## Background
This episode was published by TBPN on YouTube on June 27, 2026. Diet TBPN provides a 30-minute digest of today's live tech talk show hosted by John Coogan and Jordi Hays, which covers technology trends, Silicon Valley insights, and major industry news.
## TL;DR
Hosts John Coogan and Jordi Hays discuss current developments across the creator economy, hardware trends, and major AI industry news. They analyze shifting creator business models following their trip to Cannes Lions, examine Meta's smart glasses market dominance and new $399 Kylie Jenner collaboration, and unpack Meta's aggressive internal push into AI coding data labeling and prediction markets. Additionally, the episode covers OpenAI's restricted release of GPT-5.6 Soul due to government national security directives, its custom Jalapeño silicon co-designed with Broadcom, and the $500 million 'Intercept' health initiative aimed at eliminating respiratory viruses.
## Atmosphere & style
The show carries a fast-paced, banter-filled, and witty tone driven by the natural chemistry between hosts John Coogan and Jordi Hays. Combining playful personal anecdotes—such as joking about nicotine withdrawal symptoms, French cigarette packaging, and the health realities of parenting—with sharp tech commentary, the conversation feels like a high-energy yet candid insider discussion. The overall mood balances sharp industry critique and tech skepticism with lighthearted humor and conversational warmth.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=0s) Historically focused on traditional advertising, the Cannes Lions festival has rapidly transformed into an event centered on the creator economy.
- [00:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=0s) Creators face monetization hurdles with high-production vertical videos, as they lack standard mid-roll breaks and force reliance on sponsored posts.
- [00:07](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=7s) Meta dominates the emerging smart glasses sector with over 80% market share, generating billions in gross revenue through partnerships like EssilorLuxottica and Ray-Ban.
- [16:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=960s) Meta has forcefully reassigned 30% to 50% of core engineering teams away from product development toward manual data labeling to feed its code generation AI models.
- [17:12](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1032s) Public markets are showing skepticism over Meta's massive capital expenditure to pivot into enterprise AI solutions against entrenched rivals.
- [18:46](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1126s) Meta is internally developing a standalone prediction markets app codenamed "Arena" to directly compete with Polymarket and Kalshi.
- [20:11](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1211s) Meta and Snap are investing heavily in high-profile celebrity partnerships, including an estimated $50M Meta campaign with Kylie Jenner and a reported $100M Snap deal with Robert Downey Jr.
- [21:47](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1307s) US government intervention restricted OpenAI's preview of GPT-5.6 "Soul" to only 20 cleared companies due to global API distillation and security concerns.
- [24:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1441s) A cyber vs. biological safety asymmetry exists where AI cyber vulnerabilities can be rapidly patched in software, whereas bioweapon threats face a slow physical manufacturing lag for countermeasures.
- [29:08](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1748s) OpenAI launched its custom "Jalapeño" ASIC co-designed with Broadcom and secured 40% of the global raw DRAM wafer output through 2029.
- [30:09](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1809s) OpenAI, Anthropic, Stripe, and Bill Gates pooled $500 million to launch "Intercept," an initiative aimed at eliminating respiratory viruses.
Fresh off a trip to France, hosts John Coogan and Jordi Hays return to the studio, lightheartedly describing the "brutal" physical withdrawal symptoms—sweating and shaking—they feel by mid-day whenever they miss going live. Beyond the banter, the trip carried high stakes: Jordi recounted past clashes with French President Emmanuel Macron, while John felt personally targeted by France's recent total ban on oral nicotine products and nicotine pouches.
Reflecting on their time at the Cannes Lions festival, the hosts observed a structural shift in the event's core focus. Historically an arena for awarding traditional creative advertising, Cannes Lions has rapidly transformed into a hub centered on the creator economy. While not entirely overrun by A-list Hollywood celebrities, the festival felt like a gathering where nearly every attendee commanded a dedicated following.
The central debate at the festival—brought to the forefront by CNBC *Squawk Box* and *DealBook* anchor Andrew Ross Sorkin—focused on the trade-off between independence and consolidation. Media experts are questioning where the true economy of scale lies within the creator economy, debating whether a roll-up of independent Substack writers or creator collectives makes more sense than total independence or corporate attachment.
- **High-Production Pivot:** Independent creators are moving away from quick solo video clips toward heavily produced shows, such as *Subway Takes* or *Keep the Meter Running*.
- **Monetization Bottlenecks:** Highly produced vertical videos lack logical mid-roll ad breaks, forcing creators into awkward 1-to-1 ratios of organic content paired with dedicated 30- to 60-second sponsored posts that often perform worse algorithmically.
- **Forbes Creator Top Earners:** MrBeast topped the Forbes top creators list with $300 million in gross revenue, though his obsession with reinvesting capital into production means his net margins differ drastically from low-overhead setups like *The Joe Rogan Experience*.
- **Legacy Media Strategy:** Traditional outlets like *The New York Times* are successfully adapting to YouTube using high production values, crisp packaging, and optimized titles/thumbnails for flagship shows hosted by Ezra Klein and Ross Douthat.
[00:49](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=49s) Brian Johnson on French Cigarettes and Creator Business Models
Transitioning between topics, the hosts shared a bizarre visual discovery from their trip: French cigarette packaging featuring a image of a man remarkably resembling longevity entrepreneur Bryan Johnson. Laying eyes-closed inside what looks like either a sleeping pod or a body bag, the image sparked jokes that French health authorities inadvertently endorsed his "Don't Die" ethos.
Despite the shocking graphic warnings typical on European tobacco products, the hosts joked that getting tucked in for a night of rest after ripping a couple "darts" and crushing cold ones actually looked quite peaceful compared to standard black-lung diagnostic photos.
Closing out the discussion on creator business models, the hosts concluded that the industry is undergoing a sorting process rather than a one-size-fits-all shift. Some top creators will continue to command lucrative 80% to 90% EBITDA margins on niche, hit-driven content, while others will use the threat of going independent to negotiate larger contracts with traditional media firms.
- **Sorting Mechanism:** The blanket narrative that "independent creators are the future" is giving way to a nuanced ecosystem where scale, overhead costs, and business goals dictate structure.
- **Margin Variances:** Low-overhead commentary shows maintain massive profit margins, whereas high-concept, hit-driven shows operate on tight cash margins due to continuous production reinvestment.
[00:07](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=7s) Meta's Smart Glasses Surge and the $300 Kylie Jenner Collaboration
Turning to hardware, the hosts examined a *Wall Street Journal* article by Christopher Mims covering the growing market for smart glasses. While Mims characterized the market as experiencing a "deluge" of products, the hosts viewed it as a steady, material build-up led predominantly by Meta Platforms, which holds over 80% market share.
Addressing the hardware metrics, the hosts clarified that Meta and partners like EssilorLuxottica have moved millions of units since 2025, generating between $2.1 billion and $3 billion in estimated gross retail revenue. They highlighted Meta's hardware lineup—ranging from standard $299 Ray-Ban Meta glasses (which qualify for vision benefit coverage under insurance) to a newly launched $399 Kylie Jenner collaboration featuring a subtle jewel-like aesthetic.
```
Smart Glasses Market Snapshot (Post-2025 Data)
├── Total Industry Units Sold: 7M+ (Across Meta, Oakley, & partners)
├── Ray-Ban Meta Standalone Sales: 2M+ units
├── Estimated Gross Retail Revenue: $2.1B – $3.0B
└── Meta Market Share: >80%
```
To illustrate the hardware vs. software trade-off, Jordi noted that while Meta's hardware scale is impressive for a new device category, software initiatives like Threads offer far higher profit margins. The hosts bantered about potential business models, including hypothetical ad-supported free glasses that display localized popup offers as wearers walk past stores—or external-facing directional speakers that "bark ads at everyone around you while noise-canceling them for the wearer."
- **Unit Comparisons:** Meta's 7 million smart glasses sold compare to a global smartwatch market of over 100 million units annually and over a billion smartphones sold per year.
- **Competitor Pricing:** Snap's Spectacles target a higher price point at $2,200, featuring a chunkier design likened to high-fashion Prada runway accessories.
- **Privacy Skepticism:** The article notes public resistance toward face-worn devices, driven by screen-time fatigue, camera-privacy concerns in public, and dubious utility around real-time visual AI identification.
Here is the detailed walkthrough of this portion of the transcript:
[15:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=900s) Smart Glasses and the AI Consumer Arms Race
The discussion opens with skepticism regarding the practical utility of computer vision in daily life, referencing the classic "Hotdog / Not Hotdog" trope. The hosts argue that constantly identifying everyday objects in the real world—whether asking what building or tree you are looking at, or recognizing a rare South African meerkats or a can of Diet Coke—isn't a massive everyday pain point. What users actually need is an AI agent on their phone capable of looking up information and executing tasks on their behalf. Despite this, smart glasses are viewed as a technological inevitability—a fresh consumer hardware arms race vying for the final, prime real estate on human bodies. This trajectory is echoed by Zed, the head of Qualcomm's division that manufactures the chips powering Meta, Snap, and Samsung smart glasses, who envisions a future where all glasses will eventually be smart glasses.
[16:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=960s) Meta’s Aggressive Pivot to Coding and Forced Data Labeling
Shifted attention turns to Meta’s massive internal AI push. Meta appears heavily focused on code generation, leading what is likely the world's largest effort to generate coding training data. To feed this pipeline, Meta has forcefully reassigned 30% to 50% of core engineering teams away from building products used by hundreds of millions of people to manual data labeling—specifically providing human feedback on AI-generated GitHub repositories. One internal engineer likened the sudden reassignment to *The Hunger Games*, where tributes are randomly selected from their normal environments, except at Meta, up to three to five people on a ten-person team are affected. The hosts joke that while it mirrors the movie's disruption, it at least comes with air conditioning and free snacks. To support these capital-intensive AI ambitions and build out MSL, Meta is openly seeking innovative financing structures and partnerships to potentially raise billions more.
[17:12](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1032s) Meta's Enterprise vs. Consumer AI Strategy and Market Skepticism
Analyzing Meta's long-term strategy, the hosts express surprise at Meta building an API model to sell enterprise coding solutions rather than doubling down exclusively on consumer applications. Given Meta’s massive base of billions of social users and lack of background in B2B enterprise software—aside from ambiguous acquisitions like Manus—it was expected that Mark Zuckerberg would spend heavily to push the Meta AI consumer chat app to the top of the charts and keep it there to gather targeting data. Instead, the Meta AI app sits at #17 in the app store. Public capital markets are showing skepticism over Meta spending hundreds of billions of dollars to pivot into enterprise competition against entrenched giants like AWS, Google, Microsoft, OpenAI, Anthropic, and Chinese AI labs.
[18:46](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1126s) Prediction Markets App "Arena" and Meta's Brand Sentiment
Zuckerberg has directed Meta to build an experimental, standalone prediction markets app internally codenamed **Arena**, operating independently of Facebook and Instagram to compete directly with platforms like Polymarket and Kalshi. The hosts debate public sentiment surrounding prediction markets, noting that while general sentiment hit a high point during political elections, it has since cooled down into a sports gambling vibe. However, exploding revenues prove users still love the product and value it as a real-time data source. Given that capital market sentiment around Meta's spending is near an all-time low, the hosts quip that Zuckerberg might simply feel sentiment can't get any worse, so he might as well experiment.
- **Arena**: Meta's experimental standalone prediction market app designed to compete with Polymarket and Kalshi.
[20:11](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1211s) Kylie Jenner Meta Campaign vs. Snap’s $100M Deal with RDJ
In contrast to Wall Street's skepticism, initial public reception to the Kylie Jenner Meta Smart Glasses partnership has been wildly positive. Silicon Valley appears to have finally recognized who controls consumer spending by pairing with one of the world's most influential figures on Instagram. The hosts estimate Meta likely spent around $50 million to secure the top-tier influencer for the product launch. Meanwhile, journalist Alex Heath ("Heathnator") scooped that Snap is working on a staggering $100 million deal with Robert Downey Jr. for a partnership. The hosts find Snap's move confusing compared to Meta’s natural alignment, jokingly speculating that RDJ's deal might involve an *Iron Man 4* tie-in where Iron Man defeats Thanos for the fifth time using only Snap Spectacles.
- **Meta x Kylie Jenner**: Major campaign launch estimated around $50M to drive smart glasses adoption.
- **Snap x RDJ**: Reported $100M deal in development for Snap Spectacles.
[21:47](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1307s) OpenAI Releases GPT-5.6 "Soul" Under Strict Government Limits
OpenAI announced a preview of its next-generation model line led by **GPT-5.6 Soul**, part of a family that includes Soul (the massive frontier model), Terra (a balanced model for everyday efficient work), and Luna (a fast, affordable option for high-volume tasks). However, by US government directive, access to Soul is strictly limited to just 20 pre-approved companies. This tight restriction comes in the wake of "distill-gate," where Anthropic revealed that Chinese lab Alibaba had heavily distilled Claude via API usage, creating a staggering volume of resold tokens that industry observers characterized as "professionalized fraud." Because of these systemic leaks, labs are locking down API access tightly, even if it prevents prominent tech commentators like Dan Shipper from running routine "vibe checks" on the new model.
- **GPT-5.6 Model Lineup**:
- **Soul**: Top-tier frontier model.
- **Terra**: Balanced model for daily enterprise workflows.
- **Luna**: Lightweight, fast model for high-volume execution.
- **Access restriction**: Limited to 20 government-cleared companies following widespread API distillation concerns.
[23:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1380s) Theo Von's Data Center Rant and the AI Energy Gap
Comedian Theo Von sparked online chatter with a viral rant against data centers, claiming "nobody wants a data center" and suggesting the operators seem evil for controlling all human information, even joking about an impending "emotional credit score." The hosts note that while data centers need far better public messaging regarding water and power usage, Theo's anti-data center stance is ironic given his reliance on podcast distribution infrastructure. However, they acknowledge a counter-argument: streaming podcasts or YouTube videos consumes a minuscule fraction of the power required by AI data centers, where running a single AI agent prompt can consume as much compute as ten years of streaming video.
[24:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1441s) White House Intervention and Safety Risks: Cyber vs. Bio Capabilities
OpenAI has limited access to its newest models following direct interventions by the Trump administration, though OpenAI warned that case-by-case White House restrictions based on national security shouldn’t become the long-term norm. Safety advocates view this high-touch political intervention as the "worst of both worlds" because it grants the executive branch sweeping arbitrary authority over tech releases rather than setting clear, broad-reaching rules. The intervention was triggered after Anthropic’s **Mythos** model—and OpenAI's **GPT-5.5 Cyber**—demonstrated unprecedented autonomous hacking capabilities during red-teaming exercises, with the NSA confirming Mythos breached several of its systems.
This escalation highlights a critical asymmetry between cyber and biological safety risks:
* **Cyber Security**: When a frontier model discovers a system vulnerability (e.g., "Hack Machine 9000"), that same model can be used for legal economic reasons to write and deploy a patch within an hour, closing the gap before malicious actors catch up.
* **Biological Security**: Open-source models typically lag frontier models by roughly six months (evidenced by open models like GLM-5.2 approaching current daily-driver performance). If a frontier model designs a novel bioweapon ("Bioweapon 9000"), creating physical countermeasures, early-warning detection, and manufacturing medicine takes over a year. Because physical manufacturing takes longer than six months, open-source malicious actors could gain access to the bioweapon designs before society finishes deploying the physical biosecurity shield.
> "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." — OpenAI official statement
[29:08](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1748s) OpenAI’s Custom "Jalapeño" Chip and Raw Wafer Monopoly
OpenAI made major supply chain moves by officially launching **Jalapeño**, an ASIC designed in partnership with Broadcom. Purpose-built specifically for LLM workloads, Jalapeño will power ChatGPT, Codex API, and upcoming agentic products. The chip was developed in record time, with reports indicating it was among the first processors designed with AI agents directly in the loop to write instruction sets faster. Beyond custom silicon, OpenAI executed massive supply chain deals to purchase **40% of the global raw undoped DRAM wafer output until 2029**, locking up millions of unprocessed wafers deep in the supply chain.
- **Jalapeño Chip**: Custom ASIC co-designed with Broadcom for LLMs and agentic workloads.
- **Wafer Lockup**: OpenAI secured 40% of global raw undoped DRAM wafer production through 2029.
[30:09](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1809s) GTA 6 Cost Comparisons and Intercept Health Initiative
The segment wraps with two notable updates across gaming and healthcare. Comparisons are being made between the release of *Grand Theft Auto 6* and the construction of the Burj Khalifa: Dubai's landmark skyscraper cost $1.5 billion and took six years to build, whereas GTA 6 has reportedly amassed a $2 billion budget over a 12-year development cycle—effectively making it "two Burj Khalifas" in scale.
On the scientific front, OpenAI, Anthropic, Stripe, and Bill Gates have jointly pooled **$500 million** to launch a new organization called **Intercept**. The ambitious initiative aims to develop preventative treatments to eliminate the common cold, flu, and eventually all respiratory viruses entirely.
[30:57](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1857s) The Parental Illness Curve and Podcaster Readiness
The hosts share a lighthearted discussion about the realities of parenting, specifically mapping out the relationship between the number of children a person has and the percentage of time the parents spend sick with a viral illness. They joke that the dynamic follows a "fast takeoff curve," where having just one child leaves a parent sick with a cold, flu, or virus roughly 25% of the year. Once a family grows to two, three, or four kids, that rate dramatically ramps up to around 50% of the entire year spent ill—a stat that parents will immediately recognize and validate if they notice their own health declining.
The banter then turns toward "podcast optimizers"—the type of hyper-curated hosts who struggle to even drink two alcoholic beverages while continuing to record. The hosts poke fun at how woefully unprepared these health-conscious podcasters are for the chaotic germ-fest of raising children. They note with a chuckle that, fortunately, most podcasters do not go down the path of starting families. They joke that if podcasters actually *did* start having kids and getting sick all the time, it could lead to a serious decline in the "strategic podcast reserve."
* Having 1 child results in being sick roughly 25% of the year.
* Having 2 to 4 children causes parental illness time to scale up to 50% of the year.
[31:56](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1916s) Sign-off and Show Call to Actions
Wrapping up the episode ahead of the weekend, the hosts deliver their standard closing call to action for the audience. They ask listeners to leave a five-star review on Apple Podcasts and Spotify, and remind them to subscribe to the daily show newsletter at TBPN.com before wishing everyone a wonderful weekend and saying goodbye.
* Ratings requested on both Apple Podcasts and Spotify (5 stars).
* Newsletter sign-up located at TBPN.com.
## Notable quotes
> [24:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1441s) We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
## Mentioned
John Coogan, Jordi Hays, Emmanuel Macron, Andrew Ross Sorkin, CNBC, Squawk Box, DealBook, Substack, Subway Takes, Keep the Meter Running, MrBeast, Forbes, The Joe Rogan Experience, The New York Times, YouTube, Ezra Klein, Ross Douthat, Bryan Johnson, Meta, Wall Street Journal, Christopher Mims, Meta Platforms, EssilorLuxottica, Ray-Ban Meta, Kylie Jenner, Threads, Oakley, Snap, Spectacles, Prada, Qualcomm, Samsung, GitHub, AWS, Google, Microsoft, OpenAI, Anthropic, Arena, Polymarket, Kalshi, Facebook, Instagram, Alex Heath, Robert Downey Jr., GPT-5.6 Soul, GPT-5.6 Terra, GPT-5.6 Luna, Alibaba, Dan Shipper, Theo Von, Trump administration, GPT-5.5 Cyber, Mythos, NSA, Jalapeño, Broadcom, ChatGPT, Codex API, Grand Theft Auto 6, Burj Khalifa, Stripe, Bill Gates, Intercept, TBPN, Apple Podcasts, Spotify
TBPN hosts John Coogan and Jordi Hays detail a structural evolution across digital media and artificial intelligence, where creator business models are splitting between high-margin niche commentary and capital-intensive productions, while frontier AI labs face escalating supply chain constraints and direct federal intervention.
Creator Economy Bifurcation and Monetization Bottlenecks (0:00)
Discussions at the Cannes Lions festival signaled a major transition as traditional advertising gives way to creator-led media, prompting debates highlighted by CNBC's Andrew Ross Sorkin over whether independent creators should consolidate or remain autonomous. The hosts argue the market is experiencing a sorting process: low-overhead commentary formats retain lucrative 80% to 90% EBITDA margins, while ambitious productions like Subway Takes and MrBeast—who generated $300 million in gross revenue—face significantly compressed net margins due to continuous capital reinvestment.
Simultaneously, highly produced vertical video creators face substantial monetization bottlenecks. Because vertical clips lack organic mid-roll breaks, creators must rely on dedicated 30- to 60-second sponsored posts that frequently underperform algorithmically, complicating the financial viability of expensive independent shows.
Meta’s Hardware Dominance and Enterprise Coding Pivot
In personal devices, Meta commands over 80% of the smart glasses market, with Meta and its partners shipping more than 7 million units and generating an estimated $2.1 billion to $3.0 billion in gross retail revenue. To solidify consumer appeal, Meta rolled out a $399 Kylie Jenner collaboration supported by an estimated $50 million promotional campaign, contrasting with Snap's higher-end $2,200 Spectacles and reported $100 million partnership development with Robert Downey Jr.
Internally, Meta is executing an aggressive reallocation toward code generation, reassigning 30% to 50% of core engineering teams to manual data labeling for AI-generated code. Rather than focusing strictly on its consumer AI assistant—which currently sits at #17 in the app store—Meta is courting outside financing to enter enterprise B2B coding services against entrenched cloud providers, alongside testing an independent prediction markets app codenamed Arena.
Model Governance and Asymmetric National Security Risks (21:47)
OpenAI announced its GPT-5.6 model family—comprising Soul, Terra, and Luna—but restricted access to the flagship Soul model to only 20 government-approved enterprises under White House directives. This clampdown follows widespread industry concerns over API distillation, highlighted by Anthropic's disclosure that Alibaba distilled Claude tokens at scale, as well as cyber capabilities demonstrated when Anthropic's Mythos model breached NSA systems during red-teaming.
The hosts emphasize a dangerous temporal asymmetry between cyber and biological threats. In cyber defense, a frontier model that identifies an exploit can autonomously deploy a patch within an hour. In biosecurity, however, open-source models lag frontier performance by only six months, whereas developing and manufacturing physical countermeasures requires over a year, leaving society exposed if weaponizable designs leak before physical defenses exist.
Deep Hardware Lockups and Capital-Intensive Moonshots (29:08)
OpenAI is securing deep infrastructure advantages by launching Jalapeño, a custom ASIC developed in collaboration with Broadcom for large language models and agentic workloads. To secure future capacity, OpenAI executed an aggressive supply chain contract locking up 40% of the world's raw undoped DRAM wafer production through 2029.
This trend of unprecedented capital expenditure mirrors broader industrial scale: OpenAI, Anthropic, Stripe, and Bill Gates pooled $500 million to fund Intercept, a research initiative aimed at eliminating respiratory viruses. Meanwhile, entertainment budgets have expanded to rival major infrastructure projects, with Grand Theft Auto 6 amassing a reported $2 billion development budget over 12 years—surpassing the $1.5 billion required to construct Dubai's Burj Khalifa.
Across digital media, consumer hardware, and frontier AI, commercial success increasingly requires immense capital scale and vertical integration, yet companies face unresolved questions regarding how to monetize high production overheads and navigate arbitrary executive interventions as national security concerns rise.
## Background
本期内容由科技播客频道 TBPN 于 2026 年 7 月 23 日发布在 YouTube 平台。节目属于其 30 分钟精简栏目《Diet TBPN》，由 John Coogan 与 Jordi Hays 主持，旨在快速盘点硅谷科技圈当天的核心动态，涵盖创作者经济转型、Kylie Jenner 与 Meta 联名硬件以及 OpenAI GPT-5.6 的受限发布等重磅话题。
## TL;DR
节目首先从戛纳创意节切入，探讨了独立创作者转向重度节目制作后遭遇的内容变现断层与利润困境；随后剖析了 Meta 智能眼镜凭借时尚联名拿下 80% 市场份额的突破，以及其内部调动大批工程师攻坚代码大模型、孵化预测市场应用 Arena 的激进动作；紧接着聚焦因防范违规蒸馏与国家安全审查而仅向 20 家企业白名单限额开放的 GPT-5.6 Sol，并深度对比了网络与生物安全的防御时间差危机；最后揭示了 OpenAI 联合博通研发专用芯片 Jalapeño 并锁定全球 40% 晶圆原料的底层供应链野心，以及科技巨头联合注资 5 亿美元消灭呼吸道病毒的 Intercept 科研新项目。
## Atmosphere & style
整场节目充满了典型硅谷圈内人犀利、高频且极度松弛的脱口秀氛围，两位主持人在严肃的硬核商业推演与密集的无厘头吐槽之间切换自如。从开场调侃停播戒断反应、香烟盒上的长寿极客，到讽刺科技博主脆弱的健康优化人设，对话充斥着反差强烈的冷幽默与讽刺意味，让复杂的地缘 AI 监管和算力军备竞赛话题听起来既生动又极具现场代入感。
## Bulletin
- [09:02](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=542s) Meta 智能眼镜自 2025 年以来出货量突破 700 万副，斩获全球超 80% 市场份额；同时 Meta 内部展开激进调岗，将 30% 至 50% 核心工程师抽调至代码生成大模型的数据标注流水线。
- [21:47](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1307s) OpenAI 下一代旗舰模型 GPT-5.6 Sol 仅向 20 家白宫预审企业极度受限地开放，主因在于防范海外黑产对旗舰权重的规模化“数据蒸馏”及潜在生化风险。
- [29:08](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1748s) OpenAI 联合博通通过 AI Agent 协同设计自研芯片 Jalapeño，并激进锁定了截至 2029 年全球 40% 的未切割原始 DRAM 晶圆产能。
- [30:34](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1834s) OpenAI、Anthropic、Stripe 与比尔·盖茨联合注资 5 亿美元设立 Intercept 科研机构，目标是从阻断感冒与流感入手，彻底消除人类呼吸道病毒。
## Key takeaways
- [00:43](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=43s) 创作者经济正在从碎片化内容向高规格节目转型，但由于竖屏短视频缺乏中插广告机制，优质节目难以变现，而纯商业广告视频流量低迷，导致行业陷入盈利断层。
- [03:31](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=211s) MrBeast 等顶流网红数亿美元的年收入仅为未扣除制作成本的总流水，而乔·罗根式低成本演播室模式展现出更健康的实际净利润率。
- [04:50](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=290s) 创作者生态正在经历分化，主流机构内的专业人才更倾向于将“单飞可能性”作为与机构博弈更高保障合同的筹码，而非贸然投身竞争惨烈的内容红海。
- [09:02](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=542s) Meta 与合作方智能眼镜出货已破 700 万副、零售额预估达 21 亿至 30 亿美元，虽受限于庞大体量未被资本市场充分重估，但已展现出强劲的产品市场匹配度（PMF）。
- [10:45](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=645s) Meta 采取兼顾医保报销（299 美元起）与潮流女性联名（399 美元 Kylie Jenner 款）的价格策略，与 Snap 标价 2,200 美元厚重概念眼镜拉开巨大市场差距。
- [16:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=960s) Meta 内部掀起 AI“饥饿游戏”，核心团队多达 30% 至 50% 的程序员被强行转岗至代码标注与 RLHF 任务，以全力攻坚企业级代码大模型。
- [18:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1081s) 面对企业端 AI 追赶乏力与估值疑虑，扎克伯格内部立项代号“Arena”的独立预测市场应用，正面抗衡 Polymarket 与 Kalshi。
- [21:47](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1307s) OpenAI 确立了 Sol（旗舰）、Terra（均衡）、Luna（高并发）的三层模型矩阵，但受白宫行政令管控，顶级前沿模型 GPT-5.6 Sol 的访问权限被严格卡死在 20 家企业名单内。
- [23:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1380s) AI 任务能耗与传统流媒体存在量级差异，单个自主 Agent 执行复杂逻辑所消耗的算力与电量，相当于单个用户连续播放 10 年 YouTube 视频。
- [24:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1441s) 网络安全漏洞可由 AI 在数小时内生成补丁并修复，而生化病毒从开源模型具备合成能力到现实世界实体疫苗研制存在一年以上的严重防御时间差，构成巨大安全隐患。
- [29:08](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1748s) OpenAI 凭借 AI Agent 协同设计架构大幅压缩了硬件研发周期，推出专为 LLM 设计的 Jalapeño 芯片，并深度控制上游生 DRAM 晶圆供应链。
- [30:09](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1809s) 《侠盗猎车手 6》（GTA 6）预估开发成本超 20 亿美元、研发跨度达 12 年，在工期与资金量级上已全面超越迪拜哈利法塔（15 亿美元/6 年）。
- [30:34](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1834s) 5 亿美元成立的 Intercept 瞄准育儿家庭的真实痛点，数据显示拥有多孩的父母年均处于病毒感染期的高峰比例逼近 50%。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=0s) 重返直播间与法国之旅的戒断反应
节目开场，两位主持人 John Coogan 与 Jordi Hays 宣告正式回归。他们坦言离开直播间去法国出差的这几天格外难熬，笑称如果到了太平洋时间下午一点还不开播，整个人就会感到极度不适，“不管身在哪个时区，冷汗直流、浑身发抖的戒断反应简直疯了”。两人调侃道，如果连续停播几周或许能稍微缓解，但骤然停播的身体症状确实非常折磨人。
除了直播带来的心理戒断，这趟法国之旅本身也充满戏剧性。Jordi 曾在节目中公开抨击过法国总统埃马纽埃尔·马克龙（Emmanuel Macron），而法国近期又全面禁止了尼古丁袋（几乎禁绝了所有经口摄入的尼古丁产品），这让 John 感觉法国政府简直是在针对他个人出台禁令。在一番插科打诨后，话题迅速转向了他们在法国参加的核心活动——戛纳国际创意节（Cannes Lions）。
- 主持人调侃每日下午一点不直播就会出现严重的生理戒断症状（出汗发抖）。
- 法国近期通过了针对口服尼古丁袋（Nicotine Pouches）的全面禁令。
---
### [00:43](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=43s) 戛纳创意节与创作者经济的转折：从“做内容”到“做节目”
谈到戛纳国际创意节，John 指出这个过去主要为传统创意广告颁奖的盛会，如今已将重心大规模转向创作者经济。现场虽然缺少一线好莱坞巨星，但几乎每个人都自带一定规模的粉丝受众。CNBC《Squawk Box》与《DealBook》的主持人安德鲁·罗斯·索尔金（Andrew Ross Sorkin）在现场提出并引发热议的核心问题是：创作者在“独立经营”与“抱团整合”之间的权衡。创作者经济是否也存在最佳的规模效应？比如独立 Substack 作者是否应该被整合打包？
索尔金在返程航班上发文感慨，一些原本看似最成功的头部网红向他坦白，为了维持热度与关注，他们必须投入巨额资金制作内容，导致实际利润远没有外界头条报道的那么惊人。John 深入剖析了这一现象背后的两大动力：其一，独立创作者正在经历从单纯“做内容”到“制作高规格节目”的重大转型，代表作如《Subway Takes》和《Keep the Meter Running》。这类节目的制作成本、剪辑标准和工业化程度大幅提升，但随之而来的变现难题却愈发尴尬。
目前竖屏短视频领域尚未形成成熟的中插广告机制。像《Subway Takes》这类时长约三分钟的精良视频，完全可以容纳 5 到 10 秒的中插广告，但目前行业闸门尚未开启。绝大多数创作者不得不采取“一明一暗”的妥协模式：先发布一条高制作水准的纯内容爆款，紧接着再发一条风格相似但纯属商业推广的 30 至 60 秒竖屏视频。然而赞助视频往往流量惨淡，陷入了“受众喜爱的内容无法变现，变现的广告内容没人要看”的恶性循环。John 还顺势现场调侃了自己节目的赞助商思科（Cisco），直言如果 TBPN 也必须做一期纯讲思科、下一期再讲科技新闻，节目模式根本行不通。
```mermaid
flowchart TD
accTitle: 创作者向重度制作节目转型后的变现困境
accDescr: 展示独立创作者转向重度制作节目后，内容与商业化脱节的恶性循环路径。
subgraph Production ["制作层级提升"]
A["传统个人碎片内容"] --> B["高成本精编节目<br>(如 Subway Takes / Keep the Meter Running)"]
end
subgraph MonetizationIssue ["商业化断层"]
B --> C{"短视频缺乏合理中插机制"}
C -->|优质内容| D["用户喜爱但无法自然植入变现"]
C -->|硬广方案| E["制作纯商业赞助视频"]
end
subgraph Outcome ["结果失衡"]
D --> F["单片制作成本高企，利润承压"]
E --> G["赞助视频完播与传播量大幅下滑"]
end
```
---
### [03:31](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=211s) MrBeast 的高成本重投与传统媒体（纽约时报）的逆袭
福布斯（Forbes）最近发布了全球顶尖创作者收入榜单，MrBeast 以 3 亿美元高居榜首。但两位主持人强调，媒体用“收益（Earnings）”一词极具误导性，这充其量只能算是总营收（Gross Revenue）。鉴于 Jimmy（MrBeast）对把每一分钱重新砸进下一部视频制作的狂热执念，谁也不知道他在 3 亿美元流水下的实际净利润率究竟还剩多少。相比之下，乔·罗根（Joe Rogan）手握数亿美元天价合约，却满足于在一个简陋极简的演播室里录音，坐拥惊人利润率。
与此同时，传统主流媒体在被网红统治的平台上开始找到属于自己的节奏。以《纽约时报》为例，其旗下的播客节目在 YouTube 上大获成功。无论是代表左翼视角的《The Ezra Klein Show》还是偏右翼视角的 Ross Douthat 专栏，都极好地掌握了 YouTube 的平台逻辑：精心调优的封面图与吸睛标题（Packaging）、极高的制作与剪辑水准，使得传统媒体制作的深度长视频在社交平台上爆发，打破了过去主流媒体难以在创作者原生平台上立足的魔咒。
- 福布斯创作者收入榜 MrBeast 居首（3 亿美元），但这反映的是未扣除高昂制作成本的总营收。
- 乔·罗根的低制作成本、极简演播室模式与 MrBeast 的“激进再投资”模式形成鲜明对比。
- 《纽约时报》凭借《The Ezra Klein Show》等栏目，成功将高质量采编与 YouTube 的包装运营结合，打入流媒体受众。
---
### [04:50](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=290s) 烟盒上的“Bryan Johnson”与创作者归宿的再思考
节目中插入了一则令人啼笑皆非的法国见闻：法国香烟盒上印有一张醒目的健康警示图，画面中的男子紧闭双眼躺着，外观神似硅谷著名的抗衰长寿极客布莱恩·约翰逊（Bryan Johnson）。John 专门拍下照片，调侃称这画面简直让人联想到约翰逊的“别死（Don't Die）”哲学，搞不好约翰逊本人还会对这种反吸烟宣传表示赞许。Jordi 则戏谑道，这张图既像一个人被装进装尸袋，又像是一个刚抽完两根烟、喝了几听冰啤酒后心满意足被盖被子入睡的中年男人，比起法国烟盒上那些血淋淋的黑肺和癌变 X 光片，这张图的视觉冲击力其实已经相当温和了。
笑谈过后，讨论重新回到创作者的未来形态。这实际上存在一种“围城”效应：体制外的独立创作者看着大机构的高薪与成熟保障，觉得自己单打独斗太辛苦；而主流媒体机构内部的记者和主播又觉得如果单飞做自媒体，赚的钱会翻倍。事实证明，并不存在所谓“独立创作者就是未来”的放之四海而皆准的套路，行业正经历一种分化与分类（Sorting Process）。
一些特定垂直细分赛道的创作者确实能在机构之外找到极其优雅的商业模式，享受 80% 至 90% 的超高 EBITDA 利润率；但对另一些人而言，内容产业天然是低利润率、高竞争、完全依赖爆款（Hit-driven）的残酷行当。越来越多机构内部的专业人才，其最优解是将“单飞的可能性”作为筹码，去向传统媒体机构争取更丰厚、更有保障的合同待遇，而不是真正赤身肉搏投入创作者红海。
---
### [07:07](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=427s) 智能眼镜狂潮与机舱内的 Apple Vision Pro 体验
节目的最后一个议题切入硬件领域。《华尔街日报》科技专栏作家克里斯托弗·米姆斯（Christopher Mims）针对 Meta 最新发布的智能眼镜撰文发问：“智能眼镜不可避免，但它们到底是为了什么人、为了解决什么问题而存在？”米姆斯声称智能眼镜的洪流（Deluge）已经开启。John 对此观点持保留态度，认为现状更像是涓涓细流（Trickle）而非狂潮，但承认该品类的技术与普及度正在变得不可忽视。
Meta 的雷朋（Ray-Ban）联名智能眼镜出货量相当亮眼，已售出数百万台，营收达数亿美元规模。该产品的实际本质并非替代 AirPods，而是一副搭载了高品质摄像头与音频模块的优秀日常眼镜，并且依托 Meta 历史上最强大的分发平台与独家广告位持续推流。
在谈到智能眼镜与头显时，Jordi 分享了他在这次出差航班上首次试用苹果 Vision Pro 的滑稽经历。John 一直极力安利他体验，尽管在狭窄的机舱里设置访客模式十分繁琐，Jordi 还是戴上了设备。结果刚进入标志性的恐龙互动演示（Dinosaur Demo）不到 20 秒，沉浸感正达到顶峰时，空乘人员就突然走过来拍他的肩膀，打断了整个初体验，引发录播间一阵大笑。
- Meta Ray-Ban 智能眼镜已经取得数百万台销量及数亿美元的实质性商业体量。
- 智能眼镜目前的成功路径是“以优质眼镜为底座，附带摄像头与音频”，而非激进取代手机或耳机。
- Jordi 在飞行途中首次试戴 Apple Vision Pro，却在体验恐龙演示仅 20 秒时被空乘人员拍肩打断。
### [09:02](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=542s) 沉浸式体验的尴尬与智能眼镜销量突破
节目录制前的一段趣事引发了主持人们的调侃：戴着头显沉浸在虚拟世界中时被外界强行打断，是件极其扫兴的事，甚至让人觉得有种被冒犯的“放肆”（the audacity）。有人突然过来问“这玩意儿酷吗”，而不是提醒系安全带这种正经事，瞬间把那种完全沉浸的体验破坏得一干二净。相比之下，Apple Vision Pro 主打的“眼睛透视”功能号称无需摘下头显就能与外界眼神交流，但现实场景下依然无法消解这种隔阂。闲聊间，几人还打趣了团队成员 Tyler 去法国戴高乐机场（CDG）吃 Raising Cane's 炸鸡被落下的八卦，随后迅速切回硬核数据——Meta 智能眼镜到底卖得怎么样？
数据显示，自 2025 年以来，Meta 与 EssilorLuxottica 合作的智能眼镜出货量已突破 700 万副，其中 Ray-Ban 联名款就占了 200 万副，估计零售总额在 21 亿至 30 亿美元之间。主持人指出，如果是 Oura Ring 或 Whoop 这样单独的硬件创业公司取得 700 万副新硬件品类的销量，外界早就惊呼其拥有“不可思议的 Product Market Fit（产品市场匹配度）”，甚至足以支撑一家独立的上市公司；然而正因为它是庞然大物 Meta，市场反而习惯性忽略这一成就，甚至拿它和 Apple Vision Pro 相比，抱怨在眼镜上看《怒海争锋》（Master and Commander）等老电影的视觉体验不如头显。
- **Meta / 合作方智能眼镜总销量**：2025 年以来累计突破 700 万副。
- **Ray-Ban Meta 系列**：单款销量突破 200 万副。
- **零售额估算**：总零售额达到 21 亿至 30 亿美元区间。
---
### [10:45](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=645s) 凯莉·詹娜联名款、80% 市场份额与可穿戴生态
主持人们引用了《华尔街日报》科技专栏作家 Christopher Mims 的最新分析。Meta 针对智能眼镜的定价策略十分灵活：低端版本起售价仅为 299 美元，并且能走视力健康保险报销；而加价 100 美元（399 美元），就能买到专门针对凯莉·詹娜（Kylie Jenner）粉丝打造的联名款。这款针对年轻潮流女性的眼镜在外观上极尽隐蔽，普通人路过根本看不出是智能设备，镜框角落那颗闪烁的白色微光其实只是装饰性的珠宝点缀，而非显示屏或录像指示灯，这种纯美学设计在目标客群中大受欢迎。
作为对比，高端市场的定价则迥然不同：Meta 的高阶探索款售价约 800 美元，而 Snap 的 Spectacles 则高达 2,200 美元，外观极其笨重，被戏称为“只会出现在 Prada 时尚大片里、现实中绝不会有人戴的怪东西”。硅谷推动智能眼镜已有十余年，自 Google Glass 失败后，扎克伯格的 Meta 如今已经稳稳吃下了该领域超过 80% 的市场份额。尽管与年销量超过 1 亿只的智能手表以及数十亿部智能手机相比，700 万副看似微不足道，但这已经证明智能眼镜正在稳健步入正轨。
```mermaid
flowchart TD
accTitle: 智能眼镜产品定位与市场格局
accDescr: 对比Meta与Snap等品牌在智能眼镜上的产品分层及市场份额。
subgraph MetaLineup["Meta 智能眼镜产品矩阵"]
direction TB
Entry["入门基础款 ($299)<br/>支持视力医保报销"]
Fashion["Kylie Jenner 联名潮流款 ($399)<br/>珠宝修饰/时尚外观隐蔽"]
HighEnd["高阶探索款 ($800)"]
end
subgraph SnapCompetitor["竞争对手"]
SnapSpecs["Snap Spectacles ($2,200)<br/>造型厚重/偏向概念实验"]
end
subgraph MarketShare["整体市场体量与占比"]
MetaShare["Meta 占据全球 80%+ 份额<br/>累计 700万+ 副出货量"]
CompareScale["对比参照：<br/>智能手表年出货 1亿+ 只<br/>智能手机年出货 10亿+ 部"]
end
Entry --> MetaShare
Fashion --> MetaShare
HighEnd --> MetaShare
SnapSpecs -.-> CompareScale
```
---
### [12:28](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=748s) 广告变现模式脑洞与人脸计算机的隐私争议
面对投资者对 Meta 硬件利润率的疑虑，主持人们探讨了硬件与核心广告业务的联动可能性。硬件即便带来数十亿美元收入，也远不及高毛利的软件广告；与其指望硬件利润去盖数据中心，不如想象类似亚马逊“广告补贴版 Kindle”的廉价乃至免费智能眼镜。有主持人设想，用户戴着眼镜走在街上路过咖啡馆，视野里会自动弹出一张“免费抹茶拿铁优惠券”；而另一位主持人则脑洞大开地反向调侃，称自己更想要一种“向外狂吼广告”的眼镜——佩戴者享受主动降噪，眼镜外置扬声器却在周围人群走近时自动大声播报广告词，佩戴者瞬间化身行走的广告牌。
借着这个由头，现场插播了一段幽默的假想广告：“你的智能云平台是哪家？是 Railway 吗？Railway 是一站式智能云服务商，让你用偏好的 AI 智能体轻松部署 Web 应用、服务器和数据库，自动处理扩展、监控与安全。”
玩笑归玩笑，把计算机戴在脸上仍面临巨大的社会阻力和隐私反弹。现代人本就极力想摆脱屏幕依赖，将镜头和显示屏直接架在眼球前方难免引起大众反感与警惕；人们对于身边随时有人用连网摄像头对着自己感到极度不安，这被视作对残存个人隐私的严重侵犯。
---
### [15:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=900s) 鸡肋的“视觉识别”与人体终极地盘争夺战
智能眼镜搭载的多模态 AI 也是讨论的焦点。然而，主持人直言目前眼镜主打的“实物识别”功能纯属鸡肋，宛如美剧《硅谷》里的“热狗 / 不是热狗（Hot dog, not hot dog）”检测器。现实生活中，谁会闲着没事指着一栋楼问“这是什么建筑”，指着一棵树问“这是什么树”，或者在马路上偶遇一只罕见的南非猫鼬（meerkat）需要 AI 来科普？普通人一眼就能认出眼前是一听无糖可乐，根本不需要视觉识别。智能眼镜真正有价值的核心，在于能随时与随身 AI Agent 语言对话、调度手机任务和检索复杂信息。
尽管外界充满疑虑，可穿戴智能硬件依旧被视作技术演进的必然趋势。高通负责为 Meta、Snap 和三星智能眼镜供应芯片的高管曾直言，“未来所有的眼镜最终都将是智能眼镜”。这是硅谷科技巨头在智能手机之后，针对人类身体上“最后也是最黄金的地盘”展开的新一轮消费级硬件军备竞赛。
---
### [16:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=960s) Meta 全员大转岗：内部的 AI“饥饿游戏”
除了硬件，Meta 在底层人工智能上的押注更加激进。来自外部科技观察者和内部传出的消息表明，Meta 目前正全力攻坚代码生成大模型（Coding Models），并在掀起全球规模最大的代码训练数据生成与标注战役。为了实现扎克伯格在 AI 领域的领跑野心，Meta 内部正在经历剧烈的人力重组：核心业务部门中有 30% 到 50% 的软件工程师被强势调离原有岗位，强制转向代码数据标注（Data Labeling）以及为 AI 生成的代码库提供人类强化反馈（RLHF）。
一位 Meta 工程师向媒体将这种局面形容为电影《饥饿游戏》（The Hunger Games）：各个团队像贡品一样被随机挑选出来，直接从服务数亿用户的日常业务中拔出，塞进完全陌生的人工反馈流水线；十人规模的小组里往往有三到五个人被抽调去给 GitHub 代码打标签。虽说环境依然是开着冷气、有免费零食提供的办公室，但战略重心转向的力度之大可见一斑。Meta 目前不仅在算力基础设施上投入巨资，还在探索通过创新的融资架构募集数十亿美元资金。外界此前普遍预计 Meta 会倾尽全力将消费者端的 Meta AI 聊天机器人推至榜首以强化其广告生态，如今其在企业级代码模型与底层合成数据上的激进动作，标志着扎克伯格正在开辟全新的技术战场。
### [18:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1081s) Meta 的企业级转型疑虑与独立预测市场应用 Arena
节目开场继续审视 Meta 的当前处境。目前 Meta AI 独立应用在 App Store 仅排在第 17 位，这绝非一个能让市场放心的数字。主持人们指出，资本市场显然没有对 Meta 释放出积极信号；Meta 试图转向企业级市场，并准备砸下数千亿美元与 AWS、微软、Google、OpenAI、Anthropic 以及一众中国大模型实验室正面竞争，这如果能赢，将是一场不可思议的“绝地反击战”。然而现实很骨感，市场根本没有把这种胜算计入估值之中。
雪上加霜的是，据《纽约时报》记者 Mike Isaac 爆料，马克·扎克伯格已指令 Meta 内部开发一款独立的预测市场（Prediction Markets）应用，内部代号为“Arena”。该应用完全独立于 Facebook 和 Instagram，意图直接与当红的 Polymarket 和 Kalshi 争夺用户注意力。主持人们对此评价道，预测市场整体的大众舆论目前几乎跌入谷底——早先人们还觉得它能用来预测大选结果，现在却越来越像纯粹的体育博彩，“整个氛围彻底变味了”。虽然很多重度用户和机构依然把预测市场当作有效的数据源，而且相关平台的营收正在爆炸式增长，甚至有人靠 DraftKings 赌球赚了 600 万美元盖了带保龄球道的豪宅，但公众对其的道德好感度极差。主持人们打趣说，Meta 目前在资本市场的口碑也跌到了谷底，扎克伯格此举或许纯粹是抱着“破罐破摔，反正情况也不能更糟了（F it, I don't think it can get any worse）”的心态在放手一搏。
- **Meta AI 榜单表现**：在 App Store 免费榜仅位居第 17 位，缺乏绝对的 C 端统治力。
- **竞争版图**：企业级云端与模型面临 AWS、微软、Google、OpenAI、Anthropic 及中国头部实验室的围堵。
- **内部代号 Arena**：脱离 FB/IG 体系独立运作的预测市场应用，正面迎战 Polymarket 与 Kalshi。
---
### [20:11](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1211s) 顶流网红带货与明星天价联名：Kylie Jenner 对标 Snap 与小罗伯特·唐尼
与企业端和预测市场的低迷形成鲜明对比的是，市场对 Kylie Jenner（凯莉·詹娜）代言合作的 Meta 智能眼镜给出了极其热烈的正面反馈。硅谷似乎终于开窍，弄明白了究竟是谁在真正支配全球的消费支出。不过主持人也纠正道，Meta 长期以来一直善于同顶级名人深度合作，拉来 Kylie 并不令人意外，因为她本身就是 Instagram 上的绝对顶流，双方有着极深的历史渊源。
主持人们讨论起这笔代言费的天价数字：
> “要请动全球排名第一的超级网红来搞这种级别的战役并为你首发硬件，得花多少钱？我盲猜差不多要 5000 万美元。”
然而，知名科技记者 Alex Heath（被节目戏称为“独家爆料机器 Heathnator”）随后抛出了一记更震撼的重磅猛料：Snap 正在洽谈与小罗伯特·唐尼（Robert Downey Jr.）达成一项高达 1 亿美元的天价合作。这一消息让主持人们感到匪夷所思——Kylie 与 Meta 的受众完全契合，但 Snap 花 1 亿砸唐尼究竟图什么？两人随即开始脑洞大开地调侃起来：也许这不是单纯的代言，而是深度内容绑定，“我们搞不好能看到一部《钢铁侠 4：Snap Spectacles 特别篇》，整部电影的核心剧情就是托尼·斯塔克脱下战甲，只戴着一幅 Snap 智能眼镜，第五次去把灭霸给收拾了”。
---
### [21:47](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1307s) GPT-5.6 Sol 限额发布与“蒸馏门”：白宫管控下的 20 家企业白名单
随后话题切入重磅产品动态：OpenAI 正式小范围预览下一代大模型 GPT-5.6 “Sol”。根据命名规则，Sol（太阳）代表最高规格的顶尖前沿模型（Frontier Model），Terra（地球）代表面向日常任务的高能效均衡模型，而 Luna（月亮）则是面向高并发、大吞吐场景的快速平价模型。然而，受美国政府行政指令的严格限制，目前全美仅有 20 家预先批准的公司获得了 GPT-5.6 Sol 的访问权限。这一名额极其苛刻，以至于像 Dan Shipper 这样的核心科技博主都无法第一时间上手进行“氛围测试（vibe check）”。
这种空前的管控并非凭空而来，其背后很大程度与近期闹得沸沸扬扬的“蒸馏门（distill gate）”密切相关。Anthropic 此前公开指控阿里巴巴多次对其旗舰模型 Claude 进行违规数据蒸馏，甚至有图表披露海量非法转售 API Token 的惊人黑产链条，被业内形容为“职业化的诈骗体系”。由于担心前沿顶尖权重与模型输出被海外竞争对手瞬间蒸馏复制，美国监管机构与各 AI 巨头达成了高度防范的共识，严格收紧核心模型的下发口径。
- **GPT-5.6 矩阵命名体系**：
- **Sol**：前沿旗舰模型，性能天花板，受最严格安全管控。
- **Terra**：高能效平衡模型，主打企业日常办公与生产力任务。
- **Luna**：轻量高速且价格低廉，用于海量自动化任务。
- **白名单限制**：仅限 20 家白宫预审通过的企业使用，访问门槛创历史新高。
- **蒸馏争端**：Anthropic 披露违规蒸馏套壳行为已规模化，迫使美国对顶尖模型采取物理防线。
---
### [23:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1380s) Theo Von 炮轰数据中心与 AI 惊人的能耗代差
科技基建在民间引发的民意反弹也成了热议焦点。知名播客主持人 Theo Von 在节目中公开炮轰数据中心建设，直言：“老兄，根本没人想要数据中心，那些想要这玩意儿的人在我看来简直邪恶至极。没人需要这东西，到头来某一家公司会占有我们所有的隐私信息。”甚至他还荒诞地提出了“情绪信用分（emotional credit score）”的阴谋论担忧。
针对外界“Theo Von 既然自己录播客发在流媒体上，怎么还反对数据中心”的批评，主持人驳斥了这种浅薄的反驳逻辑。他们解释道，支撑 YouTube 视频播放或播客音频分发所消耗的数据中心能耗与算力，与 AI 相比简直是九牛一毛；“你随意调动一个自主 AI Agent 跑复杂任务，其消耗的电量和算力，恐怕相当于一个人连续看 10 年 YouTube”。当前数据中心建设对水资源、电力网络基础设施的巨大索求，与传统互联网流媒体完全不在一个数量级，行业亟需向公众更清晰地解释这种技术代价与真实用途。
---
### [24:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1441s) 国家安全审查常态化、开源追赶与“网络 vs 生物”防御时间差危机
节目重点探讨了 OpenAI 限制新模型发布的深层内幕。在与特朗普政府进行闭门商讨后，OpenAI 正式限制了最新 AI 模型的权限。不过 OpenAI 也向白宫发出警告：白宫以“国家安全”为由对 AI 行业进行逐案审查的干预模式，绝不能成为长期的常态。安全界对此同样深感忧虑，认为这种“一事一议”的行政干预是最糟糕的监管方案，既缺乏统一透明的合规标准，又赋予了政府过大的裁量权。与此同时，智谱开源的 GLM-5.2 等模型正在飞速缩小差距，许多开发者已将其作为主力生产工具；各大实验室负责人均承认，开源模型目前仅落后闭源前沿约 6 个月。这意味着 6 个月前顶尖模型的超强网络攻防能力，如今正在开源社区全盘涌现。
主持人们由此推演出了一个发人深省的安全悖论模型：**网络安全（Cyber）与生物安全（Bio）存在致命的“防御时间差”不对称性**。
在网络世界中，如果前沿模型掌握了强大的漏洞入侵能力，开发团队可以在数小时内用该模型生成防御补丁并完成代码合并（Merge），防御的部署几乎是瞬时的。然而在生物安全领域，情况完全颠覆：如果一个顶尖模型能发现高致死率的新型病毒结构，虽然它理论上也能设计出早警系统与解毒药物，但现实世界中疫苗和药物的临床试验、物理合成、实体供应链制造与全社会接种防御至少需要一年甚至更久。
```mermaid
flowchart TD
accTitle: 网络安全与生物安全的防御时间差对比
accDescr: 对比网络入侵修复与生物病毒防御在开源追赶周期下的时间不对称性
subgraph Cyber["网络安全攻防闭环 (小时级)"]
C1["前沿模型发现高危漏洞"] --> C2["AI 生成防护补丁并部署代码"]
C2 --> C3["数小时内完成社会级网络修复"]
end
subgraph Bio["生物安全时间差危机 (一年以上)"]
B1["前沿模型发现潜在致命病毒"] --> B2["开源社区 6 个月后追平合成能力"]
B2 --> B3["恶意行为者获取生化武器制造能力"]
B1 --> B4["实体药物研发、审批与疫苗生产"]
B4 --> B5["社会免疫防御建立 (耗时需 1 年以上)"]
B3 -.->|"致命空窗期：进攻先于防御落地"| B5
end
```
如果开源黑客在 6 个月后获得了制造生物武器的能力，而实体社会的生物防御网络尚未制造完成，就会形成致命的安全真空期。主持人们回忆道，此前 GPT-5.5 Cyber 版本展现出了惊人的软件漏洞挖掘实力，而 Anthropic 的 Mythos 模型在红队演练中更是直接黑进了美国国家安全局（NSA）的大批内部系统，直接引发了白宫对 AI 行业监管政策的彻底重构。这场关于进攻速度与实体防御速度的赛跑，已然成为悬在整个行业头顶的达摩克利斯之剑。
### [27:44](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1664s) GPT 5.6 限制发布与政府审查：开源蒸馏与“封禁光环”
为了遵守新规，相关访问权限已被暂停。此前 Anthropic 就曾与政府合作，对其早期版本的 Mythos 模型进行过有限规模的测试推广。OpenAI 方面则表示，希望在未来几周内让 GPT 5.6 面向公众全面开放（GA），并指出目前的政府审批流程只是配合落实特朗普总统近期关于模型监管行政令的过渡阶段。OpenAI 在声明中明确表态：“我们不认为这种政府准入审批程序应该成为长期的默认机制。它把最优秀的工具挡在了急需它们的用户、开发者、企业、网络安全防御人员以及全球合作伙伴之外。我们之所以在短期内妥协配合，是因为我们坚信这是通往更广泛开放的最有力路径。尤其是随着开源前沿模型的不断推进，继续将闭源模型深锁高阁就显得越来越不合理。”
随后两位主持人探讨了为何开源前沿模型能以如此快的速度追赶。其核心驱动力正是模型的普及程度：一旦闭源尖端模型通过日常消费者订阅全面开放，外界就会有人建立由成千上万个账户组成的网络，大规模抓取输出并对模型进行蒸馏（Distillation）。这种做法实际上等于把最顶尖的能力拱手让人，至于这是否会最终引发某种所谓的“末日场景”，目前仍未可知。谈及监管限制，主持人还开玩笑地调侃起扎克伯格（Mark Zuckerberg）和 Meta：如果 Meta 的新模型也能被政府来一张初步禁令，Meta 的股价会不会因此大涨？毕竟被政府列入“过于强大而需要封禁”的清单，简直就像是一种绝妙的“气场收割”（aura farm），这种反向营销眼下似乎非常奏效。
- Anthropic 此前曾就早期版本 Mythos 模型与政府开展有限推送合作。
- OpenAI 计划在未来数周内全面开放 GPT 5.6，视当前政府审核为行政令过渡期。
- 大规模账号网络正通过日常消费者订阅端对闭源模型进行“蒸馏”，直接助推了开源模型的快速演进。
---
### [29:08](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1748s) OpenAI 硬件野心：自研芯片 Jalapeño 与 DRAM 晶圆供应链垄断
OpenAI 在硬件领域迎来了重大突破，正式推出了与博通（Broadcom）联合设计的专用芯片“Jalapeño”。该芯片专为大语言模型（LLM）工作负载量身定制，旨在为 ChatGPT、Codex、底层 API 以及未来的智能体（Agentic）产品线提供算力支撑。令人惊叹的是其研发推进的速度——业内许多人透露，这是首款在研发流程中深度引入 AI 智能体（Agents in the loop）协同设计的芯片，使得编写和调试指令集架构的周期大幅压缩。尽管双方团队的人类工程师功不可没，但这标志着以往长达五年的芯片研发周期正在被彻底颠覆，芯片从立项到推向市场的节奏正在急剧收紧。
除了自主造芯，OpenAI 在底层半导体供应链上的操作也极其激进。消息透露，OpenAI 已经达成协议，直接锁定了截至 2029 年全球 40% 的未切割原始晶圆（raw undiced wafer）产能，涉及数以百万计的生 DRAM 晶圆。这些晶圆在深加工之前完全无法直接使用，但 OpenAI 已经选择将触角深深扎入制造业最底层的原材料供应链。主持人们感叹，OpenAI 这位居于顶层的“交易操盘大师”（指 Sam Altman）总能下出让人意想不到的大棋。
```mermaid
flowchart TD
accTitle: OpenAI 自研算力与供应链战略布局
accDescr: 展示 OpenAI 如何通过 AI 智能体辅助设计芯片 Jalapeño，并锁定上游晶圆原料来支撑未来产品矩阵。
subgraph SupplyChain["底层供应链锁定"]
A["锁定全球 40% 未切割原始晶圆<br/>(至 2029 年数百万 DRAM 晶圆)"]
end
subgraph ChipDev["芯片协同研发"]
B["AI 智能体协同设计<br/>(极速编写与验证指令集)"]
C["与 Broadcom 联合研发芯片<br/>'Jalapeño'"]
B --> C
end
subgraph ProductMatrix["终端算力赋能"]
D["ChatGPT 与 Codex"]
E["开发者 API 与 Agentic 智能体产品"]
end
A --> C
C --> D
C --> E
```
- **Jalapeño 芯片定位**：与博通联合开发，专为 LLM 工作负载打造，深度支持 Codex 与智能体架构。
- **研发范式转变**：采用“AI Agent in the loop”协助设计指令集，打破传统芯片 5 年研发周期的行业定律。
- **供应链掌控**：锁定全球 40% 的未切割原始 DRAM 晶圆产能直至 2029 年。
---
### [30:09](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1809s) 游戏界的哈利法塔：GTA 6 成本与工期对比
话题随后转向开发周期与资金投入的极致案例。目前业内纷纷将备受期待的游戏巨作《侠盗猎车手 6》（GTA 6）与迪拜的哈利法塔（Burj Khalifa）进行横向对比。现实中的世界第一高楼哈利法塔总耗资约 15 亿美元，耗时 6 年便拔地而起；而《GTA 6》的预估开发成本已经突破 20 亿美元，开发周期更是跨越了整整 12 年。
面对这组悬殊的数据，主持人戏称这部跳票多年的游戏巨作简直堪比“两座哈利法塔”的工程量。虽然耗时耗资巨大，但无论从体量还是工业化复杂度来看，它确实已经被大众公认为现代电子游戏工业皇冠上的“哈利法塔”。
| 对比项目 | 迪拜哈利法塔 (Burj Khalifa) | 游戏巨作《GTA 6》 |
|:--- |:--- |:--- |
| **总投入成本** | 约 15 亿美元 | 约 20 亿美元 |
| **建设/开发周期** | 6 年 | 12 年 |
| **产业象征** | 全球建筑工程奇迹 | 游戏工业的体量极限（“两座哈利法塔”） |
---
### [30:34](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1834s) 5 亿美元成立 Intercept：消灭感冒病毒与播客圈的“育儿荒漠”
OpenAI、Anthropic、Stripe 以及比尔·盖茨（Bill Gates）正联合注资 5 亿美元，共同创立一家名为“Intercept”的新型科研机构。Intercept 的终极愿景非常具体且极具雄心：从阻断普通感冒和流感入手，最终彻底消灭人类所有的呼吸道病毒。主持人们对这一举措大加赞赏，认为比起抽象的概念，这种能够带来切肤之痛改善的实体项目才是科技最酷的落地方式。
```mermaid
flowchart LR
accTitle: 父母年患病时间随育儿数量增长曲线
accDescr: 展示随着孩子数量增加，父母每年处于病毒感染状态的时间比例从无孩极低快速拉升至多孩 50% 的曲线变化。
A["0 个孩子<br/>(常规健康状态)"] --> B["1 个孩子<br/>(年患病周数约 25%)"]
B --> C["2 至 4 个孩子<br/>(年患病周数飙升至 50%)"]
```
随后，主持人分享了一组关于家庭育儿与患病率的真实数据图表：随着家里孩子数量的增加，父母每年处于病毒感染期（感冒、流感等）的时间比例呈现出一条极其陡峭的“快速上升曲线”。家里只要有 1 个孩子，父母每年就有大约 25% 的时间处于生病或带毒状态；而一旦孩子增加到 2 个、3 个甚至 4 个，父母每年处于生病状态的时间将直线飙升至近 50%。如果 Intercept 的计划能够成功，普通家庭将能立刻感受到“自己真的不怎么生病了”。
在节目的尾声，主持人借此大肆调侃科技播客圈里那些热衷于自我健康管理与极限优化的博主（podcast optimizers）。这帮人平时喝上两杯酒就挣扎得录不动播客，面对养育孩子这种高强度挑战简直毫无招架之力；好在大多数播客主根本不会选择结婚生子，否则播客圈的“战略播客储备”（strategic podcast reserve）恐怕要面临断崖式下跌。在一片欢笑声中，两位主播提醒听众在 Apple Podcasts 和 Spotify 上打五星好评，订阅官网通讯，并互道周末愉快，结束了本期节目。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=0s) 不管身在哪个时区，冷汗直流、浑身发抖的戒断反应简直疯了 — John Coogan / Jordi Hays
> [00:43](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=43s) 受众喜爱的内容无法变现，变现的广告内容没人要看 — John Coogan
> [10:45](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=645s) 只会出现在 Prada 时尚大片里、现实中绝不会有人戴的怪东西 — Jordi Hays
> [15:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=900s) 未来所有的眼镜最终都将是智能眼镜 — 高通高管
> [16:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=960s) 各个团队像贡品一样被随机挑选出来，直接从服务数亿用户的日常业务中拔出，塞进完全陌生的人工反馈流水线 — Meta 工程师
> [18:01](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1081s) 破罐破摔，反正情况也不能更糟了（F it, I don't think it can get any worse） — John Coogan
> [20:11](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1211s) 要请动全球排名第一的超级网红来搞这种级别的战役并为你首发硬件，得花多少钱？我盲猜差不多要 5000 万美元。 — John Coogan
> [20:11](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1211s) 我们搞不好能看到一部《钢铁侠 4：Snap Spectacles 特别篇》，整部电影的核心剧情就是托尼·斯塔克脱下战甲，只戴着一幅 Snap 智能眼镜，第五次去把灭霸给收拾了 — Jordi Hays
> [23:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1380s) 老兄，根本没人想要数据中心，那些想要这玩意儿的人在我看来简直邪恶至极。没人需要这东西，到头来某一家公司会占有我们所有的隐私信息。 — Theo Von
> [23:00](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1380s) 你随意调动一个自主 AI Agent 跑复杂任务，其消耗的电量和算力，恐怕相当于一个人连续看 10 年 YouTube — John Coogan
> [27:44](https://www.youtube.com/watch?v=Ju5Xx5ej4As&t=1664s) 我们不认为这种政府准入审批程序应该成为长期的默认机制。它把最优秀的工具挡在了急需它们的用户、开发者、企业、网络安全防御人员以及全球合作伙伴之外。我们之所以在短期内妥协配合，是因为我们坚信这是通往更广泛开放的最有力路径。尤其是随着开源前沿模型的不断推进，继续将闭源模型深锁高阁就显得越来越不合理。 — OpenAI
## Mentioned
John Coogan, Jordi Hays, Mark Zuckerberg, Sam Altman, Andrew Ross Sorkin, Emmanuel Macron, MrBeast, Joe Rogan, Ezra Klein, Ross Douthat, Bryan Johnson, Christopher Mims, Kylie Jenner, Robert Downey Jr., Alex Heath, Theo Von, Mike Isaac, Bill Gates, Meta, OpenAI, Anthropic, Apple, Snap, Broadcom, Qualcomm, EssilorLuxottica, Stripe, 智谱 AI, 阿里巴巴, 纽约时报 (The New York Times), CNBC, Polymarket, Kalshi, DraftKings, 美国国家安全局 (NSA), Ray-Ban Meta 智能眼镜, Apple Vision Pro, Snap Spectacles, GPT-5.6 (Sol / Terra / Luna), Jalapeño 芯片, Mythos, GLM-5.2, Arena, Subway Takes, Keep the Meter Running, 侠盗猎车手 6 (GTA 6), Intercept, 哈利法塔 (Burj Khalifa)
## 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=Ju5Xx5ej4As)
Source quotes
[0:00] John Coogan / Jordi Hays: 不管身在哪个时区，冷汗直流、浑身发抖的戒断反应简直疯了
[0:43] John Coogan: 受众喜爱的内容无法变现，变现的广告内容没人要看
[10:45] Jordi Hays: 只会出现在 Prada 时尚大片里、现实中绝不会有人戴的怪东西
[15:00] 高通高管: 未来所有的眼镜最终都将是智能眼镜
[16:00] Meta 工程师: 各个团队像贡品一样被随机挑选出来，直接从服务数亿用户的日常业务中拔出，塞进完全陌生的人工反馈流水线
[18:01] John Coogan: 破罐破摔，反正情况也不能更糟了（F it, I don't think it can get any worse）
[20:11] John Coogan: 要请动全球排名第一的超级网红来搞这种级别的战役并为你首发硬件，得花多少钱？我盲猜差不多要 5000 万美元。
[20:11] Jordi Hays: 我们搞不好能看到一部《钢铁侠 4：Snap Spectacles 特别篇》，整部电影的核心剧情就是托尼·斯塔克脱下战甲，只戴着一幅 Snap 智能眼镜，第五次去把灭霸给收拾了
[23:00] Theo Von: 老兄，根本没人想要数据中心，那些想要这玩意儿的人在我看来简直邪恶至极。没人需要这东西，到头来某一家公司会占有我们所有的隐私信息。
[23:00] John Coogan: 你随意调动一个自主 AI Agent 跑复杂任务，其消耗的电量和算力，恐怕相当于一个人连续看 10 年 YouTube
[27:44] OpenAI: 我们不认为这种政府准入审批程序应该成为长期的默认机制。它把最优秀的工具挡在了急需它们的用户、开发者、企业、网络安全防御人员以及全球合作伙伴之外。我们之所以在短期内妥协配合，是因为我们坚信这是通往更广泛开放的最有力路径。尤其是随着开源前沿模型的不断推进，继续将闭源模型深锁高阁就显得越来越不合理。
Source quotes
[24:01] We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
Original transcript
[0:00] We're back.
[0:00] >> France.
[0:01] >> We miss you guys. We miss you guys a
[0:03] lot.
[0:04] >> Yeah, it was rough.
[0:04] >> It's genuinely brutal. We feel deeply
[0:07] unwell by around 100 p.m. if we haven't
[0:10] gone live. Regardless of the time zone,
[0:12] the
[0:12] >> sweating and shaking is crazy.
[0:14] >> Yeah,
[0:14] >> it is a full mean.
[0:16] >> Yeah. And the withdrawals. I mean, maybe
[0:18] it would get better after a few weeks,
[0:20] but yeah, the the symptoms are
[0:22] >> brutal. Anyway, uh it was a high stakes
[0:24] trip. Of course, Jordy has battled with
[0:27] French President Emanuel Mcronone in the
[0:29] past. And then if you're to be a direct
[0:33] shot at me, France recently banned
[0:36] nicotine pouches fully. Uh very
[0:38] controversial.
[0:39] >> Pretty much all oral nicotine products
[0:41] banned.
[0:41] >> Yeah. So, uh
[0:42] >> as of April,
[0:43] >> can Lions uh it used to be mostly about
[0:47] uh creative advertisements giving awards
[0:49] for creativity and there's a big focus
[0:51] on the creator economy. It's not packed
[0:53] with A-list celebrities, but it feels
[0:54] like everyone has a following of some
[0:56] sort. And that was a big point of
[0:58] discussion. Like there was one level of
[1:00] conversation that was just like creators
[1:01] are the future. Everyone's heard this
[1:02] pitch a million times. But the bigger
[1:04] discussion was something that uh Andrew
[1:06] Ross Sorcin from uh Squawkbox CNBC
[1:09] Dealbook mentioned talking about the
[1:11] trade-offs between independence and
[1:12] consolidation. You've talked about this
[1:14] before, just the idea of a roll up of
[1:16] independent substackers. Where is the
[1:18] right economy of scale in the in the
[1:20] creator economy? Maybe not everyone
[1:22] should be independent. Maybe not
[1:23] everyone should be with a large
[1:26] organization. Uh, and this trade-off
[1:28] sort of interesting to dig into as we
[1:30] see both creators go bigger and produce
[1:33] more expensive shows and then also the
[1:35] legacy media companies are getting
[1:37] better at social media type content. So,
[1:39] we can kind of dig into this. I'm flying
[1:41] back from Can. Uh, something I've been
[1:43] thinking about. I was taken aback by how
[1:46] some of what I thought were most success
[1:49] the most successful influencers say that
[1:51] they need to spend so much money to keep
[1:53] up that they're not making the kind of
[1:55] headline popping numbers we often read
[1:57] about and he asks for thoughts. I think
[1:59] there's I think there's a few things
[2:00] happening. Uh one is the big transition
[2:02] from independent creators pivoting from
[2:04] making content to making shows and you
[2:06] can look at subway takes in this in this
[2:09] way. uh the other show is keep the meter
[2:12] running much more high much more
[2:14] produced much more edited much more
[2:16] costly to uh to produce but you're still
[2:19] faced with some of the awkwardness
[2:21] around how do you actually monetize that
[2:23] there isn't a logical ad break although
[2:25] I think there should be in a threeinut
[2:27] subway takes video a threeinut episode
[2:30] uh I think that it would be fine to have
[2:32] a 5 to 10 second midroll in the middle
[2:34] of that but that has not been done like
[2:37] that floodgate hasn't opened yet so Most
[2:39] creators are doing so either it's either
[2:42] a really polished vertical video about
[2:44] their own content and then they'll do a
[2:46] promoted post that's another 30 second
[2:48] 60-cond multi-minute vertical video in
[2:51] their same style but purely sponsored
[2:53] content and a lot of those videos don't
[2:55] go as viral and so you have this awkward
[2:57] trade-off where the content that's great
[2:59] that people want can't be monetized so
[3:01] you have to do one and one it'd be a
[3:03] little bit tricky like part of the way
[3:05] our show works is that we're just
[3:07] talking about whatever and and then
[3:09] we'll tell you about Cisco critical
[3:10] infrastructure for a the AI era unlock
[3:12] seamless real-time experiences and new
[3:14] value with Cisco. It would be different
[3:16] if we had to do every other every other
[3:19] episode is just purely about Cisco and
[3:21] then we go back to about just the news
[3:24] and talking and so that's that's
[3:26] something that's like starting to evolve
[3:28] and those higher cost structures. I
[3:29] think that they still have more
[3:30] enduring.
[3:31] >> Yeah, Forbes came out with their top
[3:33] creators list and earnings. Mr. Beast
[3:36] was at the top with 300 million. I would
[3:39] imagine that's revenue. That Yeah. Like
[3:41] earnings is is not the right word here.
[3:44] It should just be gross revenue.
[3:46] >> Yeah.
[3:47] >> Uh I I have no idea what what uh Jimmy's
[3:50] margins actually look like on that 300.
[3:52] It would appear he's spending like
[3:54] >> he's obsessed with reinvesting for the
[3:56] >> that's always been a part of his
[3:57] >> as opposed to Joe Rogan who has these
[3:59] big deals and he seems completely
[4:02] content to be in a tiny set and yes
[4:04] people are still breaking through but
[4:06] they're breaking through with new more
[4:07] polished products that of course have
[4:08] higher costs. Simultaneously traditional
[4:10] media companies are starting to dial in
[4:12] content strategies for traditionally
[4:14] creatorled platforms. I think the New
[4:16] York Times has done a fantastic job
[4:17] here. You have the Ezra Klein show and
[4:19] Ross Dowet who are sort of like the left
[4:22] and right perspectives and their shows
[4:25] have become really successful on YouTube
[4:28] a lot in due to the packaging. They they
[4:31] understand how to get a good title and
[4:32] thumbnail together. The editing, the
[4:34] production values are really really high
[4:36] quality. And so those podcasts do very
[4:39] well on YouTube uh in a way that other
[4:42] shows from legacy media companies it's
[4:45] been hard for them to break through
[4:46] traditionally. Uh, but they certainly
[4:48] have and it's interesting that it's
[4:49] recent.
[4:50] >> Uh, the other thing that was
[4:51] interesting, France appeared to put
[4:54] Brian Johnson on
[4:56] >> a pack of cigarettes.
[4:58] >> Pull up this image.
[5:00] >> That does look like Brian Johnson.
[5:02] >> Uh, really out of pocket
[5:04] >> to to take uh to take an asset.
[5:07] >> It works pretty well though. Don't die.
[5:09] Don't smoke. You know, it's his it's his
[5:12] message. I think he would approve this.
[5:14] But I saw I saw this and I had to take a
[5:16] picture because I was like, is it
[5:18] actually uncanny? He has like a lot of
[5:20] pictures of himself down like that, eyes
[5:23] closed.
[5:24] >> Also, this looks like a man being put in
[5:25] a body bag, but at the same time, it
[5:27] looks like just a man who got a getting
[5:29] a beautiful nights of sleep getting
[5:30] tucked in after a nice couple heaters.
[5:33] He ripped a couple darts, had a couple
[5:35] beers, crushed a couple cold ones, and
[5:38] now he's ready to get put to bed by
[5:40] someone. Yeah, this is actually much
[5:42] more tame than some of the other.
[5:44] >> Some of them are gruesome. Like so
[5:46] gruesome showing like the x-rays of
[5:48] cancer patients and like black lungs and
[5:50] stuff like this is pretty mild. If you
[5:52] look like this guy after smoking a whole
[5:54] pack, like you're doing pretty well. The
[5:55] end state of uh of this debate. It's
[5:58] it's interesting. there's a little bit
[5:59] of grass is always greener on the other
[6:00] side going on with independent creators
[6:02] thinking ah maybe it'd be better if I
[6:04] was with a larger organization and then
[6:06] vice versa with a lot of the folks
[6:08] inside organiza big organizations saying
[6:10] ah if I left I could probably make so
[6:12] much more money I don't think it's a
[6:15] one-sizefits-all approach here I think
[6:16] it's more of a sorting process some
[6:18] creators will find elegant business
[6:20] models outside of large organizations
[6:22] like those 80 to 90% IBIDA margins still
[6:25] exist for certain categories certain
[6:27] Obsession, right? It's like a low
[6:30] margin, highly competitive, but it's hit
[6:32] driven. So, you can have one breakout
[6:34] hit and, you know, Yeah.
[6:36] >> have insane margin.
[6:38] >> Yeah. So, I think it's more of this like
[6:39] sorting process. Some creators will find
[6:41] elegant business models outside of
[6:43] traditional organizations. Others will
[6:45] probably wind up leveraging the
[6:47] possibility of going independent into
[6:49] more meaningful contracts within
[6:51] traditional media organizations. But
[6:53] it's in it's increasingly clear that a
[6:54] one-sizefits-all approach uh perhaps
[6:56] with a tagline like independent creators
[6:58] are the future. I don't think that quite
[7:00] fits anymore. Uh to each their own. I
[7:02] think everyone sort of needs to find
[7:04] their own way and uh figure out what's
[7:06] right for them. Smart glasses are
[7:08] inevitable, but what or who are they
[7:10] for? Says Christopher Mims in the Wall
[7:12] Street Journal. Uh this is on the back
[7:14] of the new meta glasses that launched.
[7:18] Chris Christopher Mim says the deluge of
[7:21] smart glasses has begun. Has it? It
[7:25] feels like I don't know if it's a
[7:26] deluge. It feels more like a like a just
[7:28] a a trickle or like a steady stream. But
[7:31] maybe I've been following it too
[7:32] closely. I think that's all of my
[7:34] problems with analyzing this category
[7:35] and we'll get into that. Uh my
[7:37] criticisms are like very odd technical.
[7:39] >> Yeah. And it's and it's and it's so Meta
[7:41] has been selling a lot of glasses. The
[7:43] products are the products
[7:44] >> dozens hundreds. No, they've been
[7:46] selling. I believe they've already I
[7:48] believe they're already in the in the
[7:49] mill in the millions of units of notions
[7:51] of not the
[7:52] >> hundreds of millions of dollars.
[7:53] >> I believe so. Okay.
[7:54] >> Uh Ben,
[7:55] >> I think it's getting I I I agree with
[7:58] you. I think it's getting material. I
[7:59] think it's getting material.
[8:00] >> What I was going to say
[8:01] >> is that it's for products that are
[8:03] functionally like a great pair of
[8:04] glasses with a camera attached, right?
[8:06] These are not like I I don't see a lot
[8:09] of people talking about like, oh, I'm
[8:11] replacing AirPods entirely with these.
[8:14] Even if even if that is one that just
[8:16] and that is also with like having being
[8:19] able to sell through a oneofone ad unit
[8:22] on the greatest distribution platform in
[8:24] history which
[8:25] >> they're going up against stiff
[8:26] competition like the Apple Vision Pro is
[8:28] so good that like you know it's like we
[8:30] job's finished like we have the Apple
[8:33] Vision Pro why
[8:34] >> it was so it was so funny so I had never
[8:36] tried the Apple Vision Pro until this
[8:38] trip and uh I asked John I was like hey
[8:41] can I can I take him for a spin and you
[8:43] were like It's like kind of a hassle to
[8:45] set up. It is.
[8:46] >> And you have to put it in. But John was
[8:48] like, "No, I want Jordy to experience
[8:51] this."
[8:51] >> Yeah.
[8:52] >> And so we go through this whole like
[8:53] hassle of setting it up and I'm like 20
[8:56] seconds into the dinosaur demo and the
[8:59] flight attendant comes up and is like
[9:01] tapping me on the
[9:02] >> Yeah. Like actually getting
[9:04] >> out of virtual reality and it was like
[9:05] the audacity.
[9:07] >> And I think the question was just like
[9:08] are those cool or something like it
[9:10] wasn't it wasn't like sir you need to
[9:12] put on your seat belt or something. It
[9:13] was like
[9:14] >> I was I was fully immersed and it and it
[9:16] completely ruined it for me.
[9:18] >> And what's funny is like the whole pitch
[9:19] for the Apple Vision Pro was like you
[9:20] don't need to take it off to talk to
[9:22] somebody because it will show the eyes.
[9:24] And no, Tyler's on his way back
[9:26] >> because you guys went to France for
[9:27] raising
[9:28] >> raising canes. Yes. Raising can. Raising
[9:32] can is good.
[9:33] >> Yes. Tyler got left at CDG.
[9:36] >> Yeah. No
[9:37] >> shity. But uh
[9:38] >> so I have the answer for the
[9:41] >> timu timu Tyler
[9:46] >> the I have the answer for the meta
[9:47] glasses.
[9:48] >> Okay. Yeah. How how we doing?
[9:49] >> So since 2025 there's been 7 million
[9:52] plus sold across OP and Meta. $2 million
[9:55] was just the Ray-B bands.
[9:57] >> Okay.
[9:57] >> And 2 million units.
[9:59] >> 2 million units. And that's the
[10:01] estimated gross retail sales was 2.1 to
[10:03] three billion estimated.
[10:05] >> Yeah. Yeah, John, you the way you were
[10:06] framing it, it was like, "Yeah, we've
[10:08] sold 7 million."
[10:10] >> And you're like, "$7 million?"
[10:12] >> And it's like, "No, it's seven 7
[10:14] million." Like, it's a lot of units.
[10:15] It's a lot of units.
[10:16] >> Yeah. It's getting there.
[10:17] >> Uh, and if it wasn't Meta, we'd be like,
[10:20] "Wow, this company has like insane
[10:22] product market fit."
[10:23] >> Yeah.
[10:24] >> Selling 7 million units of a new
[10:26] >> If it was Oring or Whoop or something
[10:28] like that, we'd be very impressed. Yeah.
[10:30] >> I I I think you're 100% correct.
[10:32] >> And it And it would be like It could
[10:33] probably be like a standalone. Yeah,
[10:35] >> public company,
[10:36] >> but people would still be slamming it
[10:37] for it for being like way worse than the
[10:39] Apple Vision Pro in terms of watching
[10:41] watching old films. I watched Master and
[10:43] Commander on it. It was great. Anyway,
[10:45] let's go back to Christopher Mims in the
[10:47] Wall Street Journal. He says, "At the
[10:50] low end, there's a new line of Meta
[10:52] Glasses, which start at $299 and play
[10:54] well with vision benefit plans."
[10:56] Interesting. So, you can buy them with
[10:57] insurance. Uh, for another $100, there's
[11:00] a pair aimed at fans of Kylie Jenner.
[11:02] saw these in action and uh they looked
[11:05] great. They they I and I think reviews
[11:08] are very very high quality.
[11:09] >> If if I hadn't seen the launch, I
[11:12] wouldn't have noticed that they were
[11:14] meta glasses at all.
[11:15] >> Yes. I would just like it has this
[11:17] jewel. You can see it. It's that small
[11:20] uh white like glint right there. And I
[11:23] was so confused by this. But this is
[11:25] purely a MI problem because I assume
[11:28] that that was a recording light because
[11:30] or or some sort of display because if
[11:34] you look at the Meta Rayband displays,
[11:35] you'll see that there's slight grooves
[11:37] in the glass for forming the beams to
[11:40] show you the display. And so I was like,
[11:43] "Oh, that's a that's a smart feature,
[11:44] but it's in fact not. It's purely
[11:46] aesthetic." And people who are in the
[11:49] market love this. Uh, I was confused by
[11:52] it, but I think I might be the only
[11:53] person that is confused by it. So, I I
[11:55] don't think it's a problem whatsoever.
[11:58] But Metas is uh $800 and Specs from Snap
[12:02] is $2,200. They are so chunky, they look
[12:05] like something you might see in a Prada
[12:07] ad and never in real life. Silicon
[12:09] Valley has been trying to make smart
[12:11] glasses happen for more than a decade.
[12:13] Remember Google Glass, Facebook, and
[12:14] Instagram, parent Meta Platforms, has
[12:16] captured more than 80% of this market.
[12:18] Not bad. But in 2025, that meant selling
[12:21] just 7 million pairs of glasses. By
[12:23] comparison, the world buys more than 100
[12:25] million smart watches a year and more
[12:28] than a billion.
[12:28] >> What does it what does it say about like
[12:30] the broad uh investor views on smart
[12:34] glasses overall? If Meta has 80% of this
[12:37] like emerging category that's selling 7
[12:40] million units and they get like
[12:41] basically zero credit for it. When I see
[12:43] seven million pairs of glasses compared
[12:45] to 100 million smart watches, I feel
[12:47] like this is perfect trajectory. Like
[12:50] this is great. Like
[12:50] >> do you think that Meta ever makes a free
[12:53] pair of glasses that's ad supported?
[12:56] >> I don't know. I think that an ads
[12:57] business would actually be much more
[12:59] attractive when you when you say, "Oh,
[13:01] they're making a billion dollars in
[13:02] revenue. That's not 100% margin." I
[13:06] think about, well, what is Thread's
[13:08] revenue going to be? And that's going to
[13:10] be much higher margin. like threads is
[13:12] probably more exciting to the core
[13:14] business and to the financial outcomes
[13:16] of
[13:16] >> Yeah. I just think I I think if you
[13:18] >> I understand it's exciting. I like
[13:19] hardware. Like it's cool. It's just like
[13:21] it's it's it's hard to underwrite to
[13:23] like some oh it's moving the needle on
[13:24] cash flow. This is going to pay for
[13:26] another data center.
[13:26] >> I would expect to see just like we saw
[13:28] the ad supported Kindle. I think we'll
[13:30] see the ad supported Meta Glasses.
[13:32] >> Are the ads on the inside for wearer or
[13:34] on the outside?
[13:35] >> For the for the user. So, you're walking
[13:37] on a street and and you walk past a
[13:40] restaurant, it pops up like, you know,
[13:41] get get a free matcha.
[13:43] >> I don't like that. I like the other way.
[13:45] I want to be wearing the glasses and I
[13:46] want it to be barking ads at everyone
[13:48] around me while noiseancelling those ads
[13:51] to me so I don't get I don't have to
[13:53] hear that.
[13:53] >> You become a a loud billboard.
[13:55] >> Exactly. I pay for the no ads version,
[13:58] but you didn't pay and you're sitting
[13:59] next to me. You dress up as a guy on the
[14:01] street and you're pretending to do like
[14:03] what do you do for a living interviews,
[14:05] but then right when you get up instead
[14:06] of saying what do you do for a living,
[14:08] you just activate the ads and you just
[14:10] try to bark like four or five.
[14:12] >> What is your intelligent cloud provider?
[14:13] Is it Railway? Which one is it? Let me
[14:15] tell you about Railway. Railway is the
[14:17] all-in-one intelligent cloud provider.
[14:19] Use your favorite agent to deploy web
[14:20] apps, servers, databases, and more while
[14:22] Railway automatically takes care of
[14:23] scaling, monitoring, and security. That
[14:25] would be a great ad to bark out of your
[14:27] glasses. The idea of a faceworn computer
[14:30] has garnered skepticism, even hostility,
[14:32] which is understandable. When most of us
[14:34] are trying to reduce our time in front
[14:36] of screens, it seems absurd to mount
[14:38] them right in front of our eyeballs. Why
[14:40] should we be okay with everyone we meet
[14:41] pointing internet connected cameras at
[14:43] us? It feels like an assault on what
[14:45] little privacy we have left. This is why
[14:48] you carry a flashbang around. Uh, then
[14:50] there's the built-in AI. Artificial
[14:52] intelligence let these glasses identify
[14:54] objects in our field of view. That is
[14:58] not a very compelling use case for me.
[15:00] >> Hot dog, not hot dog.
[15:02] >> It's just like I'm looking at a
[15:03] building. What is this building? It's a
[15:05] building. Like, what is this tree? What
[15:07] What do you see? It's a tree. I don't
[15:09] know. Do you care if it's a fus tree or
[15:11] not? It's pretty rare. South African
[15:13] mircat. How often are you running into
[15:15] an unidentified mircat? Like, this is
[15:18] not that big of a problem. The bigger
[15:20] one is that you're just able to talk to
[15:21] an AI agent and have it look up things
[15:23] for you. Have it do whatever you want in
[15:25] on your phone in your world on in your
[15:28] agent. But it's very rare that you're
[15:30] seeing things in the real world. Oh my
[15:31] god, what can is this? What is this? I
[15:34] need to know. It's like come on, I know
[15:35] what a Diet Coke is. And yet there will
[15:37] be smart glasses and you and I might
[15:39] even choose to wear them. They're
[15:40] technological in inevitability. A new
[15:43] consumer tech arms race for the last and
[15:45] best real estate on our bodies. I
[15:47] envision that someday all glasses should
[15:50] could be smart glasses, said a man who
[15:52] should know. Uh Zed, who heads the
[15:54] division of Qualcomm, responsible for
[15:55] making the chips that go into the meta
[15:57] the Meta, Snap, and Samsung smart
[15:59] glasses. Uh well, Meta is working on a
[16:02] bunch of other stuff. They're working on
[16:04] artificial intelligence. You may have
[16:06] heard. Zephr says they seem to be pretty
[16:08] focused on coding right now. I'm pretty
[16:09] sure this is the largest coding trading
[16:11] training data generation effort in the
[16:13] world. They have the compute. 30 to 50%
[16:16] of engineers on core teams have been
[16:18] forcibly forcefully reassigned to data
[16:20] labeling. Meta's radical plan to become
[16:22] a leader in AI will require vast amounts
[16:25] of capital and her door was open to
[16:27] discuss innovative financing financing
[16:30] structures and partnerships. Uh they
[16:32] might wind up raising billions more to
[16:35] fund the buildout of MSL and Meta's AI
[16:39] efforts. One engineer told this reporter
[16:43] that the whole situation feels like the
[16:44] movie The Hunger Games when tributes are
[16:47] randomly selected and then removed from
[16:48] their envir environment to something
[16:51] completely different. Except at Meta,
[16:52] many more folks are being affected with
[16:55] between three to five from a 10erson
[16:57] team going from building products used
[16:59] by hundreds of millions to giving human
[17:00] feedback on AI generated GitHub repos
[17:03] and over. So a wider impact than in the
[17:05] Hunger Games with less drastic
[17:07] consequences. Yeah, still air
[17:08] conditioned with free snacks, which they
[17:10] are improving apparently.
[17:12] >> So, Duck is uh we're going to have Meta
[17:15] code.
[17:16] >> Yeah, I mean it seems like they're
[17:17] trying to have a model that they would
[17:19] sell on the API or something.
[17:21] >> I was expecting Meta to spend enough
[17:24] money to get the Meta AI, you know, chat
[17:27] GBT Gemini Claude competitor. Yeah. to
[17:29] the top of the charts and keep it there
[17:31] >> because I expected Zuck to think like,
[17:33] hey, if I get people asking questions
[17:35] all day long, I can just target them
[17:37] much better charts, not
[17:39] >> Yeah. So, I expected them to go super
[17:41] hard on consumer. They don't have
[17:42] background in enterprise. They bought
[17:44] Manis, but
[17:45] >> even Manis was like sort of unclear what
[17:48] they were going to do over time. maybe
[17:50] they would integrate it into ads
[17:52] manager, but I expected them to just say
[17:54] like Zack say like, "Hey, I'm going to
[17:55] spend a lot of money
[17:57] >> and I have billions of users already.
[18:00] I'm going to do whatever it takes to get
[18:01] a consumer chat app to the top of the
[18:04] charts, keep it there, and over time,
[18:07] >> uh, just focus on this one thing." And
[18:09] the Meta AI app sitting at 17 in the app
[18:12] store right now. um you know not not uh
[18:16] not nothing like the market is not is
[18:19] clearly not like sending signals like
[18:21] hey we're really excited for you guys to
[18:23] like pivot into enterprise and like
[18:25] spend hundreds of billions of dollars
[18:27] competing against
[18:28] >> Google
[18:29] >> AWS Microsoft Google openai anthropic uh
[18:33] >> you know all the Chinese labs like this
[18:36] would be an absolutely incredible come
[18:39] from behind story
[18:41] >> but I think it's reasonable that the
[18:43] market is saying like no, they're not
[18:45] pricing it in.
[18:46] >> Certainly. Um well, there's other news.
[18:48] Zuckerberg directed Meta to build a
[18:50] prediction markets app, which is a wild
[18:53] shift. Uh the experimental app
[18:55] internally called Arena would be
[18:56] independent of Facebook and Instagram.
[18:58] It could compete for attention with Poly
[19:00] Market and Kchi, the biggest prediction
[19:02] market says Mike Isaac.
[19:04] >> I think um sentiment is at near rock
[19:09] bottom. Yeah.
[19:10] >> Sentiment on poly on on prediction
[19:12] markets is like I don't know if it's at
[19:14] an all-time low. I think users actually
[19:16] like generally
[19:17] >> Yeah.
[19:18] >> like they they like
[19:19] >> way lower when it was like oh we're
[19:20] going to find out who the next president
[19:22] is and now it's like it's a gambling
[19:24] thing for sports and it's like the vibe
[19:26] has very much shifted.
[19:26] >> Yeah. And I think I think a lot of
[19:28] people have had like have benefited from
[19:31] using prediction markets as a data
[19:32] source. Like
[19:33] >> a lot of people have benefited from
[19:35] sports betting. There's a story about a
[19:37] guy who won $6 million on DraftKings and
[19:39] built a mansion with a bowling alley in
[19:41] it.
[19:41] >> So, what I was saying is like prediction
[19:43] markets are saying, you know, you know,
[19:46] revenue is exploding. I
[19:47] >> think that's a sign that users like the
[19:49] product. I appreciate them as a data
[19:51] source. I always have, but sentiment
[19:53] around them broadly is just like bad.
[19:55] >> Uh, and then sentiment around Meta,
[19:58] >> yeah,
[19:59] >> is at almost an all-time low at least in
[20:02] the capital markets. people are, you
[20:04] know, very very unhappy. And so maybe
[20:07] Zuck is just saying like, "F it. I don't
[20:10] think it can get any worse."
[20:11] >> But feedback to the Cali Jenner Meta
[20:14] Glasses has been way way better. Uh,
[20:16] wow. Silicon Valley finally figured out
[20:18] who controls consumer spending. Uh, this
[20:21] is not finally Meta's been partnering
[20:23] with great celebrities for a really long
[20:26] time. Uh, this does not come as a shock
[20:28] to me, but it is the right person to
[20:30] partner with. So, uh, it's fascinating
[20:32] cuz we were trying to estimate like how
[20:34] much Meta spent on this like partnership
[20:38] campaign.
[20:39] >> Yeah.
[20:39] >> Uh, I had I was putting it at like
[20:42] somewhere around like 50. I have no
[20:44] information on it. I was like 50
[20:46] million. Yeah. Yeah. How much does it
[20:47] cost to get the top influencer?
[20:49] >> Yeah.
[20:50] >> In the entire world to do a campaign
[20:52] like this
[20:53] >> and launch a product with you, you know,
[20:55] maybe something like uh $50 million. Uh
[20:58] then Alex Heath Heathnator who we were
[21:01] with this
[21:02] >> scoop athlete
[21:03] >> scoop athlete comes out and scoops that
[21:06] uh Snap is working on a deal with Robert
[21:09] Downey Jr. to spend $100 million uh
[21:12] potentially for a partnership there
[21:14] which is deeply confusing.
[21:17] >> Okay.
[21:17] >> Um for like you know a lot of reasons
[21:20] like the Kylie Meta campaign makes a lot
[21:23] of sense. Kylie's like one of the like
[21:25] probably the or one of the top
[21:27] influencers on on Instagram, right? So,
[21:29] it's a deep existing partnership.
[21:31] >> You don't know that this couldn't be
[21:32] this isn't deep. It could be a super
[21:34] deep integration. We could be looking at
[21:35] like Iron Man 4 Snap Spectacles and the
[21:38] whole movie could revolve around Iron
[21:41] Man taking off the suit and only using
[21:44] the Snap Spectacles to stop Thanos for
[21:46] the fifth time in a row.
[21:47] >> Previewing today, GPT 5.6 Soul, next
[21:50] generation model
[21:52] >> named after Soul, bro. That's crazy.
[21:54] >> Yeah, I don't know about that one
[21:56] >> because I would assume Soul is the
[21:58] biggest, then Terra is the medium, and
[22:01] then Luna is the smallest in terms of
[22:03] the model size.
[22:04] >> Of course, uh, by US government
[22:07] directive, access is limited to only 20
[22:10] pre-approved companies.
[22:11] >> 20. That's so small.
[22:13] >> Which is brutal because our very own Dan
[22:15] Shipper will not be able to do a vibe
[22:17] check on it. At least
[22:18] >> he's got to get access soon. But a lot
[22:20] of this comes in the back of the distill
[22:23] gate that's happening. Anthropic says
[22:24] Alibaba has distilled clawed many many
[22:27] times and I saw a chart of the amount of
[22:30] tokens that are getting resold uh off
[22:33] the back of that API and it is
[22:35] staggering. It's a true somebody was
[22:37] like they professionalized fraud which
[22:40] seems accurate and so uh having really
[22:42] tight control
[22:44] >> it makes sense in the distilling world.
[22:46] It's still very unfortunate. uh to go
[22:48] back to your other question. So, Soul
[22:50] >> uh is the frontier model. Terra is a
[22:53] balanced model for efficient everyday
[22:55] work. And then Luna is a fast and
[22:57] affordable model for high volume work.
[22:59] >> Okay, so
[23:00] >> cool. There's uh there's more data
[23:01] center backlash. Data centers, how do
[23:03] they work? The von says, "Nobody wants a
[23:05] data center, dude. And the people that
[23:07] want them to me, they seem kind of evil.
[23:09] Nobody wants this. Uh one of these
[23:10] companies going to own all this
[23:11] information." He had some very uh funny
[23:14] takes about
[23:14] >> He said emotional credit score.
[23:16] emotional credit score would be weird.
[23:18] Um, yes, uh, the data center issue needs
[23:21] to be messaged much more clearly and,
[23:24] uh,
[23:24] >> and and I think
[23:26] >> there's such a wide gap between the
[23:27] water and the power. I'm assuming I'm
[23:29] assuming he starts this way but then he
[23:30] says I'm I'm going to be switching to
[23:32] distributing my uh podcasts on
[23:34] >> that's such a bad that's such a bad
[23:36] counter take because truly like the data
[23:38] centers that are required to power the
[23:41] YouTube and podcast stream it's like so
[23:44] so small compared to what AI is doing in
[23:48] terms of data center construction and
[23:50] and and the power requirements and the
[23:52] buildout requirements. I mean like you
[23:54] fire off one AI agent and you're using
[23:56] like you it's equivalent to YouTube for
[23:59] like 10 years probably. Anyway, we
[24:01] should move on to other stuff. Uh open
[24:04] AI limits access to new models, citing
[24:06] government security concerns. Company
[24:07] says White House review of AI releases
[24:10] shouldn't become long-term default. Ban
[24:13] on mythos model remains. Um, Opening I
[24:16] said it's limiting access to its newest
[24:18] artificial intelligence models after
[24:20] discussions with the Trump
[24:21] administration, but warned that the
[24:22] recent trend of the White House
[24:24] restricting industry activity on
[24:25] national security grounds on a case-by-
[24:27] case basis shouldn't become the norm.
[24:29] The maker of Chatty PT. Yeah, I saw a
[24:31] lot of like safety people being like
[24:32] this is the worst of both worlds. It's
[24:35] not some sort of like broadreaching
[24:38] apply, you know, just like the rule
[24:40] exists and then everyone can easily
[24:42] comply. It's like the highest touch
[24:44] government intervention and it makes the
[24:45] government have much more power over AI
[24:48] which is something the safety folks I
[24:50] think at least the ones that I saw
[24:51] talking about this were not a fan of but
[24:53] >> yeah it's also just like you know
[24:55] there's the new it's GLM 5.2 too. Yes,
[24:58] that a lot of people saying like, hey, I
[25:00] could use this as a daily
[25:01] >> and all the lab leaders have said open
[25:03] source is like six months behind. And so
[25:05] now we're like approaching six months
[25:07] past like the last jump in paradigm uh
[25:10] and like massive cyber capabilities. So
[25:12] those cyber capabilities are coming to
[25:15] fruition in the open source world.
[25:16] There's an interesting dynamic between
[25:19] the risk of like bio stuff and cyber
[25:23] stuff. So if you have a model that can
[25:25] hack a system, it can usually patch that
[25:28] system for that hack in far less than
[25:31] six months. Like probably in like an
[25:32] hour, like it can it can see, oh,
[25:34] there's a there's a there's a exploit
[25:36] here and we ship the code, merged it in,
[25:39] and now that exploit is closed. And so
[25:41] the gap between like you create the hack
[25:43] machine 9,000, you it can hack
[25:46] everything, but instead you use it for
[25:48] legal reasons, for economic reasons to
[25:50] patch all the systems. That's great. The
[25:53] the problem that I'm worried about now
[25:56] is if there is a if there is a world
[25:58] where building the counter building the
[26:02] counter like the good version of
[26:04] something takes longer than six months.
[26:07] So the example I'll give is is in bio.
[26:08] So let's say that you have a next
[26:11] generation model. It's really good at
[26:13] finding viruses that if created would
[26:16] wipe out a ton of people. It's also
[26:18] really good at fi at building early
[26:21] warning detection systems and you know
[26:24] counteractive like take this pill and
[26:25] you'll no longer you know be susceptible
[26:27] to that deadly virus that is now
[26:28] possible. Uh that's fine as long as you
[26:31] continue to roll out the biocurity
[26:34] initiatives first. you should be good
[26:36] because by the time the open- source
[26:38] hacker nefarious actor state actors get
[26:41] access to the the bioweapon 9000 you've
[26:45] already patched your society so that you
[26:48] have biorisk security early detection uh
[26:51] all the different medicines that you
[26:53] need to fight back. But if manufacturing
[26:57] all of that and actually deploying the
[26:59] biocurity, the biorisk, the the the like
[27:03] the good version takes a year, but the
[27:06] frontier is only 6 months behind, you
[27:08] could wind up in a situation where
[27:10] manufacturing the virus is much shorter
[27:13] than manufacturing the antivirus. Uh
[27:14] well, let's go back to the OpenAI news.
[27:17] OpenAI followed a similar process with
[27:18] versions of GPT 5.5 like cyber which
[27:22] demonstrated an ability to discover
[27:23] vulnerabilities in software that could
[27:25] be used in cyber attacks. The similar
[27:27] capabilities of anthropics mythos models
[27:29] prompted the White House to increase its
[27:31] oversight of the industry and overhaul
[27:33] its light touch approach to AI. There
[27:35] was of course that news that the NSA uh
[27:38] commented that mythos was able to hack
[27:39] into a bunch of their systems. That was
[27:41] during a red teaming exercise. The
[27:43] unprecedented intervention led Anthropic
[27:44] to halt all access to uh comply with the
[27:47] new rule. Anthropic had previously
[27:49] worked with the administration on a
[27:50] limited roll out of an earlier version
[27:52] of Mythos. Um and uh OpenAI said it
[27:55] hoped to make 5.6 generally available in
[27:57] the coming weeks and that the
[27:58] government's current approval process
[27:59] represents a transition period while
[28:01] President Trump's recent executive order
[28:03] on model oversight is implemented.
[28:05] Quote, "We don't believe this kind of
[28:07] government access process should become
[28:09] the long-term default. It keeps the best
[28:11] tools from users, developers,
[28:12] enterprises, cyber defenders, and global
[28:14] partners who need them. We are taking
[28:16] this short-term step because we believe
[28:18] it is the strong strongest path to
[28:21] broader availability. Especially as the
[28:23] open source frontier uh advances, it
[28:26] feels less and less uh reasonable to
[28:29] keep closed source models locked away.
[28:31] >> Well, why is the open- source model
[28:33] frontier advancing? Is because the
[28:35] models are widely available. If you make
[28:37] these models available through everyday
[28:39] consumer subscriptions, there will be
[28:41] groups that will create networks of, you
[28:44] know, thousands of accounts and distill
[28:45] the models and then you're giving away
[28:47] this like advanced capability. Whether
[28:50] or not that causes like, you know, these
[28:52] sort of like doomsday
[28:54] >> scenarios, you know, not not entirely
[28:56] clear. You think Meta's stock would pop
[28:58] if Zuck could go get a preliminary ban
[29:03] >> on some of their new models? the aura
[29:04] farm of the ban. It does seem to be
[29:06] working. It's too powerful. There's
[29:08] other OpenAI news. I mean, the the chip
[29:10] Jalapeno launched uh designed with
[29:13] Broadcom.
[29:14] >> Jalapenño is purpose-built for LLM
[29:15] workloads powering Chach Codeex API
[29:18] future agentic products. Uh very
[29:20] exciting. Uh
[29:22] >> crazy
[29:23] crazy timeline here.
[29:25] >> What
[29:25] >> uh just how quickly they were able
[29:27] >> Oh yeah. Yeah. A lot of people are
[29:28] saying like this is the first chip
[29:30] designed uh with AI agents in the loop.
[29:33] So you can write the instruction set
[29:34] much quicker. Sort of unclear exactly.
[29:36] >> Got to give a lot of credit to the
[29:38] humans on both sides
[29:39] >> probably involved. But uh it is just
[29:41] cool because for a long time it was like
[29:42] oh it'll take five years to launch a
[29:44] chip and now companies are clearly doing
[29:45] it faster. And so the the the go to
[29:47] market cycle is tightening. Also
[29:49] apparently OpenAI made deals to purchase
[29:51] 40% of the global raw undic wafer output
[29:54] until 2029. Millions of raw DRAM wafers
[29:58] that cannot be used until they're
[30:00] processed. smart move uh going deeper
[30:02] and deeper into the supply chain. Always
[30:04] always interesting moves from the deals
[30:07] guy at the top.
[30:09] >> Speaking of timelines, yes,
[30:10] >> people are comparing GTA 6 with the
[30:13] Burge Khalifa.
[30:15] >> Burj Khalifa was built for $1.5 billion
[30:19] and took six years.
[30:20] >> Wow.
[30:20] >> Apparently GTA 6 is $2 billion and it's
[30:24] taking 12 years.
[30:27] But at the same time, people really have
[30:29] said like it is the Burj Khalifa of
[30:31] video games.
[30:32] >> It's kind of two Burj Khalifas.
[30:34] >> Opening Tropic Stripe and Bill Gates are
[30:36] putting $500 million into funding a new
[30:38] organization called Intercept.
[30:40] Intercept's goal is to prevent the
[30:42] common cold and the flu and eventually
[30:44] eliminate all respiratory viruses
[30:46] completely. Incredibly cool project.
[30:50] Yeah, just like an incredibly cool uh
[30:52] like tangible thing. I saw a stat that I
[30:55] saw this chart of what percentage of
[30:57] weeks are parents sick with something
[31:00] viral generally as a function of how
[31:03] many kids they have. And it's like a
[31:04] fast takeoff curve. So it's like one kid
[31:07] and you're sick like 25% of the year you
[31:10] have like some virus in your system, a
[31:11] cold, a flu, something. And then as you
[31:14] get to like two, three, four kids, it
[31:16] ramps to like 50% of the year you are
[31:19] sick. And people will feel that if
[31:22] they're like, "Yes, I'm getting less
[31:23] sick. It's working."
[31:24] >> Yeah. When when the podcast, you know,
[31:26] optimizers that uh struggle to have, you
[31:29] know, two drinks and continue
[31:30] podcasting, like they are so woefully
[31:33] unprepared for having children.
[31:36] Fortunately, most podcasters, they don't
[31:38] tend to go down that path of of of
[31:41] having of having families. Uh
[31:45] uh but
[31:47] uh
[31:47] >> it's it's a monass. If they did,
[31:50] >> yes,
[31:51] >> we could potentially see the strategic
[31:53] podcast reserve decline meaningfully.
[31:56] >> Anyway, leave us five stars on Apple
[31:58] podcast and Spotify. Sign up for a
[31:59] newsletter at tbpn.com and we will see
[32:01] you.
[32:02] >> Have a wonderful weekend.
[32:05] >> Goodbye.
