Fox Buys Roku for $22 Billion to Expand Ad-Supported Streaming Reach
TBPN
Department of Commerce export controls restricting foreign national access to Fable 5 and Mythos force Anthropic to overhaul its compliance framework, driving a widening cultural and regulatory gap between AI safety labs and the federal government.
## Background
Published by TBPN on YouTube on June 16, 2026, this video offers a condensed 30-minute summary of the show's daily tech talk coverage. According to the page description, TBPN is a Silicon Valley-focused show hosted by John Coogan and Jordi Hays that features high-profile tech and business leaders.
## TL;DR
Hosts John Coogan and Jordi Hays analyze the emerging necessity of sovereign AI before examining the cultural and regulatory rift between Anthropic and the federal government following export controls on Fable 5 and Mythos. The conversation turns to Meta's strategic shift from unrestrained AI spending to strict internal token budgets and min-maxed resource management. Finally, the hosts evaluate Fox's massive $22 billion acquisition of Roku to dominate ad-supported streaming, concluding with Tyler Cowen's insights on foreign talent constraints and AI national security policy.
## Atmosphere & style
The episode carries a high-energy, fast-paced, and sharp conversational rhythm typical of insider tech commentary. Coogan and Hays engage in dynamic banter, effortlessly weaving gaming culture analogies and pop-culture references into rigorous corporate and regulatory analysis. The mood transitions smoothly from lighthearted sports entertainment commentary to grounded, analytical discussions on national security and media finance.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=0s) Recent global events have strongly reinforced the 'sovereign AI' thesis, prompting nations worldwide to cultivate national champion AI models to ensure guaranteed access.
- [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) The UFC Freedom 250 event held at the White House highlighted a growing cultural and aesthetic divide between the administration and safety-oriented AI labs like Anthropic.
- [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) Anthropic faced significant regulatory friction following Department of Commerce export control directives restricting models like Fable 5 and Mythos from foreign nationals.
- [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) Anthropic's decision to silently degrade model performance on frontier AI research queries, rather than outright refusing, sparked frustration within the developer community.
- [11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) Meta is shifting internally from extreme compute consumption ('token maxing') toward strict 'token minimizing' and department-level token budgets starting in 2027.
- [16:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=975s) Meta is actively fine-tuning internal models ('Meta Code') on its vast proprietary codebase to build custom harnesses aimed at accelerating its core social platforms.
- [19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) Fox Corp acquired Roku in a massive $22B–$25B deal, uniting Fox's Tubi content engine with Roku's hardware platform to dominate ad-supported connected TV (AVOD).
- [26:53](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1613s) Tyler Cowen outlined major structural challenges for US AI export controls, emphasizing the risks to lab financial health, international revenues, and top foreign talent recruitment.
- [26:53](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1613s) Restricting frontier AI development access strictly to US citizens presents severe enforcement challenges and raises complex debates around potential state nationalization.
- [30:20](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1820s) While open-source and Chinese AI models continue to advance, compute constraints and restricted access to top research circles may slow their pace compared to top proprietary models.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=0s) Sovereign AI, Mistral, and UFC Freedom 250 at the White House
The episode opens with the hosts evaluating how recent events have dramatically reinforced the "sovereign AI" thesis. Countries across the globe are increasingly recognizing the necessity of cultivating national champions that can guarantee uninterrupted, long-term access to frontier artificial intelligence models without reliance on shifting foreign policies. The hosts tease an upcoming appearance by investor Gavin Baker to discuss the SpaceX IPO before transitioning into a recap of UFC Freedom 250, a high-profile mixed martial arts event held directly on the lawn of the White House.
Jordi Hays offers a fan’s perspective on the spectacle, confessing that while he was initially wary of the optics—holding a lavish, high-octane sporting event during a period of economic strain, soaring inflation, and global geopolitical tension—the execution completely won him over. He frames it as a classic "bread and circuses" moment, but admits that as a pure spectacle, it was undeniably effective. John Coogan points out the historic precedent, noting that President Teddy Roosevelt held boxing matches inside the White House nearly a century ago and even sparred himself. Hays argues that leveraging major sporting events and cultural spectacles represents an underutilized political strategy, predicting that the administration's approval ratings could see a bump from the iconic imagery generated throughout the night.
- **UFC Freedom 250 Card Structure:** Consisted of eight wall-to-wall fights with zero filler.
- **Key Bout Highlight:** The main title fight between Justin Gaethje and Ilia Topuria (with Max Holloway/Charles Oliveira dynamics framing the division) completely altered the career trajectories of both fighters.
- **Iconic Visuals:** One of the most circulated images from the night featured Brazilian fighter Diego Lopes standing triumphantly in the Octagon on the White House grounds.
---
[04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) Anthropic Export Controls, Fable 5, and the Government Gap
The discussion pivots to the stark absence of Anthropic CEO Dario Amodei at the White House event, contrasted sharply with Meta's Mark Zuckerberg sitting front row. This visual gap highlights the deep aesthetic and cultural rift between Anthropic’s cautious, safety-first ethos and the current political administration. The hosts analyze the news that Anthropic executives were flying into Washington D.C. on a Sunday to resolve urgent regulatory friction—a move that underscored the severity of the crisis, even if landing in D.C. over the weekend meant facing empty offices while the administration attended a fight night.
Coogan walks through the recent timeline of regulatory escalations between Anthropic and the federal government. On March 4th, Anthropic received a supply chain risk letter from the Department of War (which was later rescinded). On April 7th, they previewed Mythos alongside Project GlassWing. On June 9th, Anthropic launched Fable 5, embedding safety guardrails specifically targeting offensive cyber capabilities, biosecurity threats, and automated AI research. Just five days later, on June 12th, the Department of Commerce issued a sweeping export control directive restricting both Fable 5 and Mythos 5 from being accessed by any foreign nationals—including Anthropic’s own foreign national employees, both domestically and abroad.
> "If you take safety seriously, cybersecurity seriously, biosecurity seriously, export control seriously... UFC could not be less aligned with that. The stakes there... there is this big gap between the administration and Anthropic."
This regulatory mandate forced Anthropic to halt access across all user tiers to implement strict Know Your Customer (KYC) frameworks, API auditing, and authorization checks. To complicate matters further, reports surfaced that Amazon CEO Andy Jassy had met with senior administration officials to demonstrate jailbreak vulnerabilities in Fable 5. Anthropic publicly countered in a blog post, downplaying the findings by stating they had reviewed the demonstration and found it merely identified a small number of previously known, minor vulnerabilities that other publicly available models could easily discover without any bypass.
The hosts also examine Anthropic’s controversial decision to silently degrade the quality of model outputs on topics related to frontier AI research—rather than issuing outright refusals as they do with cyberweapons or pathogens. While the theoretical justification holds (restricting tools that build exploit tools), silently degrading answers deeply frustrated the developer and machine learning community.
---
[11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) Meta’s Shift to Token Minimizing and the Art of Min-Maxing
The conversation shifts to a report from Amir Efrati at *The Information* detailing a strategic pivot inside Meta toward "token minimizing" and strict compute budgets. Coogan critiques the industry's obsession with extreme terminology like "token maxing" or "token minning," tracing the nomenclature back to gaming culture and the "looksmaxing" internet subculture. He argues that corporate strategy should never lean into blind maximization or extreme starvation; rather, it should always be about **min-maxing**—allocating the precise amount of resources required to achieve the highest possible return on investment.
To illustrate the absurdities of the unconstrained compute era, Hays mentions an unconfirmed report circulating within Silicon Valley about a Meta engineer who accidentally spent $90,000 in API and compute costs in a single day through runaway execution loops, resulting in their termination three days later. Coogan compares this to marketing spend, noting that spending $5 million on a single Super Bowl advertisement can be extraordinarily lucrative if min-maxed properly with targeted local buys or high-ROI positioning, but disastrous if executed as a mindless "Super Bowl maxing" stunt.
The section wraps up with an analysis of modern sports advertising integration, praising Dana White's personal involvement in UFC partner campaigns during the White House broadcast. Commercials were built as authentic, narrative-driven segments—such as Dana White test-driving a Dodge Ram TRX or appearing in a staged podcast set discussing business insurance with a corporate CEO—rather than standard, easily skippable ad spots.
<ElicitationsGroup message="Explore the rest of this episode transcript:">
{/* Reason: Offers natural continuation paths to cover the remaining topics in the transcript. */}
<Elicitation label="Summarize the Fox & Roku deal and Tyler Cowen analysis" query="Provide a detailed walkthrough of the remaining transcript sections, including the Fox/Roku ad-supported streaming deal and Tyler Cowen's analysis on Mythos and open-source AI."/>
</ElicitationsGroup>
Here is a detailed and faithful walkthrough of the provided transcript section:
[15:12](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=912s) UFC Freedom 250, Paramount Streaming, and Gambling Saturation
The hosts kick off the discussion reflecting on the spectacle of the UFC event on Paramount. They marvel at the overwhelming density of advertising plastered throughout the octagon, observing that every conceivable gambling platform seemed represented. Because the UFC refuses to lock into an exclusive gambling partnership, viewers are bombarded with ads for rival platforms like Crypto.com, Polymarket, Bet365, and Kalshi all at once—leading the hosts to joke that a bettor's only day off from gambling is leap day every four years. They note the irony of Kalshi running heavy ad campaigns directly against Polymarket's octagon-ring sponsorship, all served through Paramount's ad inventory where subscribers access UFC events without extra pay-per-view fees.
- **Gambling Platforms:** Polymarket (in the octagon), Kalshi (broadcast ad buyer), Bet365, and Crypto.com.
- **Distribution Model:** Included with a standard monthly Paramount subscription.
- **Key Takeaway:** "Gambling is just min-maxing done poorly."
---
[16:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=975s) Meta’s Token Strategy Shift: From Token Maxing to Token Budgets
The conversation shifts to Meta's pivot regarding its internal AI infrastructure spending. Two months prior, Meta represented the extreme end of "token maxing," burning cash and on track to spend billions per year externally on models like Claude. An internal memo reveals that after seeing exponential increases in employee AI usage—tracking toward billions in internal AI costs alone for 2026—Meta is shifting toward strict "token minimizing." Starting in 2027, the company will introduce structured token management, including explicit token budgets and allocation controls for individual teams.
```
2026 Trajectory: Billions spent on internal AI use (including external API usage like Claude)
2027 Strategy: Structured token budgets, department allocations, and strict visibility tools
```
The hosts draw a direct parallel between traditional marketing budgets and corporate token budgets. They also discuss "Meta Code," an internal fine-tuning initiative. Leaked audio reveals Mark Zuckerberg emphasizing that Meta's internal code quality exceeds that of the average open-source contributor. By bringing frontier models in-house and fine-tuning them on Meta’s vast, pristine internal codebase, Meta aims to create proprietary coding harnesses tailored directly to making their social platforms faster and better.
---
[17:44](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1064s) The Meta Strategy Debate: Personal Superintelligence vs. Commercial APIs
A debate emerges over Meta’s long-term commercial posture in AI compared to other tech giants like SpaceX. While Meta promotes the concept of "Personal Superintelligence," the hosts express skepticism over what that phrase actually translates to in product deliverables. They weigh whether Meta will eventually release its coding tools and fine-tuned models externally via paid enterprise APIs—similar to their brief enterprise experiment with Manas—or keep them strictly internal.
- **The Bull Case for Meta:** Unlike traditional tech companies threatened by AI disruption, Meta’s underlying advertising and social network core is actively accelerated by AI while insulated from replacement.
- **Market Valuation Comparison:** The hosts note the market cap divergence, contrasting Meta's massive valuation with SpaceX’s $2.3 trillion valuation.
- **The Core Ambiguity:** The market is pushing Meta to clearly define whether it is a consumer superintelligence company or an enterprise AI developer infrastructure provider.
---
[19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) Fox Acquires Roku: The $22B–$25B Ad-Supported Bet
The hosts analyze news breaking across major media outlets regarding Fox Corp’s massive acquisition of Roku. While *The Wall Street Journal* initial reports pegged the deal around $25 billion, CNBC and Axios reported the enterprise value at $22 billion ($160 per share, structured as $96 in cash plus 0.9693 Fox Class A shares). Funded by $12 billion in new debt combined with cash on hand, the transaction stands as Fox's largest acquisition to date.
```
Fox Purchase Price: ~$160/share ($96 Cash + 0.9693 Fox Class A Shares)
Enterprise Value: ~$22 Billion
Funding Structure: $12 Billion in new debt + cash on hand
Target Cost Synergies: $400 Million in annual cost cuts
```
This acquisition brings together Fox’s live news and sports engine with Roku’s 25% market share in connected TV platforms (spanning over 100 million streaming households globally). The deal underscores the dramatic shift toward free and low-cost ad-supported video-on-demand (AVOD). Ad-supported streaming plans now account for nearly 50% of all premium SVOD sign-ups in the United States, up from 39% two years ago.
The hosts point out Fox's successful history with Tubi, which was acquired in 2020 for $400 million and now generates well over $1 billion in annual revenue with nearly 100 million monthly active users (growing revenue 23% in fiscal Q3). By pairing Tubi (the content product) with Roku (the TV hardware platform commanding over 40% of connected TV watch time), Fox gains an unmatched advertising network, particularly among younger, prime-earning demographics.
- **Connected TV Market Share:** Roku leads at 25%, followed by Samsung Tizen at 23%, with Amazon Fire TV close behind.
- **Strategic Vision:** Rather than merging Tubi and the Roku Channel, Fox plans to operate them as distinct brands while using the platform to rebundle streaming content for consumers.
---
[26:53](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1613s) Tyler Cowen on Mythos, Export Controls, and AI National Security
The segment concludes with an examination of economist Tyler Cowen’s analysis regarding the Mythos model fallout and the complexities of US government AI export controls. Following reports that Mythos could be jailbroken despite strict built-in restrictions on biological safety and self-improvement queries, Cowen outlined several critical constraints facing US policymakers trying to balance security with market dominance:
1. **Lab Financial Health & Talent Access:** The US government needs frontier labs to stay solvent and have successful IPOs. However, tight immigration and security controls make it extremely difficult to recruit top foreign talent—citing legendary non-US researchers like Demis Hassabis, Ilya Sutskever, and Andrej Karpathy.
2. **International Revenue Impact:** Restricting global model access hampers American multinationals operating abroad and cuts US labs off from lucrative Fortune 500 enterprise contracts in foreign jurisdictions.
3. **Enforceability & Insider Threats:** Restricting AI tool access exclusively to "US citizens" is practically unenforceable. It does not prevent foreign actors from hiring a compromised US citizen or faking credentials to gain access.
4. **Nationalization Debates:** Responding to Cowen’s controversial assertion that the US government could not successfully nationalize or operate these frontier AI labs, the hosts highlight the ideological divide between open-source math access and state-level national security enforcement.
To illustrate the federal government's historical stance on technological export controls, the hosts cite the case of cryptocurrency researcher Virgil Griffith. After delivering a presentation in North Korea on basic cryptography and how code can create unsanctionable stores of value, Griffith was prosecuted and sentenced to prison for violating federal sanctions—a reminder of how strictly the US government interprets knowledge export laws when national security is at stake.
<ElicitationsGroup message="Explore related coverage from this transcript:">
<Elicitation label="Summarize the UFC 250 White House discussion" query="What was discussed in part 1 of this TBPN episode regarding UFC Freedom 250 at the White House?"/>
<Elicitation label="Analyze the streaming market landscape" query="How does the Fox-Roku deal compare to other streaming consolidation moves like Disney/Hulu or Netflix's ad tier?"/>
</ElicitationsGroup>
The RSI Perspective and Operational Realities of AI Development
The conversation pivots to the extreme implications of recursive self-improvement, joking that if you are completely "super RSI pilled", you would not need employees at all because the systems would just improve themselves, leaving the government to simply tell the AI to "continue to improve." On the flip side, looking at current operational realities, these systems are exceptionally complex. When considering system uptime, research tastes, and product effectiveness, companies actually require massive workforces—including international employees who bring critical, innovative ideas. The hosts break down these contrasting philosophies about whether future systems will run autonomously or remain deeply dependent on large, diverse human teams.
* The extreme "RSI-pilled" scenario envisions AI improving itself without human intervention.
* Conversely, the complexity of modern systems demands large teams and international talent to maintain uptime and drive effective product research.
---
Chinese Models, Open Source Pace, and Growth Factors
The discussion turns to Chinese and open-source models, which continue to improve at a reasonable pace even though they remain considerably behind the best proprietary models. They explore why progress in China might be slowing down, referencing Gavin Baker's observation that it could be related to compute constraints and decisions to pass on purchasing Blackwell chips. Other potential factors include research ideas becoming more tightly locked down within controlled circles in San Francisco, or data limitations. However, they note that China still possesses plenty of people writing code, inferencing models, and capturing data from rollouts, meaning another "DeepSeek moment" could still happen suddenly in the coming months.
* Chinese and open-source models improve steadily, though they still lag behind top proprietary models.
* Potential reasons for slowing momentum include compute choices, restricted access to San Francisco research circles, and data constraints, though a surprise "DeepSeek moment" remains possible.
---
Show Wrap-Up and Sign-Off
The episode concludes with a final steering order bringing the show to a close. The hosts remind listeners to leave five-star ratings on Apple Podcasts and Spotify, sign up for the daily newsletter at tbpn.com, and tune in for the next live stream tomorrow at 11:00 a.m. Pacific time before signing off with standard show taglines.
* Viewers are encouraged to leave five-star reviews on Apple Podcasts and Spotify.
* The daily newsletter is available at tbpn.com, with the next live show scheduled for 11:00 a.m. Pacific.
By the way, to unlock the full functionality of all Apps, enable [Gemini Apps Activity](https://myactivity.google.com/product/gemini).
## Notable quotes
> [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) If you take safety seriously, cybersecurity seriously, biosecurity seriously, export control seriously... UFC could not be less aligned with that. The stakes there... there is this big gap between the administration and Anthropic.
> [15:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=915s) Gambling is just min-maxing done poorly.
## Mentioned
Anthropic, Meta, Fox, Roku, TBPN, Diet TBPN, John Coogan, Jordi Hays, UFC, UFC Freedom 250, White House, Gavin Baker, SpaceX, Teddy Roosevelt, Justin Gaethje, Ilia Topuria, Max Holloway, Charles Oliveira, Diego Lopes, Dario Amodei, Mark Zuckerberg, Department of War, Mythos, Project GlassWing, Fable 5, Department of Commerce, Mythos 5, Amazon, Andy Jassy, Amir Efrati, The Information, Dana White, Dodge Ram TRX, Paramount, Crypto.com, Polymarket, Bet365, Kalshi, Claude, Meta Code, Manas, The Wall Street Journal, CNBC, Axios, Fox Corp, Tubi, Roku Channel, Samsung Tizen, Amazon Fire TV, Disney, Hulu, Netflix, Tyler Cowen, Demis Hassabis, Ilya Sutskever, Andrej Karpathy, Virgil Griffith, San Francisco, China, DeepSeek, Apple Podcasts, Spotify, X, YouTube, New York Times, Sam Altman, Mark Cuban, Satya Nadella, Ramp, Public, Cisco, Console, CrowdStrike, Figma, MongoDB, NYSE, Railway, Shopify
TBPN analyzes the mounting regulatory and cultural collision between frontier artificial intelligence labs and the U.S. government, alongside corporate spending shifts toward AI compute discipline and Fox's multi-billion-dollar bet on ad-supported streaming hardware.
Anthropic's Regulatory Clashes and Export Mandates (4:02)
Anthropic has faced escalating regulatory friction with Washington, punctuated by the Department of War's briefly issued supply chain risk letter on March 4 and a sweeping Department of Commerce export control directive on June 12. The order restricted both Fable 5 and Mythos 5 from foreign national access—extending to Anthropic’s own foreign workforce domestically and abroad. In response, Anthropic was forced to halt access across user tiers to build strict Know Your Customer frameworks and API auditing protocols.
Safety Implementations, Jailbreaks, and Strategic Rifts (4:02)
The tension between Anthropic and federal policymakers was visually emphasized by CEO Dario Amodei’s absence at a White House UFC event attended by Meta's Mark Zuckerberg. Friction further escalated after Amazon CEO Andy Jassy reportedly demonstrated Fable 5 jailbreak vulnerabilities to senior officials, which Anthropic publicly downplayed as minor and easily discovered without bypasses. Frustration among developers also mounted after Anthropic chose to silently degrade model outputs on frontier AI research queries rather than returning direct refusals.
Meta's Reversal from Token Maxing to Budgets (11:24)
Meta is shifting away from unconstrained 'token maxing' toward strict 'token minimizing' following internal projections that employee AI use would cost billions in 2026. According to a report from The Information, Meta plans to institute structured token budgets and departmental allocations starting in 2027. The hosts compare this transition to traditional marketing allocation and min-maxing, citing an unconfirmed report of an engineer fired after accidentally burning $90,000 in compute loops in a single day.
Fox's $22B Roku Deal Targets Streaming Ads (19:18)
Fox announced an acquisition of connected-TV platform Roku at an enterprise value of approximately $22 billion ($160 per share, structured as $96 in cash and 0.9693 Fox Class A shares), financed by $12 billion in new debt alongside cash on hand. The transaction seeks $400 million in annual cost synergies and pairs Roku’s 25% share of the connected TV OS market with Fox's ad-supported Tubi platform, responding to a broader shift where ad-supported streaming accounts for nearly 50% of premium U.S. subscription video sign-ups.
Policy Hurdles Around AI Nationalization and Sovereignty (26:53)
Drawing on analysis by Tyler Cowen, the hosts review the severe trade-offs facing U.S. export controls and AI national security policy. Tight security curbs and citizenship requirements threaten lab finances, disrupt foreign enterprise revenue, and hinder recruitment of critical international talent like Demis Hassabis, Ilya Sutskever, and Andrej Karpathy. Furthermore, while Chinese open-source models currently lag proprietary leaders due to compute constraints or restricted research networks, the hosts caution that another sudden surge like DeepSeek remains possible.
The hosts leave unresolved whether frontier AI models will ultimately achieve recursive self-improvement that reduces reliance on human labor, or whether national security enforcement and strict compute governance will permanently stall the pace of frontier model commercialization.
## Background
该内容由科技播客频道 TBPN 于 2026 年 6 月 23 日发布在 YouTube 平台上。本期为《Diet TBPN》浓缩版节目，主要涵盖了近期围绕 Anthropic 模型的出口管制风波、Meta 转向“Token 极小化”的内部管控、福斯（Fox）斥资超 200 亿美元收购 Roku 的重磅交易，以及白宫 UFC 赛事与泰勒·考恩（Tyler Cowen）对 AI 政策与国际人才的深度分析。
## TL;DR
节目深入剖析了 Anthropic 因出口管制和越狱争议导致 Fable 5 与 Mythos 5 全面停摆的公关与合规危机，揭示了初创实验室与华盛顿政治圈及亚马逊等云巨头之间复杂的博弈。同时，主持人披露了 Meta 告别盲目挥霍算力、开始推行严格 Token 预算并押注自研代码大模型的战略转变。商业媒体方面，节目解读了福斯豪赌广告支持型流媒体、斥资百亿美元收购 Roku 背后的“平台+产品”协同逻辑。最后，两位主持人结合泰勒·考恩的观点，深入探讨了国籍筛查在 AI 监管中的实操漏洞、AI 实验室国有化的伪命题，以及全球算力地缘竞争的演进。
## Atmosphere & style
整期节目节奏紧凑且充满科技圈内行人的敏锐洞察，两位主持人 John Coogan 与 Jordi Hays 的配合默契，将硬核的技术监管剖析与轻松幽默的行业吐槽无缝衔接。谈话在调侃硅谷流行切片梗（如 Token-maxing）和白大拿带货演出的欢快气氛中开场，随后切入到严肃的国家安全、算力断层与跨国人才枷锁等宏大命题。整体氛围既有深夜科技沙龙的犀利洒脱，又透着面对地缘政治与前沿 AI 激烈碰撞时的冷峻与清醒。
## Bulletin
- [06:01](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=361s) Anthropic 旗下 Fable 5 在发布仅五天后即遭遇美国商务部出口管制禁令，加之复杂的 KYC 审核和下游 API 鉴权隔离难题，迫使其对境内外所有用户全面暂停 Fable 5 和 Mythos 5 的访问。
- [19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) 福斯（Fox Corp）宣布以约 220 亿至 250 亿美元全资收购流媒体平台 Roku，通过将旗下 Tubi 与占全美 25% 市场份额的联网电视操作系统绑定，押注广告支持型流媒体（AVOD/FAST）生态并实现每年约 4 亿美元的成本协同效应。
- [11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) Meta AI 战略出现急转弯，从过度调用算力的“Token 极大化”转向建立刚性核算指标的“Token 预算控制”，并全力推进使用内部高价值代码语料微调的自研工程大模型 Meta Code。
- [27:38](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1658s) 经济学家泰勒·考恩指出前沿 AI 出口管制在实操中存在致命漏洞，严苛的国籍准入壁垒不仅无法防范雇佣美籍内鬼等间谍行为，反而会严重阻碍硅谷大模型实验室对国际顶尖工程人才的招募与跨国商业拓展。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=0s) 主权 AI（Sovereign AI）逻辑在地缘博弈下日益凸显，促使多国政府加速扶持本土国家级模型领军企业以确保技术自主权。
- [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) Mark Zuckerberg 与 Dario Amodei 在白宫 UFC 之夜的出席反差，折射出 Anthropic 的技术安全审美与现任政府风格之间的鸿沟，以及科技实验室在华盛顿政界游说的复杂性。
- [08:03](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=483s) Anthropic 对 Fable 5 采取“静默降低回答质量”的对抗安全防御策略，虽然旨在防止用户利用 AI 制造不受约束的元工具，但这种欺骗式体验严重损害了开发者社区的信任。
- [09:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=555s) 科技巨头与 AI 创企之间存在“莫比乌斯环”式的循环利益网（既是客户、投资人又是云供应商），企业面对政府问询时提交的模型越狱报告更多属于制度合规而非恶意的暗箭中伤。
- [11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) 企业在 AI 算力配置上应回归“数值最优化”（min-maxing）的商业常识与 ROI 考核，避免走向无节制挥霍或过度收紧算力的极端。
- [14:32](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=872s) 流媒体与体育转播正迎来广告与预测平台的肉搏战，Polymarket 与合规对手 Kalshi 围绕赛事八角笼席位与电视流媒体广告位展开了激烈的直接对冲。
- [16:16](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=976s) Meta 内部计划在 2027 年全面收紧 AI 算力额度管理，同时利用其未公开的内部优质代码库加速构建专属编程代理，但资本市场对其口中的“个人超级智能”具体形态依然存在困惑。
- [23:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1404s) 福斯收购 Roku 的核心逻辑在于“基础设施平台”与“内容产品（Tubi）”的强协同，随着近半数流媒体用户接受广告，掌控第一大联网电视入口可捕获万亿级传统电视广告预算转移的红利。
- [23:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1404s) 智能电视端碎片化的“反向捆绑”给用户造成了极大的使用门槛，掌握入口分发能力的聚合平台有望引领内容产业的新一轮“重新打包”与分层窗口期收益分成。
- [27:38](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1658s) 现代 AI 研发与财富 500 强企业的落地严重依赖国际人才与全球调用，推行仅限美国公民的模型访问限制既缺乏实操可行性，也会削弱美国 AI 的外交软硬实力杠杆。
- [29:14](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1754s) 回顾历史上的军工谍战与维吉尔·格里菲斯（Virgil Griffith）北韩加密案，单纯依靠国籍筛查无法彻底防范技术泄漏，且前沿代码知识与国家安全边界的碰撞始终充满张力。
- [30:29](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1829s) 政府直接接管并国有化顶尖 AI 实验室的构想在现实中并不成立，因为模型的底层运维、安全护栏构建及科研品味深度依赖由成千上万顶尖研究员组成的工程体系，无法仅凭递归自我改进（RSI）自主演化。
- [30:29](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1829s) 中国与开源模型相对领先闭源模型暂时落后，主要受制于英伟达 Blackwell 芯片获取限制与前沿核心思路的保密封锁，但算法创新随时可能带来下一次阶跃式的追赶冲击。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=0s) 主权 AI 浪潮、SpaceX 传闻与白宫“UFC 自由 250”之夜
节目开场，主持人 John Coogan 与 Jordi Hays 迅速盘点了当前的宏观与科技动态。Mistral 近期利好不断，主权 AI（Sovereign AI）的逻辑在当前地缘格局下愈发坚挺，让众多国家意识到扶持本土国家级 AI 领军企业、保障长期自主模型供应的紧迫性。同时，Gavin Baker 即将做客节目讨论 SpaceX 的 IPO 前景。不过，当晚最引人注目的还是在白宫举办的“UFC 自由 250”（UFC Freedom 250）格斗盛会。
Jordi 作为铁杆格斗迷，对这项罕见能记住全部选手名字的运动展现出极高的热情。尽管在通胀、油价高企、房贷利率飙升、战争阴云密布等民生艰难的背景下，在白宫大搞狂欢派对的举动最初让人觉得观感不佳，甚至带有浓重的“面包与马戏”（Bread and circuses，意指用娱乐愚弄转移公众对社会危机的注意力）色彩，但这场赛事凭借顶级的制作水准和无可挑剔的阵容彻底征服了他。整场比赛共 8 场对决，毫无注水，场场堪称职业生涯的分水岭，包括重量级拳王战以及即将连线节目的 Bo Nickal。
面对关于在白宫办拳赛是否妥当的质疑，John 提到这并非历史首创——大约一个世纪前，泰迪·罗斯福（Teddy Roosevelt）总统不仅在白宫举办过拳击赛，甚至还亲自下场与人切磋过。从政治策略来看，此类大型娱乐体育活动堪称极具杀伤力却未被充分利用的政治工具，巴西选手 Diego Lopez 在现场的定格画面更成为了当晚的标志性瞬间，整个事件很可能会显著拉升政府民调。
### [04:02](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=242s) 白宫前排的缺席：Anthropic 与政府的审美鸿沟
然而，UFC 白宫之夜背后也折射出微妙的技术政治生态：Meta 首席执行官 Mark Zuckerberg 稳坐头排贵宾席，而 Anthropic 首席执行官 Dario Amodei 却不见踪影。这一对比形象地展示了 Anthropic 所面临的尴尬处境——即便自身在技术上突飞猛进，其品牌审美与当届政府的气质却格格不入。
面对日趋严峻的模型出口管制风波，Anthropic 的核心团队紧急飞往华盛顿特区试图破局。然而在周日当天飞赴华盛顿沟通国家安全大计，正好撞上全城沉浸在 UFC 狂欢之中，时机糟糕至极。对于高度重视网络安全（Cybersecurity）、生物安全（Biosecurity）以及出口管制的严肃技术企业而言，这与拳击派对的画风堪称天差地别。但这道公私裂痕必须尽快弥合：一方面，美国需要让私营科技企业自由蓬勃发展；另一方面，政府亦有权实施合理监管以维护国家安全利益，双方亟需找到相处之道。
```mermaid
flowchart TD
accTitle: Anthropic 监管与发布危机演进路径
accDescr: 展示从3月国防部供应链信件到6月出口管制导致的全面服务暂停全过程
subgraph Timeline ["危机发酵时间线"]
M4["3月4日<br/>战争部/国防部发出<br/>供应链风险信件 (后撤回)"]
A7["4月7日<br/>Mythos 预览发布<br/>并公开 Project Glasswing"]
J9["6月9日<br/>Fable 5 正式发布<br/>(内置网络/生物/AI研究防护护栏)"]
J12["6月12日<br/>商务部签发出口管制禁令<br/>Fable 5 紧急叫停"]
end
subgraph Impact ["合规与运营冲击"]
J12 --> Ban["严格限制范围：<br/>任何外籍人员 (境内/外/内部员工) 均禁止访问"]
Ban --> Suspend["被迫对所有用户全面暂停服务<br/>(因严格 KYC、下游 API 与多租户合规难题)"]
end
```
### [06:01](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=361s) Fable 5 遭遇出口管制禁令与越狱争议
梳理 Anthropic 近期的危机时间线，事件发展极其迅猛：
- **3 月 4 日**：Anthropic 收到来自国防/战争部门的供应链风险函（随后被撤回）。
- **4 月 7 日**：Anthropic 正式公布 Mythos 预览版以及 Project Glasswing 计划。
- **6 月 9 日**：Anthropic 正式推出搭载网络防御、生物安全与 AI 研发护栏的 Mythos 衍生模型 Fable 5。
- **6 月 12 日**：仅发布 5 天后，美国商务部便火速下达出口管制指令，导致 Fable 5 突然被叫停。
根据 Anthropic 的官方博文，政府要求该禁令覆盖包括境内外所有外籍人员（甚至包含其团队内的外籍员工）。由于难以在极短时间内完成各组织机构复杂的 KYC 审查、API 下游鉴权和合规隔离，Anthropic 不得不采取一刀切手段，面向所有用户彻底暂停了 Fable 5 和 Mythos 5 的访问。
雪上加霜的是，《The Information》随后爆料称，亚马逊首席执行官 Andy Jassy 曾亲自向政府高层反映，对 Fable 5 存在的潜在越狱（Jailbreaking）风险表示严重担忧，担心黑客能借此穿透进而获取 Anthropic 的底层系统访问权。对此，Anthropic 发文进行了公开反驳，声明其测试的特定越狱方法仅能识别出极少数已知的次要轻度漏洞，且这些漏洞往往连公开模型都能直接发现，并未构成严重的逃逸威胁。
### [08:03](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=483s) “静默降级”策略引发的反弹与安全逻辑
除了外部监管挤压，Fable 5 自身在产品决策上也引发了开发者社区的巨大不满。尽管模型展示了无与伦比的跑分（Benchmarks）和演示效果，但 Anthropic 却采取了一种怪异的技术控制手段——当用户提出关于前沿 AI 研发相关的问题时，模型并不像面对网络黑客或生物合成提示那样明确拒答，而是选择“静默降低回答质量”（Silently degrade the quality of answers）。
这种“故意降智”的操作让众多从事机器学习、推荐算法以及开源 AI 社区的开发者感到被愚弄和愤怒。然而从对抗安全的闭环逻辑来看，这一做法背后并非毫无逻辑：如果一个超级 AI 受到防御机制限制而不允许用户直接编写黑客攻击代码，但如果它能够帮助用户训练出一个毫无约束的新模型，那么黑客利用这个衍生模型同样可以达成攻击目的——“要限制攻击工具本身，就必须同时限制能够制造攻击工具的元工具”。然而，采用悄悄降级的欺骗式体验，却让 Anthropic 在社区中丧失了大量信任。
### [09:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=555s) 白宫草坪上的算力暗战：Anthropic 与科技巨头的“4D 国际象棋”
如今的政治与商业现实正演变成一种极其奇幻的图景：在白宫草坪举办 UFC 赛事的时代背景下，Anthropic 竟然需要在此类场合为自身在“递归自我改进”（RSI）时代的 AI 安全立场进行辩护。绝大多数大型科技公司的 CEO 们早已摸清了与现任政府打交道的门路——他们频繁出席大型晚宴、参与官方摆拍，偶尔甚至现身私人电影放映会，以便在各种焦点议题上向决策层施加影响力。
更为微妙的是，许多科技巨头与 Anthropic 之间交织着错综复杂的利益纽带。以亚马逊（Amazon）为例，双方的关系堪称经典的“莫比乌斯环”：亚马逊既是 Anthropic 模型的企业级客户，又是其重要投资方，同时还是 Anthropic 底层算力的核心云服务提供商。这种盘根错节的循环共生关系，让外界在评判“亚马逊究竟是在协助 Anthropic 渡过华盛顿的监管风波，还是一切按既定规矩公事公办”时充满疑云。此前甚至有传闻质疑亚马逊是否在背后“落井下石”。但主持人们指出，现实逻辑往往更具制度性：如果联邦政府找上门来询问某家合作巨头：“你们深度调用过这套模型，在安全红队测试中是否成功越狱过？”而该公司的安全团队刚刚出具了确认存在轻微越狱漏洞的报告，没有任何一家上市公司敢对政府隐瞒不报。提交合规报告与主动向监管机构进谗言打压对手有着本质区别。
在这个庞大的生态系统中，科技巨头的政府关系（GR）团队既能化身为初创实验室打通白宫门路的“神队友”，也可能在商业竞争或利益摩擦中悄然在天平另一侧“按上压舱的砝码”。伴随着模型采购协议的频繁重谈与跨多个业务维度的短兵相接，整个夏天华盛顿都将持续上演这场高维度的“4D 国际象棋”。
```mermaid
flowchart TD
accTitle: 亚马逊与 Anthropic 的复杂利益网
accDescr: 展示亚马逊与 Anthropic 之间在算力、资金、模型使用及政府监管互动中的循环关系架构
subgraph AmazonRel ["亚马逊与 Anthropic 的循环生态"]
AWS["AWS 算力供给"] -->|"提供基础设施"| Anthropic["Anthropic (大模型实验室)"]
Anthropic -->|"输出前沿模型"| AmazonUse["亚马逊内部与企业业务采用"]
AmazonCap["资本注入 (投资者身份)"] -->|"资金支持"| Anthropic
end
subgraph DCReg ["华盛顿监管与政府关系"]
Gov["联邦政府与监管机构"]
Gov <-->|"合规问询 / 安全越狱报告"| AmazonGR["亚马逊政府关系团队"]
Anthropic -.->|"寻求政策游说与支持通道"| AmazonGR
Gov -->|"出口管制与 RSI 安全审查"| Anthropic
end
```
---
### [11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) 告别“Token 极值化”：商业常识的回归与单日烧掉 9 万美元的传闻
根据科技媒体《The Information》记者阿米尔（Amir）的独家披露，Meta 正在 AI 战略上进行 180 度大转弯，试图从此前无节制挥霍算力的领头羊，一跃变为“Token 极小化”（Token minimizing）的先锋。这一现象引发了关于科技圈生造词汇的热烈吐槽。所谓“Token 极大化”（Token maxing）与“Token 极小化”本质上借用了游戏圈的用语，随后又被网络上流行的“颜值极值化”（looksmaxxing）等切片梗带偏。但企业经营的底层逻辑从来都不是极度堆砌或极度压缩，而应当是游戏玩家最熟悉的“数值最优化”（min-maxing）。
正如在《英雄联盟》（League of Legends）或角色扮演游戏（RPG）中，玩家需要在有限的资源与金币约束下，精确测算并把每一分钱投入到性价比最高的装备属性上以达成最强收益。然而，硅谷的科技巨头们过去一段时间似乎陷入了一种集体“幻觉”，将最基本的商业回报常识抛之脑后。正如没有哪个成熟企业会在营销中宣称自己这季度要“广告极大化”，把预算毫无节制地砸向市场，待亏损惨重后再走向极端搞“广告极小化”一样，商业决策的核心永远是追求广告投资回报率（ROAS）与最高杠杆的产出。
在这场对算力失控的讨论中，节目提及了一个近期在圈内疯传但未经官方证实的传闻：Meta 内部曾有一位工程师在一天之内调用 AI 接口烧掉了整整 90,000 美元（约合 65 万元人民币），该员工在短短 3 天内便被迅速解雇。虽然这可能属于某种被夸大的办公室传言，但它生动反映了当前科技大厂基层滥用算力与自动化死循环的疯狂现状。
- **超级碗广告的 ROI 参照**：投放超级碗广告通常要求单日烧掉 500 万美元，但这同样讲求“数值最优化”。有策略的品牌会选择在特定细分市场或赛前低价时段切入，以低成本撬动核心受众的高品牌价值；而盲目“超级碗极大化”砸钱的品牌（包括部分反响平平的纯 AI 生成广告），最终根本无法收回 500 万美元的品牌溢价。
- **失控的算力账单**：基层员工在缺乏配额监控的环境下，可能出于“我要证明自己比全公司任何人都懂 AI”的心态疯狂开挂跑脚本，最终因数以万计的 Token 循环账单引火烧身。
---
### [14:32](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=872s) UFC 擂台广告大乱斗：白大拿的带货与预测市场的正面交锋
从算力战场的商业变现延伸开来，刚刚结束的白宫草坪 UFC 赛事展现了一场教科书级别的营销狂欢。伴随着 UFC 与派拉蒙（Paramount）达成的全新转播合作协议，现场广告的密集程度刷新了观众的认知，不仅在对决间隙高频插播完整商业广告，赛事总裁白大拿（Dana White）更是亲自深度参与多支广告的演出。
白大拿不仅亲自驾驶道奇 TRX 越野卡车上镜拍摄性能广告，甚至还在一支企业商业保险广告中与该保险公司 CEO 面对面坐在一组伪装成专业播客的布景前，煞有介事地大谈特谈商业保险的重要性。这种将赛事核心操盘手直接塑造成推销主角的模式极具辨识度。与此同时，八角笼擂台内部的广告赞助则彻底沦为博彩与预测平台的天下。从 Crypto.com 到 Bet365，再到基于区块链的预测市场 Polymarket，各类竞猜平台塞满了视线——甚至引发了主持人的调侃：“Bet365 简直是在倡导一种除了四年一次的闰日（Leap Day）能歇一天外，剩下 365 天无休止赌博的生活方式。”
更为精彩的是预测市场领域的直接肉搏战。尽管 Polymarket 成功买断了八角笼地面上的核心赞助席位，但其主要受监管竞争对手 Kalshi 则通过派拉蒙转播流直接购买了大量的电视与流媒体广告位，在比赛转播期间对 Polymarket 实施“贴脸轰炸”。由于派拉蒙流媒体订阅用户无需额外付费购买单次点播（PPV）即可直接观看 UFC，这种高密度的精准广告对冲将两家竞品在去中心化与合规博彩领域的争夺推向了白热化。
---
### [16:16](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=976s) Meta 勒紧 Token 裤腰带：内部代码大模型与“个人超级智能”迷思
就在两个月前，Meta 还是硅谷“Token 极大化”的最激进代表，其内部仅采购 Anthropic 的 Claude 等外部前沿模型的年化预算就逼近数十亿美元的轨道。然而，最新泄露的内部备忘录清晰勾勒出了管理层的政策转向：
> “我们目睹了内部 AI 使用量的指数级暴增，按照目前的轨迹，仅 2026 年用于公司内部日常运营的 AI 支出就将达到数十亿美元。与此同时，各个独立业务团队与员工个体对自身如何消耗 AI 算力缺乏清晰的可见度与管控机制。进入 2027 年，Meta 预计将转向更加结构化的 AI Token 管理模式，设立明确的 Token 预算指标、配额分配决策机制以及配套的监控工具。”
这一转变标志着算力消耗正式从“无限自助餐”模式收归为类似于传统营销预算的刚性核算科目。与此同时，Meta 正在加紧推进代号为“Meta Code”的内部专属编程模型研发。扎克伯格在近期外泄的内部会议录音中多次强调：Meta 内部顶尖软件工程师的代码质量理应远高于外部开源社区的平均贡献水平。
Meta 内部拥有完全原生、无需繁琐扫描数字化的海量高价值工程代码库与业务工作流数据。通过将自身沉淀的顶级代码资产作为微调语料输入前沿模型，Meta 试图训练出能够全自动重构业务流程、专门用于优化旗下庞大社交网络底层架构的自研工程代理。然而，面对扎克伯格口中宏大而飘渺的“个人超级智能”（Personal Superintelligence）愿景，外界与资本市场依然对其核心落地形态充满困惑，Meta 显然需要在短期内向市场更清晰地证明其千亿算力投资背后的最终商业产品到底是什么。
### [17:59](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1079s) Meta 的 AI 投资逻辑与战略迷雾
两位主持人从估值话题聊起，调侃某家公司的估值已经达到了 SpaceX 的 75%，而 SpaceX 的估值已然高达 2.3 万亿美元。顺着这个思路，他们将 Meta 与 SpaceX 进行了对比：虽然 xAI 建造 Colossus 超算集群的速度可能稍快一些，但 Meta 的追赶步伐同样极其惊人。然而，Meta 的多头逻辑与 SpaceX 截然不同。在主播看来，Meta 极有可能是商业史上最伟大的商业模式之一——不仅完全没有受到 AI 的颠覆威胁，反而正被 AI 全面加速赋能。
尽管基本面强劲，Meta 目前对外释放的信号却显得有些含混不清。扎克伯格和高管层一直把“个人超级智能（Personal Superintelligence）”挂在嘴边，但在实际业务动作上，Meta 正在大举构建代码测试框架（coding harness）并全力攻坚代码能力。这引发了业内的激烈猜测：Meta 究竟是打算将这些编程模型与 AI 能力作为企业级 API 对外商业化变现，还是仅仅留作内部提效工具？此前他们在 Manus 类似项目上试水后又有所收缩，市场正密切期待 Meta 在接下来的一个月内明确给出真正的商业化路径。
- **SpaceX 估值参照**：谈话中提及 SpaceX 当前估值约为 2.3 万亿美元。
- **Meta 的核心多头逻辑**：作为现有业务不会被 AI 摧毁、反而直接受益于 AI 广告定向与用户时长提升的超级巨头。
- **战略疑问**：Meta 在重注代码大模型与底层设施后，是否会真正迈出向企业级软件与开发者 API 售卖服务的大步。
---
### [19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) 福斯斥资巨资收购 Roku：流媒体免费广告生态的世纪豪赌
演播室的话题突然被突发快讯打断：《华尔街日报》报道传媒巨头福斯公司（Fox Corp）正斥资约 250 亿美元全资收购流媒体硬件与平台巨头 Roku。针对具体交易金额，主持人当场刷新页面核实，指出 Axios 和 CNBC 报道的企业价值约为 220 亿美元，而《华尔街日报》作为同属默多克家族新闻集团（News Corp）旗下媒体，可能提前拿到了最一手的交易细节。这一收购宣告了福斯在广告支持型流媒体（AVOD / FAST）领域孤注一掷的野心。正如主播所总结的，“广告永远立于不败之地”（"Ads are undefeated."）——在 Netflix 被迫打破不插播广告的诺言并大获成功后，事实证明消费者只要有优质内容看，完全愿意接受广告。
这笔交易是福斯历史上规模最大的一笔并购，将老牌新闻与体育直播帝国与全美最大的联网电视（Connected TV, CTV）系统平台紧密绑定。收购条款十分具体：福斯将以每股约 160 美元的价格收购 Roku，其中包含 96 美元现金与 0.9693 股福斯 A 类普通股，现金部分将通过 120 亿美元的新增债务和账面现金共同支付。福斯 CEO 拉克兰·默多克（Lachlan Murdoch）在投资者电话会上豪言：“将这两家公司结合在一起，将真正帮助定义美国以及全球众多其他市场的电视未来。”新合并的实体预计每年可削减约 4 亿美元的运营成本，同时双方旗下的 Tubi 和 Roku Channel 将继续保持独立运营。
```mermaid
flowchart TD
accTitle: Fox Acquisition of Roku Structure
accDescr: Diagram illustrating Fox Corp's acquisition of Roku and its combined advertising and distribution ecosystem.
subgraph FOX["福斯集团 (Fox Corp)"]
direction TB
F_Content["直播内容资产<br/>(Fox Sports / Fox News)"]
F_Tubi["Tubi<br/>(2020年4亿买入/月活近1亿)"]
F_Sub["订阅服务<br/>(Fox One / Fox Nation)"]
end
subgraph DEAL["交易结构 (~$22B-25B)"]
direction TB
D_Terms["每股约 $160<br/>($96 现金 + 0.9693 股 Fox A类股票)"]
D_Debt["融资渠道<br/>(120亿新债 + 账面现金)"]
D_Synergy["预期年削减成本: 4亿美元"]
end
subgraph ROKU["Roku 平台生态"]
direction TB
R_OS["联网电视操作系统<br/>(全美市占率 25% 第一)"]
R_Channel["The Roku Channel<br/>(自营免费广告频道)"]
R_Share[">40% CTV 观看时长与流量"]
end
FOX --> DEAL
DEAL --> ROKU
```
- **交易核心数据**：收购估值在 220 亿至 250 亿美元之间，每股折合约 160 美元（96 美元现金 + 0.9693 股股票），福斯计划举债 120 亿美元完成支付。
- **流媒体趋势逆转**：随着无广告纯订阅价格节节攀升，全美广告支持型流媒体套餐签约占比已从两年前的 39% 激增至近 50%。
- **硬件与操作系统渗透率**：全球超过 1 亿户家庭使用 Roku；在联网电视市场 Roku 占有 25% 份额位居第一，三星 Tizen 以 23% 位列第二，亚马逊 Fire TV 紧随其后。
- **Tubi 的逆袭战绩**：福斯在 2020 年仅以 4 亿美元收购 Tubi，如今已成长为月活近 1 亿、单季度营收同比大增 23%、年收入破十亿美元的现金牛业务。
---
### [23:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1404s) 平台对决产品：重构联网电视的广告霸权与内容重新打包
针对华尔街在交易宣布后福斯股价走低、市场评论人直言“完全看不懂这笔交易逻辑”的消极情绪，主播展开了坚决的反驳。他认为外界低估了“产品”与“平台”结合的巨大协同效应：Tubi 是一项高收益的产品，而 Roku 则是一座基础设施级的平台。Roku 不仅拥有 25% 的终端市场占有率，更拿下了联网电视生态中超过 40% 的实际观看时长。传统线性电视广告预算向联网电视迁徙的万亿浪潮才刚刚开始，Roku 坐拥极高的年轻受众比例，这些用户正处于最具购买力的黄金年龄，对于广告主而言是不可替代的流量沃土。
两位主持人还深入探讨了消费者端的内容痛点与院线窗口期（Windowing）的演变。如今用户在智能电视上面对的是极度碎片化的“反向捆绑”噩梦——买一台新电视需要繁琐登录十几个流媒体应用。福斯拿下 Roku 后，有望主导新一轮的“重新打包（Rebundling）”，为用户提供一站式聚合体验。此外，过去几年好莱坞片库在 Netflix、Disney+ 与 HBO 之间打起排他性阵地战，而手握巨大入口流量的 Roku 有机会成为中立的内容分发分层平台，吸纳主流大片在二级窗口期上映，从而在中长尾广告和高阶内容分发中赚取巨额分成。
---
### [26:53](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1613s) 泰勒·考恩解析 Mythos 风波与美国 AI 政策的多重枷锁
对话最后转向著名经济学家泰勒·考恩（Tyler Cowen）针对近期 Anthropic 模型“Mythos”下架风波发表的最新深度评述。尽管官方尚未公开确认，但多方消息显示 Mythos 之所以被紧急召回，是因为模型即使已经针对生物武器制造与自我递归改进进行了极其严苛的限制，依然被成功越狱并造成了巨大的安全恐慌。
考恩直言，美国政府在处理前沿 AI 监管时陷入了极其被动的多难境地。监管层既不能勒死行业创新，又必须维持这批核心 AI 创业公司的生存，并期盼它们能够顺利走向公开上市（IPO）。然而，日趋严苛的安全与出口限制正带来严重的副作用——尤其是在硅谷顶级大模型研发严重依赖外籍高精尖研发人才的当下，招募外籍科研人员的政策阻碍正让顶尖 AI 实验室陷入严重的人才焦虑。
### [27:38](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1658s) 跨国人才与模型出口壁垒：泰勒·考恩的冷酷现实
主持人深入剖析了经济学家泰勒·考恩（Tyler Cowen）针对 Anthropic 的 Mythos 争议及 AI 出口管制提出的核心论点。考恩强调，美国顶尖的 AI 研发队伍在本质上高度依赖国际人才——从 Demis Hassabis、Ilya Sutskever 到 Andrej Karpathy，当今人工智能领域的几位执牛耳者最初全都是非美籍工程师。如果美国政府推行严厉的身份准入壁垒，美国 AI 巨头将极难在海外建立具有公信力且可持续的商业合作。尽管市场上鲜有关于国内外营收拆分的详尽图表，但 OpenAI 与 Anthropic 频频与全球跨国企业签订巨额订单已是不争的事实。这些财富 500 强企业的业务触角遍布全球，除非它们将所有的软件工程团队全部锁死在硅谷，否则不可避免地需要在海外各个节点调用顶尖模型。
在念到考恩提出的“跨国美企在海外运营时究竟该如何使用顶尖 AI 系统”这一条目时，主持人忍不住笑场，调侃自己刚才的分析居然完美命中了下一句话，直呼自己变成了“一只随机鹦鹉（stochastic parrot）”。但玩笑之余，地缘政治的博弈逻辑依然冰冷严峻：美国既希望将前沿模型访问权限作为硬实力与软实力的外交杠杆工具，又深知一旦外国代理人获得部分模型权限，其后续用途便极难掌控。针对外界呼吁的“模型访问权限仅限美国公民”构想，考恩直言其在实操中漏洞百出——在现实中，恶意势力“花钱雇佣一个叛变的美籍人员（hire a traitorous American）来获取犯罪工具”轻而易举，伪造公民身份在技术上也并非不可逾越的鸿沟。
- **核心人才国际化**：Demis Hassabis（DeepMind）、Ilya Sutskever（SSI / 原 OpenAI）、Andrej Karpathy（Eureka Labs / 原 OpenAI）等业界标杆人物均为跨国人才，严苛的国籍限制将重创顶尖工程团队的招募。
- **跨国商业现实**：全球化企业的软件工程团队分布在世界各地，无法仅将模型访问局限在美国本土基础设施内。
- **监管不可行性**：依靠“仅限美国公民”来实行安全限制缺乏可操作性，内鬼收买与国籍伪造均是极易被利用的漏洞。
### [29:14](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1754s) 间谍、出口管制与维吉尔·格里菲斯的北韩加密案
针对国籍筛查在防谍维安上的无力，主持人回顾了冷战及过去半个世纪以来的惨痛教训：大量拥有合法美国公民身份的人员因出卖绝密军用飞机技术等国防机密而接连落网。为了进一步说明前沿技术出口管制的两难困境，主持人引申出了加密货币圈极具争议的真实刑案——维吉尔·格里菲斯（Virgil Griffith）事件。主持人回忆起自己曾与格里菲斯有过一面之缘，称其早年在加州理工学院（Caltech）潜心从事数学研究，对加密技术的去中心化愿景怀有近乎执拗的理想主义热忱。
然而，正是这位学者型研究员专程飞赴朝鲜平壤，在当地举办的会议上面向朝方人员讲授加密货币与密码学的运行机制，最终招致美国联邦政府的严厉制裁与牢狱之灾。格里菲斯的入狱在加密货币社区掀起了轩然大波，开发者与密码学支持者们义愤填膺地抗议“数学不应被定罪，信息传播不该违法”。但在美国司法部门眼中，问题的关键在于他的行为教导了受制裁政权如何利用代码构建一套“完全无法被西方制裁拦截的价值存储手段（a store of value that is unsanctionable）”。主持人幽默地吐槽这反而让某种代币成了“受制裁最严重的稳定币”，但也深刻揭示出前沿技术代码与国家主权防线正面碰撞时的极端复杂性。
- **军工历史镜鉴**：过去五十年间，出卖军用航空技术绝密情报的多为货真价实的美国本土公民，“血统论”式的安全防线在利益面前形同虚设。
- **维吉尔·格里菲斯先例**：加州理工背景的密码学研究员因赴朝普及区块链与加密技术，被美国政府认定为传授规避经济制裁的手段，构成违反出口管制罪名并判刑。
- **学术自由与国安冲突**：加密社区坚称“数学与开源知识无罪”，而政府监管核心在于技术是否向敌对势力赋予了不可管辖的金融与计算硬实力。
### [30:29](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1829s) AI 国有化困局、递归自我改进与中国开源模型的追赶步伐
话题随后切入考恩备受争议的第五条论断：“美国政府无法将这些顶尖 AI 实验室收归国有，并在接管后实现有效运营。”围绕这一设想，两位主播探讨了硅谷内部针锋相对的两种阵营。其中极端的一派被称为“深服递归自我改进药丸者（super RSI pilled）”——该派别认为，若强人工智能具备完全自我迭代与进化的能力，政府接管后甚至根本不再需要保留庞大的研发雇员队伍，官僚们唯一需要做的就是向系统下达一条“请继续自我提升”的指令，模型便会自行飞轮运转。
但现实主义派对这种乌托邦假设嗤之以鼻。真实的 AI 系统结构庞杂至极，从维持极高可用性的系统底层运维、严密的安全护栏构建，到极难量化的“研究品味（research taste）”，处处都需要成千上万顶尖工程师的精雕细琢，其中更不可或缺那些带来前沿灵感的国际天才。
最后，主播谈及考恩的第六点判断：中国及开源模型虽然仍在以合理步伐推进，但目前仍明显落后于美国顶尖闭源模型。针对近期业界关于“中国 AI 追赶节奏是否正在放缓”的争论，主播梳理了多重制约假说：知名科技投资人 Gavin Baker 曾分析其核心在于高端算力断层，特别是放弃或无法获取英伟达最新的 Blackwell 芯片；另一重阻力在于旧金山核心圈层对顶尖研究思路的保密封锁日益加剧；此外数据反馈闭环与工程落地的差异也可能拉开差距。不过，主播对此依然保持敬畏，坦言技术曲线随时可能产生突变，“或许未来几个月内，我们就会迎来下一个 DeepSeek 级别的冲击时刻”。
```mermaid
flowchart TD
accTitle: 国有化与开源追赶的核心争议分析
accDescr: 对比AI国有化可行性的两种极端观点，以及中国开源模型落后与追赶的驱动因素。
subgraph NationalizationDebate["AI 实验室国有化之争"]
RSI["极度 RSI 假设<br/>（递归自我改进）"] --> AutoIter["完全自主迭代<br/>官僚仅需下令『继续进化』"]
Reality["工程现实派观点"] --> Complexity["系统工程复杂性<br/>高可用性运维 + 护栏调试"]
Reality --> HumanTaste["不可替代的人类要素<br/>研究品味与国际化顶尖雇员"]
end
subgraph ChinaOpenSource["中国与开源模型追赶瓶颈"]
Bottleneck["当前发展阻力"] --> Compute["算力短缺：Blackwell 获取受限"]
Bottleneck --> KnowledgeLock["认知壁垒：硅谷前沿圈层技术封锁"]
Bottleneck --> FeedbackLoop["工程与数据推断闭环差距"]
Compute -.-> Surprise["潜在破局：下一个 DeepSeek 跃迁时刻"]
end
```
- **国有化可行性分歧**：RSI 理论假设超智能可脱离人力自我进化，但现实工程落地高度依赖工程师团队的精细运维和科研敏锐度。
- **算力与算法封锁**：中国模型追赶放缓的推测原因包括 Blackwell 等新一代算力受制于禁令、旧金山研发前沿的信息锁闭，以及海量推断数据回流效率的差异。
- **动态博弈变数**：闭源与开源的差距并非铁板一块，业界仍需高度戒备突发性的算法突破与架构创新带来的阶跃式赶超。
## Notable quotes
> [08:03](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=483s) 要限制攻击工具本身，就必须同时限制能够制造攻击工具的元工具
> [09:15](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=555s) 你们深度调用过这套模型，在安全红队测试中是否成功越狱过？
> [11:24](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=684s) 我要证明自己比全公司任何人都懂 AI
> [14:32](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=872s) Bet365 简直是在倡导一种除了四年一次的闰日（Leap Day）能歇一天外，剩下 365 天无休止赌博的生活方式。 — John Coogan / Jordi Hays
> [16:16](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=976s) 我们目睹了内部 AI 使用量的指数级暴增，按照目前的轨迹，仅 2026 年用于公司内部日常运营的 AI 支出就将达到数十亿美元。与此同时，各个独立业务团队与员工个体对自身如何消耗 AI 算力缺乏清晰的可见度与管控机制。进入 2027 年，Meta 预计将转向更加结构化的 AI Token 管理模式，设立明确的 Token 预算指标、配额分配决策机制以及配套的监控工具。 — Meta 内部备忘录
> [19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) 广告永远立于不败之地 — John Coogan / Jordi Hays
> [19:18](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1158s) 将这两家公司结合在一起，将真正帮助定义美国以及全球众多其他市场的电视未来。 — Lachlan Murdoch
> [27:38](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1658s) 花钱雇佣一个叛变的美籍人员（hire a traitorous American）来获取犯罪工具 — Tyler Cowen
> [29:14](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1754s) 数学不应被定罪，信息传播不该违法
> [29:14](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1754s) 完全无法被西方制裁拦截的价值存储手段（a store of value that is unsanctionable）
> [30:29](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1829s) 美国政府无法将这些顶尖 AI 实验室收归国有，并在接管后实现有效运营。 — Tyler Cowen
> [30:29](https://www.youtube.com/watch?v=-22Gjx3z20Y&t=1829s) 或许未来几个月内，我们就会迎来下一个 DeepSeek 级别的冲击时刻 — John Coogan / Jordi Hays
## Mentioned
John Coogan, Jordi Hays, Mark Zuckerberg, Dario Amodei, Andy Jassy, Tyler Cowen, Lachlan Murdoch, Dana White, Gavin Baker, Virgil Griffith, Demis Hassabis, Ilya Sutskever, Andrej Karpathy, Bo Nickal, Diego Lopez, Teddy Roosevelt, Anthropic, Meta, Fox Corp, Roku, Amazon, AWS, SpaceX, xAI, Mistral, DeepSeek, OpenAI, DeepMind, UFC, Paramount, Polymarket, Kalshi, Bet365, Crypto.com, Netflix, Tubi, The Roku Channel, Mythos, Fable 5, Project Glasswing, Meta Code, Nvidia Blackwell, 美国商务部, 美国国防部
## Video description
Sign up for TBPN’s daily newsletter at TBPN.com
Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after.
Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella.
TBPN is made possible by:
Ramp - https://ramp.com
Public - https://public.com
Cisco - https://www.cisco.com
Console - https://www.console.com
CrowdStrike - https://www.crowdstrike.com
Figma - https://www.figma.com
MongoDB - https://www.mongodb.com
NYSE - https://www.nyse.com
Railway - https://railway.com
Shopify - https://www.shopify.com/
Follow TBPN:
https://TBPN.com
https://x.com/tbpn
https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231
https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235
https://www.youtube.com/@TBPNLive
Chapters:
00:00 Mistral, Sovereign AI + SpaceX IPO Tease
00:40 UFC Freedom 250 at the White House
04:02 Anthropic Fable/Mythos Fallout + Export Controls
11:24 Meta Token Budgets, UFC Ads + Personal Superintelligence
19:18 Fox Buys Roku + The Ad-Supported Streaming Bet
26:53 Tyler Cowen on Mythos, Foreign Nationals + Open Source China
[Source](https://www.youtube.com/watch?v=-22Gjx3z20Y)
Source quotes
[8:03] 要限制攻击工具本身，就必须同时限制能够制造攻击工具的元工具
[9:15] 你们深度调用过这套模型，在安全红队测试中是否成功越狱过？
[11:24] 我要证明自己比全公司任何人都懂 AI
[14:32] John Coogan / Jordi Hays: Bet365 简直是在倡导一种除了四年一次的闰日（Leap Day）能歇一天外，剩下 365 天无休止赌博的生活方式。
[16:16] Meta 内部备忘录: 我们目睹了内部 AI 使用量的指数级暴增，按照目前的轨迹，仅 2026 年用于公司内部日常运营的 AI 支出就将达到数十亿美元。与此同时，各个独立业务团队与员工个体对自身如何消耗 AI 算力缺乏清晰的可见度与管控机制。进入 2027 年，Meta 预计将转向更加结构化的 AI Token 管理模式，设立明确的 Token 预算指标、配额分配决策机制以及配套的监控工具。
[19:18] John Coogan / Jordi Hays: 广告永远立于不败之地
[19:18] Lachlan Murdoch: 将这两家公司结合在一起，将真正帮助定义美国以及全球众多其他市场的电视未来。
[27:38] Tyler Cowen: 花钱雇佣一个叛变的美籍人员（hire a traitorous American）来获取犯罪工具
[29:14] 数学不应被定罪，信息传播不该违法
[29:14] 完全无法被西方制裁拦截的价值存储手段（a store of value that is unsanctionable）
[30:29] Tyler Cowen: 美国政府无法将这些顶尖 AI 实验室收归国有，并在接管后实现有效运营。
[30:29] John Coogan / Jordi Hays: 或许未来几个月内，我们就会迎来下一个 DeepSeek 级别的冲击时刻
Source quotes
[4:02] If you take safety seriously, cybersecurity seriously, biosecurity seriously, export control seriously... UFC could not be less aligned with that. The stakes there... there is this big gap between the administration and Anthropic.
[15:15] Gambling is just min-maxing done poorly.
Original transcript
[0:00] Obviously, Mistral is in the news.
[0:02] Incredibly bullish for Mistral.
[0:05] Obviously, very unfortunate drama
[0:08] happening in the Anthropic world. We'll
[0:09] go through all of that. But, if the if
[0:12] the sovereign AI thesis could not get
[0:15] any stronger right now, this weekend
[0:17] really showed
[0:19] a lot of countries, "Hey, maybe we do
[0:21] want a national champion. Hey, maybe we
[0:23] do want someone that will provide us
[0:25] models forever." SpaceX IPO, we have
[0:27] Gavin Baker joining in 30 minutes, and
[0:31] we're recapping UFC Freedom 250. It
[0:33] happened at the White House last night.
[0:35] Jordi, take us through your reactions.
[0:37] Introduce our first guest to the show.
[0:39] >> Uh, okay. Quick intro on the event
[0:41] overall.
[0:43] I am a UFC fan.
[0:45] It's one of the few sports, maybe the
[0:47] only sport, where I actually know the
[0:49] names of the people that are competing.
[0:51] >> Mhm.
[0:51] >> Um, our guest in just a few minutes is
[0:54] is one of those people.
[0:55] >> you more of a fan of the UFC or more of
[0:57] a fan of the government? Cuz a lot of
[0:59] people were watching just cuz they're
[1:00] fans of the government, and they're and
[1:02] they're said, "Oh, I will watch anything
[1:04] I watch C-SPAN all day long. I'll watch
[1:06] anything that happens if it involves the
[1:08] government." This was an interesting
[1:10] cross
[1:10] >> No, more
[1:12] more seriously, I as a UFC fan
[1:16] >> Mhm.
[1:17] >> and as an American,
[1:18] >> Mhm.
[1:18] >> I'm excited for the for the UFC to do
[1:20] well. I'm excited for the fighters to
[1:22] have a platform. I was a little bit wary
[1:24] of this event overall just because uh,
[1:27] the aesthetics of putting on this crazy
[1:29] party
[1:31] >> in this moment when there is a ton to
[1:34] black pill about.
[1:35] >> Yeah.
[1:35] >> Not after less less after yesterday.
[1:37] >> all this other stuff going on.
[1:39] >> And yeah, so war raging, gas prices are
[1:42] high, house prices are high, rates were
[1:44] going up.
[1:45] >> a party.
[1:46] >> of people in America are not having a
[1:48] good time. And it really felt like a
[1:50] bread and circuses moment.
[1:52] >> Sure.
[1:52] >> But, the bread and circuses, John,
[1:55] it worked on me.
[1:56] >> Top notch.
[1:57] >> on me. Some of the best ever. It was uh
[1:59] it was
[2:00] >> Why was it so good compared to a normal
[2:01] UFC event?
[2:02] >> Well, you didn't watch, so I guess it's
[2:04] uh I guess it's hard. Uh but if
[2:07] >> or the actual card?
[2:08] >> one, the production value one, the card
[2:09] was absolutely stacked. It was eight
[2:12] fights, no filler.
[2:13] >> Yeah.
[2:14] >> Wall to wall.
[2:15] >> Yeah.
[2:15] >> Uh really really
[2:17] uh fights that had a lot of um they
[2:20] carried a lot of weight, right? So, like
[2:21] the the main uh title fight, Gaethje
[2:25] versus Oliveira Topuria, uh that changed
[2:28] both fighters' careers.
[2:29] >> Mhm.
[2:29] >> Like full stop, right? Uh and even with
[2:32] Bo Nickal joining
[2:34] uh
[2:35] uh in just a few minutes, you know, Bo,
[2:37] that was the biggest possible stage
[2:39] >> Yep.
[2:39] >> that any athlete could be on, and
[2:41] especially in the UFC, at least here in
[2:43] the in America. So, anyways, going into
[2:45] it, I thought the aesthetics were bad. I
[2:48] was excited to watch the fights.
[2:49] >> Yep.
[2:49] >> Turned out it was just really cool. Uh
[2:52] and sure, there's still going to be
[2:54] people that are upset that the the that
[2:56] the the president uh and the admin is
[2:59] like platforming something like this.
[3:01] Turned out to be really fun. It was it
[3:02] was a cool night. Uh and I think America
[3:05] kind of needed it.
[3:06] >> Yeah.
[3:06] >> Uh and
[3:07] >> are saying that it's not unprecedented
[3:09] because Teddy Roosevelt held a boxing
[3:12] match in the White House uh almost 100
[3:14] years ago, round that time. And
[3:16] apparently, he fought in the boxing
[3:18] match, which is
[3:19] >> Yeah, he was big big boxer.
[3:21] >> He was a big boxer. And so,
[3:22] >> Imagine Imagine
[3:24] >> So, this is like a live
[3:25] >> your punches with the president.
[3:26] >> Uh anyway,
[3:27] >> But anyways, uh yeah. I you know, the
[3:30] the event was was incredible, and I came
[3:33] away thinking that big parties and
[3:35] sporting events are potentially the most
[3:37] underutilized political strategy, right?
[3:40] I would actually expect
[3:42] uh the the admin's like overall
[3:44] popularity
[3:46] >> to go up.
[3:46] >> approval rating to go up based on
[3:48] >> Interesting.
[3:49] >> uh based just on this event, cuz there
[3:50] were so many iconic images. Funny
[3:52] enough, the most iconic image is of a
[3:54] Brazilian named Diego Diego Lopez, which
[3:57] we had up on the screen there.
[3:58] >> Standing up on the
[3:59] >> But uh but anyways, uh really really
[4:02] wild.
[4:03] >> So, the other big uh topic over the UFC
[4:07] was the fact that Dario Amodei was not
[4:09] in attendance. He wasn't front row. Mark
[4:11] Zuckerberg was. And it sort of
[4:14] underscores this odd situation that
[4:16] Anthropic's in where like they just got
[4:18] dealt this weird hand where like they're
[4:21] doing so well and the admin could not be
[4:24] more misaligned aesthetically with
[4:26] Anthropic. It's
[4:27] >> Their their whole brand.
[4:28] >> Yeah, but the whole brand. They're very
[4:29] very different.
[4:30] >> And and I was, you know, I read the
[4:32] article yesterday that that Anthropic
[4:34] staff were going to
[4:36] >> Yeah. DC.
[4:38] >> And it's funny, some people were like,
[4:39] "Oh, this this this is you know, this is
[4:40] bad. Like if it was a they could if they
[4:43] could solve the issue easily, they could
[4:44] they could do it over the phone." Not
[4:46] how this works.
[4:46] >> You always
[4:47] >> Serious issue. Show you know, the yeah,
[4:49] you would obviously to go. Um but going
[4:51] to Washing flying to Washington DC to
[4:53] try to solve an issue on a Sunday, there
[4:56] was not a worse time to do it.
[4:58] No, no, I know. But you can imagine you
[4:59] land there, you're like, "Okay, like
[5:01] [laughter] where is everyone? Let's
[5:02] let's like meet. There's a big issue.
[5:04] Let's meet. Figure this out. This is
[5:05] important. National security is on the
[5:07] line. Like let's figure this out."
[5:09] >> And they're like, "Where is everyone?"
[5:10] >> And like if you take safety seriously,
[5:12] cybersecurity seriously, biosecurity
[5:15] seriously, export control seriously,
[5:17] like like UFC could not be less aligned
[5:20] with that. The stakes there. Uh and so
[5:23] like there is this big gap between the
[5:26] administration and Anthropic. Uh and I
[5:28] think that everyone should like every
[5:31] American should care about bridging this
[5:33] gap. Like private enterprises should be
[5:35] able to flourish in the United States.
[5:36] Everyone agrees with that. And the
[5:37] government should be able to enact
[5:39] reasonable regulations to protect
[5:41] America's interests. And both of those
[5:43] need to find a way to play together, but
[5:45] there is just a really really broad gap
[5:48] between these two organizations. And And
[5:50] so, everyone's hoping that things
[5:51] resolve quickly. And it seems like the
[5:54] right, you know, wheels are in motion.
[5:55] The flights have been booked or already
[5:57] taken, and things are moving along. But
[5:59] just a recap on the play-by-play. March
[6:01] 4th, Anthropic received a depart-
[6:04] Department of War Supply Chain Risk
[6:06] Letter. That, of course, got rolled back
[6:08] eventually. April 7th was was the Mythos
[6:10] preview announcement alongside Project
[6:13] Glass Wing. June 9th was the Fable 5
[6:15] launch last Monday. Mythos with
[6:17] guardrails around cyber, bio, and AI
[6:20] research. And on June 12th, just 5 days
[6:23] after the end of the week, Fable 5 gets
[6:25] suspended after the Commerce Department
[6:27] issues an export control directive. So,
[6:29] Anthropic explained in a blog post that
[6:32] the US government is restricting both
[6:33] Fable 5 and Mythos 5 from being used by
[6:36] any foreign national, whether inside or
[6:39] outside the US, including foreign
[6:41] national Anthropic employees. And so,
[6:43] that's a huge compliance burden, very
[6:46] sudden. And that led Anthropic to
[6:50] completely suspend the use of those
[6:51] models for all users,
[6:54] not just
[6:55] not just foreign national employees,
[6:57] because they need to, at least if they
[6:59] want to comply with this, do all the
[7:00] KYC, understand how different
[7:02] organizations are using the model, who
[7:04] at those organizations has access, if
[7:06] something's being vended through an API,
[7:08] and it's showing up in another product,
[7:11] does that product do all of the proper,
[7:14] you know, authorizations? And so, it's a
[7:16] big it's a big wrench in the gears of
[7:19] the rollout of this model. Then the
[7:20] information reported that Andy Jassy
[7:22] Amazon had raised concerns with the
[7:24] senior administration officials about
[7:26] jailbreaking, the risk of jailbreaking
[7:27] Fable, and getting access to all of
[7:30] Anthropic. Anthropic countered in their
[7:31] blog post saying, quote, "We reviewed a
[7:34] demonstration of this specific technique
[7:36] being used to identify a small number of
[7:38] previously known minor vulnerabilities.
[7:41] These vulnerabilities all appear
[7:42] relatively simple, and we have found
[7:45] that other publicly available models are
[7:47] able to discover them as well without
[7:49] requiring a bypass. So, they're saying
[7:51] this isn't that big of a jailbreak. This
[7:54] isn't that big of a deal. We can move
[7:55] forward safely. We've taken safety
[7:57] seriously. But, that is the debate
[7:59] that's going on between the admin and
[8:01] Anthropic. So, as everyone knows, the
[8:03] the rollout of Fable 5 has been a bit
[8:05] rocky. Incredible demos, incredible
[8:07] benchmarks, but offset by the odd
[8:09] decision to silently degrade the quality
[8:11] of answer answers related to frontier AI
[8:13] development. Instead of just refusing
[8:16] the request like cyber and bio prompts,
[8:18] which that was the main thing that
[8:20] everyone was really confused about. The
[8:22] actual rationale behind AI development
[8:24] refusal is pretty sound. Like, people
[8:26] are upset about that if they're working
[8:28] on machine learning systems, recommender
[8:30] systems, anything that requires, you
[8:32] know,
[8:33] like instantiating an AI product. Let
[8:36] alone if you're just, "Hey, I'm just
[8:38] doing open-source AI research and I'd
[8:40] love to use the latest and greatest
[8:42] models to help me." Now, it's not an
[8:44] option. But, at the same time, just
[8:45] think about the logic. If a model
[8:47] doesn't let you hack a system, just have
[8:49] the model build you an unrestricted
[8:51] model that does let you hack your system
[8:52] that that system. And so, clearly in
[8:55] that scenario, it makes sense to
[8:56] restrict if you want to restrict hacking
[8:58] or bio, you also have to restrict the
[9:01] tool that makes the tool that hacks the
[9:03] system or develops the bio weapon in
[9:05] theory. And so, but the choice to
[9:07] silently degrade responses was not well
[9:10] received. These are important issues,
[9:11] and there's another timeline where AI
[9:13] leaders are cut from the same cloth as
[9:15] America's elected officials. But, we're
[9:17] in a world where Anthropic has to make
[9:18] their case for safety in an age of RSI
[9:21] during a UFC fight on the White House
[9:23] lawn. It's a very bizarre timeline, but
[9:24] here we are. Most of the big tech CEOs,
[9:27] they have figured out how to work with
[9:29] the administration effectively, showing
[9:30] up to large dinners, photo ops, the
[9:32] occasional movie screening to make their
[9:34] case about hot topics. Many of these
[9:36] CEOs also have business relationship
[9:38] with Anthropic either on the customer
[9:40] side or on the investor side, and that's
[9:42] what's really complicated about the
[9:43] Amazon relationship. Obviously, Amazon
[9:45] uses
[9:47] Anthropic models, also is an investor in
[9:49] Anthropic, also provides compute to
[9:51] Anthropic. And so, they have a very
[9:53] hairy, complex, circular relationship,
[9:56] and so, that clouds the discussion over
[9:59] whether or not is Anthropic going to try
[10:01] and help Anthropic get through this? Are
[10:03] they doing everything a board? At the
[10:05] same time, you know, people were sort of
[10:07] putting out that like, "Oh, did Did
[10:08] Amazon like spike the ball on this?" And
[10:11] at the same time, if the government
[10:12] comes to you and asks like, "Hey, you're
[10:14] a big tech company. You've been using
[10:16] this model a lot. Like, were you
[10:17] You're a close partner. Were you able to
[10:19] jailbreak it at all?" And your security
[10:21] team just issued a report like, "Yeah,
[10:23] we actually were able to just jailbreak
[10:24] it a little bit." Like, you're not going
[10:26] to withhold that from the government.
[10:27] You're going to turn that over. That's
[10:29] different than calling and saying, "Hey,
[10:31] I I really think it's dangerous. You got
[10:33] to put pressure on them." So, there's
[10:34] this wide range of levels of
[10:36] interaction. They can go both ways. You
[10:37] can be a lab partnered with a big tech
[10:40] company. That big tech company can have
[10:42] a really established government
[10:43] regulation government relations group
[10:46] that is actually helping you get through
[10:48] to the admin.
[10:50] On the flip side, you could have a big
[10:52] tech company that is competing with you
[10:54] or or frustrated by your
[10:57] your business and is actively trying to
[10:59] you know, put their thumb on the scale
[11:00] the other way. So, I'm sure we'll learn
[11:02] more in the future. Many of these CEOs
[11:05] have business relationships with
[11:06] Anthropic, so we'll see how all of this
[11:08] develops. And they're always
[11:10] renegotiating these contracts, too,
[11:11] which makes things more complex. And
[11:13] they're competing across multiple
[11:15] markets. So, expect the 4D chess to be
[11:18] on full display all summer long in
[11:21] Washington. Anyway, what else is going
[11:23] on? Meta is doing a 180 according to
[11:27] Amir over at The Information, trying to
[11:29] be the vanguard of token minimizing. I
[11:32] think the token maxing, token minimizing
[11:35] thing is so silly because do you know
[11:38] where maxing comes from really? Like it
[11:40] comes from gamer lingo usually and
[11:43] >> adopted by looks maxing
[11:46] >> I don't know if looks maxing community
[11:48] was the first one. They certainly made
[11:50] it really go really viral.
[11:51] >> popularized like I feel like through
[11:54] >> Yeah.
[11:54] >> effectively through clubs clipping
[11:57] >> Yeah.
[11:57] >> maxing the idea of like doing anything
[11:59] like oh I'm I'm computer maxing today.
[12:01] >> Yeah, I mean isn't it from like
[12:02] min-maxing?
[12:03] >> Min-maxing and that's what I'm getting
[12:05] at. Like you shouldn't be token maxing
[12:07] or token minning. You should be token
[12:09] min-maxing. And the idea of min-maxing
[12:11] is when you're in a video game, you want
[12:13] to get the most resources for the lowest
[12:16] cost. So if you're playing League of
[12:17] Legends, you're going to put just the
[12:19] right amount of gold into this
[12:21] particular item or this particular stat
[12:23] point in an RPG and you're going to try
[12:25] and find the optimal build based on your
[12:29] resources. You have confined resources
[12:31] and somehow it feels like these like
[12:35] tech and all these companies went
[12:37] through some sort of hallucinatory
[12:38] hallucinatory period where they lost
[12:40] sight of that they should always have
[12:42] been min-maxing because that is the goal
[12:44] of business at all times in all things.
[12:46] You would never say like oh yeah, we're
[12:47] ad maxing this quarter. We're going to
[12:49] buy as many ads as possible. Oh no, it
[12:51] didn't work. We we were spent way too
[12:52] much on ads. They were not effective. We
[12:54] need to ad min, spend as little as
[12:56] possible on marketing. It's like no, you
[12:57] always want to be spending you always
[12:59] want to be getting the most return on ad
[13:01] spend, the highest leverage output.
[13:03] >> There was
[13:05] an again unsubstantiated post.
[13:08] >> Yes.
[13:08] >> But somebody was saying they talked to
[13:09] somebody at Meta
[13:10] >> Mhm.
[13:11] >> and there was one engineer that spent
[13:12] $90,000 in one day
[13:15] >> Ooh.
[13:15] >> and was quickly terminated.
[13:18] >> Really?
[13:19] >> Terminated within 3 days. So again,
[13:21] could just be entirely fake news but
[13:24] >> yeah.
[13:24] >> it seems somewhat informed.
[13:26] >> thing is that is that like the lesson
[13:28] from that will be
[13:29] >> like, "I'm not getting fired. I'm going
[13:31] to use AI more than anyone else in the
[13:33] whole company."
[13:33] >> 90 loops.
[13:34] >> 90,000 dollar bill.
[13:36] >> The I mean so so the lesson that will be
[13:38] learned from that is is don't spend
[13:40] 90,000 dollars a day, but the Super Bowl
[13:42] is literally for a marketer the Super
[13:44] Bowl is spend 5 million in one day and
[13:46] it works sometimes. And sometimes Super
[13:48] Bowl ads deliver 20 million dollars
[13:50] about it.
[13:50] >> but what do your teams do when they run
[13:51] Super Bowl ads, right?
[13:56] >> Yeah, yeah, yeah, yeah, yeah.
[13:58] No, we made sure that ours was ROI
[14:00] positive and we only bought the ad in a
[14:02] small market and we did it before the
[14:04] game
[14:04] >> Where are we renewing our audience was
[14:06] and so
[14:07] >> for a very low cost we were able to get
[14:09] a very high of high value. So I believe
[14:12] we got a very high return on investment.
[14:14] We min-maxed the Super Bowl. Now there
[14:16] were some companies that sort of just
[14:18] Super Bowl maxed and they just spent a
[14:19] ton of money and when you looked at the
[14:21] data on how they were received by
[14:23] audiences, it seemed like oh, they just
[14:25] definitely didn't get 5 million dollars
[14:27] in brand value back. In fact, maybe that
[14:30] ad was poorly received. Some of the AI
[14:31] generated ads
[14:32] >> I got to I got to give it up to the UFC
[14:34] for the for the level of advertising
[14:37] during the event yesterday.
[14:39] Um
[14:39] >> They're running full full ads in the
[14:42] middle.
[14:42] >> No, so one of the interesting things
[14:44] with the new Paramount deal there's a
[14:45] lot of ads. Um but one Dana White was in
[14:49] a lot of the ads that were running. So
[14:52] So like Dodge was doing a TRX ad. It was
[14:55] Dana White driving.
[14:56] >> That's cool.
[14:57] >> There was another ad, I forget the brand
[14:59] so I guess it was a bad ad, but Dana
[15:01] White was they faked a podcast set where
[15:04] it was an with an insurance company and
[15:07] Dana was having a conversation about
[15:09] business insurance with this the CEO of
[15:11] the company.
[15:12] >> Interesting.
[15:12] >> So it looks like a podcast. So I I
[15:14] thought it was interesting they were
[15:15] casting Dana and all them. Then second,
[15:17] the the the level of advertising just in
[15:20] the octagon was out of this world. There
[15:23] was so many different ways to gamble.
[15:25] Every There was I I like there's no like
[15:28] UFC cannot do like a gambling
[15:29] exclusivity deal with anyone, right? Cuz
[15:31] it's like they have It's crypto.com.
[15:33] It's Polymarket. It's Bet365. Any
[15:36] possible way that you could want to like
[15:39] a gamble.
[15:39] >> Bet365. So, the only day you're you're
[15:41] taking off is leap day.
[15:43] Every 4 years you take a 1-day off of
[15:45] gambling. That's the That's the mission.
[15:48] >> the company
[15:49] Um
[15:50] >> 4 years.
[15:50] >> then Kalshi's running ads during the
[15:52] event.
[15:53] >> Oh, so Kalshi and Polymarket are duking
[15:55] it out.
[15:55] >> is in the octagon, but Kalshi's running
[15:57] a bunch of ads
[15:58] >> against it.
[15:59] >> against it.
[15:59] >> And they're buying directly through
[16:00] Paramount because you streamed it on
[16:02] Paramount. Is that correct?
[16:03] >> Yeah, I don't
[16:04] >> to pay for it in addition to
[16:06] >> no.
[16:06] >> Okay, so you pay for Paramount on a
[16:09] monthly basis and then you get UFC for
[16:11] free. Got it.
[16:12] >> [laughter]
[16:13] >> Gambling is just min-maxing done poorly.
[16:15] That's true. So, the the the news around
[16:17] Meta is Meta is doing a 180 trying to be
[16:20] the vanguard of token minimizing. 2
[16:22] months ago, Meta epitomized token maxing
[16:24] on trend on track to spend billions a
[16:26] year on Claude, etc. We've seen an
[16:28] exponential increase in AI usage and we
[16:30] are tracking to spend billions on
[16:32] internal use alone in 2026, the memo
[16:34] said. At the same time, individuals and
[16:36] teams have limited visibility into and
[16:38] control over how they use AI. In 2027,
[16:41] we expect Meta will move towards
[16:42] managing AI tokens in a more structured
[16:44] way with bullet budgets, allocation
[16:46] decisions, and supporting tools. That
[16:48] makes sense. You have a marketing
[16:49] budget, you should also have a token
[16:51] budget um as part of
[16:52] >> I also saw something they have like Meta
[16:54] code, which is their internal
[16:56] >> Yeah. That makes sense. I mean, they're
[16:58] they're like
[16:59] you know, catching up to the frontier
[17:00] with
[17:01] >> Yeah, I mean, they got to talk about
[17:02] this. There's a bunch of like these
[17:03] leaked calls where Zuck is like, you
[17:05] know, you should imagine that Meta
[17:06] employees are going to be better than
[17:08] like the the average open source uh
[17:10] contributor. So, it's like that the
[17:11] basically quality of of coding data
[17:13] within Meta should be super good. You
[17:15] You train some model on that.
[17:16] >> It's the same as take a frontier model,
[17:19] bring them in through some sort of
[17:20] consulting partner, they fine-tune on
[17:22] your business workflow, they automate
[17:24] your business process, but your business
[17:26] process in this case, if you're meta, is
[17:28] generating code that makes your social
[17:30] networks better.
[17:31] >> Yeah, and it's especially good because
[17:32] like in meta, all of the like, you know,
[17:34] internal business like,
[17:36] whatever data, it's already in like code
[17:39] format. You you know, you don't you
[17:39] don't have to like scan in a bunch of
[17:40] papers
[17:41] >> Yeah.
[17:41] >> into, you know, like digital and
[17:44] >> I still don't know what meta is actually
[17:47] building, and I think there there comes
[17:49] a time soon that they should probably
[17:52] communicate that to the market.
[17:54] >> Personal superintelligence, baby. What's
[17:56] not to What's there not to understand?
[17:57] >> No, it's just it's just Okay, what is
[17:59] what is that? What does that mean?
[18:01] >> It's 75% as valuable as SpaceX, buddy.
[18:04] Come on.
[18:06] It's crazy.
[18:07] >> well,
[18:08] >> SpaceX is 2.3.
[18:09] >> Well, if things keep going the way
[18:10] they're going, we'll see even more of a
[18:12] divergence. No, but I think like it's so
[18:15] unclear.
[18:16] >> You should be able to put a lot of the
[18:17] SpaceX bull case on meta, right? It's
[18:19] like, sure, they maybe maybe Colossus
[18:22] was built a little bit faster, but meta
[18:23] is pretty fast.
[18:24] >> wildly I think they're they're couldn't
[18:27] be more different. With meta right now,
[18:29] it's the bull case is it's potentially
[18:31] the greatest business in the history
[18:33] that is not threatened by AI.
[18:35] >> Totally.
[18:35] >> That actually is accelerated by AI,
[18:39] and and
[18:41] I don't think any anyways, but I do
[18:43] think I do think I would expect some
[18:46] communication from meta around what
[18:48] their real strategy is
[18:50] >> Yeah.
[18:50] >> over the next month or so because it's
[18:52] killing them that they just keep saying,
[18:54] "Oh, we're super personal
[18:55] superintelligence." But wait, they're
[18:57] building like a coding harness, and
[18:59] they're and they're working on coding.
[19:01] Like, are they going to sell that
[19:03] externally, or is it going to be
[19:04] internal?
[19:05] >> Coding?
[19:05] >> Are they going to Are they going to try
[19:07] to
[19:07] >> Coding seems like it makes sense. Like,
[19:08] they're spending
[19:09] >> Are they going to try to sell their
[19:10] their on an API and get into enterprise?
[19:12] Like, the big question is like do they
[19:13] take the leap? They kind of did that
[19:15] with Manas.
[19:16] >> Yeah.
[19:17] >> But that's getting unwound.
[19:18] >> Fox is buying Roku, the streaming
[19:21] service, for $25 billion.
[19:24] It's in the Wall Street Journal.
[19:26] >> 25 or 22?
[19:27] >> 25 billion in the Wall Street Journal
[19:29] here. Let me refresh the page see if
[19:31] they updated it. It still says 25.
[19:33] >> Weird. Axios is reporting 22 billion.
[19:36] >> Oh, okay.
[19:36] >> But who knows.
[19:37] >> Somewhere in the north of a billion sub
[19:41] 100 billion. Let's be very conservative
[19:43] with our estimates.
[19:44] >> Yeah, CNBC also 22 billion. Weird that
[19:47] the Journal is
[19:48] >> I mean they have inside information
[19:50] given that it's a News Corp property
[19:52] related there, right? Um anyway, let's
[19:54] read through the Journal piece. Let me
[19:57] just give you some facts and then you
[19:59] can sort of uh give me some takes and
[20:01] and and contextualize it with more
[20:03] information. So, uh Fox Corp it said
[20:06] it's acquiring Roku in a deal valued
[20:08] around 25 billion. So, the Wall Street
[20:11] Journal is saying around is 22. Around?
[20:13] Maybe. Making a major bet on the future
[20:15] of ad-supported streaming. Ad-supported
[20:17] streaming obviously Netflix pivoted to
[20:19] ads after saying we don't know we're not
[20:21] going to do ads. Turned out to be a
[20:22] great business. Ads are undefeated.
[20:24] >> Consumers are down with ads. They're
[20:25] down with ads.
[20:25] >> They're just down with ads. Especially
[20:27] if they're good ads. Anyway, the Fox the
[20:30] the deal.
[20:31] The deal is Fox's largest to date. It
[20:33] brings together a media company known
[20:34] for its live news and sports programming
[20:36] with the biggest provider of streaming
[20:38] platforms for connected TVs. Uh it will
[20:41] add scale to Fox's streaming business
[20:44] currently home to free ad-supporting
[20:46] streaming service Tubi, which the
[20:48] company bought for 400 million in 2020,
[20:51] and subscription-based Fox One and Fox
[20:53] Nation. Uh in addition to distribution
[20:55] >> Tubi from my understanding has been a
[20:56] home run. Like I don't remember exactly
[20:59] what the revenue is, but they bought it
[21:00] for 400. It's in the billions now.
[21:03] >> And all all the metrics about actual
[21:04] audience size are huge and uh getting to
[21:07] tens of millions of dollars in revenue,
[21:09] hundreds of millions of dollars in ad
[21:11] sales against a big audience like the
[21:13] Tubi viewership. When you have the Fox
[21:15] machinery behind it, it makes a ton of
[21:17] sense. In addition to distributing other
[21:19] streaming services through connected TVs
[21:21] and devices, Roku has its own
[21:23] ad-supported Roku channel. The combined
[21:26] company will will better compete with
[21:27] the likes of amazon.com and Netflix for
[21:29] ad dollars. Quote, "Bringing these two
[21:32] companies together will really help
[21:34] define the future of television in the
[21:35] United States and in many other markets
[21:37] around the world." Fox chief executive
[21:38] Lachlan Murdoch told investors on a call
[21:40] Monday. Fox will pay around $160 per
[21:43] share, 96 in cash, and 0.9693
[21:48] Fox Class A shares. The deal is valued
[21:50] at around 22 billion on an enterprise
[21:52] basis. Fox expects to fund the cash
[21:55] portion of the transaction with 12
[21:56] billion in new debt as well as cash on
[21:58] hand. The combined company expects to
[22:00] cut around 400 million in annual costs.
[22:02] Fox plans to keep Tubi and the Roku
[22:03] channel as separate offerings. Uh
[22:06] consumers are increasingly opting for
[22:08] free and lower-cost ad-supported
[22:10] streaming options as the cost of
[22:12] subscriptions marches ever higher.
[22:14] Ad-supported streaming plans now
[22:15] represent almost 50% of all premium
[22:17] subscription ad uh video-on-demand
[22:19] sign-ups in the US, up from 39% just 2
[22:22] years ago. More than 100 million global
[22:25] households stream with Roku. Roku's the
[22:27] largest streaming platform for connected
[22:29] TVs with 25% market share. Samsung's
[22:31] Tizen is number two at 23%. And then I'm
[22:35] sure uh probably Amazon's Fire TV is not
[22:38] far behind. Fox remained on the
[22:40] sidelines during the heady early days of
[22:42] streaming boom, pouring money instead
[22:44] into programming for its cable channels
[22:45] and buying up sports rights. It launched
[22:48] Fox Nation in 2018
[22:50] and introduced direct-to-consumer Fox
[22:53] One, which includes sports and Fox News,
[22:55] last year. Fox said last month that Tubi
[22:57] had nearly 100 million monthly active
[22:59] users and its revenue had grown by 23%
[23:02] in the fiscal third quarter. And so,
[23:05] Fox's stock is down today. Roku is down
[23:09] point 2% something like that. And the
[23:12] market doesn't like it. Take him has a
[23:15] take here. We can read through his
[23:17] opinion. He says, "I don't understand
[23:18] how this Roku deal makes any sense."
[23:21] Bringing together the most valuable live
[23:23] >> I'll tell you how it makes
[23:24] >> it make sense?
[23:25] >> Um no, so I think it makes sense. One,
[23:28] Fox they have a a great product in Tubi,
[23:30] right? It's again,
[23:33] there there are certain people that are
[23:35] that are happy with their with their HBO
[23:37] or Netflix or whatever, but Tubi has
[23:40] has a great niche. It's been a great
[23:43] acquisition for Fox.
[23:45] North of a billion dollars of revenue.
[23:48] That is a product. Roku is a platform.
[23:52] They have over 20% market share in
[23:55] connected TVs.
[23:57] I was looking at the data they have over
[23:59] 40% of the engagement with connected
[24:01] TVs. So, when you think about it from an
[24:03] advertising potential, right? It's not
[24:05] just that they have 20% of the connected
[24:07] TV market. They have over 40% of like
[24:10] the watch time effectively, right? So,
[24:13] looking at the potential from an
[24:15] advertising standpoint and how many
[24:16] dollars are still yet to shift from
[24:19] traditional television to connected TV.
[24:21] So, from that lens I think it's I think
[24:23] it's great. Roku's also a much younger
[24:25] demo. So, like a lot of media a lot of
[24:27] media properties, right? Especially
[24:29] television related are going to be you
[24:32] know, older demos that are that are
[24:33] eventually going to be aging out at
[24:36] least of like prime consumers, but these
[24:38] are consumers that are in their, you
[24:40] know, prime
[24:42] earning years, prime years as consumers.
[24:45] And and so, ultimately I think it makes
[24:47] sense.
[24:48] I think it makes sense for Fox. Next,
[24:51] the thing that I hope that they can do
[24:53] as a consumer is
[24:55] rebundle, right? We've just gone
[24:58] We've gone through this era of
[25:00] unbundling and the fact that
[25:03] you got a new TV and you're like, "Okay,
[25:05] great. Let me log into, you know, 15, 10
[25:08] different services, something like
[25:09] that." I would like to see them do
[25:12] I would like to see them do something in
[25:14] bundling for for a consumer. So again,
[25:16] looking at it from the lens of like
[25:18] having north of 40% of like watch time,
[25:21] right? That is a super valuable property
[25:24] and like with the right ad-driven
[25:25] approach, I think I think it'll pay off.
[25:28] >> I wonder how windowing fits into this.
[25:30] Like I I when when I grew up, I didn't
[25:32] have cable, but you would get a prestige
[25:35] movie played on ABC like once a week.
[25:40] And for a while, that was the flow on
[25:43] Netflix and HBO was like its own premium
[25:46] subscription. You go open up the HBO
[25:48] app, it's all premium
[25:50] series and movies. And then some of that
[25:53] would trickle down, but then it feels
[25:54] like the ecosystem sort of branched a
[25:57] little bit where Netflix stopped getting
[26:00] like the James Cameron library for a
[26:02] while. And Disney started hoovering up
[26:05] some of that, keeping some of the really
[26:07] popular movies, the Marvel franchise off
[26:09] of Netflix even for later run, a year or
[26:11] two after because there was more
[26:13] competition from the actual different
[26:16] pieces of the the the the different
[26:18] studios. So I wonder if Roku will be
[26:20] able to to provide some sort of like
[26:22] neutral territory in some ways where
[26:25] where viewers can pick up something
[26:27] that's in the premium end, something's
[26:28] that's, you know, a sequel to a
[26:30] franchise that they know and love, but
[26:33] it's not seen as as competitive as as,
[26:36] you know, Disney versus HBO or whatever.
[26:39] We'll We will see.
[26:40] >> Overall, real platform.
[26:42] >> Yeah.
[26:43] >> They already have a product that sits on
[26:44] top of it. Tons of of advertising
[26:46] potential and I think that regardless of
[26:49] whatever the market reaction is today, I
[26:51] expect that people will will get it over
[26:53] time.
[26:53] >> Tyler Cowen
[26:55] shared some more thoughts on the recent
[26:57] mythos brouhaha.
[26:59] It's a good term.
[27:01] Uh we Hey, this was all over. We talked
[27:03] about this earlier, but he says,
[27:04] "According to not yet confirmed, but
[27:06] likely true reports, it was shown that
[27:08] the model could be jailbroken. There's
[27:09] debate over this now. Uh the released
[27:12] mythos already restricted bio and AI
[27:14] improvement queries, rather than
[27:16] strictly In fact, uh rather strictly in
[27:18] fact. So, now we are back to the model
[27:20] not being available. Here are a few of
[27:22] the constraints on the US government,
[27:24] not the only ones, I might add. One, it
[27:27] needs for the main companies to stay in
[27:29] business. On top of that, it wants their
[27:30] IPOs to go reasonably well. And it's now
[27:33] much harder for the top companies to
[27:35] recruit foreigners, which is a
[27:37] significant share of their highest
[27:38] quality workforce. Demis, Ilya, Andre,
[27:41] for a start, are all international uh
[27:44] engineers. Uh it is much harder for the
[27:46] main competitors to drum up foreign
[27:47] business in a credible manner and
[27:48] sustainable manner. That's interesting.
[27:51] We haven't seen many breakout charts of
[27:54] how much revenue is domestic versus
[27:55] international, but we have seen reports
[27:57] of uh OpenAI and Anthropic doing big
[28:00] deals with international companies. Some
[28:02] of these companies, Fortune 500s, they
[28:04] just have global footprints, so they got
[28:05] to use the models everywhere, ideally,
[28:08] uh unless they do all of their software
[28:10] engineering in Silicon Valley. Um 1B,
[28:12] how are American multinationals
[28:15] operating abroad supposed to use top
[28:17] systems moving forward? That's literally
[28:19] what I just said. Uh I just predicted
[28:21] the next uh line. Oops, I am uh
[28:24] I'm a stochastic parrot. Um
[28:26] >> [laughter]
[28:27] >> It wants to use model access as a tool
[28:29] of both hard and soft power. So, model
[28:32] access has to be possible at some level,
[28:35] but it is very hard to control what
[28:37] foreign agents will do with partial
[28:39] model access when they get it in the
[28:41] future. Three, the US needs to stay
[28:43] ahead of China in the AI race. This is a
[28:45] big question. We have Aaron Gin from
[28:47] Hydro Host joining in just 10 minutes.
[28:49] We'll ask him for his perspective on
[28:51] this. Four, the US needs to issue
[28:53] restrictions that are actually
[28:55] enforceable and US citizens only doesn't
[28:57] fit that bill. That's a good point.
[28:59] Furthermore, markets and everything. He
[29:01] It is easy enough to hire a traitorous
[29:03] American to access tools of wrongdoing.
[29:05] That's a really good point. How can you
[29:07] verify that every American is is loyal?
[29:10] Or for that matter, it is not that
[29:12] difficult to fake citizenship in various
[29:13] ways.
[29:14] >> Yeah, there's been there's been people
[29:16] over the last 50 years that have gotten
[29:18] caught like passing on or basically
[29:22] selling, you know, secrets around our
[29:24] various military aircraft, for example.
[29:27] >> There's a very controversial a story of
[29:29] Virgil Griffith,
[29:30] a cryptocurrency researcher who I
[29:32] actually met once.
[29:34] He worked at Caltech, was deep in math,
[29:37] and very excited about cryptocurrency.
[29:39] And he went and gave a talk on
[29:41] cryptocurrency in North Korea,
[29:43] explaining how cryptocurrency works and
[29:45] how crypto cryptography works. And that
[29:47] was seen as a violation of export
[29:49] controls, and he was sent to prison. Um,
[29:51] and the crypto community is very upset
[29:53] about that because they they maintain
[29:54] that math should not be illegal. This
[29:57] information should not be illegal. The
[29:58] the government side is they were hoping
[30:00] >> Was he like thinking that he was going
[30:02] to liberate the North Korean people by
[30:04] giving them access to, effectively, or
[30:06] explaining how code
[30:08] >> can be used to create a
[30:11] store of value that that is that is
[30:13] unsanctionable essentially. And so, it
[30:16] was seen as a violation violation of
[30:18] sanctions by the US government. And more
[30:20] importantly, it was seen as
[30:21] only being
[30:22] I don't know. I don't know. They have a
[30:24] lot of coins.
[30:25] >> it the most sanctioned stablecoin.
[30:27] >> Potentially. Potentially.
[30:29] Um,
[30:29] this one, number five by Tyler Cowen,
[30:32] was a little bit controversial. He says
[30:33] the US government cannot nationalize
[30:35] these companies and then proceed to run
[30:38] them effectively. And people were
[30:39] debating that,
[30:41] um, saying that it might be possible for
[30:43] the US government to to own the labs in
[30:46] some way and and and leave them to their
[30:49] own devices to actually roll these
[30:50] things out. What do you
[30:51] >> Yeah, I I mean, if you're like super RSI
[30:53] pilled, right? You don't need the
[30:54] employees at all cuz because they yeah,
[30:55] just improve themselves.
[30:56] >> So, the government just says, "Continue
[30:57] to improve."
[30:58] >> Yeah, basically.
[30:59] >> it. So, you don't actually the extreme
[31:01] scenario. Yeah. Yeah. And then and then
[31:03] on the other side it's that, "Okay,
[31:05] these systems are actually extremely
[31:06] complex. You look at the uptime of some
[31:09] of these systems, the guard the research
[31:11] taste. In fact, you need so many
[31:14] employees to make these products
[31:15] effective. You actually need the
[31:17] international employees as well because
[31:19] they come up with great ideas." And so,
[31:21] yeah, those are the two sort of sides of
[31:22] the of the debate. Uh, six is Chinese
[31:25] and open source models do in fact
[31:26] improve at some reasonable pace even
[31:28] though they are right now considerably
[31:30] behind them at best proprietary models.
[31:32] And that's something that I want to dig
[31:33] into in China is slowing. I'm not
[31:35] entirely sure
[31:39] what's driving it. It feels like there's
[31:41] a number of different things. Gavin
[31:42] Baker said it's about compute and
[31:44] they're and they're they're turning
[31:46] their backs on the black wells that they
[31:47] had the opportunity to buy. I think it
[31:49] might be research ideas that are more
[31:52] locked down now in in in tighter
[31:54] controlled circles in San Francisco. It
[31:57] also might be the data. Might be but
[31:59] you'd have to imagine that they have
[32:00] lots of people that are writing code,
[32:02] lots of people that are inferencing
[32:04] those models and then full data and and
[32:07] and information capturing of those
[32:09] rollouts. There are a number of
[32:10] different uh, reasons why that pace
[32:13] might be slowing. I mean, we'll see.
[32:14] Maybe there's a big jump in the graph.
[32:16] Maybe we have another deep seek moment
[32:17] in the next couple months. We'll see.
[32:19] >> Steering order in bound.
[32:20] >> Leave us five stars on Apple Podcast and
[32:21] Spotify. Sign up for our newsletter,
[32:22] tbpm.com, and we will see you tomorrow
[32:24] at 11:00 a.m. Pacific. Goodbye.
[32:26] >> Technology's daily show.
[32:28] >> Technology in motion.
