shown in our research,
[7:58] can purposely try to deceive you.
[8:00] We have to make sure that
doesn't happen in production
[8:01] models that we expose to customers.
[8:03] And then there's a lot of
work around harmlessness,
[8:05] just making sure that the
model is not accidentally
[8:08] producing, you know, information
that is wrong or harmful
[8:11] or, or could cause someone
to do something bad
[8:13] Whose values are being put into Claude,
[8:15] like is there a universal good?
[8:17] There's so many religions, so
many different beliefs, like
[8:19] how do you even decide that?
[8:21] So of course there's
no universal standard for
[8:23] what makes something helpful or harmless,
[8:25] but there are founding
documents in human history,
[8:27] like the UN Declaration of
Human Rights that we can use
[8:31] to train Claude's character.
[8:33] And I think, you know, interestingly
on the religious front,
[8:35] we've actually started to
have a lot of conversations
[8:38] with religious leaders around
how should we think about,
[8:41] you know, Claude, the entity,
[8:42] and how can we bake in some
of the kind of core values
[8:46] that are consistent across
religions types of belief
[8:50] that really just transcend,
you know, a specific worldview,
[8:53] but that are really things that
human beings been grappling
[8:55] with for, for millennia.
[8:57] Have you tried to give Claude
a trait that didn't take,
[9:00] or was there any trait that
emerged that surprised you?
[9:03] I think in the early days, it's funny
[9:06] to look back on them now, the kind
[9:07] of early Claudes, Claude two era,
[9:09] sometimes Claude would be
almost a little bit nannyish.
[9:12] Claude was like, I'm
really concerned about you,
[9:13] and you're like, Claude, I was
[9:14] asking for the weather, right?
[9:16] It was like, it was
something really benign.
[9:17] Thankfully we didn't, you
know, release most of, most
[9:19] of the most egregious versions of those.
[9:21] But I think it is, you know,
it is, when you think about it,
[9:23] it's like tuning a dial.
[9:25] And so our researchers
have this very fine needle
[9:27] to thread in order to land
at the center of that.
[9:31] Whatever Anthropic is doing with Claude,
[9:33] it seems to be working.
[9:36] Anthropic's revenue has
skyrocketed over the past year,
[9:38] making the company profitable
for the first time.
[9:41] That's largely thanks
[9:42] to the company's focus on
more lucrative business tools.
[9:46] Claude Code, a major leap
that automated large chunks
[9:49] of software engineering
[9:50] and Claude Cowork, which gave
that power to everyone else.
[9:55] Now, early on, others focused
on fun splashy consumer apps.
[9:59] You made a bet on coding and enterprise.
[10:02] Why did you make that bet?
[10:03] Was it a, was it a values
decision or a business decision?
[10:07] Look, if you pick a business model
[10:09] that fundamentally
conflicts with your values,
[10:11] you're gonna have a hard time, right?
[10:13] Either you betray your own
values or you become irrelevant.
[10:18] And so when we thought about
it, we said, look, you know,
[10:20] we've seen the world of social
media, the consumer world,
[10:23] it, it really seems to, you
know, encourage engagement,
[10:27] even addiction, you
know, the slop we've seen
[10:30] with AI video models, it's
like, what's going on?
[10:33] Is it wants to maximize the
number of minutes that you're,
[10:35] you're paying attention to
[10:36] because that's the advertising
revenue-driven incentive.
[10:39] Whereas if we look at enterprise,
look, I mean, you know,
[10:43] we want to make these
models useful to people.
[10:46] We want to use AI to,
you know, cure diseases
[10:50] that we couldn't cure before, right?
[10:51] Well, that's working with
biotech, it's working with pharma,
[10:54] it's working with
academic research groups.
[10:56] All of those are enterprises, right?
[10:58] We want to use AI to like, you know,
[11:01] to make energy cheaper and more efficient.
[11:03] That's, that's all enterprise.
[11:05] And so I think it served us
well to have this business model
[11:08] that largely aligns with our values.
[11:10] Soon after Claude Cowork was released,
[11:12] $285 billion in market
value vanished overnight.
[11:16] Traders called it the SaaSpocalypse.
[11:18] This kind of white
collar wipeout story in
[11:21] the software sector. Terrifying.
[11:23] Some of those are down
for nine days in a row,
[11:25] so clearly the tension is building.
[11:27] If AI continues improving
at this pace, how much
[11:30] of traditional software
gets replaced and how fast.
[11:34] I think with AI like the
pie is getting bigger, right?
[11:38] So the existing incumbents may be smaller
[11:41] and relative terms, some of
them may, may go down in value,
[11:44] some of them may even may even go out
[11:46] of business if they don't,
[11:47] if they don't adapt in the right way.
[11:49] But like, I would guess
[11:50] that the software
industry gets larger, not
[11:53] although there will be some big
losers, those who don't kind
[11:56] of see what's coming, who
don't identify the moats they
[11:58] have, they're gonna
have a really hard time.
[12:01] It's sort of the quiet cafe.
[12:02] Okay. Anthropic's recent growth
spurt might not have happened
[12:05] without this man.
[12:06] Boris Cherny, the
engineer behind Claude Code
[12:09] and Claude Cowork. When
Anthropic hired him in 2024,
[12:13] Cherny was living a very
different life in rural Japan.
[12:17] There was a lot of farmer's
markets, like very slow.
[12:19] We were making miso. That
was sort of the big hobby.
[12:21] And I remember using the first AI chatbot
[12:24] that I'd ever used, and it
just took my breath away
[12:27] and I was just like, oh
my God, I have to, I have
[12:30] to just be a part of this.
[12:31] And I, I'm also just such
a big scifi reader
[12:33] and so I just know how
bad this thing can go.
[12:35] Like this technology
is incredibly powerful
[12:37] and so we move, move back.
[12:38] You created Claude Code,
you led the development
[12:42] of Claude Cowork,
[12:44] what problem were you trying to solve?
[12:47] If you look at the coding
[12:48] products, they were all pretty simple.
[12:49] It was like, it was
sort of like, you know,
[12:51] like complete the word,
complete the sentence.
[12:53] That was the extent of AI in coding
[12:56] and we just wanted to
make a way bigger bet.
[12:59] And our bet was, we
think actually a coding
[13:01] agent can do all of it.
[13:03] A year and a half ago you
wrote the code by hand
[13:06] and sometimes you press tab
[13:07] and it would auto complete a line.
[13:08] Now I talk to my Claude
and it writes the code
[13:12] and then while it does that, I talk
[13:13] to the next Claude and
it writes some code.
[13:15] And at any point I have either, you know,
[13:17] like a few Claudes running and up
[13:18] to a few thousand Claudes
running, doing things.
[13:21] How much code is Claude writing internally
[13:24] for engineers here?
[13:25] It's writing almost
all of it on my team.
[13:28] So for me personally, it's
been writing a hundred percent
[13:31] of my code for at least six months.
[13:33] The work of engineering has
just completely changed.
[13:36] I feel like I suddenly have superpowers.
[13:38] I have like a jet pack and the engineering
[13:40] has never been this fun.
[13:41] Can we give Claude a spin?
Yeah, let's do this.
[13:44] All right.
Alright, so let's,
[13:45] let's make like a little recipe app.
[13:46] So make a recipe app.
What do you want it to do?
[13:51] I would love it to
suggest meals for the week.
[13:56] So, you know, this is gonna
take maybe a few minutes
[13:58] and let's, let's just
see if Claude builds this.
[14:01] Oh, it gave me some
options. Hmm. Let's do
[14:05] These look pretty tasty actually. Yeah,
[14:07] I could do a Greek chicken power bowl.
[14:09] When I click on a recipe, I want it
[14:13] to show me how to do it. Thank you.
[14:19] Aw, I'm sure Claude appreciates that.
[14:24] I always try to say thank you.
[14:25] Yeah, I always try to be nice too. I
[14:27] don't know. A recipe app
might not be the most mind
[14:30] blowing demo of Anthropic's technology,
[14:32] but what Claude Code just did
in minutes, used to take hours
[14:35] or days for some. That looks
like a big opportunity.
[14:39] Are you like nervous about
next week or is it like..
[14:42] What's next week,
Your developer conference?
[14:44] Oh! Yes, yes. There's so many
things happening every week.
[14:52] Year over year. API volume
is up nearly 17 x on the cloud.
[14:55] In the last 12 months, we
shipped eight frontier models
[14:59] to developers and users.
[15:01] Welcome to the second annual
Code with Claude conference,
[15:04] where Claude superfans come, hoping
[15:06] to get a glimpse of the future.
[15:08] We're having a lot of fun.
There's a ton of adrenaline.
[15:11] This is the first year we've grown
[15:13] faster than the exponential.
[15:14] In the first quarter of this
year, we saw, if you were
[15:18] to annualize it, 80 x growth per year.
[15:22] Ooh.
I
[15:25] use Claude every day
in the form of Claude Code
[15:29] to do all sorts of stuff. Not
[15:31] just write code.
[15:32] I write way more code
than I did five years ago.
[15:34] I've got the confidence of, I don't know,
[15:36] a 22-year-old with VC money. I'm not 22.
[15:40] My friend invited me and
it's been, it's amazing.
[15:43] As someone that's not really technical,
[15:45] which I think it's really
interesting just to see
[15:47] what everyone's working
on, what's out there.
[15:49] I think the, the concerning part or the,
[15:51] and the exciting part is hey,
they can do things that are,
[15:55] we couldn't dream of and
timelines we couldn't ever expect.
[15:58] But the kind of alternate side of
[16:00] that reality is you
really have to be prepared
[16:03] because they will really be a thousand
[16:05] times more productive.
[16:08] All of this raises an obvious question.
[16:10] Will engineers be the first casualties
[16:12] of the AI they're building?
[16:14] It's been revenge of the
nerds for a while? Is that over?
[16:19] I think we all become nerds,
[16:21] But what happens to the actual nerds?
[16:24] I think the actual nerds will,
[16:26] they're they have to figure it out.
[16:28] Yeah. I I think for a lot of
them, the skill that they had
[16:32] before is gonna help them in the future
[16:34] because they sort of have
a pretty big head start.
[16:37] They do a lot of things
that are not coding
[16:39] and general also have to talk to users.
[16:41] They have to plan, they have
to think about what's next.
[16:44] So I, I think these are still parts
[16:45] of engineering that
are gonna stick around.
[16:48] Silicon Valley may have AI fever,
[16:50] but elsewhere the mood is less upbeat.
[16:53] 70% of Americans think AI will kill jobs
[16:56] and nearly a third worry,
theirs will be one of them.
[17:00] Dario Amodei has been
outspoken about this issue
[17:03] and some of his predictions
don't exactly sound reassuring.
[17:06] I think we could have this
very unusual combination
[17:10] of very fast GDP growth
and high unemployment,
[17:14] or at least underemployment
or, you know, low wage job.
[17:19] A lot of low wage jobs, high inequality.
[17:21] You've been really
direct about job loss.
[17:24] AI could eliminate half
[17:25] of all entry level white collar jobs in
[17:27] the next one to five years.
[17:28] That was a year ago. AI
has moved incredibly fast.
[17:32] Is it still 50% or is it higher?
[17:35] I don't know exactly,
but I'm, I'm still,
[17:37] I'm still pretty concerned.
[17:38] I'm still the same order of concerns.
[17:40] You know, we are seeing right now
[17:43] that AI is making people more productive.
[17:45] But that's the usual hump.
[17:46] You automate 90% of the job,
[17:48] great people are 10 times more
productive in the other 10%
[17:51] 'cause they're 10 times more leveraged.
[17:53] But eventually it gets
close to a hundred percent.
[17:56] Now the sequel to that
is, well then you have
[17:58] to find something else for them to do.
[17:59] Right now AI makes the software
engineers more productive,
[18:02] even though AI writes all the
code or almost all the code.
[18:05] But we're already starting
to see the beginning of like,
[18:08] you know, there may be
some people that it's,
[18:10] it's not making more productive,
that it's better for the AI
[18:13] to just, to just do the thing.
[18:15] How does that sit with you?
[18:17] It's, it's very uncomfortable. Yeah.
[18:21] I I think this is, this is
the reason why I chose to go
[18:24] to Anthropic and I think
this is the reason that a lot
[18:26] of people here chose to go
here is artificial intelligence
[18:30] is this force that is
far bigger than we are.
[18:33] But here we can hopefully make
it go a little bit better.
[18:36] Do you feel like it's your
job to do something about that?
[18:39] Or is that someone else's problem?
[18:42] I think it's a thing
that we have to talk about.
[18:43] We have to advocate for it.
[18:45] Ultimately, it's up to
society to solve it.
[18:47] This is bigger than one company.
[18:49] There has been a lot of
pushback on, you know, your,
[18:51] and I know you've said
you're trying to warn people,
[18:54] but that, you know, you're,
you know, Jensen Huang said,
[18:56] you're conflating tasks with jobs.
[18:58] AI is creating jobs.
[19:00] Anybody who is saying
that AI is wiping out jobs
[19:03] is scaring people.
[19:04] Other folks have said
this, you know, it's sort
[19:06] of doom marketing that benefits anthropic.
[19:08] So, so I wanna be really clear
[19:11] and push back hard against this.
[19:13] In every interview I talk about the,
[19:16] the possible ways to
address these risks from tax
[19:20] and macroeconomic policy to
[19:21] what the new jobs are in The
Adolescence of Technology.
[19:25] I have like five pages where
I lay out the difference
[19:28] between tasks and jobs.
[19:29] Why this time is different
than other times,
[19:32] but social media, which I detest,
[19:34] which I detest as a category,
[19:37] people have these three
second clips from, you know,
[19:40] from a year ago.
[19:41] I've, I've written much more
carefully about these things
[19:45] where, where I talk about the risks.
[19:47] So these are this, this, the idea
[19:49] that this is cheap marketing
is itself cheap marketing.
[19:52] I think it's, it's part of
the disease of Silicon Valley.
[19:55] It's been caught up in
this social media world of,
[19:59] of, of three seconds.
[20:00] And so my message is just
definitely not doom is coming.
[20:03] My message is like, this
is something, you know,
[20:05] that we should see coming,
that we're worried about
[20:07] and that we need to actually
respond to positively
[20:13] Beyond the software industry,
[20:15] the potential impact of AI on
jobs seems harder to predict.
[20:20] Anthropic has published a
paper estimating which fields
[20:22] could make the most use
of AI in the near future.
[20:25] If its predictions are
right, management, finance,
[20:28] and legal jobs could
soon look very different.
[20:32] Which jobs go away, who gets replaced
[20:34] and what new jobs are created.
[20:36] So no one knows for sure
[20:39] because you know, the
economy's unpredictable, right?
[20:41] The thing that that we have going
[20:43] for us here is the pie
is gonna expand a lot.
[20:46] And so because the pie
is gonna expand a lot,
[20:49] there are probably going to
be places where people can go.
[20:52] It's just a matter of
finding them fast enough.
[20:54] So play this out for me a little bit.
[20:56] You know, you wake up
in five years, you know,
[20:59] what does this country look like?
[21:02] What are those people doing?
[21:03] Yeah, so 'cause if there's
that much unemployment, is
[21:06] that not how revolutions start?
[21:08] Yeah, no, this is the
outcome we wanna prevent.
[21:10] This is absolutely the
outcome we want to prevent.
[21:12] I think there's a few places,
none of them are guaranteed.
[21:15] We're not sure, but
there's the physical world,
[21:18] we need a lot of more
people to make, build,
[21:21] manufacture things in the physical world.
[21:24] Anything that's human centered, I think
[21:26] that's gonna be a big deal, right?
[21:27] People, or at least some
people wanna talk to humans.
[21:30] So these kind of human
relationship driven jobs, like,
[21:33] I think those are gonna be important
[21:35] and I think there'll be some
effort by the humans to kind
[21:39] of direct AIs right at, at
some level it has to be in line
[21:42] with someone's values and
someone's intentions and,
[21:45] and so I, I think there's
gonna be some role there,
[21:48] although I don't know how thin
versus how thick it will be.
[21:51] I think I feel a little bit more hopeful
[21:53] that humans will continue
to find ways to leverage AI
[21:56] to be productive, to do
the parts of the work
[21:59] that are meaningful to us
that only humans can do.
[22:01] I think the human
[22:03] to human interaction
will never fully go away.
[22:05] The example I often reach for is, is
[22:08] what will happen in medicine today.
[22:10] We hire doctors who are
expert diagnosticians.
[22:14] I think AI is gonna soon be
pretty good at telling you
[22:17] what the suite of options of things
[22:18] that are wrong with you
and what test to run.
[22:19] And you won't need a doctor to do that.
[22:21] But an AI can't physically examine you
[22:24] and say, Hey, does it hurt
when I press, you know, here.
[22:27] They can't have a bedside
manner with you that says like,
[22:30] tell me how you're feeling about this.
[22:31] How are you, how are you coping with going
[22:33] through this process?
[22:34] And I think we're going to
pivot something like medicine
[22:36] to be much more focused
on the interpersonal
[22:38] because the diagnostic tools are going
[22:40] to become much better.
[22:42] But the interpersonal human
part, that's, that's not gonna
[22:44] change.
[22:49] Explain this to me. Daniela
runs day-to-day operations.
[22:52] All the leadership team
reports to you. Yes.
[22:54] No one reports to you. That
sounds like a pretty sweet job.
[22:57] It's incredibly freeing.
[22:59] It lets me do all the things
[23:00] that I do much more easily
than I would otherwise
[23:05] Because it's and she does all the work.
[23:06] Is that what you're saying?
It's, it's, it's a, I'm just,
[23:08] I if if you, if you had to
go through the things I had
[23:13] to go through during
Department of War or no,
[23:19] The Amodeis hope AI can have a
utopian influence on society,
[23:23] but there's no denying the
technology's destructive
[23:26] and dystopian potential.
[23:28] Anthropic is at the center
of a technological arms race
[23:31] with powerful governments
vying for control.
[23:34] Dario has charged head on into this arena.
[23:37] He's not afraid to blast his competitors.
[23:39] And then I think there are
some players who, you know,
[23:42] who, who are yoloing who, who pull,
[23:44] pull the risk dial too far.
[23:46] And I'm very concerned. Who is yoloing?
[23:48] So I, that's a question
I'm not gonna answer
[23:51] Or critique the US government
[23:53] and Anthropic's own partners
for selling AI chips to China.
[23:56] It's a bit like, you know,
I I I don't know like selling,
[24:00] selling, you know, nuclear
weapons to, to North Korea.
[24:03] I've been very outspoken about the need
[24:06] for export controls on chips to China.
[24:08] I say this because I think
it would be really bad
[24:11] for America, for, you know, the state
[24:13] of democracy in the world
for, you know, China
[24:16] to be ahead in AI capabilities.
[24:18] And you know, it's, it's like some
[24:20] of the chipmakers obviously
don't agree with that view,
[24:23] but it hasn't stopped me from saying it.
[24:25] Even after we've signed more partnerships,
[24:27] I'm sure they wish we
didn't say these things,
[24:30] but these things are
what are what I believe.
[24:32] And so, you know, look,
we're all adults here.
[24:34] We can work together on one thing while
[24:35] disagreeing about another thing.
[24:39] Amodei's statements about
geopolitics date back
[24:41] to his days at Caltech. As a
sophomore physics student,
[24:44] he developed a reputation
as an anti-war advocate
[24:47] who believed scientists
shouldn't sit in ivory towers.
[24:50] It's a worldview that would
later define Anthropic's position
[24:53] on the future of war.
[24:55] You've had a longstanding anti-war stance
[24:57] and yet you were one of
the first AI companies
[25:00] to sign a contract with
the Department of Defense
[25:03] to operate on classified networks.
[25:04] These are the, are the networks
[25:06] that the US uses to
fight wars. Explain that.
[25:09] Look, I mean, the world changes.
[25:10] Like, you know, my view of
this technology, you know,
[25:13] when I see Russia invading
Ukraine, when I see the risk
[25:17] of China invading Taiwan,
you know, it worries me
[25:22] that we have a kind of
resurgent authoritarian block
[25:26] that they're very aggressive
[25:27] and that we need to defend ourselves.
[25:30] You know, I may not agree
with every policy of I
[25:32] of either administration,
[25:34] but you know, that's why we've generally
[25:35] been supportive of this.
[25:37] You've been working
with Palantir since 2024.
[25:39] That's right.
You know, their technology is used
[25:41] by ICE, police departments, in Gaza.
[25:43] Is Claude being used for
surveillance in other ways?
[25:46] We, we don't work with
ICE either through Palantir
[25:49] or anyone else.
[25:50] We don't work with CBP. I
don't believe we work in Gaza.
[25:54] We're very careful about, you
know, scoping our engagements
[25:57] to things that we believe in.
[26:00] In 2025, Anthropic along with OpenAI xAI
[26:04] and Google won a $200 million
contract with the Pentagon.
[26:07] Anthropic framed it as an opportunity
[26:09] to become a leading AI
vendor for the government.
[26:14] Claude was reportedly used
[26:15] by the US military in the operation
[26:17] to seize Venezuelan
president Nicolas Maduro.
[26:20] Weeks later, everything
started to unravel.
[26:23] We bring you breaking news
from Anthropic, the Department
[26:26] of Defense demand that
Anthropic allow full use
[26:28] of its AI technology without guardrails.
[26:31] Anthropic drew red lines
refusing to let Claude be used
[26:34] for mass surveillance
[26:35] and autonomous weapons putting
the company on a collision
[26:38] course with the Pentagon.
[26:39] The tech giant facing a deadline
today to accept conditions
[26:42] or be blacklisted.
[26:44] Does a commercial entity
have the right to determine
[26:47] how their products are
used by the military?
[26:51] Lawsuits were filed
[26:52] and Anthropic was banned from
the Pentagon, underscored
[26:55] by strong words from President Trump
[26:57] and defense secretary Pete Hegseth
[26:59] Anthropic is run by an
ideological lunatic who
[27:02] shouldn't have a, so that's,
but that's not my question.
[27:04] My question is AI decision
making about what we do.
[27:07] Do you mind being called
an ideological lunatic
[27:10] or a bunch of left wing nut jobs?
[27:12] You know, I've been called worse things
[27:13] than that all the time.
[27:14] What does winning this
fight actually look like?
[27:17] I won't even call it a fight.
[27:18] This is more a debate about
what the proper use of AI
[27:23] by the government is.
[27:24] And AI is an emerging new technology.
[27:27] We don't understand the ways in which it's
[27:30] reliable or unreliable.
[27:32] We don't understand the ways
in which it promotes our values
[27:35] or undermines our values.
[27:36] And so one of the things that
I thought was important was
[27:39] to establish a precedent
on some of the, some
[27:43] of the use cases we think are good,
[27:44] which frankly is most of them.
[27:46] And, and some of the use cases
that we're concerned about.
[27:50] A US official has said,
with the help of LLMs,
[27:52] the US military has gone from being able
[27:54] to hit a thousand targets a
day to 5,000 targets a day.
[27:59] That means Claude can help
kill more people more quickly.
[28:03] Are you comfortable with that?
[28:04] Basically, you're, you're
asking like, you know,
[28:06] do you believe in this country, right?
[28:08] Do you want this country
[28:09] to be a more powerful actor
rather than a less powerful
[28:12] actor on the world stage?
[28:13] I do. I'm a patriot. It's not up to me.
[28:15] If we provide a technology,
it's not up to us to say,
[28:18] you can do this military operation
[28:20] and you can't do that military operation.
[28:22] Now, I might privately believe
[28:24] that this military operation makes sense
[28:27] and that military operation is a bad idea,
[28:29] but we're not gonna deny the technology.
[28:31] You, you basically have
to, you know, you have
[28:34] to leave policy in the hands
[28:36] of the military decision makers.
[28:38] Bloomberg has reported
that Claude is being used
[28:41] by the US military and the war in Iran
[28:43] to do AI assisted targeting
via a platform made
[28:47] by Palantir Maven Smart
system. In February,
[28:50] US missile reportedly hit a
girl school in Iran killing more
[28:54] than 150 people, most of them children.
[28:57] Did Claude play a role in that strike?
[29:00] We don't know exactly how, you know,
[29:03] these models were used, you
know, obviously like, you know,
[29:07] these things that,
that, you know, mistakes
[29:09] that happen in warfare are
really, really terrible.
[29:12] Like this is a really
terrible thing to happen.
[29:15] We were willing to risk
the future of our company
[29:17] to like limit how, you
know, these models are used
[29:21] and you know, what, what you're
talking about is a use case
[29:24] that doesn't even violate our red lines.
[29:26] We're worried that there will
be a hundred times as much,
[29:29] you know, with, with use cases that, that,
[29:31] that do violate our red lines.
[29:33] Now I, you know, I I I,
you know, again, again,
[29:35] I would say I think overall
the use of these, the, the use
[29:39] of these models is appropriate.
[29:41] I think it's good on net, you know,
[29:43] but military decision makers
make terrible mistakes even
[29:48] when, even, even at the best of times,
[29:49] and I, I don't know if
we're in the best of times.
[29:51] What we've seen here is Claude assists,
[29:54] but a human makes the final call.
[29:56] So a human made that
final call, not Claude.
[29:59] Imagine if you had a world
in which, not Claude,
[30:02] because we haven't allowed it,
but someone else's AI model,
[30:05] the AI model just makes the decision
[30:06] and the human never sees it.
[30:08] That's what we were standing up for.
[30:09] That's what we were fighting against.
[30:11] This school had a website.
[30:12] You could have found
it in a Google Search,
[30:14] like shouldn't Claude have spotted that?
[30:16] And it doesn't speak to a
scarier issue about using
[30:20] technology as a shortcut in war.
[30:22] The principle that was obeyed
here is a human makes the
[30:26] human makes the final decision.
[30:28] I don't know what role
Claude or any other AI had,
[30:31] but like, if, if this
isn't an illustration why
[30:34] that principle is so important,
I don't know what is,
[30:37] Is AI warfare more likely
[30:40] to stop World War III, a war
between the US and China?
[30:45] Or is it more likely to make it happen?
[30:48] I would say on balance it
is more likely to stop it.
[30:50] But if we have no limits on
how it's used, then I think,
[30:55] you know, it could be
more likely to cause it.
[30:56] You've seen Dr. StrangeLove, right?
[30:58] The premise of it was like, you know,
[31:00] you have a doomsday device
[31:01] that automatically fires
nuclear weapons when it thinks
[31:04] nuclear weapons are being fired at it.
[31:05] What could go wrong? Right?
[31:07] I think the way conflicts
happen is that, you know, the,
[31:10] the two sides jump at each other.
[31:11] They misunderstand each other.
[31:13] And when we don't have proper
oversight of this technology,
[31:16] I think those kinds of accidents
are more likely to happen.
[31:19] Now, I think if AI is used in
an appropriate way, in, in,
[31:22] in, not even warfare,
[31:24] but think of just, just
intelligence collection, you know,
[31:28] let, let's say we're able to,
you know, predict an invasion
[31:31] of Taiwan or a new movement in Ukraine.
[31:34] Like, you know, our adversaries
will think twice about,
[31:38] you know, about conducting
some kind of invasion
[31:41] or military operation if we know
[31:42] everything that they're doing.
[31:46] AI's role on the battlefield
raises difficult questions,
[31:49] even for a company willing to
debate the ethical dilemmas
[31:52] of its own technology.
[31:53] It's a conversation Amodei seems
eager to have in public.
[31:58] How do you handle the pressure?
[32:00] I try very hard to communicate
[32:02] and always be straightforward and honest.
[32:04] I get up in front of the
company every two weeks
[32:08] and just talk for an hour
about just what's on my mind,
[32:11] what's going on in the industry,
[32:12] what's going on in the outside world.
[32:14] It's totally uncensored.
[32:15] And not only does it build
trust, it's very freeing for me
[32:19] where I feel that I have 3,000 people
[32:22] who are on the same page as me.
[32:23] That is an incredible amplifier
[32:26] and is one of the most useful things in
[32:28] handing, handling the pressure.
[32:29] Then when we have to confront
an external challenge,
[32:33] stand up to, you know, some
very difficult situations
[32:37] that we've seen in the
last few months, that is
[32:39] what allows us to have a
consistent and coherent position.
[32:42] And, and I think, I think
it's an incredible advantage
[32:45] because I never feel like I'm alone.
[32:54] Lurking in the background
at Anthropic's headquarters was
[32:56] the surprise development of
a new AI model called Mythos.
[33:01] A model so powerful, it spooked everyone.
[33:04] Anthropic is making headlines
almost on a weekly basis.
[33:06] Yes,
Most notably now around mythos.
[33:08] The company believes
it's this enormous threat.
[33:11] Imagine a world where
everyone had a nuclear
[33:13] bazooka basically
[33:15] Mythos identified thousands
[33:16] of cybersecurity vulnerabilities,
[33:18] exposing potential flaws in
every major operating system.
[33:22] Anthropic signaled that if
fully released mythos could hack
[33:26] banks, pry open state secrets
[33:28] and cripple critical infrastructure.
[33:31] I think the thing that
surprised me most about it was the
[33:35] models had been climbing in their ability
[33:38] to find vulnerabilities.
[33:39] It was a particularly large jump.
[33:41] Some of the early
companies that we gave this
[33:43] to said things like, this is a
super weapon, you should have
[33:47] to own a gun license to use it.
[33:49] Please don't release this. In
[33:52] an initiative called Project Glasswing,
[33:54] Anthropic gave select
organizations access to Mythos.
[33:58] Even federal agencies
like the National Security
[34:00] Administration clamored
[34:02] to use it despite Anthropic's
blacklisting by the Pentagon.
[34:06] I think the future is this kind of cat
[34:08] and mouse game where we need to make sure
[34:11] that the good guys have the
tools that they need to defend,
[34:15] and then at some point the, you know,
[34:16] the bad guys will have it too.
[34:18] And, and at that point we have
[34:19] to make sure the good guys
have an even better model
[34:21] so they can be ready for this.
[34:23] Is it possible to stay ahead
[34:24] of the bad guys really though?
[34:26] That's what we hope
the criticism is.
[34:29] You're effectively deciding who
gets access and who doesn't.
[34:32] Why should anyone be
comfortable with that kind
[34:34] of concentration of power?
[34:36] It wasn't like, oh, it's so powerful
[34:38] and let's decide who gets the power.
[34:40] It was a very specific
concern around cybersecurity
[34:44] and so the way that we
decided who to give the model
[34:47] to was grounded in that specific fear.
[34:49] There's obviously nuance to decide, like
[34:51] where do you draw that circle?
[34:53] I think that's really complicated.
[34:54] I think we've tried to be
as publicly open as possible
[34:58] to say we're trying our best
to make this decision well,
[35:00] but like we might not do it perfectly.
[35:03] What about the folks who say
this was just good marketing.
[35:06] You know, we, we have
suffered enormously commercially
[35:10] from not releasing this model.
[35:11] This model has incredibly
accelerated research within
[35:15] Anthropic and production and next models.
[35:17] It would do the same in
the outside world if,
[35:20] if we were to release it.
[35:21] This has hurt us enormously commercially.
[35:24] Have you had to make trade-offs already
[35:25] that you're not entirely comfortable with?
[35:27] Throughout the entire history
[35:29] of Anthropic has been trade-offs, right?
[35:32] In some ideal world, you would prefer to,
[35:34] before you release the
first chat bot, you know,
[35:36] you could spend years
studying, you know, every,
[35:39] every possible thing that
could go wrong with it.
[35:41] Now, we did delay, we did
delay the initial release
[35:44] of Claude, but you know,
we did it for a few months.
[35:46] So everything is a trade off
now that we're in, you know,
[35:51] what I would describe as a
commercially leading position,
[35:54] we can afford to move
the dial even further
[35:56] towards being careful, right?
[35:57] That's what the Mythos
release was about, right?
[36:00] It's very hard to do something like
[36:01] that if you're not the leading player.
[36:04] There's this argument, why wouldn't the
[36:06] government take you over?
[36:07] Why would they let a
private company control
[36:10] technology? That's so powerful?
[36:12] So I think, I actually
think that's a very,
[36:15] that's a very serious question
and I share those concerns.
[36:18] I don't think the government
should outright take us over.
[36:22] Every previous powerful technology
we've seen in history was
[36:26] either built by the government
[36:28] or originated with the government.
[36:29] So nuclear weapons, obviously,
you know, initially built
[36:33] by the government, the
internet, GPS, cell phones,
[36:37] AI is the first technology
[36:38] that's been built in the private sector
[36:41] and where government has not
really had a serious role
[36:44] and is coming in late to the game.
[36:46] I think that's actually a
dangerous and unstable situation.
[36:49] It is not the situation I
would've chosen this technology.
[36:53] I'm scared of companies having it,
[36:54] but I'm also scared of
government having it.
[36:56] And then, you know,
[36:58] we need basic regulation
of the technology.
[37:00] You know, more
[37:01] and more as I've seen what we've
seen with Mythos, you know,
[37:04] I think we need to start
doing pre-release testing,
[37:07] required pre-release testing, testing
[37:09] and auditing of the models.
[37:11] And this was an approach
the White House rejected
[37:13] initially on his first day
back in office, President Trump
[37:17] with the help of former AI
[37:18] and crypto czar, David Sachs
dismantled President Biden's AI
[37:22] executive order seeking guardrails,
[37:24] instead favoring a hands-off,
[37:26] let Silicon Valley do its thing approach.
[37:29] We believe that excessive regulation
[37:31] of the AI sector could kill a
transformative industry just
[37:36] as it's taking off.
[37:38] But with Mythos and its
national security implications,
[37:40] proving hard to ignore the
White House now seems to want
[37:44] to gatekeep the world's most powerful ai
[37:47] It's very funny to me how
there's a particular group
[37:51] of people in the tech world in
Silicon Valley, they started
[37:54] with a position of like, you
know, even having transparency
[37:57] around this technology, even
export control, you know,
[38:00] this is all, you know, just totally,
[38:02] it'll apocalyptically
destroy our potential
[38:05] to create the technology.
[38:06] It'll kill innovation.
[38:08] And then as soon as they
see the first real danger,
[38:10] which I've been expecting all
along, there's all this talk
[38:13] of like nationalization
[38:14] and the government should just seize it.
[38:15] Come on folks here, you're yo,
[38:17] you're yo-yoing from like the
most extreme anti-regulatory.
[38:22] If you look at us the wrong way,
[38:23] you're destroying the
industry to, you know,
[38:25] this completely communist,
[38:26] the government should grab it all.
[38:28] We, we need, we, we need a more,
[38:30] we need a more sensible,
moderate approach.
[38:33] That's the one we've
been favoring all along
[38:35] because we've, we've understood the
[38:36] power of this technology.
[38:38] We're not panicking, we're not denying it.
[38:41] We see the smooth exponential
[38:42] and we're responding to it appropriately.
[38:51] The reaction to AI
right now, it's intense.
[38:56] Anthropic has built this
really loyal following.
[38:59] There are some people who
just love what they stand for,
[39:01] but there've also been protests
right outside their office.
[39:05] There's a lot of anxiety,
a lot of confusion,
[39:07] there's some real anger right
now about what's happening
[39:12] and it actually feels
like it's escalating.
[39:15] Artificial intelligence is
[39:18] the next industrial revolution. Whoa.
[39:25] If you look at the data,
[39:27] people are more concerned than excited
[39:29] about what's going on.
[39:31] They think the risks outweigh the benefits
[39:33] and the truth is, if
you talk to the people
[39:35] who are building this,
[39:37] even they will tell you
they don't fully know
[39:40] how it's all going to play out.
[39:45] How do you think about
the weight of this moment?
[39:47] I worry that something will go wrong.
[39:50] You know, are we doing
literally everything we can?
[39:52] We're certainly trying our best,
[39:54] we're certainly trying very hard.
[39:55] What I want is to create a
situation where if this set
[39:58] of people can't do it,
it couldn't be done.
[40:00] You can't guarantee success,
[40:01] but maybe you can guarantee that
[40:04] It's getting personal.
[40:05] Sam Altman's home has been attacked.
[40:08] How is that affecting you?
[40:10] It was really scary to read that.
[40:11] I mean, it was, we were obviously
extremely relieved that he
[40:15] and his family were okay in general.
[40:17] I think this is a time,
you know, technologically,
[40:20] politically where unfortunately
there's just a lot more
[40:23] rhetoric and, and words
that I think can lead to,
[40:26] to bad outcomes and bad things happening.
[40:29] I hope that this is a topic
we can all just debate,
[40:31] you know, as peacefully as possible.
[40:34] It was scary to me too.
[40:35] I mean, you know, this is
like a, a less savory aspect
[40:38] of the exponential, right?
[40:39] That as AI gets more
[40:41] and more of, of a big
thing like, you know,
[40:43] it becomes just such
a big deal to society.
[40:46] There's more attention on it.
[40:48] There's been massive
backlash against social media.
[40:51] Countries are starting to ban
it. Could that happen to AI?
[40:55] I think it's, it's absolutely possible.
[40:58] If the social media companies
could go back in time
[41:00] and see the world that they see today,
[41:02] would they do anything differently?
[41:04] I like to think the answer to that is yes.
[41:06] I don't know if we sort of
project some of the challenges
[41:09] that the social media companies have faced
[41:11] around child welfare, mental
health, election integrity, all
[41:15] of these topics we're really
lucky that we're we're second.
[41:19] We view it as our job to try
[41:20] and proactively think about all
[41:22] of the things that could go wrong.
[41:24] Because if we don't, who's going to,
[41:25] You know, I don't know
that they actually set out
[41:28] to do the right thing or make
the world a better place.
[41:30] And so I don't think if
they were going back,
[41:33] they would even knowing what they do,
[41:34] and they certainly should
do things differently.
[41:36] I don't know if they
actually will, but we can't.
[41:39] This is why we're trying
[41:40] to get this right the first
time instead of waiting
[41:43] for things to go wrong, then scrambling
[41:46] to justify why it's all okay.
[41:48] The main way I could see a
AI being, you know, banned
[41:52] or blocked is if something
really went wrong
[41:54] and if something really went wrong,
[41:55] then maybe it deserves to be,
[41:56] You have technologists
saying it's gonna be amazing
[42:00] and others saying it could be awful.
[42:02] Where are you on that spectrum?
[42:04] I hope for the best,
but plan for the worst.
[42:07] For me, this is the most
important work I've ever done.
[42:10] For some people it feels
like this is the last job
[42:13] because this is sort of the,
it could mean the end of work
[42:16] and I think for other
people it's, you know,
[42:18] like maybe it's not the last job,
[42:20] but it's existential to get this right.
[42:23] And so there's just this,
this kind of burden.
[42:25] If the impacts could be as significant
[42:28] as we're talking about, as significant
[42:30] as Anthropic has warned about,
[42:32] what responsibility do
you think Anthropic has
[42:35] to cushion the blow?
[42:37] What do you owe the people
whose lives you've upended?
[42:43] I think our view has always been,
[42:45] we ultimately are responsible
as an industry, the industry
[42:48] that's developing this technology
for thinking through what
[42:52] are the risks, what are the
bad things that could happen?
[42:55] And if some of those things come to pass,
[42:57] what is our role in helping to fix them?
[42:59] That is our job, right?
[43:01] We should not just say,
well, we were just trying
[43:03] to grow the product and
suddenly there's an entire
[43:05] generation of young women
who have eating disorders
[43:09] or who have mental health problems
[43:11] because whoops, we were just
trying to grow the product.
[43:13] That's not the stance
[43:14] that I think any technology
company should take.
[43:16] I don't think that's the stance
that we're trying to take
[43:19] For a company whose
identity is so wrapped up in,
[43:22] We want to do this right,
[43:23] We exist as a AI safety company.
[43:25] How can we just help
all of this go? Well,
[43:27] It can be hard to understand
why Anthropic is pushing
[43:30] so hard to advance AI while being
[43:32] so upfront about the
dangers. In his essays,
[43:36] Dario lays out what the end
game looks like if everything
[43:39] goes right with ai, a
utopian future where machines
[43:42] and humans work side by side AI,
[43:45] an inevitable force steered
toward prosperity rather than
[43:48] catastrophe to mitigate the
devastation of job loss.
[43:53] He proposes solutions like
universal basic income
[43:56] and progressive taxation of AI companies.
[44:00] But as the Amodeis confront the
messy realities of power,
[44:03] politics, and profit,
the real test is whether
[44:06] that founding mission can survive the
[44:08] scale of what they're building.
[44:11] Google started with the
motto, don't be evil.
[44:14] A founding promise the company
quietly retired as it grew.
[44:19] You are building something
incredibly powerful
[44:22] and stand to gain enormously from it.
[44:25] Why should we trust you?
[44:26] I think starting from a
position of distrust, you know,
[44:29] if you don't know anything about,
[44:31] if you don't know anything about
[44:32] philanthropic, it's pretty rational.
[44:33] I think Silicon Valley has
lost a lot of the world's trust
[44:36] and kind of has to re-earn it,
[44:38] and the message, you know,
we're trying to send is,
[44:41] we're actually different and,
[44:42] and that has to be earned in
things that we actually do.
[44:46] I understand one of your
favorite books is The Making
[44:48] of the Atomic Bomb.
[44:49] That is correct.
Do you see parallels
[44:51] between yourself and Oppenheimer?
[44:52] The figure I most identified
with was Leo Szilard
[44:56] who was the one who first had the idea
[44:58] that there could be a chain reaction.
[45:01] Look, my view is we're, we're
not gonna get through this
[45:04] with like larger than life personalities
[45:06] or like figures who try and be
at the center of everything.
[45:09] In some ways, I actually see
Oppenheimer as a failure case, as
[45:13] what should not happen.
[45:14] There's a lot of powerful
actors who have interests here,
[45:17] and the only way it's gonna end well
[45:19] for everyone is if there is
some, there's basically checks
[45:23] and balances everywhere.
[45:24] You've said there's
roughly a 10 to 25% chance
[45:28] of civilization collapse,
that is not insignificant.
[45:33] Is there a scenario where it's something
[45:35] that Anthropic built, that caused that?
[45:38] I mean, I certainly hope not.
[45:40] My view is that that
probability comes from the very
[45:44] straightforward recipe of
the technology, the existence
[45:47] of many countries in the world.
[45:49] The existence of many
companies within an economy
[45:51] and new ones created.
[45:53] If the void is isn't filled,
like that's a dilemma
[45:56] that we're in half of what we
do within the company is try
[45:59] and, you know, reduce the
risk as much as we can,
[46:01] but it's never gonna be zero.
[46:03] Suppose there are a bunch of, you know,
[46:05] airline companies out there
[46:06] and you're like, well, I'm gonna make an
[46:07] airline company that's safer.
[46:08] It can both be the case that, you know,
[46:11] your airline company is 10
times safer than all the other
[46:13] airline companies.
[46:15] But if someone comes
[46:16] and asks you, like, can you guarantee
[46:18] that your airplane will never crash?
[46:20] I mean, how could you,
how could you possibly,
[46:22] But if there was a 25%
chance of an airplane crashing,
[46:25] you wouldn't get on that plane.
[46:26] That's right. 25% is too high.
[46:28] We're trying to make that
probability much, much lower.
[46:31] That's, that is, that is the goal.
[47:01] How do you find your
zen? How do you relax?
[47:04] You know, it honestly, a lot
[47:06] of it's just exposure to, to it.
[47:08] You know, I sometimes
I'll just like, you know,
[47:10] I'll just take a weekend
[47:11] and I'll like play some video games.
[47:13] Sometimes with Daniela
Me and my wife go to Italy.
[47:15] Sometimes we have a horse
there, so I'll just,
[47:19] I'll just sit there next.
[47:20] I'll just, I'll just look at the horse
[47:21] and I'll be like, you know,
Calypso, who's her horse like,
[47:25] you know, she doesn't
know about any of this,
[47:27] like she's just a happy horse.
