All-In: Anthropic’s $30B Ramp, Mythos Doomsday, OpenClaw Ankled, Iran War Ceasefire, Israel’s Influence

概览

This episode centers on whether frontier AI companies can manage dangerous capabilities without government-led regulation. The panel debates Anthropic’s decision to withhold broad access to Mythos after the model reportedly found serious vulnerabilities across major software systems.

The AI discussion then shifts to market structure: OpenClaw access, coding agents, Anthropic’s possible dominance in coding tokens, open-source competitors, antitrust risk, and whether coding becomes the gateway to the broader agent market.

The back half covers Anthropic’s reported $30B revenue run-rate, AI profitability, OpenAI’s competitive position, and where value may accrue across chips, hyperscalers, models, and applications. The episode closes with Iran ceasefire politics, Israel’s influence in U.S. foreign policy, and X’s auto-translation feature.

分段落总结

[00:45] Opening and Guest Setup

[事实] David Friedberg is absent, and Brian Gerstner joins as the “fifth bestie” alongside Jason, Sacks, and Chamath. [事实] The hosts open with banter about moderation, show dynamics, and an internet meme about not ruminating and just building. [推测] The opening sets the tone for an episode that mixes serious AI and geopolitical debates with informal in-group humor.

[04:24] Anthropic Withholds Mythos

[事实] Jason says Anthropic is withholding its newest model, Mythos, because the company claims it is too dangerous for broad public access. [事实] The transcript says Mythos found vulnerabilities in major operating systems and browsers, including old bugs in OpenBSD, FFmpeg, and the Linux kernel. [事实] A Dario clip says strong coding ability also produces cyber capability, including chaining multiple vulnerabilities into more sophisticated exploits. [推测] The discussion frames frontier coding models as dual-use tools that can improve defense while also enabling offensive cyberattacks.

[06:00] Voluntary AI Safety and Project Glasswing

[事实] Brad argues Anthropic deserves credit for not immediately releasing Mythos and for organizing Project Glasswing with major companies including Apple, Microsoft, Google, Amazon, and JP Morgan. [事实] Brad says the goal is to spend 100 days using advanced AI to find, fix, and harden software vulnerabilities before hackers exploit them. [事实] He argues this shows industry and market forces can coordinate with government without relying on top-down regulation. [推测] Brad treats Anthropic’s move as a model for voluntary governance at the frontier of AI capability.

[09:51] Sacks on Fear Marketing and Real Cyber Risk

[事实] Sacks says Anthropic has a pattern of pairing model releases with scary studies, including a previous blackmail study he calls reverse-engineered. [事实] He says the cyber case is more legitimate because better coding models logically become better at finding bugs and vulnerabilities. [事实] Sacks argues companies may have a window of a few months to use advanced models to find dormant vulnerabilities and patch them. [推测] His position is that Anthropic may be dramatizing the risk, but the industry still has to treat the risk as real.

[14:59] Chamath Questions the Theater

[事实] Chamath calls much of the Mythos rollout “theater” and compares it to OpenAI’s cautious GPT-2 release in 2019. [事实] He argues that sophisticated hackers may already be able to do similar work with existing models such as Opus. [事实] Chamath says fully patching the internet would take years, not 100 days. [推测] The disagreement is less about whether cyber vulnerabilities exist and more about whether Anthropic’s 100-day plan can meaningfully change systemic risk.

[19:03] Hardening the Internet Before Release

[事实] Brad says the Frontier Model Forum is already coordinating on adversarial distillation and related safety work. [事实] Sacks says he expects many patches over the next few months and gives Anthropic the benefit of the doubt on this specific cyber issue. [事实] Jason says the amount of code being pushed with AI tools is rising sharply and security should be embedded directly into coding tools. [事实] Chamath says if the risk is truly severe, the effort should include energy companies, nuclear operators, and aircraft companies.

[24:02] OpenClaw Access and Anthropic Pricing

[事实] Sacks asks about OpenClaw founder Peter Steinberger saying Anthropic cut off OpenClaw’s access. [事实] Jason and Brad explain that OpenClaw users were using $200 subscriptions while consuming far more tokens than average subscribers. [事实] Anthropic’s response, as described in the discussion, was to require OpenClaw usage to move to API-based pricing. [推测] The issue combines legitimate cost control with concern about whether a model provider can disadvantage third-party tools.

[27:19] Bundling and Antitrust Concerns

[事实] Sacks asks whether Anthropic’s own first-party agent product will remain bundled under flat-rate pricing while third-party tools must pay metered API prices. [事实] Brad says a bundling or anti-competitive argument could exist if that pricing difference happens, but enterprises may still prefer Anthropic’s product for security integration. [事实] The panel debates whether Anthropic has dominant market share in AI coding. [推测] The antitrust risk depends on market definition, pricing details, and whether coding agents become a durable control point.

[29:05] Agent Competition and Open Source

[事实] Jason lists competitors to OpenClaw, including Perplexity Computer, Anthropic managed agents, Hermes Agent, Alibaba’s Qwen-based work, xAI’s Grok Computer, and assistant updates from Amazon and Apple. [事实] Jason argues frontier AI companies want to kill or contain OpenClaw because open source could become highly disruptive. [事实] Brad responds that agents are an obvious product direction because consumers and enterprises want AI systems that can do tasks, not just answer questions. [推测] The agent layer is presented as strategically important because it could hold a user’s business context, personal data, and workflow.

[32:05] Coding Market Share and Enterprise Tech Debt

[事实] Sacks says that if Anthropic provides more than half of coding tokens, it has dominant share in that market. [事实] Chamath estimates Anthropic may have 50-60% share in AI coding tokens but says AI-enabled coding is still only about 5% of the broader coding market. [事实] Chamath argues current models still struggle with long-horizon enterprise-grade software and decades of legacy tech debt. [推测] Early leadership in AI coding tokens may not translate directly into durable enterprise control.

[36:29] Ridges, BitTensor, and Distributed AI

[事实] Jason discusses Ridges AI on BitTensor subnet 62 and says it reached 80% of Claude 4 capability in 45 days using Tau-based rewards. [事实] Chamath says open-source training and orchestration could be disruptive if capital markets cannot keep funding huge model-training rounds. [事实] Chamath also says real companies will not outsource production-code reengineering to an open-source project. [事实] Jason compares open-source AI tooling to enterprise adoption of Linux, Kubernetes, Apache, Postgres, and Terraform.

[39:55] Coding as the Gateway to Agents

[事实] Sacks says AI-generated code may rise from about 5% to 95% over the next few years. [事实] He argues an early leader in coding could build a flywheel from developer usage, code access, training tokens, and data-scale effects. [事实] Sacks says agents will depend heavily on code-writing ability and urges AI companies to avoid discriminatory or anti-competitive behavior. [推测] He frames coding as a gateway market that could shape who wins the broader agent platform.

[42:21] Anthropic’s $30B Revenue Run-Rate

[事实] Jason says Anthropic’s revenue run-rate topped $30B after starting paid API access in early 2023. [事实] He describes the ramp as $1B at the end of 2024, $4B in mid-2025, $9B at the end of 2025, and $30B in April. [事实] Jason says Dario announced more than 1,000 enterprise customers paying over $1M annually. [事实] Brad says Anthropic hit its planned year-end run-rate by the end of March. [推测] The panel treats the ramp as evidence that enterprise AI demand has crossed from experimentation into urgent adoption.

[44:20] Intelligence TAM and Compute Constraints

[事实] Brad says model and product capability crossed a threshold where customers see AI as labor augmentation and labor replacement, not just IT budget. [事实] He argues the TAM for intelligence is much larger than traditional software markets. [事实] Brad says Anthropic is compute constrained with roughly 1.5 to 2 gigawatts of compute and expects major compute additions. [事实] He says some large companies already consume 65-70% of tokens through open-source models. [推测] The episode presents frontier AI and open source as expanding together rather than as mutually exclusive categories.

[49:18] Profitability, Gross Margins, and CapEx

[事实] Chamath says the market discussion is still around gross versus net revenue rather than mature free cash flow or EBITDA. [事实] Brad says compute is the main cost input and that gross margins have moved from meaningfully negative to very positive. [事实] Brad says Anthropic has only about 2,500 people and may become accidentally profitable because it cannot stand up compute fast enough. [事实] Chamath argues incumbents such as Meta and Google may want AI to become a capital and compute problem because they have fortress balance sheets. [推测] Profitability remains hard to evaluate from the transcript because audited financials and exact gross-margin figures are not provided.

[55:29] AI Bubble Debate

[事实] Sacks says AI bubble skeptics argued that massive data-center capex lacked revenue justification. [事实] He says Anthropic’s coding revenue ramp is early proof that AI revenue can justify at least part of that investment. [事实] Sacks says the top four hyperscalers are expected to spend about $350B in capex this year, and Jensen has discussed $1T by 2030. [推测] Sacks uses the Anthropic example to argue that Silicon Valley’s willingness to fund massive buildouts can pay off when the technology direction is right.

[58:01] Anthropic vs OpenAI

[事实] Jason asks about OpenAI’s internal mood, employee departures, Sora, the Disney deal, and secondary-market valuation pressure. [事实] Brad says Anthropic made focused choices around coding and co-work, avoiding multimodal, video, hardware, chips, and data-center distractions. [事实] Brad says Anthropic has out-executed OpenAI recently but warns not to count OpenAI out. [事实] He says OpenAI’s Spud model is being previewed as strong and that Meta, Google, and Elon are also in the race. [推测] The segment frames U.S. frontier-model competition as strategically positive because several strong players are still competing.

[62:13] IPO Odds and Venture Positioning

[事实] The hosts discuss Polymarket odds for IPOs before 2027, including SpaceX, Cerebras, Anthropic, and OpenAI. [事实] They also discuss a Polymarket market suggesting a high probability of Anthropic being worth more than $600B at IPO, while noting low trading volume. [事实] Brad says Altimeter invested in Anthropic around a $130B to $150B valuation and holds large positions in both Anthropic and OpenAI. [事实] Brad says venture investors are nervous about software companies exposed to AI disruption and are crowding into frontier models, SpaceX, and military modernization.

[64:42] Where AI Value Gets Captured

[事实] Sacks says early AI value capture was concentrated in Nvidia and the chip layer. [事实] He says value then expanded to hyperscalers and now appears to include model companies such as Anthropic and OpenAI. [事实] He asks whether the application layer will also capture major value, citing Palantir as one example already turbocharged by model capabilities. [事实] Chamath says it is too early to know, while energy and data centers are already clear beneficiaries. [推测] The unresolved question is whether model companies absorb application value or whether application companies with distribution and workflows retain pricing power.

[67:23] Liquidity and All-In Events

[事实] The hosts say Liquidity is sold out with hundreds of people on the waitlist. [事实] They mention speakers including Bill Ackman, Andrej Karpathy, Dan Loeb, Thomas Laffont, Brad Gerstner, and Sarah Friar. [事实] Jason says the fifth annual All-In Summit will be in Los Angeles from September 13 to 15. [事实] Sacks says he has been advised he can attend California for business reasons.

[70:12] Iran Ceasefire and Escalation Risk

[事实] Jason says the discussion is taking place two weeks into a ceasefire and that JD Vance and Jared Kushner are headed to Islamabad for talks. [事实] He summarizes Trump’s social-media threats and says Trump later agreed to a two-week ceasefire if Iran opens the Strait. [事实] Sacks says he is not part of the White House foreign-policy team and is speaking only personally. [事实] Sacks praises the ceasefire and says wars tend to climb an escalation ladder once they begin. [推测] The panel treats the ceasefire as a test of whether Trump can use pressure without becoming trapped in a wider war.

[73:15] Trump Doctrine and Market Reaction

[事实] Brad describes a “Trump doctrine” of destroying military capabilities, killing people building lethal weapons, getting out, and avoiding democracy-building. [事实] Brad says the market drawdown around Iran was much smaller than the tariff-related drawdown he cited earlier. [事实] He says markets have largely bounced back and may rally if Iran, Lebanon, Ukraine-Russia, Venezuela, and Cuba developments improve. [推测] Brad’s market read assumes investors believe Trump will avoid a long entangled war.

[75:21] Netanyahu, U.S. Policy, and Israel’s Influence

[事实] Jason summarizes a New York Times account in which Netanyahu gave Trump a four-part pitch on attacking Iran. [事实] Jason says the account described JD Vance warning about regional chaos and coalition fracture, Rubio opposing regime change, Susie Wiles worrying about gas prices, and General Dan Cain saying Israelis oversell their plans. [事实] Jason says Jewish Americans he has spoken with believe Netanyahu is harming Jewish Americans and the Jewish diaspora. [事实] Chamath says the U.S. president makes U.S. decisions, not a foreign leader. [推测] This is the episode’s most sensitive political section because it links Gaza, Lebanon, Iran, antisemitism, and U.S.-Israel alliance fatigue.

[79:43] Gulf Optimism and Israel’s U.S. Standing

[事实] Brad says outcomes will judge the strategy and says doing nothing in Iran or Venezuela also carried risks. [事实] Brad says Gulf-state contacts he has spoken with are hopeful and optimistic. [事实] Sacks mentions Naftali Bennett tweeting polling showing Israel becoming unpopular in the U.S. and saying the situation needs to be addressed. [事实] Sacks says Israel getting behind the ceasefire would be a good place to start. [推测] The discussion suggests Israel’s long-term U.S. support may depend on finding off-ramps and improving public perception.

[81:53] X Auto-Translation and Cross-Border Discourse

[事实] Jason praises X’s auto-translation feature for surfacing Hebrew, Japanese, French, Chinese, and Russian posts in users’ feeds. [事实] He says users can now reply across languages and be understood through automatic translation. [事实] Brad says X is better today with 70% fewer employees than when Elon took over. [事实] The hosts also discuss bot accounts, subscriber-only replies, and future cleanup of automated replies. [推测] The hosts present AI translation as a practical consumer product that could affect international understanding and political discourse.

[86:27] Closing Banter and Athena Mention

[事实] The hosts close with jokes about followers, replies, casino banter, and Friedberg’s absence. [事实] Jason says Athena received strong interest after being mentioned previously and may hire many new assistants. [事实] The episode ends with the hosts saying goodbye and briefly joking about merchandise.

播客点评/总结

[推测] The strongest part of the episode is the AI-market discussion. The panel connects model capability, cybersecurity, enterprise adoption, pricing, open source, and antitrust in a way that gives listeners a broad map of the current AI platform fight.

[推测] The limitation is that many numbers are discussed as insider perspective, market chatter, or host assertions rather than independently verified evidence inside the transcript. Listeners should treat claims about revenue, gross margins, model capability, and IPO odds as discussion points rather than settled facts.

[推测] This episode is most useful for listeners interested in AI strategy, venture markets, model-company competition, and U.S. foreign-policy debates. It is less useful for anyone looking for a neutral or tightly sourced policy briefing, because the format is opinion-driven and conversational.