Anthropic Revenue Growth Triples as Amodei Rejects Military AI Red Lines Standoff
Bloomberg Originals
Anthropic CEO Dario Amodei warns that unbounded AI warfare risks catastrophic escalation while defending the company's decisions to hold back powerful cyber tools and resist unconstrained defense deployment.
## Background
This extended interview from Bloomberg Originals' series "The Circuit," hosted by Emily Chang, was published on YouTube on June 17, 2026. The video features an in-depth conversation with Anthropic CEO Dario Amodei, exploring his San Francisco roots, the competitive race with OpenAI, Anthropic's enterprise focus, military contract standoffs, and the ultimate societal endgame of AI.
## TL;DR
Anthropic CEO Dario Amodei breaks down his company's trajectory, emphasizing an enterprise-first model and rigorous AI safety over short-term commercial gains. He addresses key industry inflection points, including Anthropic's rapid quarterly growth, the shifting nature of software moats, and the deployment of powerful tools like Mythos and Claude. Throughout the interview, Amodei navigates complex trade-offs surrounding workforce automation, export controls against China, geopolitical military boundaries, and the necessity of maintaining human oversight as AI capabilities follow a smooth exponential curve.
## Atmosphere & style
The interview carries a measured, intellectual, and candid tone, balancing high-stakes technical debate with calm, pragmatic reflection. Emily Chang's probing style draws out Amodei's disciplined, analytical mindset—reflecting his preference for surgeon-like rationality over Silicon Valley hyperbole. Despite grappling with heavy topics like existential risk, economic disruption, and national defense, the dynamic remains groundedly human, shifting smoothly between rigorous policy arguments and moments of dry self-awareness.
## Key takeaways
- [00:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=7s) Amodei compares the accelerating pace of the AI race to traveling on a spaceship at relativistic speeds, emphasizing the importance of calm, rational risk management over panic.
- [05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) Disagreements over AI safety were not the primary cause of Amodei leaving OpenAI; rather, it was a fundamental breakdown in trust and honesty regarding values and motives.
- [10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) Anthropic focuses heavily on enterprise applications and coding tools rather than consumer social platforms to align business incentives with positive real-world outcomes.
- [15:11](https://www.youtube.com/watch?v=x2VHFgyawPE&t=911s) Traditional software moats built solely on writing complex code are rapidly disappearing; future moats will rely on deep domain expertise and direct customer relationships.
- [17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) Amodei strongly advocates for U.S. chip export controls to China to safeguard national security and democratic interests, even when it conflicts with hardware partners' short-term interests.
- [18:28](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1108s) Anthropic recorded over 3x revenue growth in Q1 2026 alone (an 80x annualized pace), far exceeding their planned 10x compute capacity growth and driving massive capital raises.
- [27:57](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1677s) Amodei warns that AI will cause major white-collar job displacement within 1 to 5 years, advocating for active policy interventions like a 'token tax' to smooth the economic transition.
- [36:41](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2201s) Anthropic established strict red lines against using its models for mass domestic surveillance or fully autonomous weapons, resisting military pressure despite commercial and political fallout.
- [46:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2809s) Amodei believes superior AI-driven intelligence gathering can serve as a powerful geopolitical deterrent, reducing the likelihood of global conflict between major powers.
- [48:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2898s) Anthropic temporarily held back its Mythos cybersecurity model after early partners called it a 'superweapon,' prioritizing defensive patching before public availability.
- [54:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3255s) To prevent unchecked concentration of corporate power, Anthropic created the independent Long-Term Benefit Trust, which holds ultimate power over board composition and the CEO.
- [01:03:24](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3804s) Amodei describes recursive self-improvement as an ongoing, smooth exponential process rather than a sudden single tipping point.
- [01:05:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3949s) Addressing an estimated 10% to 25% baseline risk of civilizational collapse, Anthropic dedicates approximately 50% of its internal resources toward risk reduction and model safety.
- [01:07:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4052s) Amodei argues that public trust must be earned through costly actions, highlighting Anthropic's choice to cut off Chinese model access at a cost of hundreds of millions in revenue.
## Detailed walkthrough
Here is the detailed walkthrough of the transcript section provided:
[00:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=7s) Relativity, Calmness, and Exponential Growth
The interview opens with Emily Chang asking Dario Amodei how much sleep he gets under the extreme pace of the AI race. Amodei admits he has never slept particularly well, but is learning to relax and sleep through periods of "unusual pressure." He compares the rapid pace of the current AI expansion to traveling on a spaceship at relativistic speeds: every time you wake up, time on Earth has accelerated exponentially, forcing you to pack several days of relative progress into a single day. Asked if he wakes up paranoid, Amodei stresses that he actively avoids panic or yo-yoing between complacency and terror, calling such volatility a "hallmark of immature decision making." Instead, he advocates for the calm, rational risk management of a surgeon or military officer.
Transitioning to Anthropic's trajectory, Chang shares an anecdote about her son asking to use her "Claude Cowork" account, illustrating how quickly consumer usage is spilling over despite Anthropic's enterprise focus. Amodei reflects on being at the center of the AI universe, framing it as the natural outcome of a "smooth exponential" curve. While he previously charted a graph predicting Anthropic would become the top revenue and valuation AI company around this timeframe, seeing the predictions materialize in full color remains surprising. Through it all, he maintains that the core mission—training top models, building good products, ensuring safety, and managing societal risks—remains unchanged, just under a much larger microscope.
[03:34](https://www.youtube.com/watch?v=x2VHFgyawPE&t=214s) San Francisco Roots and Nonconformism
Amodei traces his background back to growing up in San Francisco, raised by a leather craftsman father and a librarian mother. Despite the initial dot-com and internet revolutions happening directly around him, he had zero interest in tech at the time; instead, he was obsessed with mathematics, sci-fi, and understanding the universe. Reflecting on the spirit of San Francisco, he credits the city's culture of individualism and nonconformist "crazy ideas" with shaping his willingness to ignore conventional expert consensus and pursue long-shot, high-upside goals.
Chang brings up 2016, when Dario, his sister Daniela Amodei, and her husband Holden Karnofsky lived together in a group house. At the time, Karnofsky was launching the Open Philanthropy Project while Amodei was working as a biological scientist. Amodei recalls advising Open Philanthropy on promising avenues for developing-world health and biological research before shifting his focus fully to artificial intelligence.
[05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) The OpenAI Departure and the Philosophy of Trust
Addressing the famous Silicon Valley narrative surrounding his departure from OpenAI, Amodei breaks down the core reasons behind the split. He clarifies that disagreements on AI safety alone were not the trigger—disagreements are normal and happen even within Anthropic. Rather, the break occurred due to a fundamental breakdown in trust and honesty regarding values and motives. Seeing "disturbing patterns of behavior" made continued collaboration impossible. Rather than engaging in perpetual conflict with leadership whose vision he did not trust, he chose to walk away and build a company according to his own principles.
Amodei notes that he is entirely at peace letting market outcomes and the court of public opinion judge Anthropic’s approach against OpenAI’s. Asked about an awkward moment at an AI summit in India where he and Sam Altman appeared to refuse to hold hands on stage with Prime Minister Narendra Modi, Amodei explains that the event was chaotic and disorganized, with last-minute stage re-arrangements. Addressing the broader concern of whether hostile AI leaders can cooperate on existential risks, Amodei refutes the idea that no one trusts each other. He highlights his 15-year relationship with Google DeepMind’s Demis Hassabis, noting that their mutual trust creates a "race to the top." He argues that when trustworthy actors unite and set high ethical standards, untrustworthy actors are forced by market and social incentives to fall in line.
* Amodei describes his relationship with DeepMind's Demis Hassabis as collaborative, sharing safety research and buying compute from Google while mutually inspiring each other's work (e.g., DeepMind's AlphaFold and Anthropic's interpretability research).
[10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) The Enterprise Bet and the "SaaSpocalypse"
Amodei explains why Anthropic chose to focus heavily on enterprise and coding tools (such as Claude Code and Claude Cowork) over flashy, consumer-facing social media apps. He warns that consumer business models, especially those reliant on advertising, inherently push companies toward maximizing screen time, addiction, and low-quality content ("slop"). By contrast, enterprise AI aligns seamlessly with positive real-world impacts—such as partnering with pharma companies to cure diseases, optimizing energy efficiency, advancing education, and driving economic growth. Enterprise relationships rely on long-term trust, which mirrors Anthropic's safety-first values.
When asked if competitors like OpenAI or Google could easily replicate Anthropic’s edge given how easily developers can swap APIs, Amodei asserts that raw model quality remains the single most critical factor, and Anthropic remains well ahead. Addressing the market reaction to Claude Cowork—where $285 billion in traditional software market capitalization evaporated in a sell-off dubbed the "SaaSpocalypse"—Amodei acknowledges that AI acceleration will inevitably erode some traditional software moats, though predicting the exact pace and financial impact remains complex.
* "If you pick a business model that fundamentally conflicts with your values, you're gonna have a hard time... Either you betray your own values or you become irrelevant."
* "We want to use AI to, you know, cure diseases that we couldn't cure before... working with biotech, it's working with pharma... All of those are enterprises."
[15:11](https://www.youtube.com/watch?v=x2VHFgyawPE&t=911s) The Changing Landscape of Software Moats
Addressing the software industry's future, Dario Amodei argues that traditional software moats are rapidly evaporating as artificial intelligence strips away the advantage of writing complex code. If a enterprise's sole defense is having written code that no one else can replicate, "good luck, you're not gonna be able to defend that." Instead, durable moats will come from deep domain expertise, industry know-how, and established customer relationships. He cautions incumbent SaaS companies against complacency, advising them to rigorously inventory their advantages, drop outdated assumptions, and adapt to newly emerging moats.
Despite potential casualties among unadaptable incumbents, Amodei projects that the overall software pie will expand rather than shrink. A rising tide in AI capabilities will amplify what is possible tenfold, enabling the industry as a whole to grow significantly even if individual legacy players lose market share or fold.
* **Software moat shift:** Writing complex code is no longer a defensible moat; deep domain knowledge and direct customer relationships take priority.
* **Macro outlook:** AI is set to expand the entire software market overall, even as individual non-adapting incumbents face steep decline.
---
[17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) Balancing Backers, Big Tech, and China Policy
Navigating relationships with massive corporate backers like Amazon, Google, Microsoft, and Nvidia presents a complex web of shared interests and ideological friction. Amodei candidly addresses this dynamic, noting that while investors hold their own agendas and commercial milestones, Anthropic remains firm on its principles. He points to his vocal advocacy for strict U.S. export controls on advanced AI chips to China—a position directly opposing the immediate financial interests of major chip manufacturers.
Amodei emphasizes that mature commercial relationships allow space for disagreement on regulatory policy while continuing to collaborate on shared enterprise goals. "Look, we're all adults here," he remarks, reiterating that China overtaking the United States in frontier AI capabilities poses severe risks to global democracy and American national security.
* **Chip export controls:** Amodei actively campaigns for U.S. restrictions on chip exports to China to protect democratic interests, despite pushback from allied hardware partners.
* **Partner dynamics:** Anthropic maintains operational independence and principle-driven advocacy alongside giant tech co-investors.
---
[18:28](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1108s) Valuation, Capital Buffers, and the Compute Crunch
When asked to contextualize reports of Anthropic achieving valuations approaching $1 trillion—unprecedented for a five-year-old startup—Amodei explains that massive capital raises act as a vital buffer against extreme operational uncertainty. As compute infrastructure scales rapidly year-over-year, securing heavy financial reserves ensures the company can fund its aggressive hardware roadmap without suffering meaningful equity dilution or risking business fundamentals.
Addressing rumors of server strain and token limits, Amodei dismisses the notion that Anthropic bought too little compute or is playing catch-up. While Anthropic conservatively budgeted for a 10x annual increase in compute demand, actual business performance in the first quarter of 2026 skyrocketed. Revenue grew over 3x within that single quarter alone—a rate that translates to an annualized 80x surge.
* **Q1 2026 growth spike:** Anthropic recorded over 3x quarterly revenue growth in early 2026, far outstripping the planned 10x annual compute capacity.
* **Capital strategy:** Immense fundraising serves as a buffer to navigate hyper-growth and extreme demand volatility.
* **Infrastructure outlook:** Amodei views compute markets as liquid; temporary shortages resolve within a month or two through expanded deals with partners like Google and Amazon.
---
[21:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1275s) Outpacing Rivals and Maintaining Internal Culture
Reflecting on Anthropic's trajectory relative to rivals like OpenAI, Amodei frames leadership in the field not as a vanity contest, but as leverage to shape the broader ecosystem. Being the preeminent AI firm grants Anthropic the agency to "pull the ecosystem along," forcing competitors to emulate higher safety standards, alignment practices, and technical architectures—even when rivals copy them while launching public attacks.
However, maintaining core corporate values during hyper-growth remains Amodei’s central preoccupation. Growing rapidly and onboarding talent from legacy Big Tech brings the risk that new hires default to old corporate norms. Amodei spends nearly half his time focusing purely on internal culture and organizational unity alongside co-founder Daniela Amodei.
* **Ecosystem pull:** Dominating frontier performance enables Anthropic to set high safety and ethical benchmarks that compel the broader industry to follow.
* **Cultural retention:** Amodei allocates roughly 50% of his time to cultural stewardship to prevent new hires from reproducing legacy Big Tech habits.
---
[24:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1447s) Rapid Product Velocity and Scientific Discoveries
Amodei attributes Anthropic's rapid shipping pace to two main drivers: tight organizational alignment and aggressive internal integration of Claude itself. Claude is now routinely deployed internally to optimize model architectures, write code, and accelerate product development cycles, creating a compounding feedback loop of speed and reliability.
Beyond enterprise software, Amodei—a former biologist—highlights AI's transformative breakthroughs in medicine, computational chemistry, and drug design. He shares an anecdote where Claude correctly diagnosed a complex medical issue that high-profile human doctors had missed, including a personal case involving his sister and co-founder, Daniela. Amodei envisions AI compressing a century's worth of scientific and medical advancements into a short span, fundamentally solving historical human deprivations.
> "Think of all the problems we had in 1900s... all the material deprivation that we don't have to deal with today. Then think of another hundred years of that. I really believe this century of scientific and medical progress... we're gonna have a much, much better world."
* **Internal AI deployment:** Claude directly assists Anthropic's engineers in developing new models and writing internal code bases.
* **Medical diagnostic breakthrough:** Claude accurately diagnosed a complex medical condition that multiple specialist physicians failed to identify.
---
[26:13](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1573s) AI's Role in Human Writing and Thought
Known for his long-form essays, Amodei explains how he personally incorporates AI into his writing workflow. While he refuses to publish text generated directly by Claude due to his distinct personal style, he relies heavily on the model as a thought partner for brainstorming, structuring arguments, and pulling obscure historical references.
Addressing concerns that outsourcing writing to AI could erode human critical thinking, Amodei acknowledges the risk. For him, writing is primarily an exercise to clarify his own internal logic and create shared reference points. He advocates for nuanced usage models where AI assists research and organization without replacing the core struggle of original thought.
* **Writing workflow:** Amodei uses Claude strictly for brainstorming, structuring, and research, preserving his direct authorship over published text.
* **Cognitive retention:** Warning against full end-to-end AI text generation, Amodei stresses that human writing remains critical for deep reasoning and conceptual clarity.
---
[27:57](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1677s) Workforce Disruption, Automation, and Economic Policy
Revisiting his past prediction that AI could displace 50% of entry-level white-collar jobs within one to five years, Amodei stands by the order-of-magnitude severity of his concern. He clarifies that his statements are often taken out of context as pure "doom," whereas he actively advocates for proactive policy interventions—such as a "token tax," macroeconomic rebalancing, and enterprise job restructuring.
Amodei observes that while AI initially boosts worker productivity, as automation approaches 100% of tasks, displacement becomes inevitable. Within Anthropic itself, AI now writes almost all code, altering traditional software engineering roles. However, demand is simultaneously shifting toward hybrid roles, such as "forward-deployed engineers" or "applied AI solutions architects," who combine technical literacy with direct customer engagement.
* **Workforce impact:** Amodei reiterates high concern over severe white-collar job disruption, pushing for active policy tools like a token tax to cushion the economic transition.
* **Role evolution:** Internal coding at Anthropic is increasingly automated, driving demand toward hybrid technical-consulting roles like applied solutions architects.
[30:40](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1840s) AI Workforce Disruption and Corporate Responsibility
Dario Amodei addresses the profound disruption AI will cause in the labor market, emphasizing that while the economy will adjust, the magnitude of the shift requires proactive policymaking. Anthropic aims to put out carefully considered thoughts and concrete proposals rather than half-baked ideas. Within entry-level white-collar fields like finance and banking, AI will initially boost employee productivity, but could eventually perform entire jobs autonomously. When consulting with enterprise customers, Anthropic actively pushes companies toward a positive-sum approach: rather than using AI merely to cut costs and lay off workers, they urge companies to accomplish more with the same headcount as the economic pie expands.
When asked to envision the macroeconomic landscape five years down the line, Amodei warns against mass unemployment that could provoke widespread social unrest and revolution. While robotics progresses much slower than software AI, physical manufacturing, infrastructure, data center construction, and human-centered, relationship-driven roles (such as healthcare and personal services) will likely absorb displaced workers. Humans will also retain a crucial role in directing AI models in alignment with human values, though the exact depth of that supervisory role remains uncertain.
- Anthropic uses tools like its Economic Index and economic research grants to analyze labor market dynamics and formulate tax/policy solutions.
- Amodei identifies two distinct corporate choices for enterprise AI adoption: reducing headcount to do the same amount of work, or keeping headcount stable to dramatically expand output.
- Non-automated sectors like physical labor, construction, manufacturing, and direct human-to-human interactions will become critical labor sinks as digital information processing becomes trivial.
---
[34:38](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2078s) Pushing Back Against Silicon Valley "Doom Marketing" Accusations
Amodei strongly rejects criticisms—including comments from NVIDIA CEO Jensen Huang—claiming that he conflates tasks with jobs or uses "doom marketing" to promote Anthropic. He notes that in his essay *The Adolescence of Technology*, he spends five pages explicitly detailing the technical distinction between tasks and jobs, why current AI developments differ from historical technological shifts, and six policy solutions spanning private philanthropy to government intervention.
He passionately criticizes Silicon Valley's media ecosystem, expressing a deep disdain for social media platforms that amplify three-second out-of-context video clips over serious intellectual work. He calls these soundbite-driven attacks "cheap marketing" and a form of laziness that prevents a mature, serious national conversation about AI's societal impact.
> "The idea that this is cheap marketing is itself cheap marketing. This is laziness; this is failure to engage with serious intellectual work... It's so stupid. It's so unserious. And whenever someone says something like that, I take them less seriously."
---
[36:41](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2201s) The Pentagon Standoff and Anthropic's Military Red Lines
Addressing his trajectory from a student with anti-war stances at Caltech to heading one of the first AI companies to sign a US Department of Defense contract for classified networks, Amodei explains that changing geopolitical realities informed his perspective. The threat posed by an aggressive, autocratic bloc—specifically Russian aggression in Ukraine and threats against Taiwan by China—convinced him that democratic nations must maintain technological defense capabilities. Anthropic engaged in government work not for profitability, which he describes as a high-friction process for relatively little money, but out of patriotic duty.
However, Anthropic established strict red lines regarding military usage, refusing to allow its models to be used for mass surveillance or fully autonomous weapons. When the company refused to remove these safeguards, it faced intense pushback: reports emerged of the administration banning the firm from federal contracts, the Pentagon labeling it a supply chain risk, and competitor OpenAI stepping in to accept terms Anthropic rejected. Despite being called an "ideological lunatic" or "left-wing nut job," Amodei maintains that sticking to principles is far more important than short-term government procurement wins.
- Anthropic began working with Palantir in 2024 to deploy models on defense networks.
- Amodei explicitly denies that Anthropic or its models work with Immigration and Customs Enforcement (ICE), Customs and Border Protection (CBP), or military operations in Gaza.
- Red lines defined by Anthropic: absolute prohibition on mass domestic surveillance and human-out-of-the-loop autonomous lethality.
---
[41:47](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2507s) Military Effectiveness, AI Targeting, and the Ethics of Warfare
Responding to statements by US officials that LLM integration has allowed military operations to scale targeting from 1,000 to 5,000 targets per day, Amodei affirms his support for American military strength, arguing that robust defense capabilities act as a deterrent to global war rather than an instigator. He clarifies that while Anthropic provides the underlying technology, software vendors cannot dictate individual tactical operations to military commanders; macro-level policy and rules of engagement must remain with human decision-makers.
The interview turns to reports of AI-assisted targeting systems (such as the Maven Smart System powered by Palantir) being utilized in military campaigns, including a tragic strike in Iran that hit a girls' school. Amodei emphasizes that Anthropic does not have visibility into exact tactical deployments on classified networks, but notes that in human-in-the-loop systems, human operators ultimately make the final decision to fire. He warns that the potential tragedy of human error in warfare underscores why Anthropic fought so hard against fully autonomous weapon systems—where AI models make lethal decisions without any human oversight or sanity checks.
- A software provider cannot dictate individual operational targets, just as Microsoft cannot restrict how the military uses Excel for specific operations.
- The distinction between AI-assisted decision support (human-in-the-loop) and fully autonomous kinetic strikes represents Anthropic's primary ethical boundary.
- Broader legislative efforts in Congress seek to establish formal statutory guardrails on military AI applications based on the debate sparked by Anthropic's standoff.
[46:29](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2789s) AI Warfare and Preventing Global Conflict
Dario Amodei addresses the role of AI in defense and international security, maintaining that human control must remain absolute. While acknowledging he lacks access to classified intelligence, Amodei emphasizes the foundational principle Anthropic operates on: a human must always make the final decision in military actions. When asked whether AI warfare is more likely to prevent or trigger World War III between the United States and China, he believes that on balance, AI is more likely to stop it—provided strict limits are enforced. Unbounded usage, however, creates dangerous dynamics akin to *Dr. Strangelove*, where automated "doomsday devices" respond without human nuance. Conflict often escalates because opposing sides misunderstand each other's actions, and a lack of proper oversight dramatically increases the chance of catastrophic accidents.
Conversely, Amodei contends that superior AI-driven intelligence collection can act as a powerful deterrent. If predictive AI allows the U.S. and its allies to foresee major aggressive movements—such as an invasion of Taiwan or tactical operations in Ukraine—adversaries will hesitate knowing their actions are fully exposed. By providing accurate foresight and enhancing response readiness, AI technology can stabilize international relations and actively deter large-scale wars.
- **Core Military Principle:** A human must retain the final decision-making power in any lethal or military operation.
- **Deterrence Mechanism:** Advanced intelligence gathering exposes adversary movements early, deterring invasions in regions like Taiwan or Ukraine.
- **Analogy:** Compares uncontrolled, fully autonomous military systems to the automated nuclear launch premise in *Dr. Strangelove*.
---
[48:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2898s) The Mythos Controversy and Cyber Safeguards
Amodei addresses the intense industry reaction surrounding Mythos, Anthropic’s flagship cybersecurity-capable AI model. Mythos possesses the ability to autonomously navigate all stages of the "cyber kill chain," rapidly uncovering software vulnerabilities and turning them into working exploits. Amodei reveals that early partner companies granted preview access were shocked by its power, calling it a "superweapon" and begging Anthropic not to release it publicly without strict controls. Rather than keeping the model permanently locked away, Anthropic aims to roll it out gradually to trusted defenders while developing robust cyber safeguards.
Responding to critics who labeled the decision to hold back Mythos as a PR stunt or claimed its performance could be replicated using cheap open-source models, Amodei strongly rejects those arguments as "incredibly false." Pointing out a flaw in the critique, he notes that while an open-source model might analyze a specific line of code highlighted for it, Mythos scans entire complex codebases autonomously to identify hidden flaws from scratch—for example, discovering 271 previously unknown vulnerabilities in Firefox as well as thousands in private corporate repositories. Holding back the model has caused Anthropic immense commercial damage, but Amodei asserts that defenders need time to patch software systems before powerful exploit tools become widespread.
- **Mythos Capabilities:** Capable of executing autonomous actions across the full cyber kill chain, moving seamlessly from finding bugs to generating exploits.
- **Concrete Track Record:** Successfully identified 271 brand-new vulnerabilities in Firefox, alongside thousands in private corporate repositories.
- **Safeguard Challenges:** Traditional classification guards (used on models like Claude 3.5 Sonnet or Opus 4.7) remain vulnerable to jailbreaking, necessitating stronger defensive measures before full release.
- **Government Coordination:** U.S. government officials and security teams have requested a slower rollout pace to mitigate counterintelligence risks.
---
[54:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3255s) Governance, Nationalization, and Checks on Power
Discussing the broader societal implications of frontier AI, Amodei reflects on the ongoing trade-offs between speed and safety throughout Anthropic's history. He addresses questions regarding whether the government might eventually nationalize AI labs, calling it a serious concern. Historically, transformative technologies like nuclear energy, the internet, and GPS originated within government or military programs. AI is unique because it was developed entirely within the private sector, leaving governments scrambling to catch up—a dynamic Amodei describes as inherently unstable.
To prevent unchecked concentration of power in private hands, Anthropic established the Long-Term Benefit Trust, an independent corporate governance body with the authority to appoint and remove a majority of the company's board members (and ultimately fire the CEO). Amodei advocates for a balanced approach: legislative and judicial checks must regulate private AI firms through required pre-release auditing and red lines, while private innovation prevents state monopolization. He criticizes commentators in Silicon Valley who alternate wildly between extreme anti-regulatory stances and sudden calls for complete state nationalization, calling for a pragmatic middle path.
> "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."
- **Long-Term Benefit Trust:** An independent panel designed to oversee board composition and balance commercial motives with public safety.
- **Policy Recommendations:** Mandatory pre-release testing, independent auditing of frontier models, and enforced red lines established by Congress.
- **White House Engagement:** Amodei reports constructive discussions regarding Mythos and AI risk with top administration figures, including Treasury Secretary Scott Bessent and Chief of Staff Susie Wiles.
---
[58:57](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3537s) China, Open Source, and the Intelligence Frontier
Reflecting on his past experience at Baidu’s Silicon Valley research outpost, Amodei offers insight into the global competitive landscape and Chinese AI developments. Addressing concerns about strong open-source models emerging from China, he explains that market adoption overwhelmingly favors the top of the capability frontier. While lower-tier open-source models are useful for simpler tasks, the value created by frontier models grows exponentially—expanding by roughly $10\times$ each year—rendering previous product heuristics obsolete.
While Amodei does not view current open-source Chinese models as an immediate commercial threat to frontier labs, he voices concern about the safety risks that emerge when powerful capabilities proliferate widely over time. Capabilities on par with Mythos will likely become accessible via open-source channels within 12 months, making preemptive patching of global software infrastructure critical. Recalling his time at Baidu, Amodei notes that his primary technical takeaway was in speech recognition, while noting the stark cultural contrast in data collection practices.
- **Capability Growth:** AI progress follows an exponential trajectory, scaling roughly $10\times$ in value and capability per year.
- **Open-Source Proliferation Lag:** Advanced offensive cyber capabilities typically trickle down to freely available open-source models within a 6-to-12-month window.
- **Baidu Experience:** Amodei spent a year working at Baidu's U.S. office, focusing primarily on speech recognition technology.
[01:02:17](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3737s) High-Tech Authoritarianism and Democratic AI
Dario Amodei voices deep concern over the geopolitical threat posed by high-tech authoritarian regimes, drawing on concrete historical examples like China's persecution of the Uyghurs, its crackdown in Hong Kong, and the CCP's pressure on US corporate networks to silence critics. He warns that when cutting-edge AI is combined with an aggressive authoritarian state, it risks creating a dystopian outcome "like *1984* or worse."
To counter this, Amodei argues that the global community has a crucial opportunity to ensure AI develops as a "pro-democracy technology"—one that enhances individual freedom and supports equal justice under the law. Whether the technology skews democratic or dystopian depends directly on the choices made by AI companies, governments, and society as a whole.
- Real-world authoritarian warning signs: Suppression of dissent within domestic and international business networks, human rights abuses against Uyghurs, and the dismantling of Hong Kong's autonomy.
- Strategic goal: Steering AI development so it functions as a tool for democracy and civil liberty rather than state surveillance and control.
[01:03:24](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3804s) Recursive Self-Improvement and the Smooth Exponential
When asked about the theoretical tipping point where an AI becomes capable of recursively improving itself, Amodei rejects the notion that this will occur as a single dramatic event. Instead, he describes recursive self-improvement as an ongoing, continuous process that is already underway—for instance, AI models are currently used to suggest architectures for next-generation models. He notes that AI-driven productivity gains have risen from a 10% to 15% increase in total factor productivity a year ago to roughly 20% to 30% today.
Amodei emphasizes that AI progress follows a "smooth exponential" curve rather than a sudden moment of running out of control. He criticizes extreme, reactionary responses—such as swinging wildly from total deregulation to sudden calls for nationalization or complete shutdowns—as unhelpful and a sign of being caught off guard. The wise approach, he argues, is a steady, rational ratcheting up of safety measures and controls in tandem with the technology's steady curve.
- Productivity metrics: Total factor productivity gains from AI have climbed from 10–15% to 20–30% within a single year.
- Governance philosophy: Rejects wild "yo-yoing" between laissez-faire development and panic-induced shutdown/nationalization; advocates for proportional, continuously ratcheting safety controls along the exponential curve.
[01:05:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3907s) Historical Lessons: Leo Szilard vs. Oppenheimer
Reflecting on *The Making of the Atomic Bomb*, one of his favorite books, Amodei notes that he identifies far more with Leo Szilard—the physicist who first conceived the nuclear chain reaction—than with J. Robert Oppenheimer. He views Oppenheimer as somewhat of a "failure case," warning against relying on larger-than-life figures who position themselves at the center of critical historical moments.
Instead of trusting centralized figures or single powerful institutions, Amodei stresses the necessity of robust checks and balances across all actors in the AI ecosystem to ensure a safe outcome.
- Key historical figure: Leo Szilard (conceptualizer of the nuclear chain reaction).
- Organizational view: Rejects single-hero narratives in favor of institutional checks and balances among corporations, governments, and civil society.
[01:05:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3949s) Managing Civilizational Risk and the Airline Analogy
Addressing his past estimate of a 10% to 25% chance of civilizational collapse due to AI, Amodei argues that Anthropic’s existence actively reduces that probability rather than increasing it. The baseline risk stems from the rapid nature of the technology, geopolitical rivalries, and market dynamics where someone will inevitably fill the void if safe actors do not participate.
While Anthropic implements extensive testing and internal safety mechanisms—with half of the company's work dedicated to risk reduction—Amodei admits the danger can never be brought down to zero because AI models are inherently unpredictable. He uses an airline analogy: a new company can build an airliner that is 10 times safer than any existing alternative, but it still cannot guarantee an airplane will never crash. Responding to Emily Chang’s point that a 25% crash rate is unacceptably high for passengers, Amodei agrees entirely, stating that the primary objective of their work is to push that probability down as low as possible.
- Estimated existential risk probability: 10% to 25% baseline risk of civilizational collapse.
- Resource allocation: Approximately 50% of Anthropic's internal efforts are focused on mitigating risk and building safety mechanisms.
- Risk posture: Models currently released are not considered dangerous outside of cyber risks, with safety testing heavily prioritized prior to launch.
[01:07:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4052s) Earning Public Trust Through Costly Action
When asked why the public should trust Anthropic when the company stands to gain enormously from building such powerful technology, Amodei responds candidly: "starting from a position of distrust... is pretty rational." He acknowledges that Silicon Valley as a whole has lost public trust over recent years and must re-earn it through concrete actions rather than rhetoric.
Amodei points to Anthropic's history of making financially disadvantageous decisions to uphold its core principles. These include withholding their powerful model Mythos despite commercial pressure, proactively cutting off model access to China—a move that cost the company "several hundred million dollars" back when that represented a substantial portion of their total revenue—and delaying the release of Claude 2 to run additional safety tests. While acknowledging that Anthropic is an imperfect, complex organization prone to mistakes, he argues that its track record demonstrates a consistent, earnest effort to do the right thing.
- Financial sacrifices for safety: Cut off model access in China at a cost of several hundred million dollars in revenue; withheld Mythos and delayed Claude 2 for safety evaluations.
- Stance on public trust: Skepticism toward tech companies is justified; trust must be earned through historical consistency and voluntary adherence to principles over short-term revenue.
[01:09:40](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4180s) Lighthearted Closing on the Reality of Being CEO
The interview closes on a lighthearted note as Chang jokes about whether Amodei always wanted to be a Hollywood star. Amodei laughs and reflects on one of the most unexpected aspects of his role as CEO: the frequent requirement to wear makeup for media appearances. "That was not on my bingo card," Amodei remarks, to which Chang reassuringly replies, "Just a little powder."
## Notable quotes
> [10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) If you pick a business model that fundamentally conflicts with your values, you're gonna have a hard time... Either you betray your own values or you become irrelevant. — Dario Amodei
> [10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) We want to use AI to, you know, cure diseases that we couldn't cure before... working with biotech, it's working with pharma... All of those are enterprises. — Dario Amodei
> [15:11](https://www.youtube.com/watch?v=x2VHFgyawPE&t=911s) If a enterprise's sole defense is having written code that no one else can replicate, good luck, you're not gonna be able to defend that. — Dario Amodei
> [17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) Look, we're all adults here. — Dario Amodei
> [24:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1447s) Think of all the problems we had in 1900s... all the material deprivation that we don't have to deal with today. Then think of another hundred years of that. I really believe this century of scientific and medical progress... we're gonna have a much, much better world. — Dario Amodei
> [34:38](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2078s) The idea that this is cheap marketing is itself cheap marketing. This is laziness; this is failure to engage with serious intellectual work... It's so stupid. It's so unserious. And whenever someone says something like that, I take them less seriously. — Dario Amodei
> [54:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3255s) 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
> [01:07:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4052s) starting from a position of distrust... is pretty rational. — Dario Amodei
## Mentioned
Dario Amodei, Emily Chang, Anthropic, OpenAI, Sam Altman, Narendra Modi, Demis Hassabis, Google DeepMind, AlphaFold, Claude, Claude Code, Claude Cowork, Daniela Amodei, Holden Karnofsky, Open Philanthropy Project, Amazon, Google, Microsoft, Nvidia, Jensen Huang, Palantir, Immigration and Customs Enforcement (ICE), Customs and Border Protection (CBP), Maven Smart System, Mythos, Long-Term Benefit Trust, Scott Bessent, Susie Wiles, Baidu, Leo Szilard, J. Robert Oppenheimer, The Making of the Atomic Bomb, The Adolescence of Technology, The Circuit, Bloomberg Originals
Anthropic CEO Dario Amodei argues that frontier AI follows a smooth exponential trajectory whose severe civilizational, military, and economic risks can be managed only through principled commercial sacrifices, enterprise alignment, and institutional checks.
Enterprise Focus and the Shifting Software Moat (10:45)
Amodei explains that Anthropic prioritized enterprise tools like Claude Code and Claude Cowork over advertising-driven consumer models because consumer setups incentivize addiction and low-quality content. Enterprise partnerships in biotechnology, education, and energy rely on enduring trust, aligning with safety priorities. Following market reactions to Claude Cowork that saw $285 billion in SaaS equity drop during the so-called 'SaaSpocalypse,' Amodei asserts that merely writing complex code is no longer a defensible moat. Instead, enduring advantage will stem from domain expertise and direct customer ties, even as the overall software market expands tenfold.
Hypergrowth, Capital Buffers, and Recursive Workflows (18:28)
To handle rapid operational scaling, Anthropic relies on large capital cushions to finance its compute needs without excessive equity dilution. While the firm planned for a 10x yearly expansion in compute demand, Q1 2026 revenue increased more than 3x in that single quarter—an annualized rate of roughly 80x. Amodei views compute markets as liquid and shortfalls as temporary issues resolved within months through partners like Google and Amazon. Operationally, Anthropic deploys Claude internally to design model architectures and write code, creating feedback loops where recursive AI productivity gains have moved from 10–15% to 20–30% within a year.
Workforce Automation and Proposed Policy Offsets (27:57)
Revisiting his estimate that AI could displace roughly 50% of entry-level white-collar jobs within one to five years, Amodei defends the severity of the forecast against Silicon Valley critics who accuse him of using 'doom marketing.' He points to Anthropic's essay *The Adolescence of Technology*, which outlines distinctions between tasks and jobs alongside policy remedies such as a 'token tax' and macroeconomic rebalancing. Amodei warns that unbounded automation risks severe economic shocks, urging enterprises to expand their output rather than cutting staff, while shifting labor toward infrastructure, physical manufacturing, and relationship-driven services.
National Security Red Lines and Geopolitical Posture (36:41)
Driven by concerns over autocratic aggression from Russia and China, Anthropic took on classified defense work with the U.S. Department of Defense via Palantir, though Amodei denies working with ICE, CBP, or on operations in Gaza. The firm set strict boundaries refusing fully autonomous kinetic targeting or mass domestic surveillance, prompting federal pushback and supply chain risk labels after declining to drop these safeguards. Geopolitically, Amodei actively supports U.S. chip export bans against China despite tension with hardware allies, arguing that combined AI and authoritarian governance presents severe risks to democratic societies.
Cyber Vulnerabilities and Managing Civilizational Hazards (48:18)
Highlighting safety trade-offs, Amodei addresses Anthropic's choice to delay the full rollout of Mythos, an AI model that uncovered 271 previously unknown flaws in Firefox alongside thousands in corporate codebases. Because the model can autonomously traverse the cyber kill chain and open-source capabilities lag frontier models by only 6 to 12 months, Anthropic restricted access to give defensive teams time to patch systems. Amodei estimates baseline civilizational AI risk at 10% to 25%, noting Anthropic dedicates about 50% of its internal effort to risk reduction. To check concentrated executive power, Anthropic created the independent Long-Term Benefit Trust, which retains the authority to remove board members and dismiss the CEO.
Amodei acknowledges public skepticism toward frontier AI labs as rational, maintaining that trust must be earned through costly commercial decisions such as withholding offensive tools and absorbing revenue losses. Whether high-capability AI stabilizes geopolitics through deterrence or escalates existential danger remains dependent on maintaining strict human oversight and formal governance guardrails.
## Background
该视频由彭博原创（Bloomberg Originals）旗下由知名记者艾米莉·张（Emily Chang）主持的深度访谈节目《The Circuit》于 2026 年 6 月 21 日发布在 YouTube 平台。页面简介显示，本次长达 70 分钟的独家专访深入探讨了 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）的旧金山成长历程、与 OpenAI 的理念分歧与竞争、与五角大楼的安全对峙以及 AI 的终局推演。
## TL;DR
在访谈中，Dario Amodei 深入剖析了 Anthropic 从坚守企业级赛道到逼近万亿美元估值的指数级演进，并首次公开回应与 OpenAI 决裂的核心症结在于信任与诚实的瓦解。面对 AI 军事化与国家安全风波，他重申了拒绝大规模监控与坚守“人在回路”的道德红线，并详述了暂扣高危攻防模型 Mythos 背后的商业牺牲与治理抉择。此外，他严厉驳斥了硅谷在极端反监管与全面国有化之间反复摇摆的“悠悠球心态”，强调递归自我改进是一条连续平滑的指数曲线，人类社会必须通过长效分权制衡机制将 10% 至 25% 的文明崩溃概率降至最低。
## Atmosphere & style
整场对话在理性质感与紧绷的现实张力之间自如穿梭，Emily Chang 提问极其尖锐直接，毫不回避误击悲剧、失业海啸与同行质疑等敏感痛点。面对步步紧逼，Dario 表现出近乎外科医生般的冷静、克制与哲学思辨深度，但在痛斥社交媒体的浮躁断章取义与地缘极权风险时，又流露出极为坚定的道德信念感。全片既充斥着身处指数级历史风暴眼中的窒息感与宏大责任感，又在结尾关于 CEO 频繁化妆的花絮调侃中回归平实，给观众带来一种智识激荡且极具历史庄重感的视听体验。
## Bulletin
- [18:19](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1099s) Anthropic 估值逼近一万亿美元大关，并在 2026 年第一季度迎来营收环比超 3 倍的爆发式增长（折算年化超 80 倍），CEO Dario Amodei 明确表示巨额融资是对冲算力指数激增的防御性安全垫。
- [48:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2898s) Anthropic 绝密模型 Mythos 具备自主遍历网络杀伤链并编写漏洞利用代码的能力，内部在 Firefox 中已挖出 271 个全新漏洞，出于安全考量 Anthropic 选择暂扣发布并先向防守方提供。
- [37:00](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2220s) 在五角大楼合作争议中，Anthropic 坚持拒绝大规模监视与全自主致命武器两条红线，哪怕因此被贴上供应链风险标签且被竞对替代，仍全力推动国会立法设立军事安全护栏。
- [01:05:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3949s) Dario Amodei 认为 AI 导致文明崩溃的概率约为 10% 至 25%，并指出 AI 自我改进并非未来的突变奇点，而是当前已经在平滑推进的指数级演进过程。
## Key takeaways
- [00:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=7s) 面对指数级技术迭代带来的极端压力，成熟决策者应当保持理智与清醒，避免在自满与无端恐慌之间反复摇摆。
- [05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) 离开 OpenAI 并创立 Anthropic 的深层原因并非单纯的技术安全理念分歧，而是核心价值观与诚信基础的彻底破裂。
- [10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) 坚决押注企业级赛道而非流量至上的消费级市场，既能避开靠注意力变现的低质内容陷阱，又能让商业模式与造福人类的长期愿景高度契合。
- [13:47](https://www.youtube.com/watch?v=x2VHFgyawPE&t=827s) 大模型行业的核心护城河依然是底层模型质量，传统软件公司单纯依靠编写复杂代码的技术壁垒正被迅速瓦解，亟需转向行业特定知识与客户深度关系。
- [17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) 坚定倡导对华高性能芯片出口管制以遏制高科技威权扩张，即便这一立场与主要芯片资方的短期商业利益存在直接冲突也决不妥协。
- [26:13](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1573s) 公司内部约 50% 的研发生产力已融入 AI 赋能，但对写长文等深度思辨过程，仍坚持不把端到端撰写交给 AI 以免丧失独立推导能力。
- [27:37](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1657s) 未来经济可能出现高 GDP 增长与白领岗位高失业并存的悖论，人类不可替代的阵地将转向物理现实世界、高信任的情感连接以及终极价值意图的对齐。
- [43:28](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2608s) 在国防合作中严守“人在回路（Human-in-the-Loop）”底线，坚决抵制由 AI 全权闭环决策生杀大权，防范类似误击女子学校的灾难在无人类复核下被无限放大。
- [50:33](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3033s) 反驳外界将扣留 Mythos 视为营销噱头的论调，强调全网系统漏洞绝对数量有限，率先将工具交给防御方能建立更坚固的安全底座。
- [56:10](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3370s) 反对在“极端反监管”与“全面国有化”之间悠悠球式横跳，主张通过 Anthropic 长期利益信托（LTBT）等机制建立跨系统与企业内部的分权制衡。
- [01:03:24](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3804s) 递归自我改进绝非断裂式的奇点瞬间，目前 AI 已为 Anthropic 带来 20% 至 30% 的研发全要素生产率提升，并沿着平滑指数曲线持续成倍增长。
- [01:07:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4052s) 公众对硅谷的信任必须依靠实打实的商业牺牲来赢回，包括主动放弃中国市场营收以及推迟高商业潜力模型的发布。
## Detailed walkthrough
### [00:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=7s) 相对论式的指数加速与高压下的理性决断
面对 Emily Chang 关于在高压下能睡几个小时的提问，Dario Amodei 坦承自己从来都不是睡眠质量很好的人，但如今正在学习如何在异常巨大的压力下寻找放松与安眠的技巧。当被问及身处如此迅猛的技术浪潮内部究竟是何感受时，Dario 用了一个狭义相对论的物理学比喻来形容这种“指数级加速”带来的冲击：想象你坐在一艘以相对论速度远离地球的飞船上，当你入睡后再醒来，地球上已经过去了整整两天，你必须在一天的客观时间里处理两天积累的事物；而随着飞船持续加速，下次醒来地球上已经过去了三天，接着是四天……内部的世界正是这样在以令人目眩的节奏被压缩和倍增。
面对每天早晨是否会因未知风险而陷入焦虑与偏执，Dario 明确表示，将精力耗费在“醒来会面对什么”的无端恐慌中毫无建设性。尽管眼前总有堆积如山、切实存在的现实挑战需要应对，但他通过研读历史中身处极高压力环境下的领导者得出结论：成熟的决策者必须学会理性应对，而不是夸大危险或在两极之间剧烈摇摆。
在 Dario 看来，那种一会儿觉得“一切安好、无需担忧”，转天又高呼“天哪，我们今天必须立刻陷入恐慌”的心态，是极其不成熟的决策特征。真正成熟的态度是既不视而不见、拒绝盲目自满，认清风险正变得越来越大，同时又能像操刀重大手术的外科医生或指挥战局的军事指挥官一样，在关乎无数人命运的抉择面前，始终保持一种理智、清醒且基本的镇定。
### [01:56](https://www.youtube.com/watch?v=x2VHFgyawPE&t=116s) 意料之外的破圈与平滑指数曲线下的行业登顶
访谈中切入了一个极具生活气息的细节：Emily Chang 谈到自己的儿子前一天想要借用她的 Claude Cowork 账号，却被她一口回绝，理由是“我自己的 Token 还不够用”。Dario 对此会心一笑，指出 Anthropic 最初的战略重心完全是面向企业端（Enterprise），但如今即便没有在消费端倾注过多推广精力，C端用户的增长也在以惊人的速度自发爆发。
对于 Anthropic 如今站在整个 AI 宇宙的风暴中心，Dario 分享了一个反直觉的认知视角：在他整个职业生涯以及创办 Anthropic 的全过程中，底层始终是一条“平滑的指数曲线”。在外界看来，指数曲线的前半段仿佛“什么都没发生、风平浪静”，随后突然出现一点小动静，紧接着就瞬间爆发、冲上云霄。其实早在很久以前，他就根据这条曲线预测，Anthropic 大概会在这个时间节点成长为营收最高、估值最具分量的顶尖 AI 公司，而现实的发展也正如模型预测的那样发生了。
Dario 感慨道，虽然从宏观图表上看一切都在意料之中，但当现实真正铺展在眼前时，其丰富生动的细节与现实张力依然令人震撼。面对聚光灯，Anthropic 内部关注的命题其实从未改变：
- 如何训练出性能更优异的模型；
- 如何将这些模型转化为卓越可靠的产品；
- 如何在底层构筑坚不可摧的安全边界；
- 如何在利用技术造福人类的同时，严密管控可能引发的社会系统性风险。所有这些曾经的核心议题，如今只是被置于了一台倍率更高的显微镜下接受全社会的审视。
```mermaid
flowchart TD
accTitle: 指数发展感知模型
accDescr: 外界视角与内部曲线预测的认知差异流程图
subgraph SG1["外界的非线性感知"]
A["长期平稳/似乎无事发生"] --> B["出现微小技术突破"]
B --> C["爆发式暴增 / 舆论轰动"]
end
subgraph SG2["内部的连续性逻辑"]
D["平滑指数模型推演"] --> E["算力与模型规模稳步跃迁"]
E --> F["登顶行业营收与估值节点"]
end
C -.->|"现象映射"| F
```
### [03:34](https://www.youtube.com/watch?v=x2VHFgyawPE&t=214s) 旧金山的童年根基与硅谷的非顺从精神
谈及在旧金山的成长经历，Dario 勾勒出了一个与现代硅谷刻板印象截然不同的童年图景。他的父亲是一名皮具手艺人，母亲在图书馆工作。在第一波互联网革命席卷旧金山湾区的大潮中，年幼的 Dario 对周遭的商业狂热毫无兴趣，甚至可以说是完全绝缘。他沉浸在自己的小世界里，狂热地草拟数学公式、阅读科幻小说，心中唯一的执念是“理解宇宙底层的运转法则”，对整个世界充满了纯粹的好奇心。
旧金山这座城市独特的文化底蕴深刻塑造了他的世界观。Dario 认为，这里包容非顺从主义（Nonconformism）、推崇极致的个人特质，并认可“有疯狂想法是完全可以被接受的”。相比于欧洲部分地区或美国其他习惯于抑制异想天开、将标新立异视为怪诞的传统环境，硅谷最可贵的一点就在于对特立独行的宽容。
尽管 Dario 明确表示自己对当下硅谷的许多弊病持有严厉的批评态度，但他依然高度赞同硅谷的核心精神内核：无论外界所有所谓的权威和专家如何反对，只要你内心拥有一套自洽的愿景与对世界的连贯认知，你就应当全力以赴去探索。即便这种尝试可能彻底失败，但由于技术创新具有长尾效应（Long-tailed distribution），某些冷门的矿脉一旦被掘开，底下可能就是一座不可估量的巨大金矿。
随后，Emily 提及在 2016 年左右，Dario、他的妹妹 Daniela Amodei（Anthropic 联合创始人兼总裁）以及妹夫 Holden Karnofsky（Open Philanthropy 负责人）曾在旧金山的一处合租房共同生活。当时拥有生物物理学背景的 Dario，主要作为顾问协助 Holden 评估发展中国家医疗健康和前沿生物学研究项目，通过严谨的理性分析判断哪些科研方向真正具备突破潜力。
### [05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) 决裂 OpenAI 的真相：核心不仅在安全，更在信任与诚实
Dario 出走 OpenAI 并创办 Anthropic 的始末，长期以来一直是硅谷最具戏剧色彩的技术公案。面对 Emily Chang 追问这背后的真实分歧到底是什么，Dario 首次给出了极其直白且严肃的回应。他指出，在构建极其强大的人工智能技术时，任何人都会面临无数极其艰难的抉择，甚至连 Anthropic 每天也都在面对不知道对错的十字路口。因此，围绕技术安全路线产生合理分歧是完全正常的，Anthropic 内部也经常出现激烈辩论，单纯的安全理念差异绝不足以导致决裂。
Dario 直言不讳地指出，真正导致分道扬镳的根本原因在于“信任的破裂”：
> “当你觉得你无法信任某个人，当你感到他们的价值观与口头宣称的并不一致，当你感到他们不够诚实、投身于此的真实意图并非他们所标榜的那样，当你目睹了令人不安的不诚实行为模式时，你就会发现，继续与这家公司合作、继续去信任这家公司变得极其困难。”
在 Dario 看来，当双方愿景不同且信任基础彻底瓦解后，无休止的内部争吵毫无意义，最体面的解决方式就是各自上路——“你做你的，我做我的”。他对于由 Anthropic 走出自己的道路感到非常坦然，并坚信最终的裁判权属于市场与公众：“我们将看到谁会在市场上胜出，谁会在公众信任的法庭上胜出。我们正在身体力行地展示如何以负责任的方式部署这项技术。如果他们不认同，大可公开拿出他们的论据，除此以外没什么好多说的。”
### [07:42](https://www.youtube.com/watch?v=x2VHFgyawPE&t=462s) 印度峰会“牵手”风波与 AI 领袖间的真实信任图景
针对此前在印度举办的全球人工智能峰会上，Dario 与 OpenAI 首席执行官 Sam Altman 在台上似乎拒绝牵手合影的尴尬名场面，Emily Chang 展开了幽默却犀利的质问。Dario 笑着澄清这纯粹是一场现场混乱导致的意外。他解释称，任何参加过多国元首出席的高级别峰会的人都会知道，现场调度往往极度混乱，绝非刻意针对印度。当时他们都是在最后一刻被叫上台，站位被临时打乱，摄影师拍照时，印度总理纳伦德拉·莫迪（Narendra Modi）突然在台上指示大家一起手拉手，整个过程仓促且无序。
面对 Emily 的进一步追问——“Sam 和 Elon 正在法庭上激烈互诉，你与 Sam 之间心存芥蒂，如果这些掌握人类最重要技术命运的人甚至无法在台上牵手，公众如何相信你们能在关乎人类生死存亡的生存性风险（Existential Risk）上通力合作？”Dario 做出了冷静的区分。
Dario 指出，在当前构建前沿 AI 的核心领袖群体中，个人的品质和值得信任的程度存在着“巨大的方差”（Wide variance）。外界关于“AI 行业领袖彼此之间毫无信任”的刻板印象是不符合事实的。作为反例，他谈到了 Google DeepMind 的负责人 Demis Hassabis（Gemini 模型的掌舵人）：
- 两人在技术竞争之外，私下相识已超过 15 年；
- 双方在多项行业核心议题上长期密切协作；
- Anthropic 不仅从 Google 采购大规模计算资源（Compute），双方还经常在安全研究的最前沿无保留地交换意见。
这表明在健康的信任与共同的科学原则之上，技术对手之间完全能够构建起理性而稳固的合作机制。
### [09:17](https://www.youtube.com/watch?v=x2VHFgyawPE&t=557s) 行业信任度与“向上竞争”的胡萝卜与大棒
达里奥·阿莫代伊（Dario Amodei）在讨论 AI 行业的生态时指出，行业内部各方玩家的可信度参差不齐。他坦承在 Anthropic 之外也存在着许多自己完全信任且行事可靠的同行。在他看来，真正理想的局面是那些值得信任的机构联合起来，树立极高的安全与伦理准则，从而倒逼那些不可信的参与者也不得不遵循相同标准。达里奥根据多年的行业经验总结出一条规律：有些人绝不会自觉主动地做正确的事，但只要行业大多数人站到了正确的一边，剩下的少数人便别无选择，只能被动跟随。
这种行业驱动机制被达里奥拆解为“胡萝卜”与“大棒”两种力量。积极的一面是良性互动与相互启发——例如他与 Google DeepMind 的首席执行官德米斯·哈萨比斯（Demis Hassabis）之间的关系。DeepMind 开发了革命性的蛋白质结构预测模型 AlphaFold，促使 Anthropic 也在生物学领域发力；而当 Anthropic 在模型可解释性研究（Interpretability Research）上取得突破时，DeepMind 也会随即展开相关研究。达里奥笑称这甚至算不上针锋相对的零和竞争，而是“每个团队做出了很酷的突破，另一方看到了觉得很酷，便想着我们也能做点什么新东西”，这是典型的以正面激励推动的“向上竞争”（race to the top）。
而隐性的“大棒”机制则在于舆论与行业声誉的重压。当行业领军企业纷纷走在正确的规范道路上时，那些拒绝跟进的公司就会在公众和同行面前显得极其难看。达里奥观察到，很多同行往往是“一边心不甘情不愿地照着做，一边嘴上强装自己在搞完全不同的创新”，甚至反过来把倡导合规的企业描绘成居心叵测。达里奥对此并不感到意外，他认为这属于意料之中的反应，但正是通过这种混合了示范与施压的机制，整个行业才能被真正凝聚起来并达成合作。
### [10:45](https://www.youtube.com/watch?v=x2VHFgyawPE&t=645s) 坚守企业级赛道：拒绝“信息垃圾”与商业模式的价值观契合
面对主持人艾米莉·张（Emily Chang）关于竞争策略的提问——为何在其他公司竞相追逐吸睛、好玩的消费级爆款应用时，Anthropic 却押注于编程与企业级场景（如 Claude Code 和 Claude Cowork），达里奥给出了一个兼具价值观与商业理性的长篇解释。他强调，创立 Anthropic 时首要考虑的是“把事情做对”，但由于训练前沿大模型成本极度昂贵，公司必须拥有可持续的造血商业模式。他从过去的职业生涯中吸取到的最核心教训就是：如果选择了一个从根本上与自身价值观相冲突的商业模式，未来必将陷入两难困境——“要么背叛自己的价值观，要么被彻底边缘化”，沦为无论怎么选都痛苦万分的“第二十二条军规”（Catch-22）。
达里奥直言不讳地批评了消费级市场与社交媒体的盈利逻辑。那类产品完全依赖广告收入驱动，本质上是在不择手段地刺激用户留存与沉迷，最大化侵占用户的注意力时长；如今在 AI 生成视频领域充斥着各种毫无营养的“信息垃圾”（slop），正是这种畸形激励机制的产物。相反，企业级市场与 Anthropic 期望释放的 AI 正向潜力高度一致。达里奥始终坚信 AI 带来的积极价值将远超负面风险，而所有真正能造福人类的重大命题，几乎全部属于“企业级”范畴：
- 攻克疑难疾病：需要与生物科技公司、大型药企及高校科研机构协同协作。
- 降低能源成本与提升能效：属于典型的大型工业与公用事业企业场景。
- 改善教育与发展中国家的公共健康：依托学校系统、学术机构以及非营利组织运作。
- 驱动宏观经济增长：本质上也是由千千万万企业的效率提升所支撑。
此外，企业级市场的商业逻辑建立在长期的信任与契约关系之上。达里奥指出，消费级应用往往带有某种投机或玩噱头（gimmicky）的色彩，容易快起快落；而企业客户看重的是数年如一日的稳定交付与说到做到，这与 Anthropic 追求安全、可控部署的初衷不谋而合。尽管在商业落地过程中偶尔也会遭遇妥协与艰难抉择，但由于选对了赛道，这种根本性的价值冲突被降到了最低。
### [13:47](https://www.youtube.com/watch?v=x2VHFgyawPE&t=827s) 开发者半天即可转投竞品，AI 护城河究竟何在？
当艾米莉尖锐地指出，开发者只需一个下午就能从 Claude 零成本切换到 ChatGPT 或 Gemini，质疑大模型行业是否真能建立长期壁垒时，达里奥给出了极其务实的回应。他坚信在当前阶段，“模型本身的质量就是最核心的关键（Model quality is the most important thing）”。他自信地表示，Anthropic 当下在模型能力与质量上处于大幅领先的地位。
达里奥坦言，虽然市场上存在一定程度的用户惯性，但他和 Anthropic 从不把希望寄托在“产品黏性大、用户懒得迁徙”这种被动防守上。唯一能够维持长久优势的方法，就是始终交付更出色的模型性能与更优秀的产品体验。他透露，至少在节目录制当下，Anthropic 的业务增长曲线丝毫没有出现任何放缓的拐点，反而呈现出加速上扬的态势。
```mermaid
flowchart TD
accTitle: 模型竞争力驱动飞轮
accDescr: 展示以模型质量为核心、拒绝依赖用户惯性的产品正向循环
subgraph CoreEngine ["核心驱动"]
A["顶尖模型质量<br>(Model Quality)"] --> B["优质开发者产品<br>(Claude Code / Cowork)"]
end
subgraph MarketDynamics ["市场表现"]
B --> C["企业深度信任与口碑"]
C --> D["业务增速持续上扬<br>(Growth Inflection)"]
end
subgraph FalseMoats ["被摒弃的伪壁垒"]
E["用户迁移成本与惯性"]
F["消费级花哨噱头"]
end
D --> A
E -. "绝不依赖".-> B
F -. "主动规避".-> A
```
### [14:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=872s) Claude Cowork 引发“SaaS 大崩盘”恐慌，传统软件行业何去何从？
在 Claude Cowork 发布后不久，二级市场软件板块遭遇巨震，甚至在短时间内蒸发了 2850 亿美元的市值，被华尔街交易员戏称为“SaaS 大崩盘”（SaaSpocalypse）。针对传统软件行业是否会被 AI 快速彻底颠覆的疑问，达里奥保持了审慎的态度。他认为市场走向极其复杂，没人能精准预言具体节奏，但传统软件公司的既有护城河必将发生深刻分化：
| 护城河类型 | 在 AI 时代的演变态势 | 达里奥的判断与应对建议 |
|:--- |:--- |:--- |
| **纯代码编写能力** | **正在彻底瓦解** | 如果一家公司的核心壁垒仅在于“写出了别人写不出的复杂软件”，其防线将不堪一击，AI 正在迅速抹平这种技术实现壁垒。 |
| **客户深度关系** | **相对更加重要** | 多年沉淀的企业商务关系、跨部门信任与采购渠道难以在短期内被算法抹去。 |
| **垂直行业特定经验（Know-how）** | **不可替代性提升** | 深入理解实际业务流如何运转的特定领域知识（Domain Knowledge），是通用大模型落地不可或缺的基石。 |
达里奥告诫软件行业的从业者：切勿心存侥幸或盲目自欺，不能妄想过去的成功范式能保命至今。正确的做法是彻底清点手中的每一道护城河，果断认清哪些正在归零、哪些权重上升，并积极寻找全新的护城河。
尽管会有一批抱残守缺的企业成为时代的牺牲品甚至直接破产，但达里奥从宏观视角对整体软件行业依然持乐观态度。他认为整个技术盘子在被迅速做大——“如果 AI 让可能性放大了 10 倍，那么即便既有传统软件厂商只增长 1.5 倍，其相对比重变小了，但整个行业绝对体量依然在扩大而非萎缩”。最终陷入泥潭的，只会是那些面对变革把头埋进沙子里的被动受害者。
### [17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) 面对科技巨头资方的博弈与对华芯片管制的坚定立场
面对来自亚马逊（Amazon）、谷歌（Google）、微软（Microsoft）和英伟达（NVIDIA）等一众既是投资方又是潜在竞争对手的巨头林立格局，艾米莉追问 Anthropic 在面对附带巨额商业里程碑条款的融资时，究竟“谁在真正发号施令”。
达里奥坚定地表示，Anthropic 始终保持着极高的独立性，敢于公开表达自己的真实观点。他举了一个极具火药味的例子：他长期以来一直在极力倡导对中国实施严格的高性能 AI 芯片出口管制。达里奥直言，如果让中国在尖端 AI 能力上取得领先，将对美国的安全以及全球民主制度的稳定造成极其恶劣的打击。尽管一些芯片巨头资方因自身商业利益显然对这一主张极不赞同，但这从未阻止达里奥公开发声。达里奥态度坦然地表示，即便在与这些巨头签署了更多商业合作协议之后，他也依然会当着所有人的面坚持重申这一政治与战略立场：
> “他们很清楚我们始终在积极与他们合作，我们也一直是非常优秀的合作伙伴。但我敢肯定，他们心里一定巴不得我们闭嘴别提这些事；可这就是我内心深处坚信的原则，你又能怎样呢？”
### [18:19](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1099s) 近万亿美元估值与算力扩张的缓冲逻辑
彭博社报道称 Anthropic 的估值正迈向超越 OpenAI 的水平，对于一家成立仅五年的初创企业而言，估值已逼近一万亿美元大关。面对主持人 Emily Chang 关于“既然在算力投入上更具纪律性、盈利路径更明确，为何还需要筹集如此巨额资金”的追问，Dario Amodei 给出了非常理性的商业解释。他指出，算力需求的爬坡速度极其惊人，即便业务的基本面非常健康，一年内所需的算力也会激增三到四倍。尽管公司完全预期收入增长能够匹配甚至超越算力的攀升，但在面对未来的“不确定性锥体（cone of uncertainty）”时，巨额融资实际上是最理性的缓冲机制。
Dario 强调，在极高估值下进行融资对股权的稀释比例极小。这种防御性资金储备并不意味着底层业务存在任何问题，恰恰相反，它与强劲的商业基本面完全兼容。他和合作伙伴之间始终保持着成熟务实的态度，正如他所言：“我们都是成年人了，完全可以在一件事上紧密合作，同时在另一件事上保留不同意见。”
- 估值规模：成立五年的 Anthropic 估值接近一万亿美元，甚至被报道超过主要竞争对手 OpenAI。
- 算力与营收预期：年度算力规模正以每年三到四倍的速度快速攀升，融资被定位为对冲不确定性的关键安全垫。
- 稀释效应：极高的市场估值使巨额融资仅带来非常轻微的股权稀释。
### [19:29](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1169s) 算力瓶颈与2026年Q1爆发式增长的真相
针对外界关于服务器过载、服务可用性下降以及用户频繁耗尽 Token 的质疑，Emily 提及 Dario 曾批评竞争对手在基础设施建设上采取“孤注一掷（YOLO）”的激进做法，并询问 Anthropic 是否算力不足。Dario 澄清道，从合理规划的标准来看，公司从未少买算力。Anthropic 原本按照每年 10 倍的算力增速进行规划，但在 2026 年第一季度，公司遭遇了超乎预期的业务爆发——单季度营收环比增长超过 3 倍，折算成年化增长率竟高达 80 倍以上（$3^4 \approx 80$）。
Dario 直言，没有任何理智的团队会提前按 80 倍的年化需求去囤积算力，因为一旦实际增速只有 10 倍，公司就将承担整整 8 倍的闲置成本。当前用户遭遇的算力紧张纯粹是“局部极端的算力需求大爆炸”，这种异常增速在全行业都是不可持续的。不过，依托与 Google 和亚马逊（Amazon）等巨头建立的长期算力协议，计算资源市场本身具备充足的流动性，只要企业能够高效运用算力并具备坚实的需求，短缺问题在短短一两个月内就能得到调配与缓解。
```mermaid
flowchart LR
accTitle: 算力规划与实际业务爆发对比
accDescr: 展示常规10倍年化规划与2026年Q1实际80倍年化增速之间的缺口应对
subgraph Plan["常规基线规划"]
P1["预期年化算力增长: 10x"]
end
subgraph Reality["2026年Q1爆发"]
R1["单季度营收增长: >3x"]
R2["折算年化增长: ~80x"]
R1 --> R2
end
subgraph Response["算力流动性调度"]
C1["短期服务器承压/Token受限"]
C2["调配 Google / Amazon 算力"]
C3["1-2个月内补齐产能缺口"]
C1 --> C2 --> C3
end
Plan -.-> C1
Reality --> C1
```
### [21:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1275s) 超越OpenAI与“向上竞赛”的生态拉动
当被问及超越宿敌 OpenAI 感觉是否良好时，Dario 并没有落入简单的胜负欲叙事，而是提出了“向上竞赛（race to the top）”的行业愿景。他认为 Anthropic 的核心战略是通过自身的技术标杆与商业成功，拉动整个行业生态向上走。他笑称竞争对手有时虽然嘴上在抨击 Anthropic，私下里却在同步抄袭其做法，但这种带动效应本身具有巨大的行业价值。
在 Dario 看来，赢得商业竞争与打造顶尖模型本身并不是为了“为了击败对手而击败对手”，而是为了拥有足够的行业话语权，引导整个生态朝着更安全、更有益的方向前进。他明确表示自己绝非主张关停 AI 技术的“停滞主义者”，公司本身就身处自由企业体制之中，追求商业成功与研发顶尖技术无可厚非，关键在于必须在推进技术的同时全力化解模型带来的安全风险，两者始终处于一种动态的平衡之中。
### [22:38](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1358s) 规模扩张下的价值观坚守与文化稀释挑战
Emily 提出，在 Anthropic 还处于弱小阶段时坚守道德制高点相对容易，而在如今万亿估值的体量下，坚守初心是否变得前所未有地艰难。Dario 坦言自己伴随着公司的每一次体量跃升都保持着极度警惕甚至“偏执（paranoid）”。公司面临的威胁既可能来自于丧失商业上的求胜心，也可能来自于核心价值观的流失，而他认为两者必须协同存在——正是因为 Anthropic 能够做出世界顶尖的模型，公司才有底气去践行并推广自身的价值观。
对于内部治理，Dario 指出真正的风险不在于创始人或管理层价值观的改变，而在于人员规模快速膨胀带来的文化稀释。当公司以闪电般的速度从各大科技巨头招聘大量员工时，新员工本能地会把前东家的工作方式带进来。为了对抗这种惯性，Dario 现在将近一半的个人时间投入在向全员沟通 Anthropic 的独特文化与运作机制上。他和联合创始人 Daniela Amodei 将维护公司文化基因视为最核心的第一要务。
- 警惕偏执：在扩张的每个阶段，公司都面临丧失好胜心或丢失核心价值观的双重风险。
- 时间分配：CEO 将约 50% 的精力用于在全公司范围内维护和传授企业文化，防止大厂惯性同化。
- 协同逻辑：商业竞争力的成功是强力贯彻安全价值观的前提保障，二者互为支撑。
### [24:07](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1447s) 疯狂的产品迭代速度与Claude的自举赋能
面对 Anthropic 近期令人瞠目的产品迭代与发布节奏，Dario 将其归结为两大支柱：一是组织内部高度协同的统一文化与惊人的组织效率；二是全面将 Claude 本身引入模型研发流程所实现的“自举加速”。团队已经开始在日常工作中使用 Claude 来辅助研发后续模型、优化架构并加速软件交付。尽管这类人机协作新范式仍在摸索中，但它已经带来了巨大且日益可靠的研发加速度。
谈到 AI 展现出的最震撼能力，生物学出身的 Dario 尤为推崇模型在生物学与医药健康领域的突破。他分享了一个具体案例：Claude 曾精准诊断出一位团队核心成员（包括 Daniela）身上连多位顶尖专业医生都漏诊的医学难题。此外，模型在分子药物设计、计算化学等高门槛任务上的表现令人惊叹。Dario 满怀憧憬地表示，人类有望迎来相当于过去整整一个世纪的科学与医学进步，正如从饱受疾病和物资匮乏折磨的 1900 年跨越至今，若能安全走过当前阶段，AI 将极大改善人类的生存体验与生命质量。
### [26:13](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1573s) Dario的写作习惯与AI辅助思考的边界
Dario 因撰写深度长文与思想随笔而在科技界闻名。在被问及是否使用 Claude 协助写作时，他坦承自己确实会用，但从未允许 Claude 直接生成的文本直接出现在最终定稿中。原因在于他对个人独特的行文风格非常苛求，目前的 Claude 尚未达到比他本人写得更好的水准，不过他相信那一刻终会到来。目前，Claude 在他写作时主要承担辅助角色，负责发散头脑风暴、梳理论述逻辑脉络以及检索参考文献。
```mermaid
flowchart TD
accTitle: Dario 的 AI 协同写作工作流
accDescr: 展示人类核心思考与 AI 辅助工具在写作过程中的分工边界
subgraph HumanCore["人类主体思考"]
H1["确立核心论点与自我思维沉淀"]
H2["撰写正文与打磨独特文风"]
H3["形成团队共识与战略参考"]
end
subgraph AIAssist["Claude 辅助节点"]
A1["头脑风暴与主题发散"]
A2["梳理行文结构与脉络"]
A3["检索参考资料与背景素材"]
end
H1 --> A1
A1 --> A2
A2 --> A3
A3 --> H2
H2 --> H3
```
针对 Emily 提出的“若将写作完全交给 AI 是否会导致人类丧失批判性思考与挣扎推导能力”的担忧，Dario 深表赞同。他直言自己写作的一半目的其实是为了厘清自己的思维脉络，从而明确下一步的决策方向，并为公司内外建立共同的话语基准。他警示道，如果单纯依靠 AI 进行端到端的黑盒生成（例如直接下达“写一篇关于 AI 风险的文章”的指令），人类确实会丢失深度思辨的核心价值；当前最健康的使用方式，是将 AI 严格限定在研究和辅助思维整理的辅助位上。
### [27:37](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1657s) 写作方式与“快速增长但高失业”的经济悖论
谈及自己亲自撰写长文的习惯，Dario Amodei 解释称，如果直接把写作完全丢给 AI，AI 既写不出他内心真正构思的思想，他也会彻底丧失沉思与梳理逻辑的益处。尽管随着大模型能力日益精进，未来或许能更直接地融入创作流并保留这些核心价值，但这将是一个需要漫长探索的微妙平衡，绝非一蹴而就。
紧接着，对话切入宏观经济的严峻走向。Dario 指出，人类社会极有可能面临一个前所未见的离奇组合：一方面是 GDP 呈现极速爆发式增长，另一方面却伴随着高失业率、广泛的就业不足（underemployment）以及充斥着大量低薪岗位，进而导致财富与收入不平等急剧恶化。主持人 Emily Chang 追问，一年前 Dario 曾直言 AI 可能在 1 到 5 年内消灭半数初级白领工作（entry-level white-collar jobs），在技术日新月异的当下，这个预估比例是否已经进一步攀升至 50% 以上。
---
### [28:25](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1705s) 自动化浪潮与企业劳动力重构：以 Anthropic 为例
面对主持人的提问，Dario 澄清了他一贯的观点。他表示网络上流传的言论经常被脱离语境截取为几秒钟的片段，而他当时的真实表意是：虽然未来确切的走向不可预测，但“50%”代表了局势可能发生颠覆性震荡的数量级（order of magnitude）。更重要的是，他始终在同步强调应对方案，包括设立“Token 税”（token tax）、协助企业在技术转型中妥善安置员工、宏观经济政策调控，甚至是重新评估职业培训计划。
Dario 指出，在技术演进的经典曲线中，短期内 AI 会显著提高人类效率。例如自动化完成了 90% 的繁琐事务后，员工在剩余 10% 的核心环节上获得了 10 倍的杠杆赋能。然而，当自动化程度逼近 100% 时，就必须为人类寻找全新的分工。
为了具体说明这种企业内部的变化，Dario 拿 Anthropic 自身的工程师团队作为典型案例：
- **纯软件工程（Pure Software Engineering）**：目前 AI 已经包揽了 Anthropic 内部几乎所有的代码编写，多数工程师借助 AI 实现了生产力跃升；但团队也观察到新的苗头——对部分员工而言，AI 并未带来更高赋能，反而由 AI 独立全权完成任务效率更高。
- **前线部署与方案架构（Forward Deployed Engineers / Solutions Architects）**：随着商业客户暴增，公司对兼具过硬技术背景与客户沟通能力的岗位需求出现激增。然而，并非每一位传统纯软件工程师都能无缝转轨至这类复合型角色。
```mermaid
flowchart TD
accTitle: AI 自动化与岗位职能演进路径
accDescr: 展示 AI 从部分赋能到完全自动化过程中产生的岗位转型及企业选择
subgraph Phase1["自动化演进阶段"]
A["传统白领 / 研发工作"] --> B["AI 自动化 90% 任务<br/>人类杠杆效应提升 10 倍"]
B --> C["AI 逼近 100% 全自动化"]
end
subgraph Phase2["企业决策分支"]
C --> D{"企业选择模式"}
D -->|"削减成本（零和博弈）"| E["裁员 / 减少招聘<br/>同样产出消耗更少资源"]
D -->|"做大蛋糕（正和博弈）"| F["维持 / 增加人力<br/>开拓全新业务范式"]
end
subgraph Phase3["Anthropic 内部人员分流"]
F --> G["技术与客户复合型岗位<br/>（如 Forward Deployed Engineer）"]
end
```
在与企业级客户打交道时，Anthropic 总是极力劝导客户选择“用相同的资源做更多开拓性的事”（正和博弈），而非单纯“用更少的人力维持原样”（削减成本），力求将社会大盘的“蛋糕”做大，从而在剧烈的技术冲击下跑赢劳动力与新岗位的匹配时钟。
---
### [33:05](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1985s) 五年后的人类社会：物理世界、人际连接与价值定向
Emily Chang 追问其终局推演：如果五年后醒来社会面临如此严重的失业冲击，是否会演变为社会动乱与革命？Dario 坚决赞同并强调这正是必须全力阻止的最坏结局。对于未来人类依旧不可替代的阵地，他列举了三大核心方向：
1. **物理世界与制造业**：尽管具身智能与机器人革命正在发生，但其发展步调显著慢于纯虚拟的 AI 算法。当信息处理几乎变得零成本时，现实世界的实体建设（如建造数据中心、实体制造）将成为真正的物理瓶颈，需要大量人类投身实体劳动。
2. **以人为本的高信任人际连接**：在医疗诊断等关键场景中，哪怕 AI 能查出连资深医生都遗漏的隐疾，或者在客服领域表现得比人类更精准，人类依然强烈渴望与真实的人类对话交流。情感沟通、同理心与信任建立是机器无法替代的。
3. **人类的价值对齐与意图定向**：在最高决策维度上，AI 的目标体系与行为边界必须严格对齐人类的价值信念和长远意图，这也为人类保留了监督与掌舵的立足点。
---
### [34:38](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2078s) 痛斥社交媒体与“末日营销”论调：黄仁勋质疑之争
针对外界的非议——例如英伟达 CEO 黄仁勋（Jensen Huang）批评其混淆了“任务”（tasks）与“工作”（jobs），以及同行抨击其夸大危机是为 Anthropic 搞“末日营销”（doom marketing）以谋取利益——Dario 表达了强烈的愤慨，并予以正面回击。
Dario 指出，Anthropic 不仅设立了经济研究资助（economic grants）并编制经济指数（economic index），他本人在其长文《技术的青春期》（The Adolescence of Technology）中更用了整整五页篇幅，极其详尽严谨地剖析了“任务”与“工作”的本质区别，论证了为何此次通用 AI 革命与过往历次工业革命截然不同，并提出了涵盖从宏观税收到私人慈善行动等六大具体防范措施。
Dario 坦言自己“极度痛恨社交媒体（detest as a category）”。他怒斥那些断章取义截取一年前三秒视频片段进行攻击的批评者是“智识上的彻底懒惰”，指责这种不愿静心研读严肃文献的浮躁风气是“硅谷的一种流行病（disease of Silicon Valley）”。他直言不讳地反驳道：指责别人在做廉价营销，这本身就是最廉价的营销手段；凡是基于三秒视频就轻率下结论的人，他只会觉得对方“愚蠢且极不严肃（unserious）”，根本不配得到认真对待。
正当谈话气氛转向严肃之际，Emily Chang 将话题引入下一个高压领域：作为全球顶级 AI 公司，Anthropic 已经深深卷入美国国家安全的军事行动之中，而其与五角大楼关于军事安全护栏（military safeguards）的对峙也正在持续升级。
### [37:00](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2220s) 进军五角大楼与划定道德红线
当被问及为何一家由加州理工学院学者背景创立、以安全著称的 AI 公司，会成为首批与美国国防部签署合同以在机密网络上运行以支持作战的机构之一时，达里奥·阿莫代伊（Dario Amodei）并未回避这一反差。他直言“世界已经变了”，目睹俄罗斯入侵乌克兰以及中国对台湾构成的潜在威胁，让他深感威权主义阵营正在强势复苏且咄咄逼人，民主国家必须具备自卫能力。他表示，无论面对哪一届政府，安索罗匹克（Anthropic）在国家安全议题上的支持立场始终如一：“我们不希望看到一个中国和俄罗斯能用 AI 分析所有情报、发起进攻，而我们却无法自卫的世界。”他坦言，参与涉密网络部署绝非为了盈利——不仅要应对繁复的法律合规与诉讼风险，在军方网络上部署系统更是一件“费力不讨好且收益微薄”的苦差事。
出于对原则的坚持，阿莫代伊认为在使用该技术时必须设立严格边界。正如他在文章《技术的青春期》（Adolescence of Technology）中所写的那样：技术可以被广泛应用，但绝不能用来侵蚀民主国家自身的核心价值观。为此，Anthropic 划出了两条绝对不可逾越的“红线”：大规模监控与全自主致命武器。阿莫代伊严厉抨击了那些在国防问题上摇摆不定的科技企业，认为从“坚决不与政府合作”骤变为“毫无原则地全面承接一切军方项目”的做法令人费解。
```mermaid
flowchart TD
accTitle: Anthropic 的国防合作原则与红线
accDescr: 展现 Anthropic 在军事技术支持中的价值取向与禁止领域
subgraph DefenseSupport["支持的国防应用领域"]
A["情报分析与处理"]
B["辅助决策与军事威慑"]
C["保障民主国家防御能力"]
end
subgraph RedLines["坚决不予逾越的红线"]
D["大规模监视 (Mass Surveillance)"]
E["完全自主武器 (Fully Autonomous Weapons)"]
end
Anthropic["Anthropic 的军事合作框架"] --> DefenseSupport
Anthropic --> RedLines
```
- **核心红线**：坚决拒绝参与大规模监视系统开发，坚决拒绝部署无人类干预的完全自主武器系统。
- **合作动因**：抗衡威权阵营在军事 AI 领域的单边突破，维护战略威慑，而非出于商业利润考量。
---
### [39:17](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2357s) 被列入黑名单与国会护栏之争
针对 Anthropic 自 2024 年起与大数据公司 Palantir 的合作，艾米莉·张（Emily Chang）提出了尖锐质疑，指出后者的技术已被用于美国移民及海关执法局（ICE）、地方警察局乃至加沙冲突。阿莫代伊澄清，Anthropic 设立了严格的业务防火墙：无论是通过 Palantir 还是其他渠道，Claude 均未接入 ICE、美国海关和边境保护局（CBP），也没有在加沙战场使用。然而，正是因为坚守红线并拒绝五角大楼的某些要求，Anthropic 随后遭到白宫行政禁令封杀，甚至被五角大楼贴上“供应链风险”标签，而 OpenAI 则迅速跟进，签署了 Anthropic 拒绝接受的那份军方合同。
谈到在这场对抗中何为“胜利”，阿莫代伊并不认为这是一家私营公司的输赢博弈，而是一场关于“政府应如何正当使用 AI”的公开社会思辨。AI 仍是一项新兴技术，人们对其可靠性边界以及它究竟是捍卫还是侵蚀民主价值尚缺乏透彻理解。尽管一份商业合同无法彻底阻止竞对妥协，但 Anthropic 的退场与坚持唤醒了公众意识，并推动了美国国会展开严肃的跨党派立法努力，试图设立硬性护栏并立法禁止最危险的 AI 军事滥用。
- **机构防火墙**：坚决排除与 ICE、CBP 的合作，未介入加沙实战网络。
- **政府反制与竞争替代**：在 Anthropic 拒签协议并被列为供应链风险后，OpenAI 迅速接手填补空缺。
- **政策推动成效**：以合同争议为契机，引发美国国会两党对 AI 军事边界立法的实质性讨论。
---
### [41:12](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2472s) 面对“意识形态疯子”指责与军事威慑逻辑
随着冲突升级，批评者甚至将阿莫代伊斥为“意识形态疯子”（ideological lunatic）和“左翼偏激狂”。对此阿莫代伊显得颇为坦然，笑称自己早就习惯了更难听的骂名。他坚信自己的人生和战略目标极其单纯明晰：一是确保公司在商业与技术上取得成功，二是坚决捍卫自身的价值观，“只要锚定这两点，你的生活其实非常简单，因为你永远清楚自己站在哪里”。
针对美军官员公开宣称在大语言模型协助下，美军每日打击目标识别能力从 1,000 个激增到 5,000 个、可能导致杀戮效率倍增的现实，阿莫代伊从战略威慑的角度进行了辩护。他强调自己是一个爱国者，坚信强大的军事能力本身并不引发战争，反而能起到遏止战争的作用。针对具体的军事行动政策，他打了一个比方：正如政府广泛使用 Microsoft Excel，微软无法也无权要求五角大楼“只能用 Excel 规划这起行动、不能规划那起行动”；作为底层技术供应商，不应越俎代庖去替军方指挥官做微观战术决策，但企业必须从宏观层面划定符合国家底线价值观的技术防线。
- **打击效能提升**：美军依托 LLM 工具，将日均目标锁定能力由 1,000 个大幅提升至 5,000 个。
- **威慑逻辑**：坚信强大的常规与技术防御力量是阻止独裁政权发动侵略的核心威慑手段。
- **责任边界类比**：技术提供方需在顶层划定伦理底线，但不应干预指挥官合法的战术自由度，正如软件厂商无法限定办公软件的战场用途。
---
### [43:28](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2608s) 致命打击惨剧与“人在回路”的底线坚守
随后访谈转向了极其沉重的话题：彭博社曾报道美军在针对伊朗的军事行动中，通过 Palantir 的 Maven 智能系统（Maven Smart System）动用 Claude 辅助打击；而在今年二月的一起空袭中，美军导弹误击了伊朗一所女子学校，导致包括大量儿童在内的 150 多人丧生。面对“Claude 是否参与其中”的逼问，阿莫代伊坦言自己并不掌握前线具体操作数据，但对战争中发生的此类平民伤亡悲剧感到无比痛心。
阿莫代伊情绪严肃地指出，这起悲剧恰恰说明了 Anthropic 冒着公司生死存亡的风险设立红线有多么关键。女子学校遭遇误击发生在“有人类在回路中做最终决策”的常规场景下——AI 仅提供辅助，最终发射命令由军官下达，军事指挥官依然会犯下可怕的错误。他警告称：“设想一个不是由 Claude、而是由其他完全放开限制的竞品 AI 主导的世界：AI 模型自行裁决生杀大权，完全没有人类参与复核。这正是我们当初誓死抵抗的未来。”当艾米莉追问这所学校明明在谷歌搜索上有公开主页、为何 AI 甚至没能识别出如此浅显的事实时，阿莫代伊指出，这更加凸显出军用决策在可靠性与技术依赖上面临的深层风险与巨大挑战。
- **实战部署平台**：涉及集成于 Palantir 旗下“Maven 智能系统”（Maven Smart System）的 AI 辅助定位模块。
- **误击悲剧**：导弹袭击造成包括女学生在内逾 150 名平民丧生。
- **“人在回路”（Human-in-the-Loop）防线**：坚决抵制由 AI 全权闭环做出交战判定的全自主武器，坚持最终决策权必须留在人类手中。
### [46:24](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2784s) 致命自主武器、失控风险与 AI 威慑论
在谈及战争中的 AI 应用原则时，Dario Amodei 明确重申了底线标准：“必须由人类做出最终决定（human-in-the-loop）”。他坦言自己可能并不掌握军方涉密的具体细节，也不清楚 Claude 或其他 AI 在特定军事场景中到底扮演了什么角色，但这一原则是绝不可动摇的准绳。
针对 Emily Chang 提出的尖锐问题——“AI 战争究竟是更能阻止第三次世界大战（如中美冲突），还是更可能引爆它？”——Amodei 给出了权衡后的判断：综合来看，AI 更可能起到阻止冲突的作用，但前提是必须对其施加限制。如果毫无底线地滥用，AI 极有可能酿成灾难。他特别引用了经典电影《奇爱博士》（*Dr. Strangelove*）中的“末日装置”（Doomsday Device）作为警示：一旦设定成检测到核打击就自动反击的自主毁灭机器，任何误判都会引发灭顶之灾。
在 Amodei 看来，现实中的武装冲突往往源于双方误判与冲动冒进，而缺乏人工监督的全自主致命武器（Lethal Autonomous Weapons）会让意外升级的几率飙升。相反，如果将 AI 恰当地用于情报搜集（Intelligence Collection）——例如精准预警针对台湾的入侵行动，或是乌克兰战场上的最新军事调动——敌对方就会意识到自己的全盘动向已被彻底掌握，从而在发起进攻前“三思而后行”。Amodei 始终坚信：压倒性的情报优势与强大的应对能力本身，就是最有效的威慑手段。
---
### [48:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2898s) 内部绝密模型 Mythos 带来的震撼与发布考量
随着 Anthropic 几乎每周都登上各大媒体头条，舆论的焦点落在了其传闻中的最强模型——Mythos 上。该模型具备自主遍历整个“网络杀伤链”（Cyber Kill Chain）所有环节的惊人能力。面对“既然 Mythos 过于强大而不能公开发布，它最让你吃惊的是什么”这一提问，Amodei 深入剖析了它的跃升幅度。
让 Anthropic 团队最为震惊的，并非该模型能够寻找漏洞（大家平时都在讨论挖漏洞），而是它能以惊人的成功率直接将漏洞转化为可用的攻击性利用代码（Exploits）。这种能力跨越非常巨大。Anthropic 在早期仅向少数合作伙伴和企业开放了测试，结果对方在没有任何诱导的情况下，纷纷惊恐地反馈：“这是一件超级武器，使用它应该像持有枪支一样需要执照”、“求求你们千万不要公开发布它”。合作伙伴在实测中挖掘出了海量严重级别的未公开漏洞与配套的 Exploit，甚至主动乞求 Anthropic 压紧底牌。
Amodei 澄清，Anthropic 并不打算将 Mythos 永久雪藏，而是计划随着网络安全防护体系的完善，逐步向更广泛的人群开放，最终推向大众。但目前的现实痛点在于现有的安全防护机制依然极其脆弱。Anthropic 此前在 Opus 4.7（一款能力更弱但网络能力不错的模型）上部署了防护机制，但此类分类器和护栏仍容易被越狱（Jailbroken）或绕过。很多友商自以为目前的防御已足够稳固，但在 Anthropic 严苛的内部测试和行业评估看来，现存的防御手段远远谈不上安全。
```mermaid
flowchart TD
accTitle: Mythos能力与防护演进
accDescr: 展示从模型发现漏洞并生成Exploit，到企业反馈与逐步防御发布的过程
subgraph Capability ["Mythos 核心能力跃升"]
A["扫描整个代码库"] --> B["自主挖掘系统漏洞"]
B --> C["自动生成攻击代码 (Exploits)"]
end
subgraph Reaction ["早期测试与反馈"]
C --> D["早期合作企业震惊"]
D --> E["'这是一件超级武器'<br/>请求 Anthropic 严禁公开发布"]
end
subgraph Strategy ["渐进式安全发布路径"]
E --> F["受限提供给网络防御方 (Defenders)"]
F --> G["修复修补既有漏洞"]
G --> H["建立鲁棒的抗越狱防御墙"]
H --> I["未来面向大众全面开放"]
end
```
---
### [50:33](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3033s) 驳斥开源复现论与“公关噱头”质疑
对于外界质疑——包括部分研究员声称能用廉价的开源模型复现其成果、OpenAI 已经具备同等能力、以及 Mythos 延期发布纯属一次精心策划的“公关营销策略”（PR Play）——Amodei 给出了毫不客气的正面驳斥：
- **开源复现纯属无稽之谈**：Amodei 直指这种说法“假得离谱（incredibly false）”。他解释道，真正的网络安全难题是让 AI 面对数百万行的浩瀚代码库，大海捞针般自主发现全新漏洞。而社交媒体（Twitter/X）上有人却宣称“只要把开源模型直接指向 Mythos 已经标注出的具体代码行，开源模型也能指出问题”，这根本不是同一回事——“在草堆里把那根针指给你看，与你顺手把针捡起来，完全是两码事”。
- **实战成果的数据支撑**：Mythos 在知名开源浏览器 Firefox 中一口气挖出了 271 个此前从未被人发现的全新安全漏洞；在合作私有企业的内部代码库中，更是锁定了数千个尚未修复或无法公开披露的高危隐患。在此之前，从未有任何模型能完成如此深度的自动化漏洞挖掘。
- **商业反噬而非营销收益**：所谓“为了炒作而扣住不发”的逻辑在商业上根本站不住脚。Mythos 已经在 Anthropic 内部极大加速了研发节奏、产品生产及下一代模型演进，如果公开发布，商业变现能力将极其巨大。为了安全性而选择不发布，恰恰让 Anthropic 在商业利益上承受了极其沉重的代价。
---
### [52:08](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3128s) 攻防失衡下的“防御优先”战略与各方拉锯
面对“如果 AI 能帮防守方，也能帮攻击方，我们未来还能守住任何系统吗”的悲观论调，Amodei 阐述了他的安全逻辑：率先将 Mythos 交给网络防御者（Defenders），就是为了在攻击者拿到同等工具之前，把互联网底层的漏洞彻底修补完。
Amodei 打了个比方：系统的暴露面就如同一块布满孔洞的表面，但“漏洞的绝对数量是有限的（finite）”。如果防御者能用强力 AI 把漏洞全部堵死，同时未来的软件代码也直接由顶尖大模型编写，那么整个代码生态的防护强度将产生质的跃迁。Amodei 预期在未来 6 到 12 个月内，全球有望构建一个比以往任何时候都更加坚不可摧的网络安全底座。
与此同时，Anthropic 正与美国政府监管部门密切沟通，尊重官方因担忧反间谍与情报外泄风险而提出的减速建议。Amodei 透露了他每天面临的剧烈冲突与撕扯：
- **企业客户与多国政府**：每天都有大量客户和外国政府官员致电 Dario，迫切要求立刻获取 Mythos 的访问权限。
- **美国政府与安全团队**：安全团队与政府机构则在另一侧发出严肃预警，强调外溢风险并要求刹车。
Amodei 坦言双方的诉求都有其合理性，这是一个需要全社会共同严肃面对的权衡难题。他犀利地抨击了社交媒体上的冷嘲热讽以及某些竞品公司的借机踩踏，指责其做法“缺乏成熟度与历史庄重感（lack of gravitas and maturity）”。
---
### [54:11](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3251s) 商业领跑下的权衡妥协，以及政府是否该将 AI 收归国有
Emily Chang 问及 Dario 是否做过让自己感到不舒服的妥协。Amodei 直言：“Anthropic 的整个创业史就是一部妥协史。”在纯粹的理想主义世界里，一个模型在公开发布前可能需要数年时间来反复推敲和测试所有潜在隐患。但现实中，Anthropic 当年仅将 Claude 的初版发布推迟了几个月。Amodei 认为走极端是荒谬的，真正的领导力在于权衡。如今 Anthropic 已经在商业化上占据了领跑地位，他和妹妹 Daniela Amodei 终于有底气将安全天平往“审慎”的方向进一步倾斜——Mythos 暂缓向公众发布正是这种策略的体现；如果不是行业领军者，企业根本没有放慢脚步的资本。
对话随后切入更深层次的政治权力命题：“既然 AI 如此强大，美国政府为何不直接将其收归国有（Nationalize AI）？政府凭什么允许一家私营公司掌控如此惊人的力量？”
Amodei 坦言这是一个极其严肃且发人深省的问题，他本人完全理解这种担忧。但他并不赞同政府采取简单粗暴的全面接管。Amodei 梳理了人类技术发展的历史脉络：在此之前，人类历史上出现的每一种改变格局的重器，几乎都起源于政府或完全由政府操盘：
- **核武器**：自始至终由国家主导研发并由军方严密控制。
- **互联网、GPS 与蜂窝移动通信**：其早期底层 R&D 全部诞生于国家实验室、联邦机构及顶尖高校科研院所。
而 AI 却打破了这一历史规律：它是人类历史上第一项完全由私营部门构思、出资并自主孕育出来的颠覆性超级技术，而政府在其中自始至终未曾扮演实质性的研发主导角色。这一根本性的范式转移，使得政府与私营 AI 巨头之间的权力界限与信任架构面临着前所未有的考验。
### [56:10](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3370s) 权力制衡与反对硅谷的“悠悠球式”极端态度
达里奥·阿莫代伊（Dario Amodei）深入剖析了 AI 行业面临的危险治理真空：政府起步过晚、陷入被动，而私营部门则在加速狂奔。他坦言这绝非他理想中的局面，但现实中别无选择——这项技术具有巨大的经济价值，技术在物理上完全可行，且对手国家也在全力研发，“它必然会被造出来”。因此，核心问题不在于私营部门在推进研发，而在于政府长期缺位。面对这种失衡，他提出必须建立全方位的“分权与制衡机制”（checks and balances）。
在企业内部制衡方面，Anthropic 设立了“长期利益信托”（Long-Term Benefit Trust，简称 LTBT）。该机构有权任命和罢免公司的大多数董事会成员，这意味着在机制上它最终拥有“解雇 Dario 本人”的权力。通过引入超越单纯股东利益的公共治理要素，确保公司无论未来发生何种变化，这一监督架构都将永存。阿莫代伊不仅鼓励其他 AI 公司效仿此类治理模式，更呼吁政府内部立法部门和司法部门主动作为，设定政策红线并强制推行模型发布前测试与审计。
谈及硅谷同行时，阿莫代伊言辞辛辣地讽刺了科技圈内某些群体极不成熟的“悠悠球式”（yo-yoing）心理动荡。这群人在初期连最基础的透明度和出口管制都极力抗拒，高呼“监管会毁灭创新、摧毁造出技术的潜力”；然而一旦见识到首个真正的技术危险，立刻一百八十度大转弯，极度恐慌地鼓吹“应当将 AI 彻底国有化，让政府直接接管”。阿莫代伊直言这种从“极端的反监管”一瞬间跳跃到“彻头彻尾的共产主义式政府全面剥夺”不仅荒谬可笑，更证明了他们缺乏远见。相比之下，Anthropic 始终主张温和理性的渐进方案：既不否认力量也不盲目恐慌，而是正视技术的光滑指数级曲线并做出适当应对。
```mermaid
flowchart TD
accTitle: 双向分权与制衡治理架构
accDescr: 展示企业内部制衡、政府与企业间相互约束以及平稳应对指数级增长的治理体系
subgraph Enterprise ["Anthropic 内部治理制衡"]
LTBT["长期利益信托 (LTBT)<br/>具备公共治理属性"]
Board["董事会绝大多数席位"]
CEO["CEO / 管理层<br/>(可被依法解雇)"]
LTBT -->|任命与罢免权| Board
Board -->|监督与解任权| CEO
end
subgraph Governance ["跨部门与跨体系相互制衡"]
Gov["政府体系<br/>(立法/司法/行政)"]
AI_Corps["AI 领军企业"]
Gov -->|强制前置测试与审计| AI_Corps
AI_Corps -->|技术透明度与专业制衡| Gov
end
subgraph Response ["面对技术指数级爆发的态度对比"]
SV_Chaos["硅谷极端派 (悠悠球心态)<br/>完全反监管 ⟷ 全面国有化"]
Anthropic_Way["Anthropic 理性派<br/>承认指数增长，平滑升级对策与安全防线"]
end
Enterprise -.->|示范企业自治| Governance
```
- **长期利益信托的核心权力**：LTBT 并非单纯的财务投资人，其使命是保障技术造福人类，掌握控制董事会多数席位的实权。
- **监管诉求**：明确提出需要立法规制，重点在于强制要求开展模型发布前的安全评估（pre-release testing）与独立审计。
- **抨击极端思潮**：猛烈批评硅谷部分人士在“连透明度都拒绝”与“要求政府强行没收 AI 资产”两个极端之间反复横跳。
---
### [58:56](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3536s) 白宫之行与 Mythos 的网络安全现实警钟
在被问及近期重返白宫的会面情况时，阿莫代伊分享了他与政府接触的一贯原则：坚持自身底线与价值观，并尽可能与体制内各方理性合作。他透露，美国政府高层对 Anthropic 代号为 Mythos 的前沿模型展现出了极其重视的态度。阿莫代伊与财政部长贝森特（Scott Bessent）以及白宫幕僚长苏西·威尔斯（Susie Wiles）展开了深入对话，并评价他们真正理解了技术的潜在风险本质，而 Mythos 展现出的惊人能力恰好让这些抽象的风险变得具象而触手可及。
阿莫代伊坦言，任何一届政府内部都会存在认知差异——有些部门沟通顺畅、深谙技术，另一些部门则相对难以磨合。对此，Anthropic 的策略是在复杂的政治生态中保持平稳游刃有余地穿梭，尽最大努力引导政策走向建设性道路。
- **高层对话对象**：与财政部长贝森特（Treasury Secretary Bessent）及白宫幕僚长苏西·威尔斯（Chief of Staff Susie Wiles）举行了富有成果的高级别会晤。
- **Mythos 的警示作用**：该模型展示的网络安全与攻防能力，成为促使美国政界核心圈直观理解尖端 AI 风险的关键转折点。
---
### [59:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3589s) 远端开源模型、中国竞争与极权主义的技术反思
针对中国开源大模型迅速普及并在美国企业中被广泛免费采用的现象，阿莫代伊从底层价值和地缘政治两个维度给出了分析。他首先指出，尖端智能（Frontier Intelligence）存在极高的价值溢价，虽然落后于前沿的模型也能形成繁荣生态并处理低难度任务，但 AI 正在以每年 10 倍的速度跨越式演进。在阿莫代伊看来，很多习惯了移动互联网或社交媒体时代产品思维的人，常常依靠过往的产品直觉来下判断，然而在 10 倍指数级模型迭代面前，这些直觉全都失灵了——“智能本身的分量压倒了一切”，最顶层的商业与战略价值始终牢牢绑定在最前沿的模型上。
令阿莫代伊真正担忧的不是商业替代，而是落后开源模型扩散带来的安全威胁。他预测，目前仅存在于 Mythos 级别的前沿网络攻击能力，大约 12 个月后可能就会降维变成任何人都能随手下载的开源模型能力。即便届时系统可能已经打了补丁，这种网络攻防能力的不可控扩散依然是一个无法阻挡且令人极度不安的现实。
谈到早年在百度（Baidu）硅谷研发中心工作一年的经历，阿莫代伊澄清这并未直接塑造他的对华地缘观念。当时除了学到语音识别技术，最令他心有余悸的细节是在收集语音数据时，对方曾轻描淡写地表示“在中国我们不在乎隐私”。阿莫代伊对极权体系的根本警惕源于更宏观的现实：维吾尔族人的遭遇、对香港自由的压制，以及中国官方甚至能将触角伸入美国商业网络以压制批评声音。他警告，当这种高科技威权国家与尖端 AI 相结合，将催生出超越《1984》的真正反乌托邦。AI 最终会成为促进自由、实现法律面前人人平等的民主技术，还是沦为极权统治的超级工具，完全取决于 AI 企业、政府乃至全人类当下的具体抉择。
- **每年 10 倍增长的指数效应**：模型智能每年的数量级飞跃使得“非前沿（far from frontier）”模型的价值迅速边缘化。
- **Mythos 级别网络风险的时间差**：最前沿的高危攻击能力大约在 12 个月后便会下沉为开源公开资源，形成巨大的防护窗口期压力。
- **高科技威权主义与《1984》**：警惕算法与极权结合形成严密监控与言论审查，呼吁企业必须承担责任，确保 AI 成为捍卫民主的武器。
---
### [01:03:24](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3804s) 递归自我改进并非奇点瞬间，而是连续的平滑指数
对于业内热议的“AI 强大到足以自我改进，从而触发失控循环”的临界时刻，阿莫代伊给出了不同寻常的技术洞察：这绝不是未来的某一个具体“瞬间”，而是一个早已经启动的连续演进过程。他透露，AI 目前已经能够为下一代 AI 建议模型架构；一年前，AI 为 Anthropic 研发带来的全要素生产率（Total Factor Productivity）提升大约为 10% 到 15%，而现在这一数字已经攀升至 20% 到 30%，并且还在持续翻倍。
阿莫代伊强调，根本不存在某一天 AI 突然脱缰失控的“奇点魔法时刻”，一切都在一条加速的指数曲线上平滑推进。理解这一进程的“罗塞塔石碑”（Rosetta Stone）正是“平滑的指数级增长”（smooth exponential）。他再次批评了那些遇到自我改进就惊慌失措、叫嚣“彻底关停一切”的恐慌分子，指出这种歇斯底里的应激反应与此前盲目反监管的人如出一辙——都是因为事前缺乏预见而被现实突袭时的失态表现。明智的治理姿态是在指数上升的每一个节点冷静评估，有条不紊地同步提升防御机制和管控力度，让安全手段与模型能力实现同频平滑升级。
- **生产率实测提升**：AI 对 Anthropic 自身算法与工程全要素生产率的贡献已从去年的 10%–15% 增至目前的 20%–30%，且在持续成倍增长。
- **自我改进的本质**：从提出架构建议到优化流程，自我迭代是一个早在此刻便已平稳运行的渐进变量，而非未来的断裂式突变。
- **理性治理原则**：拒绝在极端反智与过度恐慌之间跳跃，以确定性的安全防线平稳应对指数级能力飞跃。
### [01:05:15](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3915s) 历史教训：制衡机制与“奥本海默式”的失败
讨论从曼哈顿工程时期的历史人物延伸开来。Dario Amodei 提到了利奥·西拉德（Leo Szilard）——这位最早构想出核裂变链式反应的物理学家。在 Amodei 看来，全人类不可能依赖某个“比生命更宏大”的传奇领袖或妄图成为一切中心的风云人物来安然度过眼前的 AI 变革危机。世界充满了拥有各自利益的强大主体，唯一能让这一进程善终的路径，是在所有关键环节建立起广泛的权力制衡（checks and balances）。
基于这一视角，Amodei 直言他将罗伯特·奥本海默（J. Robert Oppenheimer）的经历视为一种反面教材与“失败案例”（failure case），代表着现实中最应该极力避免的走向——即把极度危险技术的命运完全系于极少数核心人物或中心化权威之手。
### [01:05:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3949s) 10%至25%的文明崩溃概率与“航空公司”比喻
Emily Chang 追问了一个极具冲击力的数据：Amodei 曾公开表示 AI 引发人类文明崩溃的概率约为 10% 到 25%，这绝非微不足道的数字。她尖锐地提出，导致这一灾难的导火索是否可能正是由 Anthropic 亲手打造的系统？Amodei 回应称他衷心希望不会如此，并坚信 Anthropic 所采取的行动是在压低而非推高这一灾难概率。危险的底层逻辑源于技术的直白特性、全球多国博弈的格局以及自由市场中层出不穷的新公司竞争；如果 Anthropic 退出并留下真空，风险只会更大。
Amodei 强调，AI 技术的内在属性就是难以预测，因此必须在发布前进行极其严苛的测试。除了网络安全领域之外，当前发布的各类前沿模型整体上并不具备真正的现实危险性。为了抵御潜在风险，Anthropic 内部有近一半的工作都在构建无数层防御机制以尽可能削减风险，但谁也无法将其彻底归零。对此，他提出了一个生动的类比：
> “假设市面上有一批航空公司，你想创办一家更安全的新航司。你的公司完全可以做到比所有同行安全 10 倍；但如果有人跑来问你：‘你能百分之百保证你的飞机永远不会坠毁吗？’你根本不可能给出这种保证。”
Emily Chang 随即反驳道：“但如果一架飞机有 25% 的失事概率，根本没人敢坐上去。”Amodei 赞同这一说法，直言 25% 的风险确实太高了，而 Anthropic 存在的核心使命，正是竭尽全力把这一灾难概率压制到远低于此的水平。
```mermaid
flowchart TD
accTitle: AI文明风险与Anthropic缓解策略
accDescr: 展示技术固有风险与Anthropic通过安全措施压低文明崩溃概率的逻辑流程
subgraph RiskSources["风险根源 (10%-25% 文明崩溃概率)"]
A["技术固有不可预测性"]
B["地缘政治多国博弈"]
C["商业竞争与填补真空需求"]
end
subgraph DefenseLayer["Anthropic 内部防御 (投入约50%精力)"]
D["发布前极度严苛测试"]
E["限制高风险领域能力 (如网络攻击)"]
F["快速迭代与防御机制演进"]
end
subgraph OutcomeGoal["最终愿景目标"]
G["显著压低系统性崩溃概率 (追求远高于同行的安全标准)"]
end
RiskSources --> DefenseLayer
DefenseLayer --> OutcomeGoal
```
### [01:07:32](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4052s) 为什么公众应该信任 Anthropic？用行动重建硅谷信誉
Emily Chang 直截了当地质疑：Anthropic 正在打造极其强大的技术，并有望从中攫取巨大的商业利益，公众凭什么信任你们？Amodei 的回答展现出罕见的务实与清醒。他认为，鉴于过去几年整个硅谷的所作所为，任何人在不了解他或 Anthropic 的情况下，“从不信任的立场出发”其实是非常理性的反应。硅谷已经透支了全世界的信任，如今必须一步步重新赢回这份信任。
Amodei 坚信，信任必须建立在实实在在的行动与牺牲之上，而 Anthropic 一直在用商业代价践行价值观：
- **暂扣前沿模型 Mythos**：出于对滥用风险的担忧，公司选择不公开发布这款极为强悍的模型，在商业变现层面蒙受了巨大损失。
- **主动切断中国市场访问**：在没有任何官方指令强制要求的情况下，Anthropic 自行切断了针对中国的模型访问接口；当时这一决定直接损失了数亿美元的收入，而在那时，数亿美元占公司总营收的比例极其可观。
- **推迟 Claude 2 发布**：为了彻底完成安全校准，公司曾顶着市场竞争压力多次延后核心产品上线。
Amodei 坦承公司绝非完美，组织总是充满各种运转失调和失误，但他希望公众纵观 Anthropic 的整体历史轨迹来做判断：最符合这些行为的假设，就是这群人确实怀揣着真诚的愿景，正在全力以赴做正确的事。
### [01:09:17](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4157s) 尾声：指数级彼端与 CEO 的“化妆”花絮
采访在充满未来感与轻松风趣的气氛中收尾。Emily Chang 总结道：“那么，我们就在指数级增长的彼端再见吧（We will see you on the other side of the exponential then）。”Amodei 微笑着回应：“但愿如此（Hopefully）。”
随后镜头切入一段幕后花絮，Emily Chang 打趣道：“你一直想成为好莱坞明星对吧？”Amodei 笑着感叹，出任科技公司 CEO 之后最让他始料未及的一件事，就是竟然有这么频繁的场合需要化妆，这完全不在他最初的人生规划（bingo card）里。Chang 笑着安慰他：“只是稍微扑一点粉而已。”两位在轻松的打趣声中结束了这场跨越技术哲学、商业博弈与人类命运的深度长谈。
## Notable quotes
> [05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) 当你觉得你无法信任某个人，当你感到他们的价值观与口头宣称的并不一致，当你感到他们不够诚实、投身于此的真实意图并非他们所标榜的那样，当你目睹了令人不安的不诚实行为模式时，你就会发现，继续与这家公司合作、继续去信任这家公司变得极其困难。 — Dario Amodei
> [05:51](https://www.youtube.com/watch?v=x2VHFgyawPE&t=351s) 我们将看到谁会在市场上胜出，谁会在公众信任的法庭上胜出。我们正在身体力行地展示如何以负责任的方式部署这项技术。如果他们不认同，大可公开拿出他们的论据，除此以外没什么好多说的。 — Dario Amodei
> [17:18](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1038s) 他们很清楚我们始终在积极与他们合作，我们也一直是非常优秀的合作伙伴。但我敢肯定，他们心里一定巴不得我们闭嘴别提这些事；可这就是我内心深处坚信的原则，你又能怎样呢？ — Dario Amodei
> [18:19](https://www.youtube.com/watch?v=x2VHFgyawPE&t=1099s) 我们都是成年人了，完全可以在一件事上紧密合作，同时在另一件事上保留不同意见。 — Dario Amodei
> [37:00](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2220s) 我们不希望看到一个中国和俄罗斯能用 AI 分析所有情报、发起进攻，而我们却无法自卫的世界。 — Dario Amodei
> [43:28](https://www.youtube.com/watch?v=x2VHFgyawPE&t=2608s) 设想一个不是由 Claude、而是由其他完全放开限制的竞品 AI 主导的世界：AI 模型自行裁决生杀大权，完全没有人类参与复核。这正是我们当初誓死抵抗的未来。 — Dario Amodei
> [50:33](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3033s) 在草堆里把那根针指给你看，与你顺手把针捡起来，完全是两码事。 — Dario Amodei
> [54:11](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3251s) Anthropic 的整个创业史就是一部妥协史。 — Dario Amodei
> [01:05:49](https://www.youtube.com/watch?v=x2VHFgyawPE&t=3949s) 假设市面上有一批航空公司，你想创办一家更安全的新航司。你的公司完全可以做到比所有同行安全 10 倍；但如果有人跑来问你：‘你能百分之百保证你的飞机永远不会坠毁吗？’你根本不可能给出这种保证。 — Dario Amodei
> [01:09:17](https://www.youtube.com/watch?v=x2VHFgyawPE&t=4157s) We will see you on the other side of the exponential then. — Emily Chang
## Mentioned
Dario Amodei, Emily Chang, Daniela Amodei, Holden Karnofsky, Sam Altman, Demis Hassabis, Elon Musk, Jensen Huang, Narendra Modi, Scott Bessent, Susie Wiles, Leo Szilard, J. Robert Oppenheimer, Anthropic, OpenAI, Google DeepMind, Palantir, Amazon, Microsoft, NVIDIA, Baidu, Open Philanthropy, 五角大楼 (Pentagon), 美国国防部 (DoD), 美国移民及海关执法局 (ICE), 美国海关和边境保护局 (CBP), Claude, Claude Code, Claude Cowork, Claude 2, Opus 4.7, Mythos, ChatGPT, Gemini, AlphaFold, Maven Smart System, Firefox, 长期利益信托 (Long-Term Benefit Trust, LTBT), 《技术的青春期》(The Adolescence of Technology), 《奇爱博士》(Dr. Strangelove), 《1984》
## Video description
Emily Chang sits down with Anthropic CEO Dario Amodei in a wide-ranging interview discussing his San Francisco roots, the race against OpenAI, the Pentagon standoff and AI’s endgame.
Watch Inside Anthropic, the $965 Billion AI Juggernaut: https://youtu.be/v1wZwxY3CMg
00:00 Inside Anthropic
03:34 Dario background
05:51 Leaving OpenAI
07:42 India AI summit
10:45 Enterprise bet
17:22: Big tech funding
19:29 Compute crunch
21:15 Surpassing OpenAI
24:07 Product velocity
24:52 AI discoveries
26:13 Dario’s writing style
28:10 AI and the workforce
36:41 Pentagon standoff
43:29 AI warfare
48:18 Mythos
55:15 Nationalizing AI
58:57 Visit to the White House
59:47 China
1:03:24 Recursive self-improvement
1:05:07 Dario’s favorite book
1:05:49 Civilization collapse
1:07:32 Trust
Watch more of The Circuit with Emily Chang: https://www.youtube.com/show/VLPLqq4LnWs3olV2nw7GgFjXHvxVc3T5hXXT?sbp=QAE%3D
Hosted by high-profile journalist Emily Chang, The Circuit is a fast-paced, dynamic series that lives at the intersection of culture, tech, entertainment, and business. Every week, Chang will go on location to meet the world’s most fascinating founders, influencers, and innovators, conducting intimate interviews and bringing audiences behind the scenes of the most impactful stories, launches, and trends.
#anthropic #AI #DarioAmodei #tech #business
--------
Like this video? Subscribe: http://www.youtube.com/Bloomberg?sub_confirmation=1
Get unlimited access to Bloomberg.com for just $1.99 your first month: https://www.bloomberg.com/subscriptions?in_source=YoutubeOriginals
Bloomberg Originals offers bold takes for curious minds on today’s biggest topics. Hosted by experts covering stories you haven’t seen and viewpoints you haven’t heard, you’ll discover cinematic, data-led shows that investigate the intersection of business and culture. Exploring every angle of climate change, technology, finance, sports and beyond, Bloomberg Originals is business as you’ve never seen it.
Subscribe for business news, but not as you've known it: exclusive interviews, fascinating profiles, data-driven analysis, and the latest in tech innovation from around the world.
Visit our partner channel Bloomberg News for global news and insight in an instant.
[Source](https://www.youtube.com/watch?v=x2VHFgyawPE)
Source quotes
[5:51] Dario Amodei: 当你觉得你无法信任某个人，当你感到他们的价值观与口头宣称的并不一致，当你感到他们不够诚实、投身于此的真实意图并非他们所标榜的那样，当你目睹了令人不安的不诚实行为模式时，你就会发现，继续与这家公司合作、继续去信任这家公司变得极其困难。
[5:51] Dario Amodei: 我们将看到谁会在市场上胜出，谁会在公众信任的法庭上胜出。我们正在身体力行地展示如何以负责任的方式部署这项技术。如果他们不认同，大可公开拿出他们的论据，除此以外没什么好多说的。
[17:18] Dario Amodei: 他们很清楚我们始终在积极与他们合作，我们也一直是非常优秀的合作伙伴。但我敢肯定，他们心里一定巴不得我们闭嘴别提这些事；可这就是我内心深处坚信的原则，你又能怎样呢？
[18:19] Dario Amodei: 我们都是成年人了，完全可以在一件事上紧密合作，同时在另一件事上保留不同意见。
[37:00] Dario Amodei: 我们不希望看到一个中国和俄罗斯能用 AI 分析所有情报、发起进攻，而我们却无法自卫的世界。
[43:28] Dario Amodei: 设想一个不是由 Claude、而是由其他完全放开限制的竞品 AI 主导的世界：AI 模型自行裁决生杀大权，完全没有人类参与复核。这正是我们当初誓死抵抗的未来。
[50:33] Dario Amodei: 在草堆里把那根针指给你看，与你顺手把针捡起来，完全是两码事。
[54:11] Dario Amodei: Anthropic 的整个创业史就是一部妥协史。
[65:49] Dario Amodei: 假设市面上有一批航空公司，你想创办一家更安全的新航司。你的公司完全可以做到比所有同行安全 10 倍；但如果有人跑来问你：‘你能百分之百保证你的飞机永远不会坠毁吗？’你根本不可能给出这种保证。
[69:17] Emily Chang: We will see you on the other side of the exponential then.
Source quotes
[10:45] Dario Amodei: If you pick a business model that fundamentally conflicts with your values, you're gonna have a hard time... Either you betray your own values or you become irrelevant.
[10:45] Dario Amodei: We want to use AI to, you know, cure diseases that we couldn't cure before... working with biotech, it's working with pharma... All of those are enterprises.
[15:11] Dario Amodei: If a enterprise's sole defense is having written code that no one else can replicate, good luck, you're not gonna be able to defend that.
[17:18] Dario Amodei: Look, we're all adults here.
[24:07] Dario Amodei: Think of all the problems we had in 1900s... all the material deprivation that we don't have to deal with today. Then think of another hundred years of that. I really believe this century of scientific and medical progress... we're gonna have a much, much better world.
[34:38] Dario Amodei: The idea that this is cheap marketing is itself cheap marketing. This is laziness; this is failure to engage with serious intellectual work... It's so stupid. It's so unserious. And whenever someone says something like that, I take them less seriously.
[54:15] 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.
[67:32] Dario Amodei: starting from a position of distrust... is pretty rational.
Original transcript
[0:07] How much are you sleeping?
[0:10] You know, I've never been
someone who slept all that.
[0:12] Well, let's just say I'm, you know, I'm,
[0:14] I'm learning the art of, of,
you know, finding ways to relax
[0:17] and sleep through, through
moments of unusual pressure.
[0:21] It is all moving so fast.
How does it feel on the inside?
[0:25] It's this feeling of like the
exponential, like, you know,
[0:28] suppose you were to accelerate
away from earth on a
[0:31] spaceship at relativistic speed.
[0:33] The way special relativity
works is, you know, you go,
[0:35] you go to sleep and you wake up in two
[0:37] days have gone by on earth.
[0:38] And so you have to deal
with two days in one day.
[0:40] And then you go to sleep.
[0:42] And then because you've
continued to accelerate,
[0:44] three days have gone by on
earth, and then the next day
[0:46] and four days have gone by.
[0:48] And that's a little
bit what it feels like.
[0:49] I mean, do you go to bed
constantly paranoid about
[0:51] what you'll wake up to?
[0:52] There are enough clear
[0:54] and present issues that we
have to deal with that I'm,
[0:57] I'm constantly dealing with
those, while thinking about
[0:59] how we can prepare.
[1:00] But, but you know, I, you
know, I, I don't think paranoia
[1:03] or worrying about what you'll
wake up to is productive.
[1:06] You know, I've looked at people
in history who've, you know,
[1:08] who've dealt with these very
high pressure situations
[1:10] and, you know, you need to
learn to respond rationally and,
[1:14] and not put dangers out of
proportion to each other.
[1:17] This yo-yoing between, I'm not worried
[1:19] and oh my God, we need to panic today.
[1:22] I, I I, I, I think that's a hallmark
of immature decision making. And
the actual mature decision making
[1:27] is we can't ignore this.
[1:28] We can't be complacent.
[1:30] In fact, it's getting to be
a bigger and bigger risk.
[1:32] But, you know, we, we have
to respond rationally.
[1:35] You know, like a, like
a surgeon would deal
[1:37] with an operation
[1:38] or, you know, like a military
officer would, you know, deal
[1:41] with a military operation
[1:43] or, you know, someone making
decisions that affect a lot
[1:46] of people has to make
those decisions rationally
[1:49] and they have to understand the risk,
[1:50] but they, they, you know, they have
[1:51] to maintain a basic sense of calm.
[1:56] So my son yesterday was like,
[1:59] can I use your Claude Cowork account?
[2:00] And I was like, absolutely
not. I need my tokens.
[2:05] We're seeing more and more of them,
[2:06] even in the consumer space.
[2:07] We wanted to be more of
an enterprise company.
[2:09] But, you know, it's even,
[2:10] even even consumer without us putting
[2:12] that much effort is starting to go fast.
[2:15] You are at the center of
the AI universe right now.
[2:18] What does that feel like?
[2:20] The interesting thing is, is
that the experience I've had
[2:24] for my whole career
[2:25] and certainly the whole
time at Anthropic is
[2:28] that there's this kind
of smooth exponential
[2:30] and the experience, the
smooth exponential is,
[2:33] nothing's happening, nothing's
[2:34] happening, nothing's happening.
[2:36] A little things happen and
then zoom, it goes crazy.
[2:39] That's the experience of the world.
[2:41] That's the experience of the
scale of the company compared
[2:44] to the other companies
and compared to the world.
[2:47] So, you know, I was watching
this graph for a while
[2:49] and I said, oh yeah, we'll
probably become the AI company
[2:53] with, you know, the, the most revenue
[2:54] and the most valuation some
sometime around this time.
[2:57] And, and indeed, indeed it has happened.
[3:01] So in one sense, I'm not surprised
[3:03] 'cause there's just a
smooth line on the graph.
[3:05] But of course in another sense,
when things actually happen,
[3:08] you just, you see so much
more, you know, detail
[3:11] and color to it and it, you know,
[3:13] it definitely is surprising.
[3:14] And what we're just keeping in mind,
[3:17] all the things we usually
keep in mind, which are,
[3:19] which are just, you know,
how do we train good models?
[3:21] How do we put them in good products?
[3:22] How do we make sure
that everything's safe?
[3:24] How do we help people
[3:26] but also manage the societal
risks around the technology?
[3:28] It's, it's all the same
questions, just kind
[3:31] of under a bigger, under a
bigger microscope as it were.
[3:34] What were you like as a kid
growing up in San Francisco?
[3:36] I know your, your dad
was a leather craftsman.
[3:38] Your mom worked in libraries.
How did that shape you?
[3:41] You know, the whole, you
know, like first, you know,
[3:44] internet revolution
was happening around me
[3:46] and I had absolutely no interest in it.
[3:48] I was just interested in like, doing math
[3:50] and like scra, you know,
[3:51] scrawling things, I was interested in,
[3:53] like understanding the universe.
[3:55] I was interested in science fiction.
[3:57] Like that was, that was kind
of the, you know, that was the,
[3:59] that was the general, that
was the general milieu.
[4:02] I, I think I just felt a lot
of curiosity about the world.
[4:05] You grew up in the town where, you know,
[4:07] that is the center of technology
[4:09] and right now it's the
center of AI you know?
[4:11] Is there anything about
this place, this city here
[4:14] that informed your worldview?
[4:16] Yeah, I mean, I think
the general, you know,
[4:18] the general spirit of kind
of, you know, nonconformism
[4:22] and individualism and
it's okay to be crazy.
[4:25] I think, I think a good
deal of that probably did,
[4:27] probably did probably
did rub off against me.
[4:30] You know, you hear these
stories about, you know, you go
[4:33] to countries in Europe
[4:34] or you know, even other parts
of this country where it's,
[4:36] it's just, you know, it's
just kind of discouraged
[4:39] or considered weird to like,
[4:41] think about things in
some different way, right?
[4:43] Or have some set of, of
some set of crazy ideas.
[4:46] And, you know, there's a lot
of things I'm actually very
[4:48] critical about with, with Silicon Valley.
[4:50] But one thing that I think
is good about it is this,
[4:52] this encouragement of like, you know,
[4:54] it doesn't matter if all
the experts are against you,
[4:57] it doesn't matter, you know,
if you have a coherent
[4:59] vision and a coherent
world view of the world,
[5:02] you should go and pursue it.
[5:03] Maybe it just won't work at all.
[5:05] But, but if it does, there's
this kind of long tailed ness
[5:08] to it where, you know, there
are certain places you can,
[5:11] you know, you can, you can
search certain veins of war
[5:13] where, you know, you
might, you might find a
[5:15] huge goldmine there.
[5:16] I think that spirit is very important.
[5:18] You, Daniela, your sister
[5:20] and her husband, Holden
Karnofsky lived in a group house
[5:23] together back in 2016.
[5:25] What were you debating back then?
[5:27] That was, I think the
time when, you know,
[5:28] Open Philanthropy project
was, was, you know,
[5:31] first being startup, which
Holden was the lead of.
[5:33] And I was at that time, you know, like a,
[5:35] a biological scientist.
[5:37] So, you know, I was helping them with some
[5:39] of the stuff they were doing around kind
[5:40] of developing world health
or biological research.
[5:43] So, you know, I I kind
of advised on that stuff
[5:45] and, you know, what were the
areas that were promising?
[5:47] What were the areas
that were less promising
[5:49] Your decision to leave Open
[5:51] AI has become Silicon Valley lore.
[5:54] What really happened?
[5:56] Like, beyond the narrative,
what were the issues?
[5:59] What did you disagree on?
[6:01] Look, I'll, I'm gonna say it,
[6:03] I'm gonna say it very simply.
[6:05] You know, there are many
difficult issues that you know,
[6:10] you face when you're
building powerful technology
[6:12] that Anthropic faces every day
[6:13] where we don't know whether
we're making the right
[6:15] decision or the wrong decision.
[6:18] So, you know, there are
many valid disagreements
[6:20] to be had on safety.
[6:22] We certainly had some of
those disagreements with them,
[6:24] but, you know, people that, that, that,
[6:26] that alone is not sufficient to leave.
[6:29] People here have had
disagreements with me,
[6:31] people here have
disagreements with each other.
[6:33] But when you feel that
you can't trust someone,
[6:36] when you feel that their values are not
[6:39] what they say they are, when you feel
[6:41] that they're not honest, when you feel
[6:43] that they're not in it for
the reasons that they say,
[6:47] when you see disturbing patterns
[6:52] of behavior dishonesty,
that makes it very hard to,
[6:56] you know, to continue
to work with a company,
[6:59] to continue to trust the company.
[7:00] And look, at the end of the day, why argue
[7:04] with someone when you don't have the same
[7:06] vision and you don't trust them.
[7:07] Like the way the the way
to resolve it is you go off
[7:10] and do your thing.
[7:12] They go off and do their thing.
[7:14] And I am completely at peace with the idea
[7:18] that we're doing things our way
[7:19] and they're doing things their way.
[7:21] We'll see who wins in the market
[7:23] and we'll see who wins in
the court of public opinion.
[7:26] I think those things speak
louder than any drama about why,
[7:30] who left, what, you know, we're,
we're providing an example
[7:34] of how to deploy this
technology, you know, in
[7:37] what we think is a responsible way.
[7:39] If they disagree, they
should make that argument.
[7:41] And you know that, I think
[7:43] that's really all there
is to say about it.
[7:44] There was a moment at
India's AI summit where you
[7:47] and Sam Altman appeared to refuse
[7:50] to hold hands on stage.
What happened there?
[7:53] What happened is that the summit was
[7:54] extremely disorganized.
[7:56] We all came up at the last minute
[7:58] and they like changed the order
in which we were standing,
[8:00] and then like, they took a picture of us
[8:02] and then they ordered us
all to like, hold hands.
[8:05] You know, if you've ever
been to one of these summit,
[8:07] I am not saying anything bad
about India in particular,
[8:09] but like all of these kind
of international type summits
[8:12] that have like heads of state
are like super disorganized.
[8:16] Okay? But everyone
else held hands. Come on.
[8:18] I, I look, I don't know, I
don't know what to tell you.
[8:21] Okay. There was like, you know, Narendra
[8:23] Modi up there suddenly
telling everyone to, like
[8:27] suddenly telling everyone to hold hands.
[8:29] All right, all right,
well, okay, look, Sam
[8:32] and Elon are suing each other.
[8:33] You don't like Sam. I,
[8:35] it seems if the people
building the most important
[8:38] technology in the world
can't hold hands on stage,
[8:41] how can we trust you'll
cooperate on existential risk?
[8:44] So here's, here's what I will tell you.
[8:47] There is a wide variance in the quality
[8:50] and the trustworthiness of the people
[8:51] building this technology.
[8:53] I think this meme that,
you know, different,
[8:55] that no one trusts each other,
I don't think it's right.
[8:57] You know, I've known Demis Hassabis
[8:59] who builds the Gemini models. They're
[9:01] a competitor to Claude models.
[9:02] I've known him for 15 years.
[9:04] We've worked together on like,
you know, a number of issues.
[9:07] We buy compute from Google,
[9:09] we swap safety ideas all the time.
[9:12] So, you know, my my view
of this is that one,
[9:17] there are some players who are
more trustworthy than others.
[9:20] And you know, I think there are players
[9:21] outside Anthropic who,
you know, who, who I trust,
[9:25] who I see as trustworthy.
[9:27] What I think needs to happen is
[9:28] that the trustworthy actors
need to need to get together
[9:32] and, and put the untrustworthy
actors in a position
[9:36] where they kind of have to
adopt the same standards.
[9:39] With a lot of experience, I've learned
[9:41] that there are some folks who
don't do the right thing on
[9:43] their own, but if there's
a majority of the industry
[9:45] that's doing the right
thing, then I think the rest
[9:48] of the industry is, is kind
of, they're left in a position
[9:51] where there's not much
they can do that, that,
[9:53] that then come along.
[9:54] There's like the positive version of it
[9:55] where you inspire other
people, that's like Demis
[9:58] and me inspiring each other.
[9:59] You know, he does Alpha Fold.
[10:00] We're trying to do something
in bio as well, right?
[10:03] We do interpretability research,
[10:04] they start an interpretability research.
[10:06] It's not even competition,
it's just, it's just, you know,
[10:09] each company does something cool
[10:10] and the other company's like, that's cool.
[10:12] We'd like to, you know, do that too
[10:13] and see if there's something
new within that we can do.
[10:16] So that's the kind of,
you know, the carrot side
[10:19] of the race to the top.
[10:20] Then there's the stick side
[10:21] or the implicit stick
where you're like, okay,
[10:23] these guys are doing the right thing.
[10:25] Those guys will look bad if
they don't do the right thing.
[10:27] And often we see behaviors where they kind
[10:29] of grudgingly do the
right thing while trying
[10:31] to pretend they're doing
something different.
[10:33] And there's something
bad or sinister about us
[10:35] that is to be expected.
[10:36] But I think that's the way
we get the industry together,
[10:39] and that's the way we get
the industry to cooperate.
[10:41] Now, early on, others focused on fun,
[10:43] splashy consumer apps.
[10:45] You made a bet on coding and enterprise
[10:48] and Claude Code is a hit.
[10:49] Claude Cowork is a hit.
Why did you make that bet?
[10:53] Was it a values decision
or a business decision?
[10:55] When we started Anthropic,
the thing that the,
[10:58] the base thing that mattered, the thing
[10:59] that always matters is we wanna,
[11:01] we want to do this, right?
[11:02] But then you have to ask
yourself, okay, in order
[11:05] to fund the very expensive,
you know, creation
[11:08] of these models, it, it
needs to be a company,
[11:11] it needs to have a business model.
[11:12] Does the business model get
in the way of the values?
[11:14] There's always this question,
[11:16] but I think one of the things
I learned is, you know,
[11:19] just from being at other companies
[11:21] and watching other companies is, look,
[11:23] if you pick a business model
that fundamentally conflicts
[11:26] with your values, you're
gonna have a hard time, right?
[11:29] Either you betray your own
values or you become irrelevant.
[11:34] You know, you kind of end
up in a catch 22 situation.
[11:36] And there are ways out,
there are ways to dodge,
[11:39] but it's just, it's just a hard situation.
[11:41] It's far better to pick a business model
[11:44] that is compatible with your values.
[11:46] And so when we thought about
it, we said, look, you know,
[11:49] we've seen the world of social
media, the consumer world,
[11:52] it, it really seems to, you
know, encourage engagement,
[11:56] even addiction, you
know, the slop we've seen
[11:59] with AI video models, it's
like, what's going on?
[12:01] It wants to maximize the
number of minutes that you're,
[12:04] you're paying attention to,
[12:05] because that's the advertising
revenue driven incentive.
[12:09] Whereas if we look at enterprise,
look, I mean, you know,
[12:13] we want to make these
models useful to people.
[12:16] If I think of all the positive things you
[12:18] can do with AI, right?
[12:19] I warn a lot about the negative things,
[12:21] but ultimately we think
the positive things will
[12:23] outweigh the negative things.
[12:24] Many of those are basically fall under the
[12:27] banner of enterprise.
[12:28] You know, we want to use AI
to, you know, cure diseases
[12:33] that we couldn't cure before, right?
[12:34] Well, that's working with
biotech, it's working with pharma,
[12:37] it's working with
academic research groups.
[12:39] All of those are enterprises, right?
[12:41] We want to use AI to like, you know,
[12:43] to make energy cheaper and more efficient.
[12:46] That's, that's all enterprise.
[12:47] You know, we want to use
AI to help with education.
[12:50] Most of that is enterprise.
[12:51] You know, we want to use AI
to, you know, to address,
[12:55] you know, health and developing
[12:56] world. Well there are nonprofits.
[12:57] But those are basically enterprises.
[12:59] We want to increase economic growth.
[13:01] That that is basically enterprise as well.
[13:04] And then I think there's
another factor, which is
[13:06] that enterprises care a lot about trust in
[13:08] long-term relationship, right?
[13:10] Consumer can have this, you know,
[13:12] almost this gimmicky aspect to it, right?
[13:15] Where with enterprise, it's like,
[13:16] what matters is you build a
relationship where, you know,
[13:19] you work with, you work with
a company for many years,
[13:21] you know, you deliver on
what you say they deliver on
[13:24] what they say, and they
basically trust you.
[13:26] And so it's very synergistic
with our goal of, you know,
[13:30] deploying these models in
a positive and safe way.
[13:33] And so I think it serves us
well to have this business model
[13:36] that largely aligns with our values.
[13:38] Not that there aren't conflicts sometimes.
[13:40] Not that there aren't hard
choices we have to make,
[13:43] but I think the number of such choices,
[13:44] it's much lower than
it would be otherwise.
[13:47] A developer can switch
from Claude to ChatGPT
[13:49] or Gemini in an afternoon.
[13:51] Is it really possible
[13:52] to have a long-term lead in this industry?
[13:56] And, you know, how long would
it take a serious competitor
[13:59] to replicate what you've built?
[14:01] Model quality is the
most important thing.
[14:03] Like we're, we're very far ahead
right now on model quality.
[14:07] There is some amount of inertia,
[14:08] but I've never relied on that, right?
[14:10] I've never relied on
like the, you know, the,
[14:12] Anthropic has never
relied on like, oh,
[14:14] this is sticky and people won't switch.
[14:16] I think you wanna have a better model.
[14:18] You want to have a better product.
[14:19] And you know, we, we see the growth rates
[14:22] haven't inflected at all.
[14:23] If anything, they've gone up
at least, at least at the time
[14:25] of taping this interview.
[14:28] So, you know, I think I, I I tend to think
[14:30] that is the most important thing.
[14:32] Soon after Claude Cowork was released,
[14:34] $285 billion in market
value vanished overnight,
[14:38] traders called it the SaaSpocalypse.
[14:41] If AI continues improving
at this pace, how much
[14:44] of traditional software
gets replaced and how fast?
[14:48] Yeah. So, you know, this is, this is one
[14:50] of these questions, it's kind of very hard
[14:52] to predict in advance, right?
[14:53] If you could predict it
perfectly in advance,
[14:55] then people would,
[14:56] and they'd make a huge
amount of money on the market
[14:57] and they'd always be right.
[14:59] So, you know, no one,
[15:00] no one knows exactly
what's going to happen.
[15:03] But I would note a few things, right?
[15:05] All of these traditional
software companies have
[15:06] a number of moats.
[15:08] I think what's gonna happen is some
[15:10] of these moats are gonna go away,
[15:11] but others are going
to stay around, right?
[15:13] The ability to quickly write software, I,
[15:16] I definitely think
that's going away, right?
[15:18] If, if your moat is, we
wrote this complex software
[15:21] that no one else can
write, like good luck,
[15:23] you're not gonna be able to defend that.
[15:25] But I think folks have
customer relationships.
[15:29] Folks have know how of how, you know, of
[15:32] how the field works.
[15:33] Folks have unique domain knowledge.
[15:37] So I think my advice to all of
these folks is, obviously,
[15:41] you know, don't be
complacent, don't ignore it.
[15:44] Make a list of all your moats
[15:46] and be very aware that some
of them are going to go away,
[15:49] while others are going to become
relatively more important.
[15:52] Because there are limiting factors
[15:53] and there may also, there
may also be new moats.
[15:56] And I think those that deftly
respond, that, you know,
[16:00] lean into the list of moats
that are still present as well
[16:03] as the new ones will do well.
[16:04] I think those that are complacent,
that kind of, you know,
[16:08] just delude themselves at
what worked in the past will,
[16:11] will continue to work.
[16:12] They're, they're not gonna have a good
[16:14] time.
[16:15] So that is, that is
the advice I would give.
[16:17] And you know, I I, I, I
think at the end of the day,
[16:19] I would guess, I mean, it
depends what you call SaaS
[16:22] and what you don't call SaaS,
[16:23] but like, I would guess
[16:25] that the software industry
gets larger, not smaller.
[16:27] Although there, there
will be some big losers.
[16:29] Explain that.
I just think the pie
[16:32] is getting bigger, right?
[16:33] Like, I think, I think with ai,
[16:35] like the pie is getting bigger,
[16:37] the existing incumbents may
be smaller in relative terms.
[16:41] Some of them may, may
go down in value, some
[16:43] of them may even may even go out
[16:45] of business if they don't
adapt in the right way.
[16:47] But I, I, I, you know, I think you,
[16:48] I think you see this often when growth is
[16:50] really fast, right?
[16:51] If the, you know, if the,
if, if, if what's possible
[16:54] with AI grows by 10x, it's very easy
[16:57] for an existing incumbent
industry to go up by 1.5x, right?
[17:01] Just, just, you know, not as much
[17:03] as the whole big pie is growing.
[17:05] So I think that may happen that that's not
[17:07] to say we won't have such some big losers.
[17:09] I think those who don't adapt,
[17:11] who put their heads in
the sand, who don't kind
[17:14] of see what's coming, who
don't identify the moats they
[17:16] have, they're gonna
have a really hard time.
[17:18] Your biggest backers are
companies like Amazon and Google
[17:21] and Microsoft and Nvidia.
[17:24] These are companies that
all have their own agendas.
[17:26] They are partners in rivals.
[17:28] You have huge commercial
milestones tied to funding,
[17:33] who's really calling the shots.
[17:34] There have been a number of cases
[17:36] where we've really spoken our
minds about what we think,
[17:39] you know, I've been very
outspoken about the need
[17:42] for export controls on
ships to China, right?
[17:45] I, i, I say this
[17:47] because I think it would be
really bad for America, for,
[17:50] you know, the state of democracy
in the world for, you know,
[17:53] China to be ahead in AI capabilities.
[17:55] And, you know, it's, it's like some
[17:57] of the chip makers obviously
don't agree with that view,
[18:01] but it hasn't stopped me from saying it.
[18:02] I'm saying it again now, even
[18:04] after we've signed more
partnerships, what they know is
[18:07] that we always work with them.
[18:09] We've been good partners, you
know, we can work together.
[18:12] I'm sure they wish we
didn't say these things,
[18:15] but these things are
what I believe.
[18:17] What are you gonna do? You know,
they're, they're at the end
[18:19] of the day, they want the, you know, they,
[18:22] they benefit from these
deals as much as we do.
[18:24] You know, look, we're all adults here.
[18:25] We can work together on one thing while
[18:27] disagreeing about another thing.
[18:28] Bloomberg's reported
that you're at valuations
[18:31] that are higher than OpenAI.
[18:33] We're talking nearly a trillion dollars
[18:35] for a five-year-old startup.
[18:36] How do you make sense of that number
[18:40] and why do you need that
much money if you, you know,
[18:41] you're more disciplined on compute,
[18:43] you have a faster path to profit,
[18:45] The compute is ramping
up very quickly, right?
[18:48] So it, it can both be the
case that the fundamentals
[18:52] of the business look good,
[18:53] but in, you know, in a year
you'll have three times
[18:56] as much compute as, you know,
three times or four times,
[18:59] or, I'm not gonna give exact numbers,
[19:01] but like these compute
ramps are very fast.
[19:04] And we have every expectation
that the revenue, you know,
[19:07] ramp will meet and exceed those.
[19:09] But raising money is, is kind
[19:11] of the buffer against
this cone of uncertainty.
[19:14] So it's a totally rational thing to do.
[19:16] It's, it's a very small
dilution to the business,
[19:19] and it logically is not
at all the same thing.
[19:23] In fact, it's compatible with
the opposite as you know,
[19:25] that there's anything wrong
[19:26] with the fundamentals of the business.
[19:28] There have been
reports of server strain,
[19:30] reliability issues,
[19:31] people complaining about
running out of tokens.
[19:34] You've said other companies are
yolo-ing on infrastructure.
[19:37] Do you actually have what you need
[19:39] or are you playing catch up?
[19:40] So one of these things
about compute is there's a
[19:42] marketing compute, right?
[19:44] So, you know, my view is
that over a period of time,
[19:47] even longer than a couple
months, like, you know,
[19:50] we can get large amounts of compute.
[19:53] One, one thing that's worth
saying here is, you know,
[19:56] I don't think we bought too little compute
[19:58] by any reasonable standard.
[20:00] So, you know, we were planning
[20:02] for a 10X a year growth in compute,
[20:05] 10X a year is what we expect.
[20:07] That isn't what we've seen
over the first quarter of 2026.
[20:12] We saw a greater than three x
growth in revenue quarterly,
[20:17] just in not annualized three
x in the quarter, which
[20:20] of course, three to the
fourth power is 80x
[20:23] over the course of the year.
[20:25] We didn't plan for 80x annualized growth.
[20:29] It would not have been rational to plan
[20:31] for 80x annualized growth,
[20:33] because that means if you only
get 10x, you know that you,
[20:36] you have eight times less.
[20:37] So we're, we're, we're in a
locally extreme, you know,
[20:41] explosion of compute
that's not gonna continue.
[20:44] If that continued, you know,
you just get to revenue
[20:46] by the end of the year,
you get to revenue numbers
[20:48] that no company on earth, I
don't think that's gonna happen.
[20:51] It just, it just can't.
[20:53] But you can have these short
periods where it's like,
[20:55] oh my God, like, you know,
[20:56] this is faster growth than we ever,
[20:58] ever possibly anticipated.
[21:00] But I don't know, you saw the
compute deals with Google,
[21:02] you saw the compute deals with Amazon.
[21:04] You know, there are
more that we kind of can
[21:07] and will do, like, you
know, the market's liquid.
[21:09] Like if, if, if, you know, if you're able
[21:11] to use compute really well
[21:12] and there's the demand,
you'll get your compute.
[21:14] It might might just take a month or two.
[21:16] Does it feel good to
surpass your arch rival?
[21:19] Look, I, we have a lot of
difficult challenges in front
[21:24] of us, and there's this
race to the top idea
[21:26] that we're trying to pull
other companies along with us.
[21:29] And I think we've seen that.
[21:30] We have pulled them along with us.
[21:32] Sometimes they don't admit
that that's what they're doing.
[21:34] Sometimes they copy us
while they're attacking us.
[21:37] But, but this pull is very valuable.
[21:40] And so I think the value of
being the preeminent company,
[21:44] both commercially and in terms
of models, you know, it's,
[21:47] it's not about beating rivals for the sake
[21:49] of beating rivals.
[21:51] It's, it's about having the ability
[21:53] to pull the ecosystem along with us.
[21:55] And we hope that we can do
more of that in the future.
[21:58] But winning has to feel
just a little bit good.
[22:00] I mean, look, we're always
trying to succeed, right?
[22:03] Like, we're always trying
to, you know, we're not,
[22:05] we're not trying to fail here, right?
[22:07] Like, I'm not someone who
believes we should shut
[22:09] this technology down.
[22:10] We shouldn't build it like,
you know, we, we, we, we,
[22:14] you know, we, we exist within
a free enterprise system and,
[22:17] and, you know, there's, there's nothing,
[22:20] there's nothing wrong with this.
[22:22] We just have to mitigate the
risks of the models, right?
[22:24] And, and so it's always been
the balance between the two.
[22:28] Now for most of Anthropic's
history, you were the underdog.
[22:32] I imagine it's easier
[22:33] to take the moral high
ground when you have nothing
[22:36] to lose at this scale.
[22:38] How hard is it to stay
true to your values?
[22:41] What I would say is that,
you know, I've put a lot
[22:43] of time into thinking about
how, how that's the case.
[22:48] You know, as, as, as companies scale.
[22:51] You know, I've been
paranoid at every scale,
[22:54] at every scale of the company.
[22:55] There's some new challenge,
[22:56] there's some new way the company can lose.
[22:59] Either its, its kind of will
to win just commercially
[23:04] or kind of the core of its values.
[23:06] I'm, I'm worried about both
[23:07] because I see them as synergistic.
[23:09] I actually see the fact
that we've been able
[23:11] to make such good models as
the thing that, that allows us
[23:15] to assert our values in a way that works
[23:17] as the company grows, as it gets bigger.
[23:19] There are lots of pitfalls here.
[23:21] There are lots of ways to go wrong.
[23:23] Not because me or the co-founders
[23:25] or the company's leaders
values change, but
[23:27] because the composition of
the company changes very fast.
[23:30] So I spend probably half
of my time just talking
[23:34] to the company about
the culture of anthropic
[23:36] and how the culture works, right?
[23:38] When you're growing this
fast, you're hiring a bunch
[23:40] of people from big tech companies.
[23:42] If you don't tell them
how Anthropic operates,
[23:45] they'll simply recapitulate
the only thing they know,
[23:47] which is how to operate at the
[23:48] companies that they came from.
[23:50] And so this is a constant
struggle and a constant challenge.
[23:53] And, you know, it's like, you know, me
[23:55] and Daniela's, maybe
number one top priority is,
[23:58] is figuring out how to preserve this
[24:00] because we recognize
that this is the core of
[24:02] who we are in the long run.
[24:06] Your product velocity is insane.
[24:08] You're shipping so much so
fast. How are you doing it?
[24:11] I would say two things.
[24:12] The first is, you know,
we have a unified company.
[24:14] We have a unified culture.
[24:16] You know, I think we've gotten, you know,
[24:18] grown larger while still
being incredibly efficient.
[24:21] Everyone's still being on the same page,
[24:23] like just the cultural
and organizational unity.
[24:26] I would say that's the biggest factor.
[24:28] And I would say the second
biggest factor is Claude itself.
[24:30] That we're now using
Claude to help, you know,
[24:33] develop our models and, you
know, make them more efficient
[24:36] and quickly develop products.
[24:38] There's all kinds of new
practices you have to develop.
[24:40] You know, we're, we're still new, new,
[24:42] we're still new at it,
[24:43] but you know, it's producing,
it's producing a lot
[24:45] of acceleration and increasingly producing
[24:48] reliable acceleration.
[24:49] And so those are the two
factors I would point to.
[24:51] Will you tell me the most
wild thing you've seen AI do?
[24:55] I think some of the
wildest stuff I've seen is
[24:57] around biology and medicine.
[24:59] I've seen a number of cases,
including Daniela actually,
[25:01] where Claude diagnosed a
medical problem that, you know,
[25:06] a bunch of fancy doctors had missed.
[25:08] And on the biology side,
like the models are starting
[25:10] to get surprisingly good at
like, you know, you know,
[25:14] tasks like drug design or, you
know, computational chemistry
[25:17] or things like, and I'm just
like, wow, you know, as someone
[25:20] who used to be a biologist, I look at it
[25:21] and I'm like, wow, that's hard.
[25:23] Like, you need a lot
of training to do that.
[25:25] And like Claude is getting good at it.
[25:27] And that's one area where
I think we're gonna get a
[25:29] hell of a lot of benefit.
[25:30] Like that's the positive for AI.
[25:32] We're gonna get these
huge, enormous benefits.
[25:35] Life is gonna get better.
[25:37] The quality of human
experience is gonna get better.
[25:40] A century of scientific progress,
[25:42] A a century of scientific progress
[25:43] and a century of progress.
[25:45] And what it's like to be human.
[25:46] Like a go back to 1900, think
[25:48] of all the problems we had in 1900s.
[25:51] All the reasons people died prematurely,
[25:54] all the problems they had to suffer,
[25:56] all the material deprivation
[25:58] that we don't have to deal with today.
[26:00] Then think of another
hundred years of that.
[26:02] I really believe this
century of scientific
[26:05] and medical progress, if
we can get through this,
[26:08] and I, I think we will.
[26:09] I'm increasingly optimistic.
[26:11] We're gonna have a
much, much better world.
[26:13] I know how much you love writing.
[26:14] You're known for your essays. Do you
[26:15] use Claude to help write?
[26:17] I do. I have not gotten to the point
[26:19] where I actually allow text
directly written by Claude in,
[26:22] in, because I, i, i, I just
have such a specific style
[26:25] that I'm, I'm a little picky about it,
[26:27] but I basically use
Claude to like, you know,
[26:30] to help me brainstorm, to help
me think through the themes
[26:34] to help me kind of, oh, you know,
[26:36] what are some references
I could use for this?
[26:39] So it it, it kind of
plays a supportive role.
[26:41] I don't know how far we
are from Claude being
[26:43] able to write better than me.
[26:44] We're not quite there yet.
[26:46] But, but you know, I think, I
think certainly it's coming.
[26:48] I love writing too, and I,
you know, I feel like writing
[26:51] it helps you struggle through ideas.
[26:53] There is a lot of critical
thinking involved in that.
[26:55] Do we lose that if we let Claude do it for
[26:58] us?
[26:59] I, I, I'm a little worried about that.
[27:00] And in fact, that's half
the reason I write myself.
[27:02] It certainly is for external audiences.
[27:04] Many people read what I write,
[27:05] but it, it, it is just as much
to clarify my own thinking so
[27:10] that I kind of know what to do next
[27:12] and to create a common reference
point across me and others.
[27:15] I think we're still grappling
with the question of
[27:17] how exactly do we use
AI in a way that kind
[27:20] of preserves those benefits.
[27:22] I think the thing I'm doing now
does that where I use Claude
[27:26] for research and I use
Claude for kind of, you know,
[27:29] how do I help organize my own thoughts.
[27:32] I think if we just used it end to end,
[27:34] like write an essay about the
risks of ai, first of all,
[27:37] it wouldn't write the things that I think,
[27:38] but also I would, I would
exactly lose that benefit.
[27:41] There's some way, as the models
get better, I think probably
[27:45] to, to use them directly much
more directly in the writing
[27:48] and yet still preserve those benefits.
[27:50] But I think it's gonna be a subtle thing.
[27:51] It won't be all one thing.
[27:52] It won't, we'll have to kind
of figure it out over time.
[27:57] I think we could have this
very unusual combination
[28:01] of very fast GDP growth
and high unemployment,
[28:05] or at least underemployment
or, you know, low wage job.
[28:10] Lot of low wage jobs, high inequality.
[28:12] You've been really
direct about job loss.
[28:15] AI could eliminate half
[28:16] of all entry level white collar jobs
[28:18] in the next one to five years.
[28:19] That was a year ago. AI
has moved incredibly fast.
[28:23] Is it still 50% or is it higher?
[28:25] I've always said,
[28:27] and you know, if you go back
to those original clips,
[28:29] they always get like, you
know, cut out of context
[28:31] and like the three seconds.
[28:32] But like, you know, the, the
real statement was always,
[28:35] I don't know what's gonna happen,
[28:37] but this is an order of magnitude for
[28:38] how crazy things could be.
[28:40] Also, I always talk about
all the things we can do in
[28:42] response to this, right?
[28:43] I've talked about token tax
[28:46] and working with enterprises
to adjust people,
[28:48] and I'm a little skeptical
of retraining programs,
[28:50] but like, we should throw them in the mix
[28:53] macroeconomic policy,
even from the beginning.
[28:55] I always talked about solutions,
[28:57] but you know, somehow there's
this tendency in the human
[29:00] psychology to clip the three seconds
[29:02] of like, doom is coming.
[29:04] So my message is just
definitely not doom is coming.
[29:06] My message is like, this
is something, you know,
[29:09] that we should see coming,
that we're worried about
[29:11] and that we need to actually
respond to positively.
[29:14] You know, I don't know exactly,
[29:16] but I'm, I'm still pretty concerned.
[29:18] I'm still the same order of concern.
[29:20] You know, we are seeing right now
[29:23] that AI is making people more productive.
[29:25] But that's the usual hump.
[29:26] If you go back, you know, to
the kind of industrial,
[29:29] you know, I wrote about this
in Adolescence of Technology.
[29:31] You automate 90% of the job,
[29:33] great, people are 10 times more
productive in the other 10%.
[29:36] 'cause they're 10 times more leveraged.
[29:38] But eventually it gets
close to a hundred percent.
[29:41] Now the sequel to that
is, well then you have
[29:43] to find something else for them to do.
[29:45] I, I don't know about the long run,
[29:47] I'm truly uncertain about that.
[29:49] But I do think there
are types of adaptation.
[29:51] Like one thing I'll
talk about is, you know,
[29:54] software engineers within Anthropic.
[29:56] We're, we're going through this
transition right now where,
[29:59] you know, right now AI makes
the software engineers more
[30:01] productive, even though AI writes all the
[30:03] code or almost all the code.
[30:05] But still it makes people more productive.
[30:07] But we're already starting
to see the beginning of like,
[30:10] you know, there may be
some people that it's,
[30:12] it's not making more productive
that it's better for the AI
[30:15] to just, to just do the thing.
[30:17] So that's one side of it.
[30:19] The other side of it though is
[30:20] what do we need more demand for?
[30:22] You know, there's something
we call a forward deployed
[30:24] engineer or in like applied
AI solutions architect
[30:27] where their job is a mix
[30:29] of technical work and
talking to customers.
[30:31] There's a lot of demand for that
[30:32] because there's a lot of customers
[30:34] and we're growing very quickly.
[30:35] Now, does every person
[30:37] who is in the pure
software engineering quite
[30:39] work for the this other?
[30:40] No, you know, it's not perfect.
[30:42] It's not one-to-one.
That gives you a flavor
[30:45] of, there's gonna be a hell
of a lot of disruption,
[30:48] but things will also
adjust. Which wins out?
[30:52] I don't know. But the
reason it's important
[30:54] to warn about it is that
that's how we can respond.
[30:59] That's how we can make policy, right?
[31:00] Both within Anthropic
[31:01] and macroeconomically for the whole world.
[31:04] We wanna put out carefully
considered thoughts.
[31:07] We don't wanna say things
[31:08] that people don't
believe we'll actually do.
[31:10] We don't wanna say things
that are half baked.
[31:12] We want to think carefully about
[31:13] what should actually be done about
[31:15] these, these problems.
[31:16] You put out this chart
showing potential job disruption
[31:19] like sales, finance, you
know, which jobs go away,
[31:23] who gets replaced and
what new jobs are created.
[31:26] So no one knows for sure,
[31:28] because you know, the
economy's unpredictable, right?
[31:30] It's the same as the stock market, right?
[31:32] They're these kind of
decentralized processes that you,
[31:34] you don't really know ahead
of time what are the pieces
[31:37] of the job that people are
still gonna be able to do.
[31:40] But what I would say broadly
is that, you know, anywhere
[31:43] that you have, you know, these kind
[31:45] of entry level white collar,
you know, whether it's banking,
[31:48] whether it's finance, whether
it's, you know, there's,
[31:50] there's, you know, there's,
there's gonna be a lot
[31:53] of potential for AI to first
make people more productive.
[31:57] But you know, then, then, then
there's gonna be, you know,
[32:00] then, then there's gonna be
a wholesale AI can do the job
[32:03] and then we're gonna have to
think about, well, you know,
[32:06] what is it that people can do?
[32:08] And I think we need to plan
about that ahead of time.
[32:11] We're already doing it. When we talk
[32:13] to enterprise customers, we
see choices that they face.
[32:16] They face the choice of, you
know, should I save costs?
[32:19] Which often means hiring less people,
[32:22] basically do the same
thing with less resources,
[32:26] or should we do more things
[32:27] with the same amount of resources?
[32:28] And we always, when we can
try to push them to doing more
[32:32] with the same amount of resources,
[32:33] because basically that means
like, hire the same number
[32:36] of people or maybe even more people,
[32:38] but just do, do kind of,
kind of do new things.
[32:40] Pushing them towards the positive sum.
[32:42] The thing that that we have going
[32:44] for us here is the pie
is gonna expand a lot.
[32:47] And so, because the pie
is gonna expand a lot,
[32:50] there are probably going to
be places where people can go.
[32:53] It's just a matter of
finding them fast enough.
[32:55] It's the size of the disruption.
It's, it's gonna be big.
[32:59] And that's what I'm warning people about.
[33:01] But we kind of, we have to
solve that matching problem.
[33:04] And
[33:05] So play this out for me a little bit.
[33:06] You know, you wake up in five years,
[33:08] what does this country look like?
[33:10] What are those people doing?
[33:12] Yeah, so 'cause if there's
that much unemployment, is
[33:14] that not how revolutions start?
[33:17] Yeah, no, this is the
outcome we wanna prevent.
[33:19] This is absolutely the
outcome we want to prevent.
[33:22] You know, I think, I think there's,
[33:23] I think there's a few places,
none of them are guaranteed.
[33:26] We're not sure, but there's
the physical world, right?
[33:29] Like things that are in
the physical world, yes,
[33:31] there's a robotics revolution as well,
[33:33] but it's a lot slower than
what's happening in AI.
[33:36] People always talk about
building data centers,
[33:38] but like when processing information
[33:40] of any type becomes a lot easier,
[33:43] maybe the restriction is gonna be
[33:45] things in the physical world.
[33:46] And so we need a lot
of more people to make,
[33:49] build, manufacture things
in the physical world.
[33:52] Anything that's human centered, I think
[33:54] that's gonna be a big deal, right?
[33:56] I hear all these stories about
AI found something that my,
[33:59] like my doctor couldn't
find and I feel, but like,
[34:02] but there's a, people really
wanna talk to other humans,
[34:05] particularly over kind of
important things, right?
[34:08] Maybe AI can do better customer service,
[34:10] but nevertheless people,
[34:12] or at least some people
wanna talk to humans.
[34:14] So these kind of human
relationship driven jobs, like,
[34:17] I think those are gonna
be important, right?
[34:19] And I think there'll be some
effort by the humans to kind
[34:23] of direct the ais right?
[34:24] At, at some level it has to be
in line with someone's values
[34:27] and someone's intentions and,
[34:30] and so I, I think there's
gonna be some role there,
[34:33] although I don't know how thin
versus how thick it will be.
[34:36] And I think it's very hard to say.
[34:38] There has been a lot of pushback,
[34:39] and I know you've said
you're trying to warn people,
[34:41] but that, you know, you're, you know,
[34:43] Jensen Huang said you're
conflating tasks with jobs.
[34:46] Other folks have said
this, you know, it's sort
[34:48] of doom marketing that benefits Anthropic.
[34:51] So, so I wanna be really clear
[34:53] and push back hard against
this, the whole picture
[34:56] of there are risks to job
loss and here are some ideas.
[34:59] I mean, we haven't fully
fleshed out the ideas
[35:02] because I want to get them right,
[35:03] but Anthropic has come
up with lots of ideas.
[35:06] We've had economic grants,
we have the economic index.
[35:08] I talk about the, the possible ways
[35:12] to address these risks from tax
[35:14] and macroeconomic policy to
[35:16] what the new jobs are in the
adolescence of technology.
[35:19] I lay out, you know, I have
like five pages where I lay out
[35:24] the difference between tasks and jobs.
[35:27] Why this time is different
than other times.
[35:29] A list of six different
things we can do from private
[35:32] philanthropy to government action.
[35:34] I talk about the problems,
I talk the solutions,
[35:38] but social media, which
I detest, which I detest
[35:41] as a category, people have
these three second clips from,
[35:46] you know, from a year ago.
[35:47] They don't actually read the essays
[35:49] or they prey on the idea
that social media, I've,
[35:54] I've written much more carefully
about these things where,
[35:57] where I talk about the risks, the idea
[35:59] that this is cheap marketing
is itself cheap marketing.
[36:02] This is, this is laziness,
this is failure to engage
[36:06] with serious intellectual work.
[36:09] And, and I think that
is part of the problem.
[36:11] Again, I think it's, it's part
[36:13] of the disease of Silicon Valley.
[36:15] It's been caught up in
this social media world of,
[36:19] of, of three seconds.
[36:20] And so people only respond to it
[36:21] or they think they only
have to respond to it again.
[36:24] I think it's very dangerous
[36:25] and we failed to have
a mature conversation.
[36:28] Instead, people just lazily
see this like three second clip
[36:32] and, and they're like, oh,
this is what Dario was saying.
[36:34] It's, it's so stupid. It's so unserious.
[36:37] And whenever someone
says something like that,
[36:39] I take them less seriously.
[36:42] One of the leading AI
companies in the world is deeply
[36:46] embedded in many different aspects
[36:49] of US national security
across military operations,
[36:52] The standoff between Anthropic
[36:54] and the Pentagon over AI military
[36:55] safeguards is ramping up.
[36:57] You've had a longstanding
anti-war stance starting all the
[37:00] way back to your days at Caltech,
[37:02] and yet you were one of
the first AI companies
[37:05] to sign a contract with
the Department of Defense
[37:08] to operate on classified networks
[37:09] that the US uses to
fight wars. Explain that.
[37:12] Yeah, so, you know, what
I would say is, is look,
[37:15] I mean the world, the world changes.
[37:17] Like, you know, my my view
of this technology, you know,
[37:20] when I see Russia invading
Ukraine, when I see the risk
[37:24] of China invading Taiwan, it
worries me that we have a kind
[37:29] of resurgent authoritarian block
[37:32] that they're very aggressive
[37:34] and that we need to defend ourselves.
[37:36] That is something that I,
you know, have believed
[37:38] for a while now continue to believe.
[37:41] And, and that's why across
both administrations, you know,
[37:45] you know, across, you
know, I may not agree
[37:46] with every policy of I,
of either administration,
[37:50] but you know, that's why we've generally
[37:51] been supportive of this.
[37:53] We don't want a world where China
[37:55] and Russia can build, you know,
[37:57] can analyze all the intelligence
with AI, can, you know,
[38:00] can, can use AI for, you know,
[38:02] for attacking Taiwan and Ukraine.
[38:05] And, and we can't defend them.
[38:07] So that's why we worked with them.
[38:08] We certainly don't do
it for the, the money.
[38:10] It's a huge pain.
[38:12] You know, even, even even
putting aside the, the lawfare,
[38:15] it's just a huge pain to get
up on government networks
[38:17] for not that much money.
[38:18] So we, we did it because we
cared about it, but similarly
[38:23] because we're doing it
[38:24] because we cared about it, there need
[38:25] to be limitations on the
use of the technology
[38:28] and the formulation
[38:29] that I used in Adolescence of Technology.
[38:31] We should use this technology
in every way except the,
[38:36] the ways that undermine
our own values, right?
[38:38] And our red lines of mass surveillance
[38:41] and fully autonomous weapons.
[38:42] Those are things that I
believe undermine our values.
[38:45] It's not worth democracies winning if
[38:47] democracies do those things.
[38:49] And, and so that's the,
that's the balance that I,
[38:51] that I see, and that's
the stand that we took
[38:54] and it, it explains both why
we were the first to work
[38:57] with Department of War and
why there were some things we
[38:59] wouldn't do when,
[39:01] when others were willing
to do those things.
[39:03] I, I think you need to pick a
stand and stand your ground.
[39:07] This idea of, you know,
companies that seesaw from,
[39:10] we won't do anything with the government
[39:12] to suddenly we're doing absolutely
[39:13] everything with the government.
[39:15] I don't, I don't get
it. You should pick your
[39:16] principles and stick with them.
[39:17] You've been working
with Palantir since 2024.
[39:20] That's right.
You know, their technology is used
[39:21] by ICE, police departments, in Gaza.
[39:24] Is Claude being used for
surveillance in other ways?
[39:26] We don't work with ICE
either through, either
[39:30] through Palantir or anyone else.
[39:31] We don't work with CBP, I
don't believe we work in Gaza.
[39:36] You know, our, our, we're
we're very careful about,
[39:39] you know, scoping our engagements
[39:41] to things that we believe in.
[39:42] So, you know, you drew
your, your red lines,
[39:44] the president banned you
from the federal government.
[39:46] The Pentagon labeled
you a supply chain risk.
[39:48] OpenAI jumped in
[39:50] and signed the contract that you wouldn't,
[39:53] what does winning this
fight actually look like?
[39:56] You know, I don't think
there's any winning.
[39:57] This fight for a private
company like this isn't a fight.
[40:00] Anthropic is, is trying to win
[40:02] or thinks about winning or losing.
[40:04] This is more a, I won't
even call it a fight.
[40:08] This is more a debate about
what the proper use of AI
[40:12] by the government is.
[40:14] And AI is an emerging new technology.
[40:16] We don't understand the ways in which it's
[40:19] reliable or unreliable.
[40:21] We don't understand the ways
in which it promotes our values
[40:24] or undermines our values.
[40:26] And so one of the things that
I thought was important was
[40:28] to establish a precedent
on some of the, some
[40:32] of the use cases we think are
good, which frankly is most
[40:35] of them, and some of the use cases
[40:37] that we're concerned about.
[40:38] And as I've said, we've
already seen, you know,
[40:41] you can only do so much
with a contract, right?
[40:43] As we've seen someone
else can sign a contract
[40:45] that doesn't respect
your, your same red lines.
[40:49] But what it has done is raised
awareness for the issue,
[40:51] and then we have serious
bipartisan efforts in Congress
[40:55] attempting to ban some of the things
[40:57] that we're concerned about and
attempting to set guardrails.
[41:01] Again, I don't want to
talk about this as a fight,
[41:03] but that's kind of winning the effort
[41:06] to get our country to
think more carefully about
[41:10] what is appropriate
use of this technology.
[41:12] That's. Anthropic is run
[41:13] by an ideological lunatic who
shouldn't have a, sole, that's,
[41:16] But.
That's not my.
[41:17] Question. My question is
[41:18] AI decision making over what we do,
[41:19] Do you mind being called
an ideological lunatic
[41:22] or a bunch of left wing nut jobs?
[41:24] You know, I've been called worse things
[41:25] than that all the time.
[41:27] You know, people, people can call me
[41:29] or, you know, people can call me
[41:31] or anthropic whatever they want.
[41:32] The two things that matter
are, we're successful
[41:34] as a company and, you know,
we stand up for our values.
[41:37] Like, I actually, in some
ways my life is really easy
[41:40] because when those are your, you know,
[41:41] those are the two things
you're trying to do.
[41:43] It's, it's, it's really simple, right?
[41:45] Like it, you know, you just,
[41:46] you always know where you stand.
[41:47] A US official has said
with the help of LLMs,
[41:50] the US military has gone from being able
[41:52] to hit a thousand targets a
day to 5,000 targets a day.
[41:56] That means Claude can help
kill more people more quickly.
[42:00] Are you comfortable with that?
[42:02] I think there's two things here, right?
[42:03] There is, there is the
ability of the United States,
[42:07] you know, to be more effective militarily.
[42:09] I, i I am supportive of that ability.
[42:12] I think having that ability be
stronger doesn't cause wars.
[42:16] It deters wars.
[42:18] Like, you know, I basically,
you you're asking like,
[42:21] you know, do you believe
in this country, right?
[42:23] Do you want this country
[42:24] to be a more powerful actor
rather than a less powerful
[42:26] actor on the world stage?
[42:28] I do, I'm a patriot.
[42:29] There's a separate question,
which is, you know,
[42:32] are there particular policies
[42:33] that the US government is
engaged in that I might support
[42:38] or not support?
[42:39] Obviously I support some of them
[42:41] and I don't support others of them.
[42:43] It's not up to me. If we
provide a technology, you know,
[42:46] the DOW made this point and
we actually agree with them.
[42:49] If we provide a technology,
it's not up to us to say,
[42:52] you can do this military operation
[42:53] and you can't do that military operation.
[42:55] Now, I might privately believe
[42:58] that this military operation makes sense
[43:00] and that military operation is a bad idea,
[43:03] but we're not gonna deny the technology.
[43:05] You, you have to leave policy in the hands
[43:07] of the military decision makers.
[43:10] What you can do is to assert
some high level boundaries
[43:14] that, you know, for, for
us, prevent the use cases
[43:17] that seem inconsistent
with, with our values,
[43:20] with our country's values,
[43:21] and promote the use cases that
we think, you know, we think,
[43:25] we think encourage our values.
[43:27] So that's how we think about it.
[43:28] Bloomberg has reported
that Claude is being used
[43:30] by the US military in the war in Iran
[43:32] to do AI assisted
targeting via platform made
[43:36] by Palantir, Maven Smart
System in February.
[43:39] A US missile reportedly hit a
girl school in Iran killing more
[43:43] than 150 people.
[43:44] Most of them children.
[43:47] Did Claude play a role in that strike?
[43:49] We look, we don't have
access to, you know, we,
[43:51] we don't know exactly how, you
know, these models were used.
[43:56] You know, obviously like, you
know, these things that, that,
[44:00] you know, mistakes that
happen in warfare are
[44:02] really, really terrible.
[44:04] Like, this is a really
terrible thing to happen.
[44:07] If that doesn't make clear
why we have to, you know,
[44:09] stand up for use cases
that, you know, we don't,
[44:11] we don't support, like, you
know, we, we, we were willing
[44:14] to risk the future of our
company to like limit how,
[44:18] you know, these models are used
[44:20] and you know, what, what you're
talking about is a use case
[44:23] that doesn't even violate our red lines.
[44:25] We're worried that there will
be a hundred times as much,
[44:28] you know, with, with use cases that, that,
[44:30] that do violate our red lines.
[44:32] Now I, you know, I I I,
you know, again, again,
[44:34] I would say I think overall
the use of these, the, the use
[44:38] of these models is appropriate.
[44:40] I think it's good on net, you know,
[44:42] but military decision makers
make terrible mistakes even
[44:47] when, even, even at the best of times.
[44:48] And I, I don't know if
we're in the best of times.
[44:51] Like there's several
things we can talk about.
[44:52] We can talk about making
red lines that, you know,
[44:55] prevent uses of the models
that are more likely to lead
[44:58] to those problems, right?
[44:59] If we had allowed, you
know, fully at high,
[45:01] if we had just given, in which
almost every other company
[45:04] now has to fully
autonomous weapons, right?
[45:06] This is like a human, what
[45:08] what we've seen here is Claude assists,
[45:11] but a human makes the final call.
[45:13] So a human made that
final call, not Claude.
[45:15] Imagine if you had a world
in which, not Claude,
[45:19] because we haven't allowed it,
but someone else's AI model.
[45:22] The AI model just makes the decision
[45:23] and the human never sees it.
[45:24] That's what we were standing up for.
[45:26] That's what we were fighting against.
[45:27] I would, I would also say, you know,
[45:29] there's a separate thing here.
[45:30] Again, I don't think procurement
is the right way to do it,
[45:33] but like, you know, we, we,
we need to make sure that,
[45:36] you know, it's, it's a matter of interest
[45:38] to the American people,
not to me as a supplier
[45:40] of the technology, but
to the American people
[45:42] that are military decision
makers don't make these mistakes
[45:45] that they operate
reliably, that, you know,
[45:47] they choose wisely what to do.
[45:49] Again. You know, that's, that's
of concern to me as a, as a,
[45:52] you know, as a citizen, as a
supplier of the technology.
[45:54] Like, you know,
[45:56] the government uses
Microsoft Excel a lot.
[45:59] You know, if I said micro, you
can use Excel for, you know,
[46:03] this military operation,
[46:04] but not the, you can't, you
can't realistically do that.
[46:08] But hopefully that gives you a
[46:09] sense of how we think about it.
[46:11] This school had a website.
[46:13] You could have found
it in a Google search.
[46:14] Like shouldn't Claude have
spotted that, shouldn't AI
[46:17] or whatever technology they
used have spotted that.
[46:19] And does it speak to a scarier
issue about using technology
[46:24] as a shortcut in war.
[46:25] Look, I look what
I'm, what I'm, you know
[46:27] what I'm gonna say is, you know, and,
[46:29] and I, you know, I don't, I
don't know, this relies on,
[46:31] you know, maybe classified
knowledge that I don't have.
[46:34] But you know, the principle
that, that we have established,
[46:38] and I think the principle
[46:39] that was obeyed here is
a human makes the human
[46:42] makes the final decision.
[46:44] I don't know what role
Claude or any other AI had,
[46:47] but like, if, if this
isn't an illustration why
[46:50] that principle is so important,
I don't know what's, is
[46:53] Is AI warfare more likely
[46:56] to stop World War III, a war
between the US and China?
[47:01] Or is it more likely to make it happen?
[47:03] I would say on balance it
is more likely to stop it.
[47:07] But if we have no limits on how it's used,
[47:12] then I think, you know, it could
be more likely to cause it.
[47:14] You know, you've seen
Doctor Strangelove, right?
[47:16] The premise of it was like
you have a doomsday device
[47:19] that automatically fires
nuclear weapons when it thinks
[47:21] nuclear weapons are being fired at it.
[47:23] What could go wrong? Right.
[47:24] Again, i I I get to this lethal, you know,
[47:27] fully autonomous weapons thing.
[47:29] I think the way conflicts
happen is that, you know, the,
[47:32] the two sides jump at each other.
[47:33] They misunderstand each other.
[47:35] And when we don't have proper
oversight of this technology,
[47:38] I think those kinds of accidents
are more likely to happen.
[47:41] Now, I think if AI is used in
an appropriate way, in, in,
[47:44] in, not even warfare,
[47:46] but think of just, just
intelligence collection, you know,
[47:50] let, let's say we're able to,
you know, predict an invasion
[47:53] of Taiwan or a new movement in Ukraine.
[47:56] Like, you know, our adversaries
will think twice about,
[48:00] you know, about conducting
some kind of invasion
[48:03] or military operation if we know
[48:04] everything that they're doing.
[48:06] And so I think superior
intelligence really can de
[48:09] deter conflict here.
[48:10] Superior ability to
respond can deter conflict.
[48:13] I continue to be a
believer in these things.
[48:17] Anthropic's making headlines
almost on a weekly basis.
[48:19] Yes. Most notably now
around mythos. Of course,
[48:22] This is the latest and
greatest Anthropic model,
[48:25] and it is capable of going
through all the links
[48:28] of the cyber kill chain
and doing so autonomously.
[48:31] You said mythos was too
powerful to release to the public.
[48:34] What surprised you most about it?
[48:37] I think the thing that
surprised me most about it was the
[48:41] models had been climbing in their ability
[48:44] to find vulnerabilities
[48:45] and importantly turn those
vulnerabilities into exploits,
[48:48] which people only talk
about the vulnerabilities.
[48:49] They don't often talk about
turning the vulnerabilities into
[48:52] exploits, which it, it was quite good at.
[48:55] So the things that surprised
me are we saw this huge jump.
[48:58] It was a particularly large jump
[49:00] and without us really
prompting them at all, some
[49:05] of the early companies that we gave this
[49:07] to said things like,
this is a super weapon.
[49:09] You should have to own
a gun license to use it.
[49:12] Please don't release this.
[49:14] Like, the, the demand
[49:16] to do this was coming from
the companies we gave it to
[49:18] who were finding so many
critical vulnerabilities
[49:21] and exploitability around
these critical vulnerabilities
[49:24] that, you know, they, they were
basically asking us not to,
[49:28] not to not to release it.
[49:30] Now to be clear,
[49:31] 'cause things always get
distorted in the world
[49:34] of social media, the goal isn't
[49:36] to keep this locked up forever.
[49:37] We're kind of gradually trying
to open this up to a wider
[49:41] and wider set of people
[49:42] and eventually we believe that
we should release mythos to
[49:47] to, you know, to a general audience,
[49:49] but with kind of strong cyber safeguards.
[49:51] Now, a concern is
today's cyber safeguards,
[49:55] which we did release on Opus 4.7,
[49:57] which is a good cyber model,
but a substantially weaker one.
[50:01] These can be jailbroken
[50:02] and we're a little concerned about some
[50:04] of the other companies who
think this is a sufficient
[50:07] defense because yeah, it works sometimes,
[50:10] but you know, we all know
[50:12] that these classifiers can
be jailbroken or gone around
[50:16] and our own testing as well
as frankly our assessment of
[50:20] the models that other, the defenses
[50:22] that other companies have
put in place suggests
[50:25] that these defenses are not strong yet.
[50:27] And, and that's what we're
waiting for, getting the defenses
[50:30] to the point where we really
have confidence in them.
[50:33] There was a lot of pushback on it.
[50:34] You know, you have researchers
saying they were able
[50:36] to replicate it using, you know,
cheaper open source models.
[50:41] Some folks say OpenAI, you know,
[50:44] has these capabilities already,
you know, what do you say
[50:46] to folks who say this is a
grand PR play? Yeah, so the,
[50:50] the claim that could be replicated
[50:53] with open source models,
that's just incredibly false.
[50:57] So the idea is mythos looks
across the whole cold base
[51:01] and finds something.
[51:02] Some guy went on Twitter
[51:04] and said, well, if you point
an open source model at exactly
[51:06] the line of code that mythos finds,
[51:08] then it finds the same issue.
[51:10] That, that, that isn't
the, that isn't the prompt,
[51:13] that isn't the question, right?
[51:14] Like, that is not, that
is not the same thing.
[51:17] The ultimate test of this
is like, we go to companies,
[51:20] we go to open source repos.
[51:21] We found 271 new
vulnerabilities in Firefox.
[51:26] We found many thousands
within the private, you know,
[51:29] companies who haven't fixed them yet
[51:30] or can't disclose them yet.
[51:31] Like no one found those 271
[51:34] vulnerabilities with the previous model.
[51:36] So like the actual workflow of
[51:38] what actually works in practice
as opposed to, you know,
[51:42] okay, I find the exact line
that mythos found, you know,
[51:44] I found the needle in the haystack.
[51:46] Something else can now pick up the needle.
[51:48] But what
about
[51:49] the folks who say, this
was just good marketing.
[51:51] You know, we have suffered
enormously commercially from
[51:54] not releasing this model.
[51:56] This model has incredibly
accelerated research within
[52:00] Anthropic and production and next models.
[52:02] It would do the same in the
[52:03] outside world if we were to release it.
[52:06] This has hurt us enormously commercially.
[52:08] If this helps defenders,
it also helps attackers.
[52:12] Can we defend anything anymore?
[52:13] What I would say is that the reason
[52:15] that we're giving Mythos to defenders
[52:17] before we give it to attackers
is to patch all the bugs.
[52:20] I don't know, as the models
get better, there may be more
[52:23] and more bugs to be found,
[52:24] but there's only so many,
they're finite, right?
[52:26] It's like you have this surface
[52:27] and there's only so many holes in it.
[52:29] You, you patch all the holes
[52:31] and the surface becomes
very hard to attack, as well
[52:34] as the code itself is written
with the powerful models.
[52:37] So it's, it's then becomes very hard
[52:38] to find flaws in or break into.
[52:41] So I think on the other side
of this, hopefully six months
[52:45] or a year from now, we have
a much more secure internet
[52:48] ecosystem than we had in the past.
[52:50] We're trying to get to that world
[52:52] and we're doing the best
we can to open up mythos
[52:55] to new cyber defenders.
[52:58] We've been talking to the
government, we're very respectful
[53:00] of their recommendations.
[53:02] They're slowing the pace
at which we open it up
[53:04] because they're worried about
counterintelligence risk.
[53:06] I think that's sensible.
[53:08] I think all serious people here understand
[53:10] that there's real trade offs here.
[53:12] We see a lot of sniping
from people on Twitter
[53:15] and from, you know,
from other AI companies.
[53:18] You look at what they're saying
[53:20] and the inconsistency
with what they're doing.
[53:22] It's not, they're not serious people.
[53:24] They're not seriously engaging with the,
[53:27] the serious trade-offs that,
that, that, that we have here.
[53:31] Look, I have customers calling
me up every day saying,
[53:34] I want access to Mythos.
[53:36] I have countries call calling
me up saying, I want access
[53:38] to Mythos and I have the US government
[53:40] and my security team
saying, no, wait a minute.
[53:43] There's risk to it. What?
[53:46] You know, I'm not saying one
side or the other is right?
[53:48] I think it's somewhere in between.
[53:49] Both sides have valid points,
[53:51] but there's a real challenge here
[53:53] and we need to face it
together as a society,
[53:55] not accuse things of
being cheap marketing,
[53:58] not use cheap marketing to try
[54:00] and counter position, which some
[54:02] of the other companies are doing it.
[54:03] It just, it just all shows
an an incredible lack
[54:07] of gravitas and maturity.
[54:08] We need to all face these
mo this moment together.
[54:11] Have you had to make trade-offs already
[54:13] that you're not entirely comfortable with?
[54:15] Throughout the entire history
[54:16] of Anthropic has been trade-offs, right?
[54:19] The entire history of Anthropic, right?
[54:21] Where, you know, in,
in, in, in, in, in, in,
[54:23] in some ideal world, you
would, you know, you would,
[54:26] you would prefer to, before
you release the first chatbot,
[54:29] you know, you could, you
could spend years studying,
[54:31] you know, every possible thing
that could go wrong with it.
[54:34] Now, we did delay, we did
delay the initial release
[54:37] of Claude, but you know,
we did it for a few months.
[54:39] So what I'm saying is
everything is a trade off.
[54:42] You know, the, the extreme
ends of the spectrum are, are,
[54:46] are completely insane.
[54:47] And so everything is a trade off.
[54:50] What I would say is that
now that we're in, you know,
[54:54] what I would describe as a
commercially leading position, I,
[54:57] I am actually, and Daniela
are actually doing all we can
[55:01] to move the dial even further towards,
[55:04] towards being careful.
[55:05] That's what the Mythos
release was, was about, right?
[55:07] It's very hard to do something like
[55:09] that if you're not the leading player.
[55:12] And so I think you're
gonna see more things,
[55:15] more things like that.
[55:17] You know, there's this argument,
[55:18] why wouldn't the government take you over?
[55:20] Why would they let a
private company control
[55:23] technology that's so powerful?
[55:25] So I actually think that's a very,
[55:27] that's a very serious question
and I share those concerns.
[55:31] I don't think the government
should outright take us over,
[55:35] but I would put it this way.
[55:37] I would say, just to back up
[55:39] and describe the situation,
[55:41] every previous powerful technology
we've seen in history was
[55:45] either built by the government
[55:46] or originated with the government.
[55:48] So nuclear weapons, obviously,
you know, initially built
[55:51] by the government and pretty much
[55:52] built by the government after that.
[55:54] But even like the internet,
GPS, cell phones, all the R&D
[55:58] was, you know, was done in the labs
[56:00] and the federal labs and the universities.
[56:03] AI is the first technology
[56:04] that's been built in the private sector
[56:07] and where government has not
really had a serious role
[56:10] and is coming in late to the game.
[56:12] I think that's actually a
dangerous and unstable situation.
[56:15] It is not the situation I would've chosen.
[56:17] There's not really an
alternative, like, you know,
[56:20] this technology is possible to build,
[56:22] our adversaries are building,
it has economic value,
[56:25] like it's, it's going to get built.
[56:27] The, the issue is the
government not doing it,
[56:29] not the private sector doing it.
[56:32] I think we need to think about
checks and balances on power.
[56:36] So I think there need to be checks
[56:37] and balances on the power
of the AI companies, right?
[56:39] We have this thing, the
long term benefit trust.
[56:42] What that is, is it's a set
of, basically it's a body
[56:46] that can appoint the
majority of the board members
[56:49] and remove the majority
of the board members.
[56:51] So it basic, it essentially,
if you thread it through,
[56:53] has the power to fire me.
[56:55] And what we're looking at
is we're introducing some
[56:58] elements, you know, nowhere
near all the elements,
[57:01] but we're introducing a a
little bit of the elements
[57:04] of like public governance, right?
[57:05] Where, where it's like, you
know, you're accountable
[57:07] to someone who just doesn't,
[57:09] he doesn't just have stock
stock in the company.
[57:11] So that's, that's very important
[57:13] and that structure is gonna
continue no matter what happens
[57:15] to the company. That's on the AI...
[57:18] And we encourage other companies
to have similar structures
[57:22] on the government side, I think
we need checks and balances.
[57:25] You know, there are, there
are efforts in Congress
[57:27] that have been announced to
enact those red lines, right?
[57:30] So I really think the, you
know, the legislative branch
[57:33] and the judicial branch
need to exert themselves
[57:36] because this technology, I'm
scared of companies having it,
[57:39] but I'm also scared of
government having it.
[57:41] And then the companies need to
provide checks on government,
[57:44] and the government needs to
provide checks on companies.
[57:46] You know, we need basic
regulation of the technology.
[57:49] You know, I think we need
to start doing pre-release
[57:52] testing, required
pre-release testing, testing
[57:55] and auditing of the models.
[57:56] You know, it, it's very funny to me
[57:58] how there's a particular group
[58:00] of people in the tech world
in Silicon Valley who started,
[58:03] you know, they, they, they
started with a position
[58:05] of like even having transparency
around this technology,
[58:08] even export control, you
know, this is all, you know,
[58:11] just totally, it'll apocalyptically
destroy our potential
[58:15] to create the technology.
[58:16] It'll kill innovation.
[58:18] And then as soon as they
see the first real danger,
[58:20] which I've been expecting all
along, there's all this talk
[58:22] of like nationalization
[58:23] and the government should just seize it.
[58:25] Come on folks here, you're,
[58:26] yo you're yo-yoing from
like the most extreme
[58:30] anti-regulatory, you
know, you know, if you,
[58:33] if you look at us the wrong way,
[58:34] you're destroying the
industry to, you know,
[58:36] this completely communist,
[58:38] the government should grab it all.
[58:39] We, we need a more, we need a more
[58:41] sensible, moderate approach.
[58:43] That's the one we've
been favoring all along
[58:45] because we've, we've understood the
[58:46] power of this technology.
[58:48] We're not panicking, we're not denying it.
[58:51] We see the smooth exponential
[58:52] and we're responding to it appropriately.
[58:56] So how was your visit
back to the White House?
[58:58] You know, we always try to work together
[59:00] with whoever we can in government.
[59:03] You know, I, I I said we
have this simple approach,
[59:04] like we have a set of principles,
[59:06] we like follow those principles
[59:08] and we hope that folks on the
other side are reasonable.
[59:10] And you know, honestly,
[59:12] the government has taken
Mythos very seriously.
[59:15] Like we've had good conversations
with Treasury Secretary Bessent
[59:18] with Chief of Staff, Susie Wiles.
[59:21] I think they really understand, you know,
[59:23] the nature of the risks here.
[59:25] Mythos has, I think, you know, helped them
[59:27] to feel much more concretely
where these risks are.
[59:30] So, you know, again, as
with any administration,
[59:34] there are parts who we
get along with very well
[59:37] and who understand it.
[59:38] And you know, there are other parts
[59:39] that are harder to get along with.
[59:40] I think that's normal.
[59:41] That would be the case in
any, in any administration,
[59:44] and we just try to
navigate it as best we can.
[59:49] You worked at Baidu
earlier in your career,
[59:51] big Chinese tech company.
[59:52] You worked at the Silicon
Valley outpost of it,
[59:54] and you've been clear
on your views on China.
[59:57] Strong open source models
are coming out of China
[59:59] and you have US companies building on them
[60:01] for free. Is that a threat?
[60:03] So, you know, one of
the things we've seen
[60:06] with this technology is that
there's really a premium to
[60:10] how intelligent the models are.
[60:12] We very, very rarely see
that people would prefer
[60:17] to use models with lower intelligence.
[60:19] Now to be clear, there's
a thriving ecosystem.
[60:21] There are lots of challenges
[60:22] and problems that are much
easier than, you know,
[60:25] the ones we need frontier models for.
[60:27] But again, it's an exponential, right?
[60:29] Like it's possible
[60:30] that like these far from frontier
models have economic value
[60:34] comparable to what we
saw in 2023 and 2024.
[60:37] But again, we have this 10X
[60:38] a year growth and, and,
[60:40] and so what, what we find is
[60:42] that what's on the frontier is
always much, much larger than
[60:46] what is, what is away from the frontier.
[60:48] I think this is something
that people who are used
[60:50] to building products in
the previous era don't
[60:53] quite understand, right?
[60:54] As someone who's come in who,
you know, hadn't run a company
[60:57] before, who's like, you know,
[60:59] has never thought about
the previous product era
[61:02] particularly to test the social media era,
[61:05] I feel like an outsider to that world.
[61:07] And I, I, I feel that
people's instincts are wrong.
[61:11] They have all these kind
of product heuristics
[61:15] and I think the 10x per
year model exponential really
[61:19] breaks that like intelligence
is just such a huge factor
[61:22] that it outweighs everything else.
[61:24] And so we're just seeing over
[61:25] and over again that, you know,
[61:27] the value is found on the frontier.
[61:29] Now what I do worry about with some
[61:31] of these laggard models
is the risks of them
[61:34] where we have Mythos
class cyber capabilities.
[61:37] 12 months from now,
[61:38] we'll have much better cyber capabilities,
[61:41] but the Mythos class cyber
capabilities may just be
[61:44] available for, for anyone
to, to download. Now,
[61:47] hopefully we'll have patched
everything before then.
[61:49] I don't think there's
anything we can do to stop it,
[61:50] but I, I think it's a serious concern.
[61:52] Did what you saw by do
shape your views on China?
[61:55] Not really, no. I worked there for,
[61:58] I worked there for a year.
[61:59] You know, I think I probably
learned more about like speech,
[62:03] speech recognition and,
[62:04] and you know, all, all, all all of that.
[62:06] Maybe the only thing that
concerned me was, you know, part
[62:08] of how we got all the speech
recognition data was, you know,
[62:11] they're like, they said ominously, oh,
[62:12] we don't care about privacy in China.
[62:14] So we have all this, this
speech recognition data.
[62:17] But I think, I think
aside from my worries,
[62:20] here are geopolitical, you
know, I think the things
[62:22] that most worried me about
what happened in China are,
[62:26] you know, what we, what we
saw happen to the Uyghurs,
[62:29] what we saw with suppression
of criticism even in the US,
[62:33] with what happened with Hong Kong, right?
[62:35] The fact that the CCP could
reach into the US business
[62:39] network and, you know,
[62:41] and suppress criticism, that's
an authoritarian state and,
[62:45] and a high tech authoritarian state.
[62:47] And when I see how that
combines with AI you,
[62:50] you really get a, a dystopia
here like 1984 or worse.
[62:55] And, and my focus is on
trying to prevent that.
[62:59] And I think we have an
opportunity to prevent that.
[63:01] I think we have an opportunity for AI
[63:04] to be a pro-democracy
technology, you know, that kind
[63:08] of makes people freer that
delivers on the promise
[63:11] of equal justice for all.
[63:13] Or it could go the other way. And,
[63:14] and which way it goes depends on the
[63:16] actions of the AI companies.
[63:17] It depends on the actions
of the government,
[63:19] depends on the actions of all of us.
[63:21] And so I see us as having
a responsibility here.
[63:24] There's a moment that people
in your field talk about
[63:26] where AI gets good
enough to improve itself
[63:30] and then the improved version
improves itself and so on.
[63:34] Some of your researchers think
that that moment is close.
[63:37] How far away is it?
[63:38] I don't think it's a moment in time.
[63:40] I think it's a continuous process.
[63:41] We're already seeing it in
some ways where the AI is able
[63:44] to suggest architectures for the next AI.
[63:47] You know, I would say a
year ago, we were seeing 10
[63:49] to 15% kind
[63:51] of increase in total factor
productivity due to AI.
[63:54] Like that's probably up to 20
[63:55] or 30% now might, you know,
might, it might be doubling like
[63:59] as with all things we're
on the exponential,
[64:01] there's no moment where AI improves itself
[64:04] or runs out of control or becomes unsafe.
[64:06] What we have is an
accelerating exponential
[64:08] and at each point on the
exponential we have to assess is,
[64:12] is this a time to slow down?
[64:13] Is this a time to, you know,
[64:15] put more controls on on this technology?
[64:18] I think more and more of
that is gonna be required.
[64:21] But I, you know, I think
the Rosetta Stone to all
[64:23] of this is the smooth exponential.
[64:25] Again, I think there's an
object lesson in the people
[64:28] who were against all AI regulation
[64:30] and then they saw one thing
and they wanted to nationalize.
[64:33] I think there's an object
lesson in the people
[64:35] who dismissed the power of AI
[64:37] and then said, oh my God,
it's improving itself.
[64:39] It's running outta control.
We have to shut it all down.
[64:41] Yo-yoing between those extreme
reactions is incredibly
[64:45] unhelpful as a response
to this technology.
[64:48] The right response, the wise response is
[64:51] to say, we're not gonna panic.
[64:53] Our countermeasures
will smoothly ratchet up
[64:57] with the power of the technology.
[64:59] If you see someone having this
kind of crazy yo-yo reaction,
[65:02] that's a sign that they
were caught by surprise
[65:04] and that they're not serious.
[65:06] I understand one of
your favorite books is The
[65:08] Making of the Atomic Bomb.
[65:09] That is correct.
Do you see parallels
[65:11] between yourself and Oppenheimer?
[65:13] You know, the figure I most identified
[65:15] with was Leo Szilard,
who, you know, the one
[65:17] who first basically had the
idea that there could be a kind
[65:20] of chain reaction.
[65:21] Look, my view is we're we're
not gonna get through this
[65:24] with like larger than life personalities
[65:26] or like figures who try
[65:28] and be at the center of everything, right?
[65:30] There needs to be a balance
of power here, right?
[65:33] There's a lot of powerful
actors who have interests here,
[65:36] and the only way it's gonna end well
[65:38] for everyone is if there is
some, there's basically checks
[65:41] and balances everywhere.
[65:44] So in some ways I actually see
Oppenheimer as a failure case
[65:47] as what should not happen.
[65:49] You've said there's
roughly a 10 to 25% chance
[65:52] of civilizational collapse.
That is not insignificant.
[65:57] Is there a scenario where it's something
[65:59] that Anthropic built, that caused that?
[66:02] I mean, I certainly hope not.
[66:04] My view is that, you
know, the, the actions
[66:07] that we have taken lower
[66:08] that probability rather
than increasing it, right?
[66:11] That probability comes
from the, the, you know,
[66:14] the very straightforward
recipe of the technology,
[66:17] the existence of many countries
in the world, the existence
[66:20] of many companies within an
economy and new ones created.
[66:23] If the void is isn't filled, like
[66:25] that's a dilemma that we're in.
[66:27] We are trying to act to
lower that probability.
[66:30] I think we lower it a lot
more than we raise it.
[66:32] But, you know, the inherent
property of this technology is
[66:35] that it's unpredictable.
[66:36] And so, you know, we
try to build something
[66:39] and test it a lot before it's released.
[66:41] And then the models
[66:42] that are released today are not dangerous,
[66:44] or at least not, you know,
really I think dangerous
[66:46] outside of cyber.
[66:48] And then we try and iterate
and learn from that.
[66:50] So there's like a zillion defense
mechanisms. You know, half
[66:53] of what we do within the company is try
[66:55] and, you know, reduce the
risk as much as we can,
[66:57] but, you know, it's,
it's never gonna be zero.
[67:00] I guess what I would say is, you know,
[67:02] suppose there are a
bunch of like, you know,
[67:03] airline companies out there
[67:05] and you're like, well, I'm gonna make an
[67:06] airline company that's safer.
[67:07] It can both be the case that, you know,
[67:09] your airline company is 10
times safer than all the other
[67:12] airline companies.
[67:13] But you know, if, if someone comes
[67:15] and asks you, like, can you guarantee
[67:17] that your airplane will never crash?
[67:19] I mean, how could you,
how could you possibly,
[67:22] But if there was a 25%
chance of an airplane crashing,
[67:24] you wouldn't get on that plane.
[67:25] That's right. 25% is too high.
[67:27] We're trying to make
that probability much,
[67:29] much lower. That is the goal.
[67:31] You are building something
incredibly powerful
[67:33] and stand to gain enormously from it.
[67:37] Why should we trust you?
[67:38] My view of this is actually
when any company starts out
[67:41] and, and particularly,
you know, what we've seen
[67:43] with the behavior of, of just
Silicon Valley as an entity.
[67:47] It's, it's thinking over
the last couple years.
[67:50] I think starting from a
position of distrust, you know,
[67:53] if you don't know anything about me,
[67:54] if you know anything about
Anthropic is pretty rational.
[67:56] I think Silicon Valley has
lost a lot of the world's trust
[67:59] and kind of has to re-earn it.
[68:01] And the message, you know,
we're trying to send is
[68:04] or actually different and,
[68:06] and that has to be earned in
things that we actually do.
[68:08] You can agree or disagree, but
we stood up for our values.
[68:12] The thing with, you
know, Mythos, like it's,
[68:14] it's really hampered us commercially not
[68:16] to put this very powerful model out.
[68:19] And there were a bunch of smaller things
[68:20] before it, you know,
we, we, we put our money
[68:22] where our mouth is on, you know, China,
[68:24] we cut off access to, to models.
[68:26] We didn't have to do that.
No one told us to do that.
[68:29] You know, that cost us several
hundred million dollars back
[68:31] when several hundred
million dollars was a big,
[68:33] was a significant fraction of our revenue.
[68:35] You know, the, the delay of Claude two,
[68:37] like we have a long history of it.
[68:38] We aren't perfect, we make mistakes.
[68:40] But you know, what I
would ask is for people
[68:43] to look at the overall history
[68:45] and say, if you add up
that overall history,
[68:49] what is the hypothesis about us
[68:50] that is most consistent
with that overall history?
[68:53] People have to decide for themselves,
[68:54] but I think the hypothesis
[68:56] that's consistent is we are
[68:58] genuinely trying to do the right thing.
[68:59] We're imperfect
organizations are, you know,
[69:02] always dysfunctional.
[69:04] We're always trying to, you know, fix them
[69:05] and make them work better.
[69:07] Many foot faults, many
things that go wrong,
[69:10] but at basis we, we have a honest
[69:13] and earnest picture of
how to do the right thing
[69:15] and we're trying to
execute on that picture.
[69:17] We will see you on the other
side of the exponential then
[69:20] Hopefully.
[69:40] You always wanted to
be a Hollywood star, right?
[69:42] I, I, that's one surprising thing
[69:44] that I didn't understand
about the CEO job is
[69:46] how often you have to wear makeup.
[69:48] That was not on my bingo card,
[69:50] Just a little powder.
