Meta Surges Ahead in AI Race as Rivals Face Talent and Data Shortages
TBPN
Meta's massive investment in reinforcement learning engineers and custom AI infrastructure positions it to surpass rival labs in the race for frontier model dominance.
## Background
TBPN published this video on YouTube on July 11, 2026. According to the page description, Diet TBPN delivers the highlights of TBPN's weekday live tech talk show in a condensed 30-minute episode hosted by John Coogan and Jordi Hays.
## TL;DR
The episode covers major tech and business developments, starting with SK Hynix's historic $26.5 billion NASDAQ debut and the community reactions to OpenAI's GPT-5.6 launch, including its creative workflow capabilities and desktop app backlash. The hosts break down Meta's aggressive AI strategy, hardware investments, and platform battle with Apple, supported by an influential report from SemiAnalysis. Finally, they cover the 'Phia Gate' affiliate marketing hijacking controversy surrounding Phoebe Gates' shopping app and Hollywood's multi-million dollar bidding wars for viral internet horror IP.
## Atmosphere & style
The tone is fast-paced, conversational, and energetic, blending sharp tech industry commentary with casual insider banter. The hosts maintain an engaging, upbeat dynamic as they bounce between serious market breakdowns and humorous internet drama. Listening feels like catching up on Silicon Valley news with well-informed peers who balance analytical depth with lighthearted skepticism.
## Key takeaways
- [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) SK Hynix made a historic Wall Street debut on the NASDAQ, raising $26.5 billion at a valuation of approximately $1.03 trillion amidst surging HBM chip demand.
- [00:32](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=32s) The launch of OpenAI's GPT-5.6 triggered a wave of creative 3D and video pipeline experiments, highlighting the shift toward physically grounded style-transfer workflows.
- [00:32](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=32s) Cerebras hardware previews demonstrated GPT-5.6 Sol running at 750 tokens per second, enabling high-speed automated software manipulation.
- [00:09](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=9s) GPT-5.6 features integrated tool use in its image models, allowing simultaneous UI generation, coding, and asset generation in a single prompt.
- [00:09](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=9s) OpenAI faced significant community backlash over resource consumption and removed shortcuts after transitioning its macOS app to an Electron-based shell.
- [15:55](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=955s) Mark Zuckerberg's heavy investments in open AI models and hardware are driven by platform risk and the desire to circumvent Apple and Google's app platform controls.
- [16:55](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1015s) Meta is heavily betting on smart glasses as a potential successor to smartphones, leveraging them to monetize real-world local transactions and AI interactions.
- [19:48](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1188s) According to SemiAnalysis, Meta is the only tech giant on track to be world-class across data, top talent, and compute infrastructure simultaneously.
- [24:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1484s) Phoebe Gates' shopping app Phia faced intense scrutiny after reports revealed its extension secretly hijacked background affiliate codes from original content creators.
- [29:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1784s) Hollywood studios are engaged in aggressive bidding wars to acquire viral internet horror IP from platforms like YouTube, 4chan, and Reddit.
## Detailed walkthrough
### [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) SK Hynix Surges 14% in Record $26.5B NASDAQ Debut
South Korean memory chip giant **SK Hynix** made a massive Wall Street debut on the NASDAQ, raising **$26.5 billion** in the largest-ever first-time US share sale by a foreign company. Demand reportedly exceeded available shares by more than sevenfold. Despite the huge 7x oversubscription—which in smaller IPOs might trigger an 80% to 100% initial jump—the scale of the valuation meant shares opened with a clean 14% pop. The company's American Depositary Receipts (ADRs) opened at $170, up 14% from their $149 offering price, bringing SK Hynix's market capitalization to roughly **$1.03 trillion**.
This blockbuster debut adds another trillion-dollar titan to the ongoing AI boom. SK Hynix, Samsung, and US-based Micron dominate the global memory market, with SK Hynix benefiting immensely from skyrocketing demand for High Bandwidth Memory (HBM) paired alongside advanced AI accelerators. Over the past year, SK Hynix’s primary South Korean listing on the KOSPI index has surged over 600% as investors scrambled to get exposure to one of the primary hardware bottlenecks in AI. The new NASDAQ listing simplifies access for US retail and institutional investors who previously had to navigate foreign markets. Early participants in the transaction included major institutional players like Situational Awareness and Baillie Gifford.
* **Temporary Ticker (`SKHYV`):** The company is temporarily trading under the symbol `SKHYV` before transitioning to its permanent ticker, `SKHY`, on Monday. The "V" denotes "when-issued" trading, a mechanism used by exchanges to let investors trade shares before final transaction settlement occurs.
* **Massive Capital Raised:** $26.5 billion raised in the offering, marking a historic US listing milestone for foreign entities.
---
[00:32](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=32s) GPT-5.6 Demos, Cinema Homages, and AI Workflow Realities
Following the launch of OpenAI's **GPT-5.6** (including the *Sol* and *Sol Ultra* model variants), the internet has been flooded with creative demos and community experiments. One viral project recreated scenes from Christopher Nolan's *Interstellar* using a pipeline from text prompts to 3D model generation and rendering. While joking that this is "the exact opposite of what Christopher Nolan wants to see" given his preference for practical effects over CGI, the hosts point out that prompt-to-3D speed opens up entirely new artistic paradigms. The real potential lies not in duplicating existing box-office hits, but in combining lightweight 3D engines like Three.js or Blender with style-transfer video models to generate photorealistic, physically grounded cinema.
> "You can go and recreate *Interstellar* perfectly, but at the end you just have *Interstellar*, which already exists and is available to stream... you have to bring something special to the table."
The hosts highlight the contrast between abstract capabilities and bizarre real-world applications. OpenAI's launch video featured a broccoli farmer running his entire farm on GPT-5.6, sparking viral memes across the timeline. This led to a friendly debate over AI's learning curve: while some argue AI tools have no friction because they present a simple text box, others contend that users require an educational process to understand the "fringes of their imagination" and figure out how to apply open-ended models to their daily routines.
* **Workflow Mechanics:** Modern AI video pipelines are moving toward blocking out low-fidelity, physically accurate 3D animations (in Blender or via raw iPhone footage) and applying AI style-transfer models over them to maintain motion determinism.
* **Cerebras Acceleration:** Community member Chris highlighted GPT-5.6 Sol running at **750 tokens per second** via Cerebras hardware previews, enabling real-time computer use where the AI manipulates software environments like Blender at superhuman speeds.
---
[00:09](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=9s) Code Generation, Ecosystem Blending, and the Desktop App Backlash
The capabilities of GPT-5.6 are blurring the lines between modal boundaries. Packy McCormick shared a web design test where the model "one-shotted" a clean cosmos-themed site layout and generated a matching preview logo without even being asked. OpenAI team member Nick noted that the image generation models now feature integrated tool use, allowing them to scaffold out raw HTML and generate visual preview graphics simultaneously.
The episode recounts a humorous story from user Michael, who found himself caught in an automated loop with OpenAI's moderation systems:
1. Michael was suddenly banned from OpenAI for alleged "cyber abuse."
2. He pasted the ban notification directly into Codex (which was still functioning).
3. Codex discovered that the trigger was an API key request for Michael's own local server.
4. Codex drafted a formal legal/appeals letter and submitted it automatically.
5. Minutes later, an OpenAI automated system approved the appeal and restored his account.
*He was banned by AI, convicted by AI, defended by AI, and pardoned by AI in under 10 minutes.*
Despite the model upgrades, OpenAI faced swift community backlash over its unified macOS desktop application. The update replaced the original native Mac app with an Electron-based shell derived from the Codex app. Power users complained that the Electron version consumed more resources, lacked native OS shortcuts (like `Cmd + Space` triggers), and hid familiar sidebar features like past chats and project folders.
OpenAI's Thibault ("Tibo") responded publicly on X, acknowledging the team's missteps during the rollout:
> "The short version is that there's a lot of excitement for 5.6 Chat, but we also didn't get everything quite right... We made it too easy to use the highest compute settings without making the impact on usage limits sufficiently clear."
* **Usage Limit Resets:** To compensate for the confusing compute settings at launch, OpenAI reset power-user usage caps twice in 24 hours to let developers continue testing.
* **Codex Brand Clarification:** OpenAI reiterated that Codex is not being phased out into a generic "ChatGPT Work" tool, emphasizing that "Codex is here to stay."
---
[00:14](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=14s) The Meta Bull Case: Ad Models, In-App Purchases, and Platform Taxes
The hosts unpack a deep-dive analysis by *SemiAnalysis* regarding Meta's business strategy, MSL (Meta Superintelligence Lab), and its expansion into enterprise B2B sales. The discussion turns to the broader thesis around Mark Zuckerberg's aggressive AI capital expenditures. A popular commentary from the business community suggested that Zuckerberg is going "all-in" on open-source AI and proprietary hardware to avoid paying a "30% platform tax" to Apple and Google for the rest of his company's existence.
The hosts challenge this specific premise, noting a fundamental misunderstanding of app store fee structures. Meta’s primary revenue engine—digital advertising—is exempt from Apple's 30% in-app purchase (IAP) cut because ad views and real-world transactions are classified as real-world services rather than digital goods. While digital subscriptions (like X Premium), gaming currencies (Fortnite V-Bucks), or app-based digital upgrades trigger Apple's 30% fee, buying physical goods on Amazon or running standard web-based ad campaigns do not. However, small-business ad boosting directly inside mobile iOS apps can occasionally run into Apple's payment rails, illustrating the friction Meta constantly navigates against platform gatekeepers.
Here is the detailed, faithful walkthrough of the transcript section.
---
[16:55](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1015s) Mark Zuckerberg's Platform Strategy and Meta's AI Push
The hosts dive into the underlying platform dynamics between Meta and Apple, discussing why Mark Zuckerberg feels compelled to spend heavily on open AI models and hardware. Off the record, Apple executives harbor deep disdain for Meta, placing Zuckerberg in a perpetually uncomfortable and vulnerable position. He refuses to run a company where he must constantly ask Apple for permission to operate. A prominent example cited by the hosts (referencing an insight from Andrew) is how Meta lost billions of dollars in revenue directly due to Apple's privacy permissions prompt, which severely hindered Meta's ability to run its core targeted advertising business and track users cross-platform.
Beyond advertising, the future of consumer tech remains highly unpredictable. The hosts note that computing could shift away from discrete social media apps toward omnipresent AI agents that dynamically surface content on the fly. Rather than sitting back and hoping Meta isn't disrupted, Zuckerberg's philosophy is to aggressively participate in this massive tech cycle to secure a dominant position. While AI is getting massive capital investment, the hosts view Meta's smart glasses as an even bigger long-term platform bet than AI alone. Zuckerberg envisions a world where people wear cheap or even free smart glasses that replace phones and laptops. As users walk around, the glasses layer always-on intelligence onto the physical world—allowing a user to look at an item and say "buy me that," or popping up a deal for free boba when passing a restaurant, with Meta taking a transactional cut of every interaction.
- **Platform Risk:** Meta lost billions in ad revenue when Apple introduced app tracking transparency prompts.
- **The Glasses Vision:** Cheap or free smart glasses aim to replace smartphones and PCs, monetizing real-world transactions and local ad offers.
- **Accounting Cover:** Because overall AI infrastructure spending is so massive (e.g., $200 billion on data centers), Meta can easily hide an extra $8 billion in annual Reality Labs spending without drawing investor backlash.
---
[19:48](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1188s) Changing Consumer Habits and SemiAnalysis on Meta's AI Potential
The discussion shifts to shifts in daily screen time and an influential report from SemiAnalysis regarding the AI race. One host shares personal behavior patterns as an early adopter, noting a major shift in screen time away from traditional social feeds like Instagram and toward ChatGPT and deep research tools. Where an interesting car video on Instagram used to lead to manual Google searches and tab-collecting, it now prompts immediate, deep conversational queries with LLMs.
The hosts then review a breakdown from SemiAnalysis (highlighted on X/Twitter by Scale AI's Alexandr Wang), which asserts that building a true frontier AI model requires three core pillars: **data, talent, and compute**. According to SemiAnalysis, Meta is the only hyperscaler or NeoLab on track to be world-class across all three components, giving them the best chance to catch up with frontier labs like Anthropic and OpenAI.
<Carousel>
{/* Reason: Displays visual depictions of major AI frontier labs and compute infrastructure providers (Meta, Anthropic, OpenAI) discussed in the text. */}
<Image alt="Meta AI logo and branding" caption="Meta AI Infrastructure" src="image_agent_tag_1"/>
<Image alt="Anthropic company logo" caption="Anthropic AI Lab" src="image_agent_tag_2"/>
<Image alt="OpenAI logo" caption="OpenAI Research" src="image_agent_tag_3"/>
</Carousel>
The analysis sparked discussion by placing Meta ahead of competitors, while explaining that SpaceX is selling $26 billion a year worth of compute to Anthropic and Google. SemiAnalysis points out that while Google's Gemini has vast compute and data, it lacks the necessary top-tier research talent structure. Crucially, Meta revamped its data efforts by creating a new Applied AI Engineering organization during recent restructuring. Meta assigned 3,000 engineers—including 70% of their new hires and many senior staff—to build Reinforcement Learning (RL) tasks and environments full-time. While Anthropic previously led the market in buying expert coding data (logging 2.5 million expert hours in Q2 2026, equivalent to 5,000 full-time workers), Meta has quickly matched this scale and holds a massive internal pool of 70,000 additional employees it can draw from.
- **Data Labeling Economics:** The report dispels the myth that AI data work is low-level $10-an-hour taskwork; top RL environment design contractors at data companies now command seven-figure annual salaries ($1M+).
- **Market Reaction:** Meta's stock jumped over 6% following the convergence of the SemiAnalysis report and the release of Muse Spark 1.1.
The hosts break down a new investigative report from Bloomberg involving "Phia," the buzzy shopping app co-founded by Phoebe Gates (daughter of Microsoft co-founder Bill Gates). Positioned as a personal shopping assistant that finds consumers the lowest prices across clothing and fashion retailers, Phia operates on an affiliate marketing model, collecting commissions when users make purchases. However, independent technical testing by Bloomberg, Capital One Shopping, and Edelman revealed that Phia's mobile browser extension engaged in predatory affiliate hijacking.
During checkout processes across more than 50 tested websites, the Phia extension opened hidden background tabs without any user interaction or knowledge, injecting its own affiliate referral code to override and steal legitimate referral commissions from original publishers. The hosts compare this to the famous YouTube exposé by MegaLag on the browser extension Honey, which had accumulated 20 million views for exposing how Honey stole affiliate credit from creators (such as podcasts promoting AG1 with custom codes). However, Phia's implementation is criticized as even cruder and more predatory: whereas Honey required users to at least interact with or dismiss a popup before triggering a redirect, Phia silently hijacked commissions in the background while users simply browsed.
- **Financial Scale:** Phia has driven over $12 million in sales over ~6 months of this behavior, representing roughly $1 million in missing or redirected commission revenue.
- **Impact on Publishers:** A content site (like a niche review blog) that spends money on Google Ads to educate consumers loses its affiliate payout when Phia silently inserts its code right at checkout.
- **Public Backlash:** Because Phia raised venture capital with a celebrity investor list that "looked like a Coachella poster," tech observers had been skeptical of the hype, leading to intense online focus on this controversy as Phia's potential "villain arc."
---
[29:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1784s) Hollywood's Gold Rush for Internet Horror Memes
The final segment focuses on a Wall Street Journal report detailing Hollywood's aggressive bidding wars for online creepypasta and viral horror IP created by young internet creators. The trend follows the massive theatrical success of *The Backrooms*—a horror film originating from a 4chan post and viral YouTube channel that went on to gross over $100 million.
The latest major deal involves Trevor Henderson, a 40-year-old illustrator who drew the viral monster "Siren Head" eight years ago. Despite Siren Head inspiring millions of YouTube videos, Amazon merchandise, and indie video games, Henderson had never profited from the character until Warner Bros. recently bought the film rights for over $1 million. Major studios are now scouring Reddit, YouTube, and Roblox to find pre-built Gen Z horror IP, resulting in competitive "dogfights" for digital scripts reminiscent of 1980s and 1990s Hollywood spec script bidding wars.
<Timeline>
{/* Reason: Visualizes the chronological timeline of major internet horror IP adaptations and studio acquisition deals. */}
<TimelineEvent time="2018" title="Siren Head Created">
Trevor Henderson draws the faceless Siren Head monster, spawning years of unmonetized viral YouTube films and games.
</TimelineEvent>
<TimelineEvent time="2022 - 2025" title="The Backrooms Breakout">
Originating on 4chan and popularized on YouTube, the backroom lore gets adapted into a feature film grossing over $100 million.
</TimelineEvent>
<TimelineEvent time="2026" title="The Mandela Catalogue Bidding War">
11 Hollywood studios compete for film rights to Alex Kisser's YouTube series, resulting in a multi-million dollar deal with United Artists and Amblin.
</TimelineEvent>
</Timeline>
In another high-profile deal, 11 studios fought over the rights to *The Mandela Catalogue*, an online psychological horror series. Amazon-owned United Artists and Steven Spielberg’s Amblin Entertainment won the rights for millions of dollars, granting the series' 22-year-old creator, Alex Kisser, the director's chair for the feature film. The hosts close out the show bantering about why people are obsessed with horror, signing off with reminders to subscribe to the newsletter and leave 5-star podcast reviews.
- **Key Quote:** Warner Bros. Motion Picture Group co-chair Michael De Luca noted of internet lore: *"We're seeing them as a resource for adaptations the same way we look at books and other media."*
- **Sign-off:** The hosts joke about growing up in the country and being too terrified to watch horror movies before wrapping up the episode.
<ElicitationsGroup message="Explore related topics from this episode:">
{/* Reason: Gives the user relevant follow-up paths directly connected to Meta's AI strategy, affiliate marketing mechanics, or Hollywood digital IP trends. */}
<Elicitation label="Break down Meta's AI strategy vs. OpenAI and Google" query="Break down Meta's open-source AI strategy versus OpenAI and Google, including their compute investments and talent acquisition."/>
<Elicitation label="Explain how browser extension affiliate hijacking works" query="Explain how browser extension affiliate hijacking and attribution theft work technically in e-commerce."/>
<Elicitation label="Analyze the trend of YouTube/Roblox creators landing Hollywood deals" query="Analyze the financial and creative trend of Hollywood adapting YouTube, 4chan, and Roblox horror IP."/>
</ElicitationsGroup>
## Notable quotes
> [00:32](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=32s) You can go and recreate *Interstellar* perfectly, but at the end you just have *Interstellar*, which already exists and is available to stream... you have to bring something special to the table.
> [00:09](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=9s) The short version is that there's a lot of excitement for 5.6 Chat, but we also didn't get everything quite right... We made it too easy to use the highest compute settings without making the impact on usage limits sufficiently clear. — Thibault
> [29:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1784s) We're seeing them as a resource for adaptations the same way we look at books and other media. — Michael De Luca
## Mentioned
GPT-5.6, Phoebe Gates, Phia, Hollywood, TBPN, John Coogan, Jordi Hays, X, YouTube, The New York Times, Mark Zuckerberg, Sam Altman, Mark Cuban, Satya Nadella, Ramp, Public, Cisco, Console, CrowdStrike, Figma, MongoDB, NYSE, Railway, Shopify, Spotify, Apple, SK Hynix, NASDAQ, Samsung, Micron, High Bandwidth Memory, KOSPI, Situational Awareness, Baillie Gifford, Sol, Sol Ultra, Christopher Nolan, Interstellar, Three.js, Blender, OpenAI, Chris, Cerebras, Packy McCormick, Nick, Michael, Codex, Electron, Thibault, SemiAnalysis, Meta Superintelligence Lab, Google, Amazon, Andrew, Reality Labs, Instagram, ChatGPT, Scale AI, Alexandr Wang, Anthropic, SpaceX, Gemini, Applied AI Engineering, Muse Spark 1.1, Bloomberg, Bill Gates, Capital One Shopping, Edelman, MegaLag, Honey, AG1, Wall Street Journal, The Backrooms, 4chan, Trevor Henderson, Siren Head, Warner Bros., Roblox, The Mandela Catalogue, United Artists, Steven Spielberg, Amblin Entertainment, Alex Kisser, Michael De Luca
TBPN explores major transitions across tech, e-commerce, and entertainment, highlighted by SK Hynix's record-setting public listing, expanding GPT-5.6 workflows, Meta's defensive AI investments, an affiliate attribution scandal at Phia, and Hollywood's pursuit of viral internet horror IP.
SK Hynix Debuts on NASDAQ at $1.03 Trillion (0:02)
South Korean chipmaker SK Hynix completed the largest first-time US share sale by a foreign entity, raising $26.5 billion in an offering oversubscribed sevenfold. Trading under the temporary ticker SKHYV prior to transitioning to SKHY, American Depositary Receipts jumped 14% at open from the $149 offer price to $170, placing the company's valuation at approximately $1.03 trillion. The surge reflects intense institutional demand for High Bandwidth Memory essential to frontier AI accelerators, following a 600% rise in SK Hynix's primary South Korean KOSPI listing over the preceding year.
GPT-5.6 Deployment, Creative Workflows, and Client Backlash (0:32)
OpenAI's rollout of GPT-5.6 (including Sol and Sol Ultra variants) spurred experimental pipelines combining prompt-based 3D generation with tools like Blender and Three.js to produce physically grounded video. Hardware previews showed GPT-5.6 Sol reaching 750 tokens per second on Cerebras chips, enabling accelerated real-time computer navigation. Simultaneously, image models gained tool-use capabilities to scaffold HTML and visual previews concurrently.
Despite creative enthusiasm, OpenAI faced community pushback after replacing its native macOS application with an Electron-based shell. Users reported elevated resource usage, missing native shortcuts, and removed sidebar features. OpenAI representative Thibault publicly conceded rollout missteps regarding unclear compute tiers and usage limits, prompting OpenAI to reset power-user caps twice within 24 hours while affirming that Codex remains an independent tool.
Meta's Push Across Compute, Data, and Hardware (16:55)
Meta's heavy capital expenditures are driven largely by platform vulnerability, having previously lost billions in advertising revenue following Apple's tracking permission changes. To hedge against future mobile platform restrictions, Meta is pursuing both AI infrastructure and smart glasses designed to monetize real-world transactions directly. Massive data center spending also provides cover for an estimated $8 billion in annual Reality Labs expenses.
A report by SemiAnalysis contends that Meta is the only hyperscaler positioned to achieve world-class standing across data, talent, and compute. Following internal restructuring, Meta assigned 3,000 engineers—representing 70% of new hires alongside senior staff—to build Reinforcement Learning environments full-time. The report also noted that demand for high-tier reinforcement learning design has pushed elite contractor compensation past $1 million annually.
Phia Browser Extension Faces Affiliate Hijacking Allegations (24:44)
An investigative report by Bloomberg, supported by technical tests from Capital One Shopping and Edelman, revealed that personal shopping extension Phia engaged in affiliate attribution theft. Co-founded by Phoebe Gates, the extension was found to silently open hidden background tabs during checkout across more than 50 e-commerce sites to inject its own referral codes without user consent. Over six months, this mechanism diverted roughly $1 million in commissions on $12 million in sales, overriding attribution credit previously earned by original content publishers and digital reviewers.
Hollywood Acquires Viral Creepypasta and Digital Horror (29:44)
Following the box-office run of the online-inspired film The Backrooms, which grossed over $100 million, major studios have initiated competitive bidding wars for internet horror properties. Warner Bros. recently acquired feature rights to Trevor Henderson's viral creature Siren Head for over $1 million, eight years after its creation. In another major transaction, an 11-studio contest for Alex Kisser's YouTube series The Mandela Catalogue ended with Amblin Entertainment and United Artists securing multi-million dollar rights and placing the 22-year-old creator as director.
Across memory hardware, generative AI applications, affiliate retail, and studio acquisitions, the program documents how digital infrastructure and viral IP are reshaping commercial gatekeeping, though technical execution and platform dependencies remain ongoing vulnerabilities.
## Background
本期视频由科技脱口秀频道 TBPN 于 2026 年 8 月 3 日发布在 YouTube 平台。节目属于“Diet TBPN”精简播客系列，由 John Coogan 和 Jordi Hays 主持，旨在用 30 分钟左右的时间浓缩当天科技前沿动态，涵盖 GPT-5.6 社区反响、菲比·盖茨旗下 Phia 应用争议以及好莱坞追逐网络恐怖 IP 等热点话题。
## TL;DR
节目首先复盘了韩国内存巨头 SK 海力士在纳斯达克创纪录的 265 亿美元巨型 IPO 及其破万亿美元市值的行业背景，紧接着深入剖析了 GPT-5.6 发布后社区掀起的影视复刻、极速 Computer Use 演示以及 OpenAI 桌面端整合所带来的体验争议。随后，主持人们驳斥了扎克伯格全力押注 AI 仅为规避“苹果税”的流行论调，并结合 SemiAnalysis 的前沿模型三要素论深入解析了 Meta 转型强化学习环境与数据工程的战略野心。最后，节目曝光了比尔·盖茨之女联合创办的购物插件 Phia 涉嫌静默劫持推广返利佣金的丑闻，并探讨了好莱坞制片厂争相以数百万美元抢购《警笛头》和《曼德拉记录》等网络怪谈 IP 的新型淘金热潮。
## Atmosphere & style
整期节目节奏紧凑、快人快语，洋溢着典型硅谷科技圈内人的敏锐机锋与黑色幽默。两位主持人默契十足，在严谨的技术架构拆解与毒舌八卦调侃之间无缝切换——既能一针见血地戳穿商业模式背后的伪逻辑，也能绘声绘色地吐槽被 AI 闭环“定罪又赦免”的荒诞现实。节目在硬核的算力与工程讨论中穿插着极客式的捧腹笑点，整体氛围轻松随性却不失深度洞察，结尾关于乡间走夜路怕鬼的闲聊更添了几分亲切真实的播客烟火气。
## Bulletin
- [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) SK海力士正式在纳斯达克挂牌上市（临时代码 SKHYV），IPO募资265亿美元创外国公司在美首发上市历史新高，开盘上涨14%并推动市值突破1.03万亿美元大关。
- [07:57](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=477s) GPT-5.6 Soul Ultra 展示出每秒高达750 Token的极速 Computer Use 界面自动化控制能力，同时 OpenAI 整合 ChatGPT 与 Codex 客户端因 Electron 架构卡顿及算力限额骤减引发老用户反弹。
- [20:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1222s) SemiAnalysis 提出顶尖前沿模型三大基石（数据、人才、算力），指出 Meta 是唯一三者兼备并有望反超竞争对手的巨头，带动 Meta 股价开盘飙升超6%。
- [24:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1484s) 比尔·盖茨之女菲比·盖茨联合创办的导购插件 Phia 被曝在后台静默劫持电商推广返利链接，涉嫌挪用约100万美元归因佣金，波及逾1200万美元交易额。
- [30:16](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1816s) 好莱坞掀起对 Z 世代网络恐怖怪谈 IP 的争夺狂潮，华纳兄弟超百万美元买下《警笛头》改编权，安培林娱乐与联艺电影数百万美元签下《曼德拉记录》并钦定原作者执导。
## Key takeaways
- [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) SK海力士以265亿美元创下外国公司在美上市募资纪录，市场超额认购超7倍，凭借HBM在AI算力中的供应链瓶颈地位跻身万亿美元俱乐部。
- [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) 临时交易代码 SKHYV 末尾的“V”代表处于“发行时交易”（when-issued）状态，允许在股票宣布上市至股份最终结算交割之间先行交易。
- [03:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=202s) AI影视制作当前最有效的工作流并非纯文本直出，而是利用真人动捕或三维低模构建物理确定性底层素材，再通过AI模型进行风格迁移与神经渲染。
- [07:57](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=477s) GPT-5.6 Soul Ultra 在 Fast 模式下结合硬件加速达到每秒750 Token的极速输出，且直接通过键盘鼠标控制三维建模软件完成操作，展现出“AGI降临感”。
- [09:15](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=555s) 多模态模型的图像生成、前端代码构建与工具调用能力正走向彻底消融，模型已能根据外部网页风格参考自动设计高水准Logo与网页架构。
- [10:14](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=614s) 新版 ChatGPT 桌面端将 Codex 与原生客户端整合为 Electron 架构，因割裂原有 macOS 系统级快捷键交互习惯及信息架构混乱遭到社区批评。
- [12:40](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=760s) 用户对 AI 工具的终极诉求是一个无所不能的“全能输入框”（Omni box），OpenAI 将工作协同与自动化开发区分为不同工具的做法增加了认知摩擦。
- [14:29](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=869s) 外界所谓扎克伯格押注 AI 是为了逃避苹果30%抽成的论调存在商业逻辑漏洞，Meta 绝大多数广告营收结算于 Web 端，且广告归类为现实服务并不适用内购税。
- [14:29](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=869s) 苹果对 Meta 的真正重创是 ATT 应用跟踪透明度框架，核心定向广告机制受制于人促使扎克伯格必须借由智能眼镜等新硬件开辟独立计算平台。
- [17:13](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1033s) 千亿级 AI 算力资本开支为 Meta 硬件研发提供了财务掩护，智能眼镜一旦普及将通过视觉实时导购与本地商业抽成重构流量入口。
- [20:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1222s) SemiAnalysis 认为顶尖前沿模型取决于数据、人才与算力三要素，Meta 凭借庞大算力与战略执行力成为唯一全维度具备反超潜力的选手。
- [22:43](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1363s) 数据工程已告别低端密集型外包时代，Meta 调集3,000名工程师全职构建强化学习交互环境，高水准复杂任务设计承包商年薪已达百万美元。
- [24:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1484s) 导购插件 Phia 在用户结账时静默弹出后台标签页替换 Referral 追踪代码，完全绕开用户交互劫持原作者佣金，行为性质比 Honey 更加隐蔽恶劣。
- [29:43](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1783s) 继《后室》院线票房过亿后，好莱坞正将 Reddit、YouTube 与 Roblox 等网络平台成熟恐怖怪谈模因视作类似当年争抢热门剧本的新黄金赛道。
## Detailed walkthrough
### [00:02](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=2s) SK海力士纳斯达克挂牌：巨型海外IPO与首日破万亿美元市值
韩国内存芯片巨头SK海力士（SK Hynix）正式登陆纳斯达克市场，本次首次公开募股（IPO）一举募资265亿美元，创下了外国公司在美上市首发募资的历史最高纪录。在认购阶段，市场需求超过流通股配额的7倍以上。通常面临高达7倍超额认购的IPO，投资者普遍预期首日会出现80%甚至100%的暴涨，但在如此庞大的募资体量和估值基数下，SK海力士首日依然斩获了14%的涨幅，开盘表现相当强劲。该公司的美国存托凭证（ADR）以170美元开盘，相比149美元的发行价上涨14%，推动SK海力士的整体市值一举跨过1.03万亿美元大关，成为当前AI淘金热潮中诞生的又一家“万亿美元俱乐部”成员。
这场史诗级上市的背后，是全球AI系统对高带宽内存（HBM）近乎疯狂的饥渴需求，特别是与前沿AI加速芯片紧密配套的存储硬件。当前全球内存市场几乎被SK海力士、三星电子（Samsung）以及美光（Micron）三家瓜分。在过去一年中，由于投资者争先恐后希望切入这一AI基建中最关键的供应链瓶颈，SK海力士在韩国综合股价指数（KOSPI）挂牌的原生股票已累计飙升超600%。此前欧美普通投资者若想参与，必须费心去投资韩国本土市场，如今只需在纳斯达克直接交易ADR即可。早期的参与方包括了苏格兰著名投资机构柏基（Baillie Gifford）以及Situational Awareness等知名基金，首日账面浮盈均达14%。
面对《华尔街日报》等媒体关于SK海力士为何使用临时股票代码交易的疑问，节目中解释道，该股目前挂牌代码为“SKHYV”，随后将于下周一恢复为永久代码“SKHY”。代码末尾的“V”代表其目前处于“发行时交易”（when-issued）状态，即股票已经宣布上市并允许投资者在首个交易日先行买卖，但实际股份结算与交割尚未彻底完毕。这种机制并非所有ADR都会采用，而是一种帮助市场平稳过渡、填补公布与交付之间时间差的工具。虽然看走势图在开盘点位附近显得平缓，但相对于149美元的发行价，这场跨洋IPO已是一场彻底的大捷。
- **募资总额**：265亿美元（外国公司在美首发上市历史最大规模）
- **超额认购**：超过7倍需求
- **发行与开盘价**：发行价149美元，开盘价170美元（涨幅14%）
- **公司总市值**：约1.03万亿美元
- **交易代码变动**：首日以临时代码 `SKHYV`（when-issued 发行时交易）挂牌，后续转为正式代码 `SKHY`
---
### [03:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=202s) GPT-5.6亮相：复刻《星际穿越》与西兰花农夫的AI新现实
前一天GPT-5.6及其衍生版本5.6 Soul与Soul Ultra正式上线，社区里各类脑洞大开的Demo迅速在社交网络刷屏，同时也伴随着针对Mac桌面客户端的诸多争议。其中最博人眼球的Demo之一是有人尝试用AI复刻整部克里斯托弗·诺兰的科幻巨作《星际穿越》（Interstellar）。主持人调侃道，这简直是在精准踩中诺兰的雷区——毕竟这位大导演出了名地讨厌电脑特效（CGI），甚至尽量实拍，看到纯AI生成的粗糙3D模型，大概会气得七窍生烟。
尽管目前的粗模在短时间内不可能在院线拿到10亿美元票房，但用户仅凭一段提示词就能迅速走通“Prompt生成3D模型再直接渲染”的全流程，依然极具震撼力。这种体验正在催生一种全新的迷因文化和艺术门类。单纯完整复刻一部已经能在线流媒体观看的原片并无实际意义，未来的真正看点在于创意拼接和恶搞，比如“把男主角换成康纳·麦格雷戈（Conor McGregor）重新主演《星际穿越》”。
```mermaid
flowchart LR
accTitle: AI视频制作工作流
accDescr: 展示从低保真物理确定性参考到最终高保真AI视频的生成流程
subgraph S1["粗模与动作捕捉 (Deterministic Base)"]
A["创意构思与剧本"] --> B["iPhone实拍动捕 / 简易朗读台词"]
A --> C["Blender/Three.js 低模搭建与机械运动"]
end
subgraph S2["神经渲染与风格迁移 (AI Enhancement)"]
B --> D["AI视频生成模型"]
C --> D
D --> E["风格迁移层 (Style Transfer)"]
end
E --> F["高保真、写实影视级成片"]
```
目前AI影视制作最有效的工作流并非纯文本直出，而是先用确定性、符合物理规律的素材做底。创作者要么用iPhone随手录制真人表演来锁定台词节奏和神态，要么在Blender里用简单几何方块搭出运镜和车辆行驶路径，然后再用AI视频模型进行风格迁移与精细化渲染。由于底层参考具有严谨的空间物理逻辑，最终生成的AI成片质感远超凭空想象的幻觉片段。
Demo展示中另一个走红的案例是一位“用GPT-5.6打理农场的西兰花农夫”。知名科技创作者Daniel Tenreiro发梗图讽刺道：“我们原本被许诺能开上飞天汽车，结果现实里却只得到了一个用GPT-5.6种西兰花的农夫。”然而对于普通人而言，这种天马行空的非典型应用恰恰是AI最具魅力的地方。大家早就对AI写代码、做行业研报感到审美疲劳，反而迫切想知道AI如何渗入意想不到的日常琐事。针对彭博名记Joe Weisenthal所谓“AI毫无学习门槛”的论调，节目嘉宾反驳称，AI虽然看似只有一个输入框，但面对这片空白，绝大多数人根本不知道自己想象力的边界在哪，理解这个工具到底能做什么本身就需要一个漫长的启蒙过程。
---
### [07:57](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=477s) 每秒750 Token的极速操作与计算机控制的“AGI时刻”
社交媒体上另一段引发广泛讨论的视频展示了模型惊人的自动化操作速度。在视频未作任何加速处理的情况下，模型直接在三维建模软件Blender中以极高的手速构建起一门三维大炮，熟练地调用并组合球体、圆环等基础几何图元。社区用户Chris分析指出，这很可能是5.6 Soul Ultra在Fast模式下跑出的惊人表现，借助Cerebras硬件加速，实际输出速度推测高达每秒750个Token。
最让观众感到震撼的在于，这并不是通过专门开发的软件API或内部插件完成的后端调用，而是纯粹的“Computer Use”（直接控制计算机界面与鼠标键盘）。看着一台机器以完全超越人类生理极限的高速，噼里啪啦地直接操纵屏幕上的软件界面，带给人一种难以言喻的超现实感。除了那些把“软件极速盲打”当作专属才艺的网络极客之外，没有任何人类能达到这种操作频率，这一幕被主持人们戏称为切切实实体会到了“AGI降临的时刻”。
---
### [09:15](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=555s) Packy McCormick的意外Logo：前端建站与图像生成能力的界限消融
著名科技商业专栏作家Packy McCormick也分享了他测试GPT-5.6 Soul的惊艳体验。他原本只是给模型投掷了一个Cosmos设计风格的链接，让其参考生成网页前端，没想到模型在完全未收到指令的情况下，竟“自作主张”地直接一枪（one-shot）为他重新设计了一整套全新Logo。模型仿佛在对他说：“兄弟，看你这现成的Logo实在不行，我顺手给你重做了一个。”主持人笑称这个自动生成的Logo相当酷炫，Packy甚至真应该直接采纳。
长期以来，GPT系列模型在前端构建、排版审美与UI设计能力上一直被诟病落后于竞争对手，但5.6版本的迭代展现出了显著的跨越式跃升。这一案例甚至引来了OpenAI Codex团队的Nick在评论区询问这究竟是不是纯图像生成出来的。如今底层的技术逻辑正在经历快速融合：当用户要求系统搭建一个网站时，多模态模型不仅能直接输出高保真的预览效果图，而且其图像模型本身已经具备了工具调用（Tool Use）能力，甚至能根据需要现场生成底层HTML/CSS代码进行架构搭建。生成式图像、代码构建与系统控制之间的界限，正在这些前沿模型中彻底消融。
### [10:14](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=614s) AI 的闭环裁决与 ChatGPT 桌面端整合风波
节目中分享了一则颇具黑色幽默的荒诞轶事：团队里的 Michael 突然被 OpenAI 封号，封禁理由竟是所谓的“网络滥用”（cyber abuse），而他本人完全摸不着头脑。令人啼笑皆非的是，尽管账号被封，他不知通过什么途径依然打开了 Codex，并将封禁通知直接粘贴进去，让 Codex 帮他排查到底哪里踩了红线。Codex 迅速查明，原来是程序此前曾向他自己的服务器请求过 API 密钥而触发了风控。随后，Codex 自行撰写了一份申诉信并提交上去；仅仅几分钟后，OpenAI 端的审核 AI 就自动批准了该申诉并完成解封。整个过程堪称当代 AI 生态的绝妙写照：“他在大约 10 分钟内，被 AI 封禁，被 AI 定罪，被 AI 辩护，最后又被 AI 赦免。”
```mermaid
flowchart TD
accTitle: AI自主封禁与申辩闭环
accDescr: 展示用户被AI封禁后通过Codex排查、申诉并被审核AI秒速解封的全流程
subgraph 封禁阶段 ["风控触发"]
A["OpenAI 风控 AI 判定'网络滥用'"] --> B["Michael 账号被即时封禁"]
end
subgraph 申辩阶段 ["自主排查与反击"]
B --> C["将封禁通知发给 Codex"]
C --> D["Codex 查明原因为请求自身服务器 API Key"]
D --> E["Codex 自动撰写并提交申诉信"]
end
subgraph 赦免阶段 ["全自动结案"]
E --> F["OpenAI 审核 AI 自动秒级通过"]
F --> G["10 分钟内完成定罪到平反闭环"]
end
```
与此同时，新版 ChatGPT 桌面客户端的上线在社区激起了大量反弹。OpenAI 试图将其打造成一个大一统的“超级应用”，直接把 ChatGPT 客户端与 Codex 整合在一起——实质上现在的 ChatGPT 应用就是换皮的 Codex 客户端。虽然安装时允许保留旧版作为“ChatGPT Classic”，但老用户依然怨声载道。症结在于老版 ChatGPT 是深度集成的原生 macOS 应用，拥有极其流畅的系统级快捷键交互；而 Codex 客户端则是基于 Electron 架构打包的，本质上就是套壳浏览器与 Web 视图，既笨重又丢失了大家早已习惯的原生体验。不过主持人也坦言，虽然有同事曾把原生快捷键当成各种场景下的“作弊利器”，但普通人日常究竟有多少比例会常驻打开 Mac 桌面客户端本身也存疑。
- **快捷键惯性**：macOS 原生客户端可通过便捷快捷键直接呼出模型交互界面，而改用新版后若需自行重新绑定热键，对用户而言构成了额外的迁移摩擦成本。
- **工作流割裂**：不少用户的标准习惯是在桌面端常驻 Codex 负责重度任务，在浏览器新标签页里进行常规对话，再通过手机端移动应用随时与桌面端 Codex 协同，原本就极少专门打开独立的 ChatGPT 桌面应用。
### [12:40](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=760s) OpenAI 官方回应、计算限额重置与“全能输入框”之争
针对铺天盖地的负面舆情，OpenAI 桌面产品负责人 Thibault 迅速发文救火，承认团队过去 24 小时内密集翻阅了社区反馈并调取了后台使用数据。他坦言虽然外界对 GPT-5.6 和 ChatGPT 的新特性充满兴奋，但团队在多个产品细节上确实操之过急。首先，产品界面让用户过分轻松地滑向了最高算力档位，却未能清晰提示该操作对使用限额（usage limits）的剧烈消耗；其次，桌面端的大幅激进改版打乱了信息架构，导致用户最熟悉的对话列表与项目目录变得极难寻找；更关键的是，发布会过度强调“工作场景”的定位，甚至引发了 Codex 核心铁粉的恐慌，误以为 Codex 未来会被边缘化乃至彻底砍掉。Thibault 强调团队深爱 Codex 且绝无此意，并当即宣布当天将两度重置用户的使用额度，供大家尽情测试。
对于普通用户来说，目前最大的困惑莫过于分不清“ChatGPT Work”与“ChatGPT Codex”到底有何本质区别。主持人们指出，普通人根本不想在两个独立的工具之间反复横跳，大家真正想要的是一个无所不能的“全能输入框”（Omni box）——就像 Google 一样，从来不存在所谓的“企业级 Google 搜索”与“家庭 Google 搜索”的分野，所有人都在同一个框里检索一切。尽管将工作流、数据集成点与深度代理融合到单一入口需要经历“爬、走、跑”的渐进周期，但终极形态必然是合二为一。现场更调侃起将 Codex 誉为“思想之马”（A horse for the mind）的业界赞誉，笑称其定位已超越了乔布斯当年的“思想自行车”。
- **限额补偿措施**：为平息由于算力消耗提示不清引发的额度熔断焦虑，OpenAI 在发文当天主动执行了两次使用上限刷新。
- **产品认知分歧**：用户期望 AI 呈现为零学习成本的单一超级入口，而 OpenAI 现阶段的产品矩阵仍在企业协同（Work）与全能自动化工作空间（Codex）之间摇摆。
### [14:29](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=869s) 扎克伯格的 AI 豪赌：苹果 30% 税率的误读与反追踪真相
讨论转向了 Meta 的长期战略以及 SemiAnalysis 近期针对 Meta（MSL）算力与新业务线撰写的反向看多深度分析。针对社交网络上诸如“Boring Business”等博主声称“扎克伯格全力押注 AI 是因为他誓死不愿余生继续向平台方缴纳 30% 苹果税”的流行论调，主持人展开了犀利的辩驳。这种观点认为 Meta 抓住了 2010 年代初的移动浪潮，却因缺乏自研硬件生态而不得不向苹果长期进贡数十亿美元的应用内购抽成。
主持人直接指出这一逻辑漏洞百出：“没有任何一个广告主是通过 iOS 应用内购来采购 Meta 广告的。”现实中几乎所有的广告投放与结算均在 Web 端管理后台完成，而且广告曝光本质上是购买人类的眼球与注意力，在商业法规上被界定为现实物理世界的服务（Real-world service）。就像用户在 iPhone 上的 Amazon 购买实体 T恤、在 DoorDash 点外卖或购买电影票一样，苹果根本不可能抽取 30% 的佣金（只有订阅 Netflix、X Premium 或是购买《堡垒之夜》V-Bucks 等纯虚拟数字商品时才适用抽成）。虽然 Instagram 移动端提供给小型商户的“一键推广贴文”（Boost Post）功能确实被苹果强行划入内购范围收税，但这在 Meta 千亿规模的广告大盘中完全无足轻重。
- **真正的数亿美元血亏**：苹果真正对 Meta 造成致命打击的从来不是 30% 的抽成，而是隐私框架中的“应用跟踪透明度”（ATT）授权弹窗，此举直接掐断了跨平台的归因与转化数据流，重创了定向广告引擎。
- **寄人篱下的根本危机**：私下里苹果高管对 Meta 毫无好感已是公开的秘密，一家全球市值名列前茅的巨头却在核心生死线上始终受制于敌对平台的规则与权限审核，这才是扎克伯格的心头之患。
### [17:13](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1033s) 摆脱寄人篱下的困局：智能眼镜平台与被 AI 吞噬的屏幕时间
扎克伯格之所以不惜代价狂砸 AI，核心驱动力在于他坚决不愿再扮演一个“凡事都要向别人请求准许”的受害者角色。在未来的技术变革中，传统多款社交 App 共存的模式随时可能被颠覆——个人专属的 AI 代理可能随时根据用户意图在后台动态组合内容，用户根本不再需要打开 Instagram 或 Facebook。面对这种平台范式重构的未知风暴，主动参与并力争成为周期的定义者，远比被动祈祷丧钟不会敲响更安全。
更深层的破局点在于智能眼镜这一新型硬件载体。在扎克伯格的宏大蓝图里，极度廉价甚至免费普及的智能眼镜将成为彻底终结 iPhone 与 PC 垄断的下一代个人计算平台。依托全天候在线的环境智能（Always-on intelligence），用户走在街上看中某样物品随口一句“帮我买下这个”，或是路过餐厅镜片上弹出“消费满 20 美元赠送波霸奶茶”的即时优惠，Meta 都能直接在本地交易与电商闭环中抽成。此外，当下 AI 军备竞赛那动辄 2000 亿美元级别的数据中心资本开支，反而为硬件研发提供了极佳的财务掩护：扎克伯格完全可以名正言顺地将 Reality Labs 的巨额投入重新打包藏进总预算里，上演一出华尔街之狼式的“我绝不退场”；即便下一代 Meta Ray-Ban 智能眼镜的屏幕研发再多烧掉 80 亿美元，在天文数字般的算力基础设施开支面前也只会被外界视作一笔无关痛痒的“四舍五入抹零”。
```mermaid
flowchart LR
accTitle: 扎克伯格摆脱手机生态的战略跃迁
accDescr: 展示Meta如何通过智能眼镜与生成式AI重构商业底座，从被动防御转向主动掌控入口
subgraph 历史阵痛 ["移动时代受制于人"]
A["依赖 iOS/Android 宿主平台"] --> B["苹果 ATT 弹窗掐断跨平台跟踪"]
B --> C["核心定向广告每年遭受百亿级冲击"]
end
subgraph 硬件战略破局 ["下一代独立硬件平台"]
D["Meta Ray-Ban / AR 眼镜"] --> E["极低成本乃至免费普及终端"]
E --> F["替代手机成为常驻环境计算入口"]
end
subgraph 商业模式重构 ["闭环商业新变现"]
F --> G["视觉'所见即买'交易抽成"]
F --> H["LBS 线下商户动态优惠导流"]
H --> I["彻底摆脱对应用商店权限的乞求"]
end
```
- **资本掩护效应**：动辄数千亿美元的 AI 基础设施支出极大地拉高了投资人的心理阈值，使得元宇宙与硬件实验室（Reality Labs）原本刺眼的巨额亏损能够轻松隐匿在大规模 Capex 之中。
- **屏幕时间的迁移信号**：作为早期技术采纳者，主持人的个人屏幕使用时间正在剧烈向 ChatGPT 等 LLM 转移；过去在 Instagram 刷到感兴趣的汽车视频会引发一系列繁杂的 Google 搜索，如今则演变为直接切屏向 AI 进行深度提问与研究报告生成，注意力经济的底层正在被重塑。
### [20:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1222s) AI 争夺用户注意力与 SemiAnalysis 的前沿模型三要素论
随着用户行为的变化，人们花在 AI 系统上的时间正急剧增加。主持人指出，自己如今每天花在 AI 对话应用上的时间已经达到一小时，明显超过了刷 Instagram 的时间；无论这是早期采用者的特例，还是普遍趋势的开端，科技巨头们都清楚地意识到，凡是消费者停留时间最长的地方，都必须牢牢占据立足点。正是在这一背景下，半导体与 AI 分析机构 SemiAnalysis 发布重磅分析，断言构建真正的顶尖前沿模型（Frontier Model）核心只需要三样东西：数据（Data）、人才（Talent）与算力（Compute）。SemiAnalysis 认为，在所有超大规模云厂商和新兴 AI 实验室中，Meta 是唯一一家在这三项维度上均有望达到世界顶尖水平的公司，因此它拥有追赶并反超 Anthropic 与 OpenAI 的最大几率。
Scale AI 首席执行官 Alexandr Wang 在社交平台转发了这一观点并调侃道：“算力教父（Compute Daddy）Dylan Patel 发话了。”紧接着，市场迅速做出反应，Meta 股价在开盘时暴涨超过 6%。尽管有预测市场账号将股价上涨归因于前一日发布的 Muse Spark 1.1，但主持人们一致认为，资本市场真正买账的是 SemiAnalysis 勾勒出的宏大叙事。对于其他竞争者，SemiAnalysis 的评估极为苛刻：报告提到 xAI 等公司每年向 Anthropic 和 Google 出售价值数十亿美元的计算资源，因此在模型竞争格局中存在不同定位；更尖锐的是，这一观点暗指 Google Gemini 虽然坐拥庞大数据与算力，却在顶尖 AI 研发与工程人才的凝聚力上略显逊色，无法三者兼备。
- **前沿模型三大基石**：数据（Data）、顶尖人才（Talent）、基础设施算力（Compute）。
- **市场异动与评价**：Meta 股价受分析报告提振开盘大涨逾 6%；Alexandr Wang 戏称 Dylan Patel 为“算力教父”（Compute Daddy）。
```mermaid
flowchart TD
accTitle: 前沿模型竞争要素与 Meta 优势
accDescr: 展示 SemiAnalysis 提出的前沿模型三要素及 Meta 的布局优势
subgraph Core["前沿模型三大基石"]
D["优质数据 (Data)"]
T["工程人才 (Talent)"]
C["庞大算力 (Compute)"]
end
subgraph Advantage["Meta 转型落地"]
E["3,000 名工程师团队转型"]
R["全职设计 RL 环境与任务"]
Q["高水准专家级数据标注"]
end
Core --> Advantage
Advantage --> Goal["具备叫板 OpenAI 与 Anthropic 的综合实力"]
```
### [22:43](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1363s) 告别廉价标注：Meta 的强化学习环境革命与百万美元数据工程师
Meta 在数据工程上的战略转折始于其近期重组。在最新一轮裁员与架构调整中，扎克伯格（Mark Zuckerberg）主导成立了全新的应用 AI 工程团队，直接调动了 3,000 名工程师——其中包括 70% 的应届毕业生以及大量资深工程师——全职负责构建强化学习（RL）任务与交互环境。主持人们探讨认为，这展现了扎克伯格雷厉风行的执行力。相比之下，Google DeepMind 虽拥有相似规模的人才梯队，却未能果断打破既有组织惯性去“撕掉创可贴”。在数据采购方面，Anthropic 此前一直是购买强化学习环境与代码数据最激进的实验室，这也是其模型代码能力拔群的关键原因；例如，Mercor 曾披露其平台在 2026 年第二季度就记录了 250 万个专家工时，相当于 5,000 人全职投入，而 Meta 如今内部调动的研发力量已足以与此类规模匹敌，且平均数据质量更高，背后甚至还有 70,000 名员工随时待命。
主持人们特别驳斥了外界对于“让顶尖工程师去干低端数据标注”的误解。早年由低学历外包劳工画边界框（Bounding Box）或对 NSFW（不合规内容）进行粗暴分类的劳动密集型时代早已一去不复返。当前大语言模型已经足够聪明，创建具有启发性、高维度的优质训练数据本身就是极具智力挑战的工作。设计强化学习环境不仅绝非枯燥乏味的苦差，反而是极其严谨的科学探索。如今在各大头部数据公司中，能够精确定义并从头到尾规范复杂任务解决逻辑的顶级专家承包商，年薪甚至已达到七位数（上百万美元），彻底颠覆了以往“每小时 10 美元廉价零工”的刻板认知。
- **Meta 人才调配规模**：3,000 名工程师（涵盖 70% 新毕业生及众多资深骨干）全职投入 RL 环境与任务开发，后续具备动员数万员工的潜力。
- **数据复杂度升级**：数据生产进入智力密集型阶段，顶级任务设计承包商年薪突破百万美元级别（Seven Figures）。
### [24:44](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1484s) “PhiaGate”爆发：菲比·盖茨的购物插件被曝劫持返利链接
随后节目探讨了当天最具争议的商业八卦——“Phia 丑闻”（PhiaGate）。由微创创始人比尔·盖茨（Bill Gates）的女儿菲比·盖茨（Phoebe Gates）联合创办的购物应用 Phia 登上了彭博社（Bloomberg）的头条，面临严重的电商推广佣金归因劫持指控。Phia 对外标榜为一款个人购物助手，旨在帮助用户搜寻服装和时尚配饰的最低全网优惠；按行业惯例，这类导购工具主要依靠返利推广程序（Affiliate Marketing Programs）向达成交易的零售商收取销售佣金。然而，彭博社联合 Capital One Shopping 与 Edelman 进行了独立代码审计与实测，在覆盖超过 50 个主流购物网站后发现：Phia 的移动浏览器插件在用户结账付款时，会在后台静默弹出一个无感知标签页，强行注入 Phia 自己的推广返利代码，直接覆盖了内容创作者或原始出版商应得的合法推广代码。
主持人们将此事件与此前知名 YouTube 博主 MegaLag 揭露 Honey 插件的 2000 万播放量轰动调查进行了对比。过去 Honey 的争议在于诱使创作者推广插件，却在后台利用重定向手段窃取博主原本推广合作品牌（如 AG1）的佣金提成，使创作者陷入“推广自杀”的讽刺境地。然而，Phia 的行径被指责比 Honey 更加恶劣和粗暴：Honey 至少还需要用户在结账时点击交互弹出窗口（哪怕点击取消），而 Phia 则是“完全静默运行”。用户可能刚在评测网站看完了商品推荐，顺着链接进店购买，原本应归属于评测网站作者的佣金，却在用户毫不知情的情况下被后台潜伏的 Phia 插件全盘截胡。
- **技术审计发现**：Phia 插件在结账期间未经用户交互便打开后台标签页，强行劫持并重写 Referral 代码。
- **波及规模与涉案金额**：据估算，该插件至少劫持带动了 1,200 万美元以上的线上交易额，涉及约 100 万美元被挪用的推广佣金分成。
- **舆论与明星资本反噬**：Phia 此前融资时吸引了大量好莱坞名流与知名网红站台，投资阵容豪华得如同“科切拉音乐节海报”（Coachella poster）。面对长期以来网络上对其背靠巨富资源创业的冷嘲热讽，舆论迅速引爆，主持人们戏称这或许将促成其个人“反派崛起路线”（Villain Arc）。
### [29:43](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1783s) 好莱坞狂欢：网络恐怖怪谈 IP 开启吸金热潮
在节目的尾声，话题转向影视娱乐工业正在掀起的新风口：好莱坞对互联网恐怖迷因（Horror Meme）的狂热掘金。近年来，起源于 4chan 匿名论坛的都市传说与 Creepypasta 概念正被迅速商业化。典型的范例莫过于《后室》（The Back Rooms）——从论坛文字点子演变成风靡 YouTube 的超自然短片频道，最终被好莱坞知名制片厂相中拍成院线大片，票房突破一亿美元。当下，敏锐的影视资本与制片人正争先恐后地在各大社交平台扫荡类似的恐怖怪谈网络 IP，掀起了一场点石成金的造富热潮。
### [30:16](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1816s) 好莱坞掀起网络恐怖 IP 争夺热潮：从《警笛头》到《曼德拉记录》
《华尔街日报》（The Wall Street Journal）近期报道了一则引人注目的交易：知名网络怪物“警笛头”（Siren Head）的创作者特雷弗·亨德森（Trevor Henderson）成功拿下了一笔丰厚的版权交易。在好莱坞各大制片厂疯狂追逐下一个类似《后室》（The Backrooms）现象级爆款的浪潮中，这位 40 岁的插画家终于迎来了迟到多年的商业回报。早在八年前，亨德森画出了这个长着警笛脑袋、令人毛骨悚然的无面怪物形象。在此后的岁月里，这个恐怖模因迅速在互联网上疯传，催生了无数 YouTube 衍生短片、亚马逊上的山寨周边商品以及各类同人电子游戏，但亨德森本人此前却“未曾从中赚到过一分钱”。如今，他正式将《警笛头》的电影改编权以超过 100 万美元的高价出售给了华纳兄弟（Warner Bros.）。
据知情人士透露，这笔交易折射出当下好莱坞的新一轮“淘金热”——各大制片厂正全力在互联网上搜寻兼具概念和原创能力的人才，试图打造出下一个能在 Z 世代中引发轰动的恐怖爆款。好莱坞看中的正是这些 IP 已经在 YouTube、Reddit 甚至 Roblox 等平台上积累的庞大年轻受众与验证过的号召力。华纳兄弟电影集团联合主席迈克尔·德卢卡（Michael De Luca）表示：“我们看待这些网络资产作为改编资源的方式，就和我们审视书籍及其他传统媒介一模一样。”这种对热门网络数字版权的狂热争夺，甚至让人联想起了上世纪八九十年代好莱坞为争抢热门剧本而展开的激烈厮杀。
另一个极为典型的案例是 YouTube 心理恐怖伪纪录片系列《曼德拉记录》（The Mandela Catalogue）。近期有多达 11 家制片厂参与竞拍该系列的电影改编版权，最终由亚马逊旗下的联艺电影公司（United Artists）与史蒂文·斯皮尔伯格创办的安培林娱乐（Amblin Entertainment）联手胜出。两家公司不仅支付了数百万美元的高额版权费，更直接拍板同意让该系列年仅 22 岁的原创作者亚历克斯·基斯特（Alex Kister）亲自担任电影导演。
- **《警笛头》（Siren Head）交易**：由插画家特雷弗·亨德森于 8 年前创作，在长期未获得实质商业收益后，以逾 100 万美元的价格将电影版权卖给华纳兄弟。
- **《曼德拉记录》（The Mandela Catalogue）争夺战**：吸引 11 家制片厂竞标，最终由联艺电影与安培林娱乐以数百万美元拿下版权，并由 22 岁原作者亚历克斯·基斯特执导。
- **目标孵化平台**：制片厂重点监控 YouTube、Reddit 以及 Roblox 上具有 Z 世代影响力的原生恐怖内容。
### [31:51](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1911s) 观众对恐怖片的困惑、乡间童年回忆与节目收尾
对于互联网恐怖热潮席卷好莱坞的现象，观众丹尼尔（Daniel）在留言中表达了不解与吐槽：“大家对恐怖片到底有什么好着迷的？喜欢看这种东西的人多半脑子有点问题。”主持人对此表示完全赞同，坦言自己也从来不爱看恐怖电影，并分享了自己的一套独特哲学：只要平时不看恐怖片，夜里走夜路时就不会自己吓自己。主持人回忆起自己在乡间长大的经历，作为在乡村度过青春期的少年，经常需要在夜间四处游荡、冒险，如果满脑子都是恐怖电影里的怪力乱神，根本无法安心在黑夜中独自行走。
在节目的最后，主持人们照例呼吁听众在 Apple Podcasts 和 Spotify 上为节目留下五星好评，并前往官网 tbpn.com 订阅每日新闻通讯，随后以一句俏皮的“扔闪光弹了”（throwing flashbang）和大家道别，相约下周一再见。
## Notable quotes
> [03:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=202s) 我们原本被许诺能开上飞天汽车，结果现实里却只得到了一个用GPT-5.6种西兰花的农夫。 — Daniel Tenreiro
> [09:15](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=555s) 兄弟，看你这现成的Logo实在不行，我顺手给你重做了一个。 — Packy McCormick 描述 GPT-5.6
> [10:14](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=614s) 他在大约 10 分钟内，被 AI 封禁，被 AI 定罪，被 AI 辩护，最后又被 AI 赦免。 — 主持人
> [12:40](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=760s) 思想之马
> [14:29](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=869s) 没有任何一个广告主是通过 iOS 应用内购来采购 Meta 广告的。 — 主持人
> [17:13](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1033s) 凡事都要向别人请求准许 — 扎克伯格（观点引述）
> [17:13](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1033s) 我绝不退场
> [20:22](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1222s) 算力教父（Compute Daddy）Dylan Patel 发话了。 — Alexandr Wang
> [30:16](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1816s) 未曾从中赚到过一分钱 — 特雷弗·亨德森（交易背景描述）
> [30:16](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1816s) 我们看待这些网络资产作为改编资源的方式，就和我们审视书籍及其他传统媒介一模一样。 — 迈克尔·德卢卡
> [31:51](https://www.youtube.com/watch?v=SPIMqPIrLUE&t=1911s) 大家对恐怖片到底有什么好着迷的？喜欢看这种东西的人多半脑子有点问题。 — Daniel
## Mentioned
TBPN, John Coogan, Jordi Hays, SK海力士 (SK Hynix), 纳斯达克 (Nasdaq), 三星电子 (Samsung), 美光 (Micron), 柏基 (Baillie Gifford), Situational Awareness, 华尔街日报 (The Wall Street Journal), OpenAI, GPT-5.6, GPT-5.6 Soul, GPT-5.6 Soul Ultra, Codex, ChatGPT, ChatGPT Classic, 克里斯托弗·诺兰 (Christopher Nolan), 星际穿越 (Interstellar), 康纳·麦格雷戈 (Conor McGregor), Daniel Tenreiro, Joe Weisenthal, Cerebras, Blender, Packy McCormick, Cosmos, Thibault, Meta, 马克·扎克伯格 (Mark Zuckerberg), SemiAnalysis, Dylan Patel, Apple (苹果), Instagram, Facebook, DoorDash, Netflix, Meta Ray-Ban, Reality Labs, Scale AI, Alexandr Wang, Muse Spark 1.1, Anthropic, Google, Google DeepMind, Gemini, xAI, Mercor, 菲比·盖茨 (Phoebe Gates), 比尔·盖茨 (Bill Gates), Phia, Honey, Capital One Shopping, Edelman, MegaLag, 特雷弗·亨德森 (Trevor Henderson), 警笛头 (Siren Head), 华纳兄弟 (Warner Bros.), 迈克尔·德卢卡 (Michael De Luca), 曼德拉记录 (The Mandela Catalogue), 亚历克斯·基斯特 (Alex Kister), 联艺电影 (United Artists), 安培林娱乐 (Amblin Entertainment), 史蒂文·斯皮尔伯格 (Steven Spielberg), 后室 (The Backrooms), Roblox, Reddit, YouTube, 4chan
## 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
[Source](https://www.youtube.com/watch?v=SPIMqPIrLUE)
Source quotes
[3:22] Daniel Tenreiro: 我们原本被许诺能开上飞天汽车，结果现实里却只得到了一个用GPT-5.6种西兰花的农夫。
[9:15] Packy McCormick 描述 GPT-5.6: 兄弟，看你这现成的Logo实在不行，我顺手给你重做了一个。
[10:14] 主持人: 他在大约 10 分钟内，被 AI 封禁，被 AI 定罪，被 AI 辩护，最后又被 AI 赦免。
[12:40] 思想之马
[14:29] 主持人: 没有任何一个广告主是通过 iOS 应用内购来采购 Meta 广告的。
[17:13] 扎克伯格（观点引述）: 凡事都要向别人请求准许
[17:13] 我绝不退场
[20:22] Alexandr Wang: 算力教父（Compute Daddy）Dylan Patel 发话了。
[30:16] 特雷弗·亨德森（交易背景描述）: 未曾从中赚到过一分钱
[30:16] 迈克尔·德卢卡: 我们看待这些网络资产作为改编资源的方式，就和我们审视书籍及其他传统媒介一模一样。
[31:51] Daniel: 大家对恐怖片到底有什么好着迷的？喜欢看这种东西的人多半脑子有点问题。
Source quotes
[0:32] You can go and recreate *Interstellar* perfectly, but at the end you just have *Interstellar*, which already exists and is available to stream... you have to bring something special to the table.
[0:09] Thibault: The short version is that there's a lot of excitement for 5.6 Chat, but we also didn't get everything quite right... We made it too easy to use the highest compute settings without making the impact on usage limits sufficiently clear.
[29:44] Michael De Luca: We're seeing them as a resource for adaptations the same way we look at books and other media.
Original transcript
[0:02] SKH Highix is up. Successful IPO up 14%
[0:06] in the Wall Street debut. The South
[0:09] Korean memory chip giant SKH Highix made
[0:13] his NASDAQ debut today, raising 26.5
[0:16] billion in the largest ever firsttime US
[0:22] share sale by a foreign company. demand
[0:25] reportedly exceeded the available shares
[0:27] more than sevenfold. But to give an to
[0:30] give a sense of like the scale of this
[0:32] thing, normally when we hear about 7x
[0:34] over subscribed IPOs, you might be
[0:36] expecting like a 100% pop, 80% pop, but
[0:39] at at this level of price and demand, uh
[0:41] you're just seeing a 14% pop, but still
[0:44] a great debut nonetheless. The company's
[0:46] American Depository receipts opened, the
[0:49] ADR shares opened at 170, 14% above the
[0:53] 149 they were offered at, giving X SKH
[0:55] Highex a market cap of roughly $1.03
[0:59] trillion. We got another one another
[1:03] >> another one
[1:03] >> another trillion dollar company in the
[1:05] AI boom. The blockbuster list encaps
[1:09] memory manufacturers fueled by surging
[1:11] demand from AI systems, particularly for
[1:14] high bandwidth memory used alongside
[1:16] advanced AI accelerators. SKH, Samsung,
[1:19] and US-based Micron dominate the global
[1:21] memory market. And SKH is Korean listed
[1:25] shares are up over 600%
[1:28] in the past year as rushed to gain
[1:31] exposure to one of the AI boom's most
[1:33] critical bottlenecks. But it's been
[1:34] tricky to get shares and participate in
[1:36] SKHX because you got to go over to the
[1:38] Kusby. You got to go over to the South
[1:39] Korean index instead. Now you can just
[1:42] head to the NASDAQ and get your ADR
[1:44] shares in SKHEX. So congratulations to
[1:47] the whole team over there. A massively
[1:50] successful IPO. The Wall Street Journal
[1:52] asks a question. They say, "Why is
[1:54] SKHENX trading under a temporary ticker
[1:57] symbol?" Uh they're currently trading
[1:59] under the temporary ticker SKHYV
[2:03] before reverting to a permanent ticker
[2:05] on Monday. SKHY.
[2:07] The V in the temporary symbol indicates
[2:10] the shares are trading on a whenissued
[2:12] basis. They haven't been fully issued
[2:14] yet. Meaning the exchange is allowing
[2:16] investors to buy and sell the stock on
[2:18] its first US trading day before the
[2:20] transaction actually settles. They're
[2:22] sort of being eased into the American
[2:23] markets. This is not the case with all
[2:25] ADRs, all American depository receipts,
[2:27] but rather a tool used to close the gap
[2:29] when a transaction has been announced,
[2:31] but the actual delivery of the shares
[2:33] hasn't finished. So, all in process now
[2:36] to bring SKH Heinix to the American
[2:38] markets. And that looks like a lot of
[2:39] red, but the stock is up. Uh, it just
[2:42] opened uh at 170. And so, the fact that
[2:45] it is flat relative to where it opened
[2:48] does not reflect the 14% pop or 16% pop.
[2:52] What was it? 14% pop from the actual
[2:55] issuance price. Most of the folks that
[2:56] bought shares in the IPO, that 26
[2:59] billion dollars that went into the
[3:01] company, uh, of course, situational
[3:03] awareness, there were a few others. Uh,
[3:04] Bailey Gford all participated early, at
[3:07] least according to early news. Uh, I
[3:10] assume that those transactions closed.
[3:12] Uh, they are all sitting up 14% today.
[3:15] Uh, and we'll be continuing to follow
[3:17] this. Caner in the chat says CERN has
[3:19] only been off for a week and John is
[3:21] dressing casual.
[3:22] >> You think it's a conspiracy? I like
[3:25] [laughter] it. I like it. 5.6 launched
[3:28] yesterday. 5.6 Soul and Soul Ultra and
[3:31] people are having fun with it. More and
[3:32] more demos are hitting the timeline.
[3:35] There's a lot of controversy over the
[3:37] over the desktop app for Mac. We can get
[3:40] into that. But first, let's play a
[3:42] little bit of this video. Someone tried
[3:44] to recreate all of Interstellar. truly
[3:47] trying to uh infuriate Christopher Nolan
[3:50] the most. Probably this is the exact
[3:52] opposite of what he wants to see. I
[3:54] imagine considering that he doesn't even
[3:56] use uh CGI very often. But this is an
[4:00] incredible demo. Although uh I don't
[4:02] think this will be uh picking up a
[4:04] billion dollars at the box office
[4:05] anytime soon. But it is uh it is
[4:08] incredible that uh you're able to sort
[4:11] of go from prompt to 3D model to
[4:15] rendering so quickly. Uh very very cool
[4:18] demo and it feels like a new meme, a new
[4:21] a new art form. Uh you have to be
[4:23] creative with it. Merely recreating
[4:26] Interstellar is not going to do it. It's
[4:28] not going to do anything other than just
[4:30] be an an impressive demo of the
[4:31] technology. But uh when you look at what
[4:34] people have done with Blender, you can
[4:36] imagine that with creativity added to
[4:38] this and a new story line and something
[4:40] that leans into the rough edges of this
[4:42] particular
[4:43] >> Well, there's also going to there's also
[4:44] going to be a whole meta of like I
[4:46] recreated Interstellar, but I cast
[4:49] >> Conor McGregor as Yeah. You know,
[4:51] >> and I imagine that this gets even more
[4:54] powerful when you use uh style transfer
[4:57] and and video models on top. I've been
[4:59] seeing a lot of people block out uh AI
[5:02] videos either just using their iPhone.
[5:04] So they'll just have like you're playing
[5:06] an actor who's playing like an old
[5:08] character and then you're reading the
[5:09] lines, you're giving the delivery and
[5:11] then that is that is used as the driving
[5:13] video for some of these uh some of these
[5:16] AI video models and you get much better
[5:18] results with that. Uh, and I've also
[5:20] seen people block out little uh examples
[5:23] of driving scenes or different uh
[5:26] mechanical scenes in Blender and then
[5:29] use an AI video model to take it to high
[5:31] fidelity. And because the video model
[5:34] has an underlying reference that is
[5:36] deterministic and physically accurate
[5:38] even though it's low fidelity and very
[5:40] like blocky, you wind up with a better
[5:43] AI video at the end of the project. So
[5:45] you could imagine someone taking an
[5:47] idea, puppeteering a 3JS model like
[5:51] what's happening here and then adding a
[5:53] basically a style transfer layer on top
[5:55] with AI video to get to something that's
[5:57] a lot more photoreal and more striking.
[5:59] But again, you wind up in this situation
[6:02] where uh if you you can go and recreate
[6:05] Interstellar perfectly, but at the end
[6:07] you just have Interstellar which already
[6:09] exists and is available to stream and so
[6:11] you didn't really get anything in the
[6:12] process. you have to bring something
[6:14] special to the table. Um, but I'm
[6:16] excited for the inevitable mashups of
[6:18] this, the the Conor McGregor starring in
[6:20] Interstellar or something like that.
[6:22] Also, there's a broccoli farmer that's
[6:23] running his farm on GPT 5.6 apparently.
[6:26] Uh, this was in the demo video, but
[6:28] Daniel 10 Row shares a very funny meme
[6:32] showing the uh I don't even know what
[6:35] aesthetic this is. This is the share
[6:37] zone aesthetic. That was the account
[6:39] that shared all of this. the extreme uh
[6:41] skeleton with the two revolving pistols.
[6:45] Uh very fun. Daniel Tenro shares some
[6:47] more information. He says, "We were
[6:49] promised flying cars and instead we got
[6:50] a broccoli farmer running his farm with
[6:52] GPT 5.6." I don't know. Uh it's still
[6:55] relevant to show a demo uh that is as
[6:58] far a field. Everyone knows how these
[7:00] models can be used for research, how
[7:01] these models can be used for coding.
[7:02] That gets a little boring. I want to
[7:04] imagine something completely different
[7:06] that I can then pull into my daily life.
[7:08] what is the equivalent task that I'm
[7:10] doing in my life that I can take? Uh we
[7:12] we we sort of differ on this sometimes.
[7:14] You you think that uh like a lot of
[7:16] these things are self-explanatory. They
[7:17] don't this is Joe Weisenthal position
[7:19] too. This is like AI has no learning
[7:21] curve. I I disagree. I I I I think that
[7:24] it doesn't necessarily have a learning
[7:25] curve, but there is a an educational
[7:28] process that happens when you're
[7:30] presented with a blank box. You can type
[7:32] anything.
[7:34] People need to understand the fringes of
[7:36] their imagination. what is possible and
[7:39] and then they can go and reach for that
[7:41] tool. But the tools are getting sharper
[7:43] and sometimes they get sharper without
[7:45] the box actually changing because it's
[7:46] just a box of what do you want to build
[7:48] or what do you want to do?
[7:49] >> Not not everyone has a Tyler where I I I
[7:53] tend to prompt Tyler.
[7:54] >> Yes. Yes. Goes into codeex.
[7:57] >> Yes. There was another video that was uh
[7:59] shaking up the timeline. People were
[8:00] going back and forth on how fast this
[8:03] model was building in uh Is this
[8:05] Blender? Blender.
[8:06] >> This is Blender, right? Yeah. Uh so it's
[8:08] building a 3D cannon uh modeling it and
[8:11] adding different uh different
[8:14] primitives, different spheres and
[8:16] Tauruses. Uh and uh the video is not
[8:20] sped up. And Chris says soul f uh 5.6
[8:24] soul is or and what I assume is 750
[8:28] tokens per second in the wild. That is
[8:31] the uh preview of Cerebrus. that this
[8:34] apparently was not actually uh
[8:37] >> yeah so ultra on fast mode
[8:40] >> so we should be seeing even faster
[8:43] rollouts although every time I see this
[8:45] I think
[8:45] >> and this is just computer use
[8:48] >> like this is not an like an integration
[8:51] just
[8:53] >> yeah it is an interesting experience
[8:55] watching a machine use your machine
[9:00] >> at a rate that would be completely
[9:03] impossible Unless it was like your
[9:05] entire bit online was like I just use
[9:07] software really fast.
[9:08] >> Yeah.
[9:09] >> Um it it is uh yeah it's a interesting
[9:13] sort of feel the AGI moment.
[9:15] >> Anyway, Packy McCormack is having fun
[9:16] with 5.6 Soul. He says it oneshotted
[9:19] this on its own. He didn't ask for a new
[9:21] logo. [laughter] Just like it's just
[9:23] like look bro you need a new logo.
[9:25] >> Actually I actually think you should
[9:27] adopt this.
[9:28] >> It looks cool. I like it. He says uh
[9:30] it's part of a web design based on a
[9:31] link to a cosmos. not perfect, but way
[9:34] better than any other time I've tried.
[9:35] And that's been the big uh the big
[9:36] talking point is like they were behind
[9:38] on or the GPT models were sort of behind
[9:40] on front end and design and this is a
[9:42] big leap forward in that. So I'm excited
[9:44] to see
[9:45] >> Nick on the Codex team is asking is this
[9:47] image gen? I think it might be.
[9:48] >> Well, I mean that's the thing about soul
[9:51] and and codeex
[9:52] >> but it's been generating it's been now
[9:54] if you're asking it to build a site it
[9:57] will generate you a preview image
[9:59] >> and so I think this is the preview image
[10:00] for the site.
[10:01] >> Yep. which is
[10:02] >> the really confusing thing is that the
[10:04] image model now has tool use and can
[10:06] scaffold things out in HTML if it needs
[10:10] to. And so like all of these different
[10:12] models are sort of blending together at
[10:14] this point. Michael had a very funny
[10:16] interaction where he got banned for
[10:19] OpenAI for cyber abuse. He had no idea
[10:21] what he did. uh he pasted the ban notice
[10:24] into Codeex, which I don't know how he
[10:26] was still using Codex if he was banned,
[10:29] but he said he asked it to figure out
[10:31] what triggered the ban. Codex found that
[10:33] he that it had asked for an API key to
[10:36] his own server. Codex wrote the appeal,
[10:39] submitted the appeal. A few minutes
[10:40] later, it the appeal was auto approved
[10:42] by an AI at OpenAI. So, he was banned by
[10:45] AI, convicted by AI, defended by AI, and
[10:48] pardoned by AI in about 10 minutes.
[10:50] There's a bunch of push back on the new
[10:52] chat GPT desktop app. Um because it is
[10:56] the super app. It unifies codecs. The
[10:58] the new chatbt app is just the codeex
[11:01] app. And when you install it, you can
[11:03] keep the old chatpt app as chatbt
[11:07] classic. But a lot of people don't like
[11:09] that because the codeex app is written
[11:11] in electron which means it's a bunch of
[11:13] like web views, web browsers effectively
[11:15] as opposed to uh the chat app which was
[11:19] a native app and had a bunch of
[11:20] integrations that people had sort of
[11:22] come to know and love. And so uh this is
[11:25] like a uh a change in the user
[11:27] experience if you're deeply embedded in
[11:30] a particular workflow. The flip side is
[11:32] like I just don't know how many people
[11:33] were daily driving chatpt as a Mac
[11:37] desktop app. Uh we've heard from someone
[11:40] on the team that was effective for
[11:41] cheating and we won't talk who it was
[11:44] but
[11:44] >> well so I would sometimes use it because
[11:46] it there's like a very nice shortcut. So
[11:48] it was like uh control plus or command
[11:50] plus is is the normal like uh main
[11:52] search on Mac but it was like a very
[11:54] easy way to to interact with models.
[11:55] >> Yeah. I wonder if that can just be can
[11:58] that just be um like can that shortcut
[12:01] just be created with the new codeex or
[12:04] chatbt work?
[12:06] >> Yeah, probably where you just bind that
[12:08] same key shortcut. But I mean it is
[12:10] making the user do work which is always
[12:13] always a tricky thing. But I feel like
[12:15] for most people's workflow at least for
[12:17] mine it was the desktop app that I used
[12:19] was codeex and then when I wanted to
[12:21] interact with chatbt in a chat flow I
[12:23] would do a new tab and then on my phone
[12:26] I would have the chatbt app and then the
[12:28] chatbt app on my phone could talk to
[12:30] codeex on the desktop but I never
[12:32] actually needed to open the chatbt
[12:34] desktop app. But maybe I'm just weird
[12:37] but I don't know. We we we'll see as
[12:39] this
[12:40] >> same same flow for me. Tibo says, "We
[12:42] spent the last 24 hours reading
[12:43] feedback, looking at usage patterns and
[12:45] talking with many of you." The short
[12:47] version is that there's a lot of
[12:48] excitement
[12:49] >> uh for 56 uh chatb,
[12:52] but we also didn't get get everything
[12:54] quite right.
[12:55] >> Uh we made it too easy to use the
[12:57] highest compute settings without making
[12:58] the impact on usage limits sufficiently
[13:00] clear. They reorganized the desktop app
[13:03] in one bold move, making familiar things
[13:05] like chats and projects harder to find.
[13:07] Our launch framing was focused on work
[13:09] and to some of our Codex fans it made it
[13:11] feel like Codex was going away over
[13:12] time. Absolutely not our intention. We
[13:15] love Codex and it is here to stay. Long
[13:17] story short, reset uh usage limits are
[13:19] getting reset twice today so people can
[13:22] keep uh experimenting.
[13:23] >> Hammer it uh plan out some sites. Uh
[13:27] there's also some confusion over the
[13:29] difference between Chach work and CHP
[13:31] codeex. Um, and I wonder
[13:35] >> I just want one or the other
[13:37] >> place to do things.
[13:38] >> Yeah. Yeah. The omni box that can do
[13:40] anything for you. Yeah. Uh, and I think
[13:42] that's obviously where it goes, but
[13:43] there's uh there's a walk, crawl, run,
[13:45] there's integration points and there's
[13:47] uh and people have ingrained um
[13:50] ingrained experiences and workflows and
[13:53] habits and there's all sorts of
[13:54] different tradeoffs here. There is no
[13:56] enterprise Google search like you just
[13:59] use Google at work, you use Google at
[14:01] home. It's all one box that gives you
[14:03] answers and there was never like a a
[14:05] dividing line between these two things.
[14:07] And so, uh, that's probably where all of
[14:10] this goes ultimately. Let's tell you
[14:11] about Codeex. There we go.
[14:14] >> Codex is a powerful workspace for
[14:15] getting things done [laughter] with AI
[14:17] agents. Whether you're writing code,
[14:18] analyzing data, creating content, or
[14:20] automating business workflows, Codex
[14:22] projects forward from start to finish.
[14:24] >> People are saying it's a horse for the
[14:26] mind.
[14:27] >> A horse for the mind.
[14:29] So the metabol case uh semi analysis
[14:32] wrote a long piece about uh the future
[14:34] of MSL what's going on there very
[14:37] interesting contrarian take I think a
[14:39] lot of people were confused by the
[14:41] expansion of of activities and sort of
[14:44] an entirely new business line with B2B
[14:47] sales and uh and and we've discussed the
[14:50] bull cases here the benefits obviously
[14:52] there's a ton of benefits the ad model
[14:54] they have the cash flow to draw down on
[14:56] there's a lot of different things that
[14:56] we can go through But let's read some of
[14:59] the reactions first and then we can go
[15:01] into the semi analysis piece. So boring
[15:02] business says people don't realize that
[15:04] Zuckerberg is all in on AI because he
[15:06] refuses to pay a 30% tax to platform
[15:10] owners for the rest of his life. Meta
[15:12] caught the shift toward mobile back in
[15:15] the early 2010s. But they were never
[15:17] able to create their own hardware
[15:18] platform to capture users in their
[15:20] ecosystem. As a result, they got pay
[15:22] they got stuck paying a 30% tax to Apple
[15:25] on inapp purchases, which has cost Meta
[15:28] billions of dollars in profit.
[15:30] >> That makes no sense to me.
[15:33] >> No one buys ads via inapp purchases.
[15:36] Every advertiser I know uses the
[15:39] webbased.
[15:40] >> Also, also ads, I believe, count as like
[15:43] a real world service. So there's not
[15:46] like there's there's a lot of things you
[15:48] order in the that you order on mobile
[15:50] apps like groceries
[15:51] >> kind of a digital good but yes I I I
[15:53] agree with you that it would be exactly
[15:54] >> yeah but but it's like eyeballs right
[15:56] it's like impressions uh which which
[15:59] like eyeballs human attention is not a
[16:03] digital good it is like a real world
[16:05] thing
[16:05] >> yes so if I open up the Amazon app on my
[16:09] iPhone and I buy a t-shirt Apple does
[16:12] not get 30% of that because that's real
[16:14] >> or Door Dash good example
[16:15] >> or a movie ticket.
[16:17] >> Yes. But if I go to Kindle or Netflix or
[16:23] and I subscribe or even like the X
[16:25] subscription premium, Apple will take
[16:28] 30% if I buy Fortnite V-Bucks or Roblox
[16:32] or any digital good. Yeah. And I could
[16:35] see Apple saying, "Look, if you want to
[16:38] be able to because in in the uh in the
[16:40] meta platforms, you can be an Instagram
[16:43] and you can just say I want to boost
[16:44] this post and you can put down your
[16:46] credit card and pay for that and you can
[16:49] get started on the advertising flow as a
[16:51] very small business I think through
[16:53] effectively inapp purchases." Yeah. So
[16:55] the big the bigger the bigger thing is
[16:57] just like
[16:57] >> big billions is right.
[16:58] >> You have this one of the largest
[17:00] companies in the world.
[17:01] >> Yeah.
[17:03] constantly playing
[17:13] what do Apple executives think about
[17:15] Meta if you were to ask them over coffee
[17:18] off the record and it would be like we
[17:20] don't like Meta and and so that's
[17:22] fundamentally like a very very
[17:23] uncomfortable position to be in
[17:26] >> and so he doesn't want to be in a
[17:28] position running his company where he
[17:30] has to ask for permission for things.
[17:33] And so an example with that, somebody
[17:35] brought it up, I think Andrew, Meta lost
[17:36] billions from Apple because of the
[17:38] permissions prompt.
[17:39] >> Yes. You know, making it harder to just
[17:41] run their core advertising business.
[17:44] Yeah.
[17:44] >> And also there's so much that's like
[17:46] very unpredictable about where AI is
[17:48] going. Like it's possible that we get to
[17:50] the point where you don't go to a bunch
[17:53] of different social media platforms and
[17:57] you're just like talk your agent is just
[17:59] kind of surfacing content for you on the
[18:02] fly. Like there's so much like potential
[18:04] disruption. I think the stance is like I
[18:07] would rather participate in this massive
[18:09] technology cycle and have a shot at
[18:12] being a big part of it. Yeah.
[18:13] >> Than like just like hope nothing bad
[18:15] happens to me. The other thing I've
[18:17] always felt is like, you know, with the
[18:18] with the glasses, you're just going to
[18:20] be able to like he sees a world like the
[18:23] glasses are are more of a a new platform
[18:26] bet right now than AI.
[18:29] >> And I just think he wants to be in a
[18:32] world where you're wearing your very
[18:33] very cheap pair or maybe even free pair
[18:35] of glasses around. They act as your
[18:37] phone and your computer and you can look
[18:40] at stuff and say, "Buy me that." And
[18:42] he's getting a cut of it. you're walking
[18:44] by a restaurant and it pops up an offer
[18:47] and it says like free boba for any
[18:50] orders over $20 and he's getting a cut
[18:53] of that, right? So like that will be
[18:54] like some type of like always on
[18:56] intelligence. So yeah,
[18:58] >> it all it all totally makes sense even
[19:00] if it is like deeply sort of like
[19:02] chaotic and at times confusing.
[19:04] >> Yeah.
[19:05] >> Uh at least over the last year. The
[19:07] funny thing about the glasses as a
[19:09] platform is AI spending is so huge that
[19:12] he can probably bring back a lot of the
[19:15] reality lab spending and sort of just
[19:18] tuck it in a bucket and people won't
[19:20] notice if there's an extra
[19:22] >> I'm not leaving. I'm not leaving.
[19:27] >> I'm not leaving.
[19:31] >> I'm not.
[19:31] >> Yeah. I mean, if there's an extra 8
[19:33] billion spent on the next version of the
[19:36] Meta Rayban displays or something,
[19:38] people be like, "Ah, it's a rounding
[19:40] error compared to the 200 billion on
[19:42] data centers or whatever."
[19:43] >> Ari says the tax Meta pays to Apple is
[19:45] the fact that they can't track users
[19:46] crossplatform.
[19:48] >> Yeah. Yeah. Totally. The other
[19:49] interesting thing is like I do think the
[19:51] uh the screen time is a is at least like
[19:54] a flashing warning light. Like I've
[19:56] personally noticed as an early adopter
[19:58] of LLMs that my screen time has switched
[20:02] much more towards chat GPT and LLMs and
[20:06] deep research reports just because if I
[20:08] see something interesting, oh there's a
[20:10] there's a car video on Instagram, I will
[20:14] often switch over and then start
[20:16] prompting to actually learn more.
[20:18] Whereas before that would have required
[20:20] a whole bunch of Google searches. You
[20:22] might have to sit down and like collect
[20:23] a document like you're not actually
[20:25] going to do that. Now, I've been
[20:26] noticing that I'm much more in like the
[20:28] hour a day on chat versus hour a day on
[20:32] Instagram. And so, I don't know if I'm
[20:34] just like the early adopter that will
[20:35] never do that, but you can certainly see
[20:37] that people are spending a lot of time
[20:38] with AI systems. And as you've put it
[20:41] before, like wherever consumers,
[20:43] wherever people are spending time is an
[20:45] important place to have a a foothold at
[20:47] least.
[20:47] >> Semi- analysis said at the simplest
[20:49] level, there are three things you need
[20:50] to build to need to build a true
[20:52] frontier model. Data, talent, and
[20:54] compute. We believe Meta is the only
[20:56] hyperscaler
[20:58] Neolab on track to be worldclass at all
[21:00] three and therefore has the best chance
[21:02] at catching up with anthropic openai.
[21:05] We'll explain why in full detail below.
[21:08] And so Alex Wang uh quoted this uh in a
[21:12] funny uh in a funny way. He said compute
[21:16] Daddy Dylan Patel has spoken
[21:18] >> um and yes the stock is up 6%. Wang uh
[21:23] quote quoted uh poly quote tweeted
[21:26] polyarket money which said meta jumps
[21:28] over 6% of the open as investors react
[21:31] to yesterday's muse spark 1.1 release
[21:34] again I don't really think that they
[21:36] were reacting to muse spark as much as
[21:39] semi analysis saying
[21:41] >> there's a there's a there's a chance um
[21:43] so yeah Alex Wang said nice with the
[21:45] thumbs up so he's he's [clears throat]
[21:46] pumping the stock
[21:47] >> what's interesting is like
[21:48] >> the crazy thing is that like semi
[21:51] analysis just like completely leaving
[21:53] out SpaceX and
[21:56] >> they didn't leave out SpaceX and Google.
[21:58] They said
[21:59] >> but but but saying
[22:01] >> they said SpaceX is selling $26 billion
[22:03] a year worth of GPUs to Anthropic Google
[22:06] and that's why they're not in the race.
[22:10] Uh even though it is odd because Meta
[22:12] might wind up selling a lot of GPUs.
[22:15] It's sort of unclear what the breakdown
[22:16] will be, but they're at least laying out
[22:19] a narrative that if Meta picks it up
[22:23] would differentiate them from SpaceX AI
[22:26] in some ways, I think. Yeah, I I I agree
[22:29] with you. It is.
[22:29] >> But they're essentially saying like
[22:32] Gemini doesn't have because Gemini has
[22:34] compute.
[22:35] >> Yep.
[22:35] >> And data.
[22:36] >> Yeah.
[22:36] >> They're saying they're effectively
[22:38] saying Gemini does not have talent.
[22:41] >> Yeah.
[22:41] >> Right. They're saying you need all
[22:43] three. Meta took their data efforts to
[22:45] another level in late May by announcing
[22:47] a new applied AI engineering org as part
[22:49] of their most recent round of layoffs
[22:51] restructuring. 3,000 engineers, which
[22:53] includes 70% of their new grads and a
[22:55] significant number of seniors, will now
[22:57] be making RL tasks and environments
[22:59] full-time. We think this is an extremely
[23:02] under underappreciated advantage for
[23:03] MSL. The question is if this becomes the
[23:06] the the standard playbook, you would
[23:08] imagine that Google DeepMind would be
[23:10] able to do the same thing. they have the
[23:11] same structural uh workforce, but maybe
[23:14] it's like it's a band-aid that Google's
[23:16] not willing to rip off, but Zuck is
[23:18] willing to rip rip it off at Meta.
[23:21] That's maybe the the thesis at a deeper
[23:23] level. Uh Anthropic has been the most
[23:24] aggressive lab by far when it comes to
[23:26] buying coding data from RL environment
[23:28] startups, and it's one important reason
[23:29] why their models are so good at coding
[23:31] today. Meror recently disclosed that
[23:33] they logged 2.5 million expert hours on
[23:36] their platform in Q2 of 2026. That's
[23:39] equivalent to 5,000 people working 40
[23:41] hours a week. Meta is already in the
[23:43] same ballpark and their average quality
[23:45] is likely higher. Additionally, they
[23:47] have another 70,000 people to pull into
[23:50] that organization if they need to if it
[23:52] winds up being as valuable as semi
[23:54] analysis thinks it will be. Uh it's also
[23:56] worth briefly dispelling this myth that
[23:58] 3,000 meta engineers will be doing
[24:00] mindless low-level data labeling. The
[24:01] days of undereducated contractors from
[24:03] third world countries drawing boundary
[24:05] boxes or classifying text as NSFW are
[24:08] gone. at this point the models are
[24:09] sufficiently smart that creating a good
[24:11] piece of training data is a real
[24:13] intellectual challenge. So uh they're
[24:15] sort of taking the contrarian stance
[24:16] that uh you know being an being an RL
[24:19] environment designer is not soul
[24:21] crushing and in fact can be
[24:23] intellectually rigorous. They also say
[24:25] that this the top expert contractors at
[24:28] every data company, they're all making
[24:30] seven figures a year, which is pretty
[24:32] crazy since people thought of data
[24:34] labeling as a $10 an hour task. And now
[24:37] you have people making millions of
[24:39] dollars a year because they're actually
[24:41] really great at defining what it means
[24:42] to complete a task from start to finish.
[24:44] We got to talk about
[24:45] >> talk about FIA Gate.
[24:47] >> FIA Gate, the story that is at the
[24:49] center of Phoebe Gates's life. Her
[24:52] company FIA is uh in Bloomberg today
[24:56] with some accusations of misattribution
[25:00] in the e-commerce world. FIA, the buzzy
[25:03] shopping app co-founded by Bill Gates's
[25:05] daughter, Phoebe, is claiming credit for
[25:07] online sales. It didn't actually drive.
[25:09] An advertising startup co-founded by
[25:11] Phoebe Gates, daughter of billionaire
[25:12] Microsoft Corp. co-founder Bill Gates
[25:15] builds itself as a personal shopping
[25:18] assistant that helps users find the
[25:20] lowest prices on a broad range of
[25:22] clothes and fashion accessories.
[25:24] Businesses like these known in industry
[25:27] parlance as affiliate marketing programs
[25:29] typically collect commission from
[25:31] retailers that ma that make the
[25:32] resulting sale. Bloomberg tested the FIA
[25:35] mobile browser extension across more
[25:37] than 50 websites and found that during
[25:39] the checkout process, FIA opened a
[25:41] background tab without user interaction
[25:44] and injected its own referral code that
[25:46] overrode legitimate referrals from other
[25:48] publishers. These findings were
[25:50] consistent with Capital One Shopping and
[25:52] Edelman's independent testing and code
[25:53] review.
[25:54] >> Let's get into why what is being accused
[25:56] of and why it is
[25:58] >> potentially if this is what they're
[26:00] doing. We don't know again if it's an
[26:02] accident or it's intentional, but why
[26:04] it's uh potentially even more predatory
[26:08] than uh than honey.
[26:11] >> So, Mega Lag, the YouTuber that did the
[26:13] honey expose, it was an interesting deep
[26:16] dive that got 20 million views because
[26:19] it was using a hidden redirect near
[26:21] checkout to replace the creators
[26:23] affiliate tag with Honeys. So the irony
[26:26] here was that a lot of those creators,
[26:28] they would make money from I'm promoting
[26:30] AG1, use code Huberman. And then if they
[26:33] had also promoted Honey, Honey was
[26:36] stealing their AG1 affiliate revenue.
[26:39] And so the creators were sort of
[26:41] promoting their own demise in some way.
[26:43] So it was like this this very viral,
[26:46] very ironic situation eventually got
[26:48] sorted out. HUD's broader scandal also
[26:50] involved allegations that it suppressed
[26:52] better coupons and deliver and
[26:53] deliberately evaded affiliate network
[26:55] standown checks. Nothing uh comparably
[26:58] uh comparably extensive has yet been
[27:00] established about FIA, but FIA's
[27:03] reported implementation may actually be
[27:04] crudder. Honey generally created the
[27:07] overwrite when the shopper interacted
[27:08] with its checkout prompt, even if only
[27:11] to dismiss it. So, you would see a honey
[27:14] popup. At the very least, you had to
[27:17] click I don't want honey and then it
[27:19] would do something. The Bloomberg report
[27:21] says FIA's mobile extension is firing
[27:23] without any FIA interaction at all. So
[27:27] you go from horseereviews.com over to
[27:29] Clydesdale's direct. You merely have FIA
[27:32] installed as an extension. You don't
[27:35] >> you don't you don't know that FIA has
[27:36] been a part of the interaction at all.
[27:38] >> You just click through. But you have the
[27:40] software running in your browser that is
[27:43] inserting a a effectively attribution
[27:47] for FIA without you knowing, stealing
[27:50] >> potential revenue from horsereviews.com.
[27:52] Yes. And horsereviews.com probably had
[27:55] to run ads on Google to get people to
[27:58] horsereviews.com to educate them on the
[28:00] the horse market to get them to
[28:03] Clydesdale's direct to get them to
[28:05] close. And so they're sitting here being
[28:07] like, "Hey, we did all this work to
[28:09] uh get uh Clydesdaleirect.com
[28:12] a new customer and we're not being
[28:14] compensated for
[28:15] >> Direct might say well normally I'm
[28:17] sending a check to horsereviews.com
[28:20] I
[28:21] >> I guess I'm sending a a check to FIA now
[28:25] whatever as long as I'm getting
[28:26] customers I'm fine
[28:27] >> so so one of the concerning things here
[28:29] >> is that I think this had been running
[28:31] since like January or December and so
[28:33] they had some software update that was
[28:35] pushed this started happening so it's
[28:37] been happening for let's say around 6
[28:39] months
[28:40] >> and you would assume that they would
[28:43] have figured this out but maybe it's
[28:46] possible that that it you know
[28:48] >> I don't know a lot of I mean it it it's
[28:50] not the biggest scale I think the number
[28:52] is 12 million plus in sales driven so
[28:56] the commission on that is like $1
[28:58] million that's sort of like missing and
[29:01] it's and it's so diffuse across so many
[29:05] different e-commerce providers I
[29:07] guarantee you as someone who's run
[29:09] multiple e-commerce sites, uh I would
[29:11] not notice if a couple hundred dollars
[29:14] moved from one affiliate to another.
[29:16] Now, everyone on the timeline was
[29:18] praying on Phoebe's downfall
[29:21] >> way before this.
[29:21] >> Way before this, people people wanted
[29:24] this so badly. They didn't want to hear
[29:26] this be a success. And so, they've been
[29:28] chomping at the bit for something. She
[29:30] raised money from a lot of very popular
[29:32] celebrities and it looked like a
[29:34] Coachella poster. And there's all these
[29:36] like, oh, like why are we even raising
[29:37] venture capital? So, our theory is that
[29:40] this is where the comeback begins.
[29:42] >> The villain arc.
[29:43] >> The villain arc potentially. But
[29:44] seriously, okay, let's go over to
[29:46] movies. Hollywood's horror meme gold
[29:49] rush is minting millionaires. We talked
[29:51] about this with uh Back Rooms, the uh
[29:54] originally 4chan uh creepy pasta that
[29:57] wound up being a viral YouTube channel,
[30:00] and then a full Hollywood movie that
[30:03] made over a hund00 million. I believe,
[30:05] but there are many more pieces of
[30:07] intellectual property in the horror meme
[30:10] genre and they're all getting scooped
[30:11] up.
[30:12] >> That was Horoszi, by the way.
[30:13] >> It did like herozi. [laughter]
[30:16] It did look like Hermosi. Uh, so the
[30:18] creator of a monster called Siren Head
[30:20] scores a rich deal as studios hunt for
[30:22] the next backroom style hit in the Wall
[30:25] Street Journal. Trevor Henderson drew a
[30:27] faceless monster called Siren Head eight
[30:30] years ago. That's pretty creepy. I'm not
[30:32] a fan. eight years ago and watched it
[30:34] spawn YouTube films, knockoff Amazon.com
[30:37] merchandise, and video games, but he
[30:39] never made a penny from his creation.
[30:41] Interesting. He just spawned this
[30:43] horrific meme and then uh hadn't
[30:45] profited on it until now. So, he's a
[30:47] 40-year-old illustrator and he told he
[30:49] sold the Siren Head movie rights to
[30:52] Warner Brothers for more than a million
[30:54] dollars. Pretty cool that he just
[30:56] illustrated this weird thing and then
[30:57] made so much money. According to people
[30:59] familiar with the matter, the deal was
[31:00] part of a new Hollywood gold rush to
[31:02] find concepts and talent online that
[31:04] could fuel the next horror hit like back
[31:07] rooms and obsession. Uh studios are
[31:09] trying to find the next potential horror
[31:11] hit that already has proven its appeal
[31:13] to Gen Z on YouTube, Reddit, or this is
[31:15] interesting, Roadblocks. The chair
[31:17] co-chair of Warner's motion picture
[31:19] group, Michael Duca, said, "We're seeing
[31:21] them as a resource for adaptations the
[31:23] same way we look at books and other
[31:25] media. The race for hot digital
[31:27] properties resembles the kind of dog
[31:29] fights for buzzy scripts that used to be
[31:30] common in Hollywood during the 80s and
[31:32] 90s. 11 studios recently bid on the film
[31:34] rights to the psychological horror
[31:36] YouTube series, The Mandela Catalog.
[31:38] Amazonowned United Artists and Steven
[31:40] Spielberg's Ambulent Entertainment
[31:42] ultimately agreed to pay millions of
[31:43] dollars and let the videos 22-year-old
[31:46] creator Alex Kisser direct the movie.
[31:49] That's very interesting.
[31:51] >> Daniel says, "What's the obsession with
[31:53] horror? There's something wrong with you
[31:55] if you're into that.
[31:57] >> Totally agree. I I never gravitated to
[32:00] horror movies, but my theory was always
[32:02] if I don't watch them, I will ne I won't
[32:04] be afraid walking at night. Like I grew
[32:06] up in the I grew up in the country,
[32:09] >> you know, teenager in the country, you
[32:10] know, you're going around on adventures,
[32:13] stuff like that.
[32:14] >> Watch this. You're going to be
[32:15] terrified. Leave us five stars on Apple
[32:17] Podcasts and Spotify. Sign up for our
[32:19] newsletter at tbp.com and have a great
[32:21] weekend. We'll see you Monday.
[32:24] We
[32:25] >> throwing flashbang.
