Amazon Blocks Meta Shopping Agent to Shield $76 Billion Ad Business
TBPN
Third-party AI shopping agents face structural defeat against platform monopolies because automated headless checkout destroys high-margin sponsored product auctions that generate vastly more revenue than transaction fees.
ByteDance's SeaDance 2.5 has set a new benchmark for generative video by enabling full-body dynamic subject swaps with high physical consistency across challenging camera angles and motion. (0:01)
Meta's stock surged 27% in a single month driven by the breakout reception of its Muse personal AI agent across its messaging and social footprint. (18:49)
Amazon blocked Meta's Muse agent from shopping on Amazon.com, protecting its high-margin $76 billion advertising business against automated headless browser checkouts. (20:35)
Paramount settled with California and state attorneys general, clearing the final regulatory hurdle for its $111 billion acquisition of Warner Bros. Discovery while committing to keep operations in the state. (27:55)
Detailed notes
## Background
Published on YouTube on September 21, 2026, by tech talk channel TBPN, this episode of Diet TBPN delivers a 30-minute condensed broadcast hosted by John Coogan and Jordi Hays. According to the show description, the episode rounds up the day's top tech stories, spotlighting breakthroughs in AI video face swaps, Amazon blocking Meta's new agent Muse, and Paramount settling regulatory challenges to remain in California.
## TL;DR
The hosts break down the rapid leap in generative video fidelity catalyzed by SeaDance 2.5, exploring its impact on viral culture, VFX pipelines, and stunt work while analyzing why ByteDance holds a distribution edge over American frontier labs. The discussion transitions to Meta's breakout consumer agent Muse, examining Amazon's swift decision to ban the bot in order to defend its high-margin $76 billion retail ad business. Finally, they cover Paramount's regulatory settlement with California state attorneys general, which resolved antitrust hurdles around its $111 billion acquisition of Warner Bros. Discovery and guaranteed its production facilities remain in Los Angeles.
## Atmosphere & style
The conversation carries a punchy, irreverent tech-insider tone, balancing high-speed analytical breakdowns with comedic self-experimentation. Coogan and Hays display an effortless, conversational chemistry as they laugh over absurd visual demos—such as deepfaking themselves into Steve Jobs keynotes and classic Jackass stunts—before pivoting seamlessly into sharp macroeconomic analysis of corporate moats and regulatory disputes. The experience feels like sitting in on a fast-paced Silicon Valley debrief that treats complex technological shifts with equal parts genuine awe and dry skepticism.
## Bulletin
- [00:01](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1s) ByteDance's SeaDance 2.5 has set a new benchmark for generative video by enabling full-body dynamic subject swaps with high physical consistency across challenging camera angles and motion.
- [18:49](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1129s) Meta's stock surged 27% in a single month driven by the breakout reception of its Muse personal AI agent across its messaging and social footprint.
- [20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) Amazon blocked Meta's Muse agent from shopping on Amazon.com, protecting its high-margin $76 billion advertising business against automated headless browser checkouts.
- [27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) Paramount settled with California and state attorneys general, clearing the final regulatory hurdle for its $111 billion acquisition of Warner Bros. Discovery while committing to keep operations in the state.
## Key takeaways
- [00:01](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1s) AI video replacement has shifted from early, manual DreamBooth workflows into fully automated agentic pipelines orchestrated by tools like Codex.
- [04:03](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=243s) SeaDance 2.5 achieves realistic full-body and object replacements that match posture, clothing, lighting, and camera physics rather than applying superficial face masks.
- [04:03](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=243s) Dylan Patel's viral AI swap into a new Drake music video demonstrated that unfamiliarity with recent source media significantly amplifies deepfake believability.
- [05:18](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=318s) Keynote recreation tests proved SeaDance 2.5 can simultaneously modify both live onstage figures and background projection screens while preserving identity at extreme viewing angles.
- [09:12](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=552s) Generative video tools are poised to replace tedious post-production visual effects cleanups and hazardous stunt double work in Hollywood productions.
- [10:30](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=630s) Stress-testing SeaDance on erratic Jackass stunts showed robust multi-person identity tracking through extreme motion, water, and chaotic scenes.
- [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) ByteDance maintains a decisive lead in video AI by leveraging proprietary video distribution surfaces like TikTok and CapCut rather than chasing recursive coding benchmarks.
- [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) Unlike its reasoning models, ByteDance keeps SeaDance 2.5 proprietary to defend its vertically integrated consumer and advertising ecosystem.
- [18:49](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1129s) Meta Muse represents Meta's first clear category leadership in consumer personal AI agents, powering a 27% monthly surge in company stock.
- [20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) Amazon blocked Muse for crawling without bot identification and capturing credentials, escalating the conflict over agentic commerce mediation.
- [22:41](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1361s) Amazon's $76 billion ad business disincentivizes third-party autonomous agents because automated checkouts eliminate lucrative sponsored placement auctions.
- [22:41](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1361s) Third-party agentic checkout is encountering an unstable 'Napster moment' defined by continuous front-end anti-scraping patches and platform litigation.
- [26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) Meta prioritized Muse's launch by aggressively allocating internal marketing real estate from Threads and Instagram and rolling out automated negotiation bots on Facebook Marketplace.
- [27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) The Paramount and Warner Bros. Discovery settlement mandates 30 theatrical film releases per year, $1.5 billion in production spending, and an independent newsroom board without forced asset divestitures.
- [29:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1745s) The cleared Warner Bros. merger averts Paramount's threatened flight from California, preserving thousands of entertainment jobs and historic studio lots in Los Angeles.
## Detailed walkthrough
### [00:01](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1s) 94 Days to Christmas & The Evolution of AI Video Swaps
The show kicks off with TBPN's ongoing countdown: exactly 94 days until Christmas. Hitting double digits makes the holiday feel remarkably close, prompting reminders to start hunting for presents and eyeing Christmas trees. After a whirlwind Friday trip flying back and forth to Boston in a single day—a brutal, exhausting slog that left them drained—the hosts dive straight into the tech that derailed their entire weekend: SeaDance 2.5. John describes being completely "oneshotted" by the leap in AI video quality, declaring that the clunky early era of digital face swaps is officially over.
Reflecting on how quickly the space evolved, John looks back to the release of Google's DreamBooth paper. In those early days, pulling off a convincing swap was a manual, agonizing chore. A user had to scrounge through their camera roll for 10 to 20 reference photos under diverse lighting conditions and angles, then spin up a Python notebook on Google Colab. Colab back then was pure luck of the draw: an instance might saddle you with an old GTX 1080 Ti that lacked the VRAM needed to run the fine-tune, or hand you a prized A100 GPU that meant it was time to "lock in." That friction vanished after an indie hobbyist machine learning engineer—who happened to be a guitarist—distilled DreamBooth down to run locally on commodity consumer gaming hardware atop Emad Mostaque’s open-source Stable Diffusion base. Jumping through those technical hoops gave tinkerers access to magic avatar generation roughly three months before viral apps commoditized it into every mainstream chat client.
By contrast, AI video has moved at a much more sluggish pace due to crushing compute demands. With OpenAI having pulled back access to Sora and major American frontier labs largely prioritizing reasoning and text agents over video generation, open experimentation has shifted overseas and into orchestration tools. Rather than downloading specialized software or running local training loops, John automated his entire generation pipeline simply by texting Codex. Running remotely on a dedicated PC, Codex handled the orchestration end-to-end: John didn't even provide source footage or headshots, telling the agent to scour the web for reference photography of Jordi and pair it with iconic clips.
```mermaid
flowchart LR
accTitle: Evolution of Face Swapping and Persona Replacement
accDescr: Diagram comparing the multi-step 2022 DreamBooth workflow with the automated 2026 SeaDance agent workflow.
subgraph Early ["2022: Manual DreamBooth Workflow"]
A["10-20 Camera Roll Photos"] --> B["Google Colab GPU Lottery"]
B --> C["Stable Diffusion Fine-Tune"]
C --> D["Static Avatar Photo"]
end
subgraph Modern ["Modern: SeaDance 2.5 via Agent"]
E["Codex Agent Prompt"] --> F["Web Reference Scraping"]
F --> G["SeaDance 2.5 API / Fal.ai"]
G --> H["Full-Body Dynamic Video Swap"]
end
```
[04:03](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=243s) SeaDance 2.5 and the Viral Drake / Dylan Patel Deepfake
The breakthrough of SeaDance 2.5 lies in full-body and object substitution rather than superficial face-swapping. Instead of pasting a rubbery facial mask over an existing actor, the model replaces the entire subject—matching their posture, clothing, muscle dynamics, and subtle camera physics. This fidelity extends across diverse visual styles, effortlessly blending into everything from crisp modern digital video down to degraded 1990s VHS tape footage with roughly a 90% success rate. Jordi notes that it makes strategic sense for developers and end-users to lean on meta-orchestrators like Codex and Claude; by interfacing seamlessly with third-party models, users bypass the astronomical capital expenditures and GPU cluster commitments needed to train and serve heavy video architectures.
The sheer realism of this new generation went viral over the weekend when SemiAnalysis founder Dylan Patel posted an AI edit of himself substituted into Drake's latest music video wearing a cheetah-print outfit. The underlying choreography gave the clip an uncanny, fluid realism that fooled broad swaths of the internet. Because Drake’s official video had only dropped twenty-four hours earlier, many viewers had not yet seen the original source material, leading them to believe the footage was genuine. The pipeline driving these clips typically runs SeaDance 2.5 wrapped via Higgsfield AI’s harness API or hosted instances on Fal.ai, raising eyebrow-raising questions about how Chinese labs are delivering such high-tier video models despite facing severe Western compute restrictions.
- Dylan Patel’s viral Drake swap relied on the unfamiliarity of a 24-hour-old music video, proving how fresh cultural references amplify deepfake believability.
- Higsfield AI and Fal.ai serve as the primary API harnesses delivering SeaDance 2.5 generation pipelines to developers.
- Viewers in the live stream chat quipped that the swap was an obvious fake solely because Dylan was not wearing his signature headphones.
[05:18](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=318s) Jordi as Steve Jobs, Jensen Huang, and Steve Ballmer
Experimenting with Codex’s autonomous video search, the hosts generated a highlight reel of tech lore featuring Jordi in place of Silicon Valley icons. Early attempts swapping John and Jordi into an existing viral parody of Elon Musk and Sam Altman rapping ran into odd artifacts where facial features occasionally blended back into Musk's likeness, exposing that the model was processing an already layered derivative clip. An interesting psychological quirk of AI video immediately surfaced: while the subject themselves often fixates on minor imperfections and insists "that doesn't look like me," external observers find the likeness uncomfortably spot-on.
The standout result placed Jordi on stage as Steve Jobs pulling the MacBook Air from its manila envelope during the historic 2008 keynote. SeaDance 2.5 simultaneously modified both the live onstage figure of Jobs and the towering projection screen behind him within the identical frame. Even when Jobs turned his head so that only the sliver of his mouth remained visible behind the microphone at the edge of the frame, the model accurately maintained Jordi’s features out-of-frame—a challenge that historically broke traditional face-swappers dependent on center-cropped bounding boxes.
```mermaid
flowchart TD
accTitle: Autonomous Tech Keynote AI Generation
accDescr: Pipeline showing how Codex retrieves video and applies SeaDance full-body replacement across key historical tech moments.
subgraph Prompt ["User Interface"]
P["Prompt to Codex: Find iconic tech clips"]
end
subgraph Retrieval ["Agent Execution"]
P --> R1["Jobs: MacBook Air Keynote"]
P --> R2["Jensen: Kitchen GPU Reveal"]
P --> R3["Ballmer: Developers Chant"]
end
The experiments concluded with two other tech staples: Jensen Huang and Steve Ballmer. Substituting Jordi into Jensen Huang’s pandemic-era kitchen reveal of the massive DGX A100 server—affectionately billed as the world's largest graphics card—yielded comedic gold as Jordi’s frame grunted under the physical weight of the server rack, prompting praise for Jensen's inherent theatrical showmanship recently seen at Salesforce Agentforce. The model subsequently tackled Steve Ballmer’s infamous high-energy "developers, developers, developers" sweat-drenched rallying cry. While the frantic movement caused minor facial drift, the model captured the exact sheen and texture of perspiration on Jordi's face, demonstrating immediate utility for urgent Hollywood production re-shoots and visual effects cleanup.
[09:12](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=552s) Hollywood VFX, Henry Cavill's Mustache, and AI Stunt Swaps
The conversation opens with a classic Hollywood VFX disaster: Henry Cavill's mustache debacle during the filming of *Justice League* and *Mission: Impossible – Fallout*. After wrapping production on *Superman*, Cavill grew a real mustache for his role in *Mission: Impossible*. When *Superman* required emergency reshoots, the *Mission: Impossible* production refused to let him shave it, forcing the *Superman* VFX team to promise they would "fix it in post." Using 2017 VFX technology, artists had to digitally remove the mustache and reconstruct his upper lip and mouth frame by frame. The result looked notoriously uncanny, rough, and unnatural, leading to widespread critique and mockery within the VFX community. While one host wonders aloud why Cavill couldn't simply have shaved and worn a high-end fake mustache for *Mission: Impossible*, the filmmakers clearly insisted on the real thing.
Modern AI video pipelines are poised to take over precisely these tedious, expensive post-production cleanups. While traditional actor performances will likely stay captured in-camera for the foreseeable future, generative replacements are already revolutionizing stunt doubles, hazardous action scenes, and background digital doubles.
[10:30](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=630s) Recreating Jackass Stunts with SeaDance 2.5 and Higgsfield
To stress-test SeaDance 2.5's handling of extreme physical motion and multi-person tracking, the host compiled roughly 20 classic clips from Johnny Knoxville and Steve-O's "generational run" on *Jackass* into a custom AI trailer. The raw clips depict chaotic physical comedy: jumping into a cactus patch, hand-slap bike jousts, jet ski stunts inside a luxury house swimming pool, a fire hose rodeo, the infamous live alligator loose in a house (noting the alligator was likely chilled beforehand to slow it down), and full-speed bull runs. The surreal nature of the original stunts mirrors AI prompts so closely—such as "drive a jet ski in a pool in a luxury house"—that people often mistake the underlying real footage for generative output. The only actual AI modification was seamlessly swapping the stunt performers' faces and outfits. Even through violently erratic movement, the model maintained consistent visual identity across multiple people simultaneously.
The end-to-end technical pipeline relied on several connected tools, though navigating the third-party infrastructure proved frustrating:
```mermaid
flowchart LR
accTitle: AI Video Generation Pipeline
accDescr: Workflow from mobile chat to Codex, reference images, Higgsfield API, and SeaDance 2.5 rendering.
subgraph Prompting ["1. Interface & Code"]
A["ChatGPT Mobile App"] --> B["Codex on Local PC"]
end
subgraph Generation ["2. Asset Prep"]
B --> C["ChatGPT Image Gen\n(Reference Faces)"]
end
subgraph Rendering ["3. Video Synthesis"]
C --> D["Higgsfield API"]
D --> E["SeaDance 2.5\n(Video Face Swap)"]
end
```
The user experience on Higgsfield drew sharp criticism. The service splits its billing into a confusing two-tier structure: API credits versus platform subscription plan credits. After depleting initial API credits, the host upgraded to a multi-hundred-dollar "ultimate plan," only to discover that subscription credits cannot be spent on API calls. Attempting to add credits to the plan simply locked additional funds into the web app balance, stranding money across disconnected accounts. The hosts likened the site's dark patterns and gamification—such as "countdown 10 minutes to claim 75% discount" pop-ups—to a casino platform optimized for "brain rot," noting persistent community complaints about near-impossible subscription cancellation flows.
[14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) Why ByteDance Is Dominating Video Over US Frontier Labs
Despite Western export controls leaving Chinese tech firms with roughly one-fifth the compute capacity of American frontier labs, ByteDance has taken an undisputed lead in generative video fidelity with SeaDance 2.5. This dominance raises key strategic questions: why ByteDance directs finite compute toward video rather than chasing recursive self-improvement (RSI) coding models like Moonshot AI's Kimi or Zhipu's GLM, and why they refuse to open-source the model.
ByteDance possesses a massive, built-in distribution flywheel that pure-play AI labs like Anthropic and OpenAI lack entirely:
- **End-to-End Distribution Surfaces:** ByteDance directly owns ubiquitous consumer video apps—TikTok, Douyin, and CapCut—where generative video feeds directly into recommendation algorithms, advertising products, and creator editing suites. Conversely, OpenAI's Sora lacked an organic distribution surface, requiring them to invent a consumer platform from scratch.
- **Divergent Strategic Focus:** US frontier labs (Google DeepMind, Meta, and OpenAI) are heavily "AGI-pilled," dedicating their clusters to personal agents, knowledge retrieval, and recursive self-improvement. ByteDance instead focuses on monetizing and expanding its dominant short-form video ecosystem.
- **Fair Use and Training Data:** Western claims of copyright infringement against Chinese models overlook the transformative nature of internet video. Massive quantities of Hollywood footage already live permanently on TikTok, YouTube, and Instagram in the form of film critiques, tutorials, and hyperlapse "vibe reels"—all protected under transformative fair use. These vast repositories provide an unmatched native video dataset.
- **The Open-Source Playbook:** While Chinese labs frequently open-source reasoning and coding models to foster an indigenous developer ecosystem that boosts domestic hardware (like Huawei) and challenges the US tech stack, video is kept proprietary. ByteDance follows the classic market dynamic: challengers lacking ecosystem buy-in open-source their weights to build leverage, but incumbents with complete vertical integration and established distribution lock down their frontier models.
[18:49](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1129s) AI Video, Humanoid Robotics, and Meta Muse's Breakout
Looking past the immediate wave of viral memes and comedic face swaps, the hosts briefly consider the long-term speculative value of AI video models: training embodied humanoid robots. China is advancing quickly in humanoids, and if high-fidelity AI video becomes foundational for simulation-to-real ("sim-to-real") translation—training robotic physical policies on simulated video data—it could transform the physical economy. For now, however, that horizon remains further out, leaving consumer entertainment and memes as the immediate battlefield.
The conversation turns to Meta, whose stock surged 27% over the month, driven almost entirely by the explosive reception of its new AI agent, Muse (as highlighted by investor Sheel Mohnot). The hosts note that when a tech giant pours hundreds of billions into capital expenditures with little consumer-facing payoff, delivering a genuinely beloved hit makes Wall Street sit up and realize there is a fighting chance. While competitors have claimed temporary crowns—Google with its Gemini Nano experiments, and Anthropic dominating frontier coding models—Meta has historically lagged in pure knowledge retrieval and vibes-based frontier performance. With Muse, Meta has engineered a sub-frontier base model into a polished, category-defining consumer experience that puts them firmly in first place within personal AI agents.
- Meta's stock gained 27% in a single month following the launch and viral adoption of Muse.
- Unlike previous iterations where Meta's models trailed frontier competitors in coding or benchmark vibes, Muse marks the first time Meta holds a clear category lead in agentic personal assistance.
- The product's real leverage comes from distribution: the ability to seamlessly integrate directly across Meta’s massive footprint in messaging, social apps, and social commerce.
```mermaid
flowchart TD
accTitle: The Battle for Agentic Commerce
accDescr: Diagram illustrating the clash between Meta's external agent Muse, e-commerce platforms, and internal ad networks.
subgraph ClosedEcosystem["Amazon Walled Garden"]
AMZ["Amazon.com Storefront"]
ADS["$76B In-House Ad Engine"]
RUF["Rufus Shopping AI"]
BAN{"Access Policy"}
end
subgraph OpenPartnership["Alternative Rails"]
SHOP["Shopify Network"]
end
M -- "Tries to scrape & auto-checkout" --> BAN
BAN -- "Blocks unauthorized bot credentials" --> AMZ
M -- "Integrated checkout partnership" --> SHOP
AMZ --- ADS
AMZ --- RUF
```
---
[20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) Amazon Bans Muse: The Digital Knife Fight Over Agentic Commerce
While Shopify quickly partnered with Meta to counterposition against Amazon, Amazon dropped an immediate banhammer on Muse. As venture investor Buco Capital noted, anyone celebrating the seamless future of agentic shopping is overlooking the brutal "digital knife fight" brewing over who gets commoditized. Amazon officially barred Muse from shopping on Amazon.com on behalf of users, stating it tried and failed to get Meta to voluntarily exclude the store. According to Amazon, Meta never sought permission, failed to have Muse identify itself while crawling, and appeared to capture and store customer credentials—creating severe security and privacy liabilities. Customers attempting to use Muse for Amazon purchases were greeted with a hard stop: *"Continued access by an unauthorized AI agent violates Amazon's conditions of use, to which our customers have agreed."*
The hosts contrast Amazon's scorched-earth response with Walmart's historically friendlier posture toward third-party chat apps like ChatGPT, noting that market challengers are far more willing to experiment than entrenched monopolies. However, both retail titans have since deployed their own internal, dog-named shopping assistants: Amazon has Rufus, and Walmart has Sparky. The hosts joke about colleagues "dailying" Rufus for everything from trivial product queries to complex psychological dilemmas, marveling at how aggressively Amazon defends its surface area.
- Shopify partnered with Meta's Muse to enable integrated agent checkout, directly counterpositioning against Amazon's closed garden.
- Amazon blocked Muse after discovering the agent failed to declare its bot headers, bypassed access protocols, and collected user login credentials.
- The retailer landscape has split between internal proprietary agents (Amazon's Rufus, Walmart's Sparky) and external personal assistants attempting to mediate consumer purchases.
Several tech investors initially labeled Amazon’s decision to ban Muse as "shortsighted," only to quietly delete their posts once reminded of a massive economic reality: Amazon's advertising business. Over the last 12 months through Q2 2026, Amazon generated $76 billion in high-margin ad revenue—dwarfing the combined revenue of OpenAI and Anthropic. Allowing an autonomous bot like Muse or Perplexity Comet to bypass product listings, compare prices silently, and execute transactions destroys sponsored placement auctions. You cannot serve sponsored product ads to an automated headless browser.
Analyzing the broader macroeconomic reality, the hosts reference analyst Eric Seufert's foundational thesis, *"Agentic Commerce is a Mirage."* Seufert argued the knife fight effectively ended a year prior: Amazon began blocking scrapers in July, sued Perplexity over Comet’s autonomous checkout flows, and recognized that native agents like Alexa Shopping ($12 billion in incremental sales) and Walmart Sparky drive substantial value without ceding control. OpenAI itself experienced this reality: ChatGPT launched sponsored ad units in February 2026 and shortly thereafter shuttered its native instant checkout feature, because affiliate transaction cuts are economically inferior to high-intent ad auctions.
The hosts compare attempts by external agents to scrape and check out on e-commerce sites to the "Napster moment" of consumer AI: an unstable cat-and-mouse game where platforms continually rewrite front-end scripts, patch API bindings, and litigate against autonomous clients. Meanwhile, users themselves react defensively when agents attempt monetization; when autonomous app Instinct recently blasted push notifications urging users to invite friends, consumers revolted, feeling violated that an assistant ostensibly working for them was running self-promotional marketing campaigns.
- Amazon's ad business generated $76 billion in trailing twelve-month revenue, making ad-space preservation non-negotiable over third-party affiliate conversions.
- ChatGPT abandoned instant direct checkout shortly after introducing ad units, proving that advertising economics dramatically out-earn commerce affiliate fees.
- Independent shopping agents face constant technical disruption, scraping countermeasures, and active litigation, mirroring the legal and architectural fragility of early peer-to-peer file-sharing networks.
---
[26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) Pushing Muse Everywhere and "Poisoning" Facebook Marketplace
Meta is leveraging every internal lever to guarantee Muse's dominance across the consumer internet. Market observers noted that Meta's Threads app began sliding in the app store download rankings precisely as the company reallocated internal promotional muscle to Muse. On Instagram, users are now met with full-bleed hero stories marketing the assistant.
This aggressive internal priority was captured in a behind-the-scenes anecdote shared by former Meta product lead Nikita Bier. One of Muse's premier launch capabilities was an automated negotiation feature designed to haggle with sellers on Facebook Marketplace. The product manager responsible for Marketplace protested vigorously, warning leadership that unleashing millions of automated bots to relentlessly lowball human sellers would ruin seller trust and ultimately kill the platform. When the issue was escalated directly to Mark Zuckerberg, Zuckerberg approved the rollout anyway, ruling: *"Okay, we'll poison Marketplace just a little."* The hosts laugh about the dystopian trajectory of peer-to-peer commerce, imagining a chaotic marketplace where autonomous buyer bots relentlessly badger human sellers with lowball offers, counter-bots handle incoming messages, and automated outbound spam messages ping anyone who ever clicked "like" on a Dodge Hellcat enthusiast group.
[27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) Paramount Settles with California and Clears the $111B Warner Bros. Deal
Following a joke about selling items for cash on Facebook Marketplace ("Don't ask questions"), the hosts transition to breaking Hollywood business news: Paramount is officially remaining in California. After an intense back-and-forth between the studio and state officials, Paramount announced that it reached a settlement with the coalition of state Attorneys General who had challenged its massive $111 billion acquisition of Warner Bros. Discovery. The settlement clears what had become the final major regulatory barrier standing in the way of the mega-merger.
The legal battle began in July, when California Attorney General Rob Bonta led a coalition of states in suing to block the transaction. The regulators argued that combining the two storied media empires under a single corporate umbrella would grant one entertainment giant excessive control over major film studios, cable channels, and broader media properties. Under the terms of the settlement resolving the lawsuit, Paramount agreed to several concrete concessions without having to accept broader structural breakups:
- **Editorial Independence:** The combined entity will establish an independent board of journalists designed specifically to safeguard the editorial integrity and newsroom autonomy of both CNN and CBS News.
- **Theatrical Output Commitments:** The company officially committed to releasing at least 30 feature films theatrically per year.
- **Production Spending:** Paramount agreed to invest an additional $1.5 billion into film production over the next five years.
- **Preserved Assets:** Crucially, Paramount avoided forced divestitures; it will not be required to sell off CNN or spin off any of its historic movie lots or studios.
[29:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1745s) The New Hollywood Powerhouse and the Data Center Joke
Once finalized, the combined entity will bring together an unprecedented stable of entertainment brands and IP under one roof, uniting Paramount+ alongside HBO Max, CBS, and CNN, as well as major tentpole film franchises like *Superman*, *Minecraft*, and *Sinners*. Throughout the regulatory process, Paramount defended the necessity of the tie-up by arguing that sheer scale is essential to survive against the dominant tech-backed streaming giants, most notably Netflix.
The deal's resolution also ends high-stakes political leverage between the studio and state leadership. Paramount had repeatedly threatened that if the deal were blocked, it would pack up significant production operations, cut thousands of local jobs, and exit California altogether, even going so far as scouting studio facilities in Nashville. This prompted California Governor Gavin Newsom to sound alarms publicly, warning that a Paramount exodus would strike a devastating blow to the state's status as the global entertainment capital. With the settlement filed, that immediate threat of a major studio flight from Los Angeles has backed off. The hosts sign off with a tongue-in-cheek jab at tech enthusiasts, joking that the settlement is "extremely bad news" for the pro-data-center crowd who were eagerly hoping the historic Paramount lot would get torn down and repurposed into a massive, power- and water-guzzling Los Angeles AI data center. Instead, they wryly conclude, "looks like they're going to be making movies instead," before wrapping the episode and reminding listeners to rate five stars and subscribe at TBPN.com.
## 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/
Chapters:
00:00 SeeDance 2.5 + The Moment AI Video Face Swaps Got “Solved”
04:03 Dylan Patel as Drake + Viral AI Replacements Take Off
09:08 Hollywood VFX, Jackass Stunts + The New AI Video Pipeline
14:05 Why ByteDance Is Winning the AI Video Race
19:15 Meta Muse Takes Off + Amazon Blocks Its Shopping Agent
26:43 Meta Pushes Muse Everywhere + The Facebook Marketplace Bot Problem
27:55 Paramount Stays in California + The $111B Warner Bros. Deal Clears a Hurdle
TBPN reviews how SeaDance 2.5 marks a shift from clunky face swaps to dynamic full-body video substitution, while a commercial collision erupts as Amazon blocks Meta's viral AI agent Muse to protect its advertising revenue, and Paramount settles state antitrust challenges over its $111 billion merger.
Full-Body Substitution Redefines Generative Video (0:01)
Unlike early avatar generation pipelines that relied on manual DreamBooth fine-tuning and Colab GPU lotteries, ByteDance's SeaDance 2.5 performs coherent full-body and object replacements across entire scenes. In demonstrations orchestrated remotely via coding agents, the model modified multi-person choreography, matched lighting across modern digital and degraded VHS footage, and maintained physical likeness even through extreme motion—such as recreating chaotic stunts from Jackass or swapping subjects into historical keynotes featuring Steve Jobs, Jensen Huang, and Steve Ballmer. Observers note the utility of these generative workflows for Hollywood stunt doubles, digital background doubles, and post-production fixes, avoiding past uncanny visual failures like Henry Cavill's digitally removed mustache in Justice League.
ByteDance's Distribution Moat Over Frontier Labs (14:05)
Despite facing severe export controls that leave Chinese firms with roughly one-fifth the compute capacity of American frontier labs, ByteDance has focused finite resources directly on video synthesis rather than recursive self-improvement coding models. The company benefits from native distribution through TikTok, Douyin, and CapCut, where generative tools feed directly into existing recommendation algorithms and creator suites. Because ByteDance holds vertical integration and massive built-in audiences, it keeps SeaDance proprietary, whereas labs lacking market leverage typically open-source model weights to court developer ecosystems.
Amazon Blocks Meta Muse Over Ad Revenue (20:35)
Meta's stock surged 27% in a month following the rollout of Muse, a personal AI agent that gained rapid traction across messaging, social apps, and shopping tasks. While Shopify partnered with Meta for integrated checkout, Amazon promptly banned Muse from crawling Amazon.com. Amazon cited concerns over unauthorized automated scraping, omitted bot identification headers, and captured user credentials, posting explicit error messages to customers attempting agent purchases.
The Fragility of Agentic Commerce (22:41)
Amazon's decision to ban external agents protects a high-margin advertising business that generated $76 billion in the trailing twelve months through Q2 2026. Automated shopping agents bypass sponsored product auctions, rendering digital real estate unmonetizable. Reflecting analysis from Eric Seufert, native assistants like Amazon's Rufus and Alexa Shopping protect transaction surfaces, while independent agents face a cat-and-mouse dynamic of frontend script rewrites and litigation. Internally, Meta pushed Muse aggressively across Instagram and Facebook Marketplace, where Mark Zuckerberg reportedly approved automated seller haggling despite product team warnings about degrading seller trust.
Paramount Settles Antitrust Suit to Keep Studio in California (27:55)
Paramount reached a settlement with a coalition of state Attorneys General led by California Attorney General Rob Bonta, clearing the final major hurdle for its $111 billion acquisition of Warner Bros. Discovery. The agreement unifies intellectual property across HBO Max, Paramount+, CBS, and CNN to compete against streaming giants like Netflix. To resolve regulatory concerns, Paramount committed to releasing at least 30 theatrical films annually, investing an additional $1.5 billion in film production over five years, and forming an independent board of journalists for CNN and CBS News, entirely avoiding forced asset divestitures or a threatened studio relocation out of California.
While AI video generation and personal agents are accelerating consumer adoption, their commercial viability remains tightly constrained by platform walled gardens, proprietary ad networks, and regulatory agreements across traditional entertainment and commerce infrastructure.
## Background
本视频由科技商业播客 TBPN 于 2026 年 9 月 22 日发布在 YouTube 平台，属于其每日三十分钟精炼特辑“Diet TBPN”。节目由 John Coogan 与 Jordi Hays 联合主持，重点聚焦科技与传媒领域的最新大事件，本期核心探讨了 AI 视频全人物替换技术的质变跃升、亚马逊全面封杀 Meta Muse 购物智能体背后的生态角力，以及派拉蒙与华纳兄弟 1110 亿美元合并案和解并留守加州的相关内幕。
## TL;DR
节目首先深入分析了以字节跳动 SeedDance 2.5 为代表的新一代 AI 视频模型，展示其突破简单换脸局限、实现包含身形骨骼与光影一体化的“全人物替换”突破，并剖析了字节凭借 TikTok 分发飞轮在有限算力下领跑 AI 视频的战略逻辑。紧接着，话题转向 Meta 消费级智能体 Muse 引发的行业震荡，详细拆解了亚马逊为死守 760 亿美元站内搜索广告命脉而悍然切断第三方代购通道的“数字白刃战”，并揭露了扎克伯格强推 Marketplace 自动化砍价机器人的内部决策。最后，节目梳理了派拉蒙以设立新闻独立委员会、承诺院线排片及追加 15 亿美元制作费为让步，扫清反垄断障碍并完成与华纳兄弟的世纪并购，成功消除好莱坞核心岗位迁离洛杉矶的危机。
## Atmosphere & style
整期对谈节奏轻快跳跃且充满硅谷极客独有的黑色幽默与敏锐洞察。两位主持人配合默契自然，在分享亲自利用本地 Codex 管线重构《蠢蛋搞怪秀》特技与经典发布会视频时笑场连连、气氛极具感染力；而在转向 Higgsfield 暗黑充值套路吐槽、巨头广告护城河解构及好莱坞政商博弈时，言辞又变得犀利辛辣、直击本质。整场节目既有前沿技术玩具带来的纯粹兴奋感，又兼具高密度、内幕级的冷峻商业研判质感。
## Bulletin
- [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) 字节跳动全新推出的 SeeDance 2.5 实现了逼真的“全人物替换”，在算力储备仅为美国同行约五分之一的情况下，依托 TikTok、抖音与剪映的原生分发飞轮领跑全球 AI 视频赛道。
- [20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) 亚马逊全面切断 Meta 旗下 Muse 智能体的跨平台代购与结算通道，誓死捍卫过去 12 个月高达 760 亿美元的站内竞价广告核心命脉，避免平台沦为无广告曝光的底层供货管道。
- [26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) 为了激进争夺 AI 落地场景，扎克伯格力排内部反对意见，拍板允许 Muse 在 Facebook Marketplace 上线自动化大砍价功能，甘愿冒着二手交易生态被机器人破坏的风险。
- [27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) 派拉蒙与加利福尼亚州检方达成反垄断和解协议，承诺留守加州并追加 15 亿美元电影制作预算，清除了以 1110 亿美元合并华纳兄弟探索公司的最后一个关键法律障碍。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=0s) 从当年门槛极高的 DreamBooth 照片微调演进至 SeeDance 2.5，AI 换脸技术已彻底突破面部贴图局限，升级为能协同重构骨骼体态与物理光影的“全人物替换”。
- [02:40](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=160s) Codex 等上层智能体通过自然语言调度第三方 API 与模型底座，构建起全自动的人物替换管线，令基础软件无需亲自投入昂贵的基础视频模型研发。
- [10:30](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=630s) 最新视频模型在高动态极限测试中表现惊人，即便面对《蠢蛋搞怪秀》级别的剧烈肉搏、翻滚与冲击，仍能保持人物多视角连续追踪而不产生崩溃或跳帧。
- [09:12](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=552s) 好莱坞长期依赖的人工视效修复（如亨利·卡维尔补拍超人戏份时的“胡子门”惨剧），正面临高精度 AI 面部与替身替换技术的全面渗透与降维替代。
- [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) 字节跳动在算力受限下领跑视频模型，核心在于其拥有 TikTok、抖音及剪映等天然分发场景，生成成果能迅速融入内容流与广告生态实现商业闭环。
- [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) 与字节跳动务实聚焦视频不同，OpenAI、Anthropic 等美国主流机构将绝大部分算力倾注于递归自我改进（RSI）、编程智能体及底层 AGI 赛道。
- [16:06](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=966s) TikTok 上海量好莱坞影视二创混剪天然符合版权法中的“转换性使用”豁免，为字节跳动沉淀了难以被版权方阻断的高质量训练数据护城河。
- [17:41](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1061s) 顶尖视频模型属于高度垂直整合的变现工具，作为领跑者的字节跳动选择坚决闭源；这与中国科技界在代码模型上开源构建自主生态的逻辑截然不同。
- [18:49](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1129s) 高保真 AI 视频的远期战略价值不仅在于内容娱乐，更在于通过仿真到现实（Sim-to-Real）的高精度模拟视频为人形机器人具身智能提供规模化训练数据。
- [19:15](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1155s) Meta 凭借全新消费级智能体 Muse 推动股价单月暴涨 27%，成功扭转了外界对其每个季度数百亿美元 AI 资本支出缺乏杀手级产品的质疑。
- [21:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1295s) 亚马逊强硬封禁 Muse 代购，主因在于第三方智能体绕过搜索结果页将摧毁其年营收达 760 亿美元的高利润站内广告业务，巨头绝不甘于沦为被套壳的管道。
- [21:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1295s) 第三方智能体电商在现实商业利益面前是一场幻觉，电商巨头绝不会开放代购接口，各平台内置的原生助手（如亚马逊 Rufus、沃尔玛 Sparky）才是唯一可行的代买形态。
- [25:11](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1511s) AI 智能体面临严峻的商业化与信任困境：用户视其为代表个人利益的私人秘书，一旦在工作流中植入赞助广告或裂变推广，极易引发代理人道德风险反弹。
- [26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) 扎克伯格力排产品团队顾虑，拍板放任 Muse 在 Facebook Marketplace 自动低价扫货砍价，凸显 Meta 为争夺 AI 场景不惜部分牺牲二手交易生态的激进策略。
- [27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) 派拉蒙通过设立独立采编委员会、承诺每年院线上映 30 部电影及追加制作投资，与加州检方达成和解，为 1110 亿美元合并华纳兄弟探索公司扫清反垄断障碍并留守加州。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=0s) 从 DreamBooth 到 SeeDance 2.5：AI 换脸与视频生成的演进史
节目录制时正值秋季，距离圣诞节还有 94 天，进入了“两位数倒计时”，两位主持人开玩笑说大家现在就该开始挑圣诞树和备礼了。谈到刚过去的周末，两人周五刚经历了一趟波士顿当天往返的特种兵行程，虽然极其折腾且略显疲惫，但状态依然拉满。真正让 John 彻底兴奋并感到震撼的，是刚刚发布的全新 AI 视频模型——SeeDance 2.5。在他看来，传统意义上简单的“AI 换脸”（Face Swap）时代已经成为过去，AI 视频生成技术迎来了彻底的质变。
回溯这一技术路径，John 提到了当年谷歌发布 DreamBooth 论文时的场景。在那个阶段，想要把自己的脸逼真地替换到一张照片中，流程极其繁琐：你得在手机相册里翻箱倒柜，找 10 到 20 张自己在不同角度、不同光照条件下的真实照片，然后把它们喂给托管在 Google Colab 上的 Python 脚本。当时在 Colab 上跑模型全凭运气，如同抽奖：有时候运气差，系统分配了一块显存严重不足的 GTX 1080 Ti，根本无法运行；而一旦抽中了 A100，用户就会在社区里高呼“抽到 A100 了，兄弟们赶紧冲！”。随后，一名吉他手出身的业余机器学习爱好者把这篇论文的核心代码进行了蒸馏优化，让 DreamBooth 能够在普通的消费级游戏显卡上跑起来。结合当时 Emad Mostaque 创立的 Stability AI 推出的开源 Stable Diffusion 模型作为底座，整个技术生态迅速引爆。
在那个早期阶段，只要愿意折腾这些复杂门槛，你就能比普通大众提前三个月玩上类似“魔法头像”（Magic Avatar）的效果。然而这项技术很快被彻底平民化，几周之内就变成了各种独立的移动端 App，随后更被整合进各类主流聊天软件。然而与图像生成相比，AI 视频的发展在此前一直慢得多。视频不仅计算量呈指数级上升，而且随着 OpenAI 暂停 Sora 的大范围推进，美国主流 AI 实验室似乎并没有将核心精力全力倾注在视频模型上。
---
### [02:40](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=160s) AI Agent 自动生成流：Codex 与第三方视频模型的协同
为了测试最新的视频生成效果，John 构建了一套全自动的生成管线。让人意想不到的是，用户现在甚至不需要下载专门的软件，也不用手动上传个人素材。John 直接通过远程给运行在本地 PC 上的 Codex 发送自然语言指令：“去全网搜集一些 Jordi 的公开参考照片，然后去挑几段有趣的视频，把 Jordi 替换进去。”Codex 作为底层代理，自主完成了素材检索、参考图提取以及视频调用。
Jordi 对此分析指出，当前像 Codex 和 Claude 这类前沿工具不亲自去卷视频基础模型，其实在商业逻辑上非常合理。视频基础模型的研发和推理成本极其昂贵，需要分配庞大的 GPU 算力集群；而作为上层 Agent 工作区，Codex 可以直接通过 API 接入全球最优秀的垂类视频模型（例如 Fal 或 Higgsfield 封装的接口），既满足了终端用户的实际需求，又无需承担沉重的模型研发和底层算力托管成本。
```mermaid
flowchart LR
accTitle: AI Video Generation Pipeline via Codex
accDescr: 描述用户通过Codex提示词自动抓取参考照与经典视频，并调用SeeDance生成视频的流程。
subgraph Agent["AI Agent 调度 (Codex)"]
A["用户自然语言指令"] --> B["全网检索目标人脸参考照"]
A --> C["自动检索经典视频片段"]
end
subgraph Engine["视频生成与封装"]
B & C --> D["API 封装 harness (Higgsfield / Fal)"]
D --> E["SeeDance 2.5 视频模型"]
end
E --> F["输出全身替换高保真视频"]
```
- **算力分配与模型接口**：SeeDance 2.5 极度消耗算力，目前开发者与普通创作者通常通过 Higgsfield 提供的封装 harness 或在 Fal 等模型托管平台上调用其 API 进行创作。
- **老旧画质兼容**：最新的 AI 视频模型展现出惊人的适应力，不仅能够处理 4K 高清素材，即便面对 20 世纪 90 年代的 VHS 录像带画质，在 90% 的场景下替换处理后也不会产生突兀的伪影。
---
### [04:03](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=243s) Dylan Patel 化身 Drake：全人物替换打破传统“换脸”局限
SeeDance 2.5 之所以引发全网轰动，核心在于它不再只是机械地贴一张脸，而是实现了“全人物替换”（Whole Person Replacement）。模型不仅改变了面部轮廓，还会彻底替换整个人的身形、骨骼姿态，并完美融合原始视频中的光影变化与微小动作细节。
SemiAnalysis 的知名半导体分析师 Dylan Patel 在 X（Twitter）上发布了一段视频，将自己替换成了穿着猎豹纹服饰热舞的说唱歌手 Drake，该视频瞬间骗过了大批网友。由于这支 Drake 的原版 MV 在前一天才刚发布，很多人甚至还没见过原片，第一眼看到 Dylan Patel 的视频时完全信以为真。视频中行云流水的肢体律动完全继承自底层的高水准编舞，但屏幕中的人物已经被完美替换。John 和 Jordi 强调，中国团队在高端算力储备整体承压的大背景下，反而依靠架构创新在 AI 视频赛道实现了关键突破，这一技术反超的战略意味极为深远。
---
### [05:18](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=318s) 恶搞经典名场面：乔布斯、黄仁勋与鲍尔默的“分身”实验
随后，节目展示了 John 利用这套管线自动生成的一系列科技圈经典视频恶搞片段：
1. **说唱双人组（Elon Musk 与 Sam Altman）**：原本是马斯克与奥特曼的 AI 说唱对决视频，模型将主持人替换了进去。虽然效果拔群，但细看时主持人的面容偶尔会与马斯克发生混叠，身高比例也会随之漂移，甚至引来 X 平台直播观众打趣道：“头上都没戴耳机，一眼假！”
2. **史蒂夫·乔布斯发布 MacBook Air**：原片是乔布斯当年从牛皮纸信封中抽出 MacBook Air 的发布会。AI 将大屏幕上的特写乔布斯和小舞台上的远景乔布斯，在同一画幅中同时精准替换成了 Jordi。最令两人惊艳的细节在于，当乔布斯侧身、嘴唇大部分脱离画幅边缘时，模型依然能够准确推断画外结构并实时重构 Jordi 的面部。唯一出戏的是声音仍保留了乔布斯的原声。
3. **黄仁勋的厨房烤箱显卡发布会**：将 Jordi 替换到了黄仁勋当年在厨房从烤箱里捧出巨大 GPU 服务器机架（A100/DGX）的标志性画面中。骨骼体态与肌肉发力的反差制造了巨大的喜剧效果，虽然人物换成了 Jordi，但那股搬运几十公斤沉重机架时的吃力表情和喘息动作被精准还原。
4. **史蒂夫·鲍尔默狂吼“Developers”**：替换了鲍尔默在微软大会上满头大汗、连声咆哮“Developers! Developers! Developers!”的疯狂名场面。模型精准捕捉了鲍尔默面部疯狂流淌的汗水光泽，但由于动作幅度过于剧烈，面部偶尔会出现细微的抖动与跳帧。
两位主持人由此总结，这种高精度的全身替换技术目前正在快速渗透好莱坞的后期制作流程，特别是在高成本补拍（re-shoots）陷入时间紧急的工况下，正成为颠覆性的生产力工具。比如十年前亨利·卡维尔（Henry Cavill）因补拍与档期冲突引发的特效噩梦，如今在 AI 面部与躯体替换技术面前已经可以轻松化解。
### [09:12](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=552s) 亨利·卡维尔的胡子与好莱坞视效的 AI 革命
主持人提到了好莱坞视效史上著名的“胡子门”惨剧：亨利·卡维尔（Henry Cavill）当年拍完《超人》（正义联盟）后杀青离开，转头去拍摄《碟中谍》（Mission: Impossible）。在《碟中谍》剧组里，他的角色需要留一嘴货真价实的小胡子。然而《超人》剧组突然通知镜头没拍全，必须紧急召回他补拍超人戏份。卡维尔表示合同在身绝不能刮胡子，华纳制片方只好大手一挥承诺“后期擦掉”。结果在 2017 年的技术条件下，视效团队不得不一帧一帧地用数字手段抹掉胡子、重建他的上唇与嘴部动态，最终成片效果僵硬怪异，被整个视效（VFX）界疯狂吐槽，过程既极其折磨又造价不菲。
另一位主持人困惑地反问：“粘个假胡子难道不是世界上最容易的事吗？他为什么不干脆刮了，去《碟中谍》剧组贴假胡子？”但《碟中谍》剧组显然坚持要原汁原味的实拍质感。卡维尔嘴唇被数字化抹除的惊悚画面，直接证明了传统视效在面对这类细微面部修复时的笨拙。主持人断言，像这类局部替换与修补工作，AI 工具将以惊人的速度直接渗透并接管。不过对于主流的剧情表演，只要演员能演，片方依然会依靠真人实拍，AI 面部驱动目前最大的用武之地还是替身特技动作戏。
### [10:30](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=630s) 《蠢蛋搞怪秀》特技实测：AI 面部替换的暴力压力测试
为了测试这套 AI 视频工作流在极端动作下的极限，主持人做了一个相当疯狂的实验：将《蠢蛋搞怪秀》（Jackass）里约翰尼·诺克斯维尔（Johnny Knoxville）和史蒂夫-欧（Steve-O）整整一个时代的经典名场面切出大约 20 个片段，剪成了一支全新的预告片。这段混剪在办公室播放时让全员笑疯——从直接跳进仙人掌丛、双手绑在自行车把手上对撞、在豪宅泳池里硬开摩托艇，到经典的室内放鳄鱼（为了减慢鳄鱼攻击速度剧组还提前给它降温）、高压水枪当马骑（Fire Hose Rodeo），以及与狂奔的公牛肉搏。
有趣的是，这类实拍恶搞画面天生就带着一种“AI 生成的荒诞感”——比如“在豪宅泳池里开摩托艇”，听起来简直就像一条随手写出的生图 Prompt。但片中的一切痛苦与危险全都是货真价实的实拍，AI 所做的仅仅是将原片中演员的脸和身体服装完美替换掉。即便面对极其剧烈的撞击、公牛顶撞和高频翻滚，AI 依然能高度一致地稳定追踪并替换多个人物面部，没有发生任何面部崩溃或剧烈闪烁，其表现出的连续性令人瞠目结舌。
```mermaid
flowchart TD
accTitle: AI Video Face-Replacement Pipeline
accDescr: 演示从移动端触发到最终视频生成的四步端到端流程
subgraph Input["控制与提示端"]
A["手机端 ChatGPT App"] --> B["PC 本地运行的 Codex"]
end
subgraph Generation["参考图生成"]
B --> C["ChatGPT 图像模型<br/>(生成角色参考图)"]
end
subgraph Rendering["视频换脸渲染"]
C --> D["Higgsfield API"]
D --> E["SeeDance 2.5 核心渲染"]
end
subgraph Output["最终交付"]
E --> F["高动态一致性恶搞视频"]
end
```
### [12:23](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=743s) 本地工作流与 Higgsfield 的暗黑计费套路
主持人拆解了自己制作这套特技预告片的端到端工作流：首先在 PC 本地运行 Codex，自己则躺着用手机上的 ChatGPT App 与其远程交互；接着让 Codex 自动调用 ChatGPT 的生图能力，批量生成高质量的角色参考图（Reference Images）；最后将这些参考素材传入基于字节跳动模型驱动的 SeedDance 2.5 流程中进行换脸，而底层的实际渲染调用则是通过 Higgsfield 的 API 来执行。
虽然 Higgsfield 作为 API 交付资产的技术体验相当惊艳，但其商业化设计却被主持人痛斥为“极其混乱的流氓暗黑模式（Dark Patterns）”：
- **API 额度与套餐额度完全隔离**：平台将充值拆分为“API 额度（API credits）”和“订阅套餐（Plan）”。当 API 额度耗尽提示升级时，主持人顺手花了几百美元购买了顶配 Ultimate Plan，结果系统无情提示“该套餐内赠送的额度不能用于 API 调用”。
- **套娃充值陷阱**：随后主持人尝试直接向套餐追加点数，结果追加的点数依然被锁在 Plan 账户中，无法划拨给 API，导致大量现金被无意义地锁在不同账户抽屉里。
- **博彩网站化与取消订阅困难**：整个网站界面充满了类似赌博网站的倒计时弹窗（例如“限时10分钟，享下一笔点数75%折扣”），完全是针对冲动消费设计的成瘾漏斗。更恶心的是，社区普遍反馈其订阅极其难以取消，一旦被套牢就几乎成了“终身被动扣费客户”。
### [14:05](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=845s) 算力仅有五分之一，字节跳动凭什么领跑 AI 视频？
面对 SeedDance 2.5 的强势表现，主持人抛出了核心疑问：中国团队在仅拥有美国顶级机构约五分之一有效算力的情况下，为什么能在视频模型赛道上反超领跑？而且习惯了开源一切的中国 AI 圈，为何偏偏对这款前沿模型严格保密？
深入调研后，核心原因浮出水面：
1. **天然的飞轮与分发场景（Distribution Surface）**：与 Anthropic 和 OpenAI 不同，后者空有模型却缺乏原生的视频消费场景（OpenAI 的 Sora 尽管引发轰动，但由于缺少内置分发平台，难以形成自闭环飞轮）。而字节跳动坐拥 TikTok、抖音、剪映（CapCut）、庞大的信息流广告系统与推荐算法，视频 AI 生成的成果能瞬间在其核心业务中变现，不需要额外冷启动用户需求。
2. **战略重心的分道扬镳**：美国的 Google DeepMind、Meta、OpenAI 和 Anthropic 更加沉迷于“AGI 药丸”，将压倒性的算力资源全部砸向 RSI（递归自我改进）、智能体工作流（Agentic Workflows）、知识检索以及编程代码模型等底层 AGI 赛道；而字节跳动则极其实用主义，既然在短视频领域拥有绝对护城河，便集中算力在视频生成领域建立压倒性优势。
- 算力分配的选择差异：面对有限的算力配额，中国头部阵营不仅要向 GLM 或 Kimi 等大语言模型注资，字节更是坚定地将重度算力切给视频生成。
- 视频模型极其消耗算力，哪怕无视好莱坞版权壁垒，单靠算力硬堆也必须有清晰的商业变现通道作为支撑。
### [16:06](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=966s) 训练数据与版权争议：TikTok 上的“合理使用”护城河
关于字节跳动模型为何能丝滑还原好莱坞名星（如输入“汤姆·克鲁斯”就能生成以假乱真的面部与神态），外界常常质疑其版权侵权。但主持人指出，事情并没有那么非黑即白：在模型训练端，“合理使用（Fair Use）”原则赋予了极大的操作空间。
用户每天在 TikTok、YouTube 和 Instagram 上上传无数由好莱坞电影剪辑而成的视频——例如配上快节奏音乐的《壮志凌云》（Top Gun）极速混剪（Vibe Reel）或专业的电影打光分析短片。这些二次创作属于典型的“转换性使用（Transformative Use）”，它们并非作为原片替代品而存在（观众不会因为看了 30 秒的混剪就不去看原片），因而稳稳落在合理使用豁免的范畴内。这意味着海量包含顶级影视画质与明星面容的高质量视频数据合法地沉淀在 TikTok 的服务器上，成为字节跳动取之不尽的独家训练金矿，好莱坞甚至无法对此提出任何版权下架诉求。
### [17:41](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1061s) 为何不开源？领跑者的守门逻辑与生态博弈
针对“中国 AI 偏好开源，为何 SeedDance 2.5 却坚决不开放权重”的悖论，主持人总结了两大底层商业逻辑：
- **领跑者从不主动开源**：科技界永恒的铁律是“一旦你在前沿技术上取得绝对领先，你就不会开源”。开源往往是落后者为了打破垄断、聚拢生态而采取的颠覆性策略（正如早期 Meta 开源 Llama 冲击专有模型）。
- **代码模型与视频模型的国家级战略差异**：在代码生成模型上，中国科技界倾向于开源权重以迅速打造自主可控的本土开发者生态，帮助华为等国产硬件构建软件护城河，降低对美国技术栈的依赖；但视频模型属于完全垂直整合（Vertical Integration）的商业变现工具，字节跳动从底层的算力训练、中间的模型层，一直到最前端的用户分发终端（TikTok/抖音/剪映）全部打通闭环。作为拥有绝对渠道控制力的市场现任巨头，将顶尖视频模型彻底闭源、构筑商业高墙才是最理性的算计。
### [18:49](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1129s) 视频模型的终极远景：人形机器人与具身智能
在谈及近期 AI 视频模型爆发的长远意义时，对话指出其真正的战略价值可能不仅限于内容创作，更在于具身智能（Embodied AI）与人形机器人的训练。当前中国在人形机器人领域的进展相当迅猛，如果 AI 视频生成技术能够在 Sim-to-Real（仿真到真实世界迁移）领域发挥核心作用——通过高保真模拟视频来规模化训练人形机器人——那么这将补全技术演进的关键拼图。不过，这种工业级落地尚处于早期阶段，眼下的市场和公众在接下来的时间里，注意力依然会被海量的模因（meme）和搞笑整蛊视频牢牢占据。
---
### [19:15](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1155s) Meta 凭借 Muse 强势反弹：百亿支出终见成效
Meta 近期的市场表现势如破竹，单月股价大幅飙升了 27%。投资人 Sheel Mohnot 认为这几乎全归功于全新 AI 个人智能体产品 Muse 的推出。长期以来，市场一直质疑 Meta 每个季度狂掷数百亿美元搞 AI 资本支出却缺乏落地成果；而一旦真正交出了一款具备爆发力的现象级消费级产品，投资者的信心便迅速被点燃。大量亲自体验过的投资人都对 Muse 给出了极高的正面评价。
这次突破尤为引人注目的是，Meta 破天荒地在自己此前从未占据过优势的细分子领域中拔得头筹。纵观过往的大模型竞争格局，王座频繁更迭：Google 曾凭借 Nano Banana 短暂领跑图像多模态，Anthropic 则在编程模型领域稳坐钓鱼台，而 Meta 在知识检索与实际体验质感上往往落后于最前沿模型。然而，Muse 却凭借对非前沿基座模型的工程化打磨与交互创新，在 Agent 体验层拿下了第一梯队的位置。更关键的是，Muse 能无缝嵌入 Meta 现有的庞大生态中，无论是电商导购、社交通信还是即时消息流，都有着立竿见影的协同价值。
---
### [20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) 亚马逊重拳封禁 Muse：电商领地的“数字白刃战”
伴随 Shopify 宣布与 Muse 达成战略合作、形成针对传统电商巨头的抗衡姿态，亚马逊迅速祭出了“封禁重锤”（Banhammer）。Bucco Capital 犀利点评称，尽管外界看好 Meta 与 Muse，但许多人忽略了即将爆发的残酷“数字白刃战”——在平台生态中，谁都不甘心被降级为底层商品管道或被外层套壳边缘化。
亚马逊官方宣布，已全面切断 Meta 旗下 Muse 个人 AI Agent 代替用户在 Amazon.com 上进行代购和结算的通道。在此之前，亚马逊曾尝试与 Meta 协商要求其自愿在 Muse 中排除亚马逊平台，但沟通无果。亚马逊指责 Meta 从未获得授权，Muse 在爬取和浏览时不仅隐瞒机器人身份，还会捕获并存储用户的登录凭证，带来了严重的隐私与安全风险。周日晚间开始，试图用 Muse 在亚马逊购物的用户都会收到拦截弹窗，警告“未授权 AI Agent 的持续访问已违反亚马逊用户协议”。
```mermaid
flowchart TD
accTitle: 亚马逊与第三方AI Agent的冲突机制
accDescr: 展示第三方AI Agent代购与亚马逊平台生态及广告保护之间的结构性冲突
subgraph MetaMuse ["Meta Muse 客户端"]
User["用户提出购物指令"] --> Agent["Muse 自动化 Agent"]
Agent -.->|"隐瞒身份爬取 / 代客下单"| AmazonPlatform
end
subgraph AmazonEcosystem ["亚马逊封闭生态 (防护核心)"]
AmazonPlatform["Amazon.com 交易前端"]
AdEngine["高利润广告飞轮 (LTM 760亿美元)"]
Rufus["官方内嵌 Agent (Rufus)"]
AmazonPlatform -->|"威胁广告曝光与数据闭环"| Intercept["弹出违规提示，强制封杀"]
AdEngine -.->|"优先引导站内竞价广告"| Rufus
end
```
---
### [21:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1295s) 智能体电商是幻觉：广告命脉与站内代购的壁垒
针对亚马逊的封禁，讨论指出其核心逻辑在于誓死保护自身的广告命脉，而非单纯的技术冲突。许多投资人最初天真地指责亚马逊封闭短视，但很快就删除了言论，因为他们意识到：第三方 Agent 代买会直接绕过搜索结果页，让亚马逊的站内广告沦为摆设。数据显示，在截至 2026 年第二季度的过去 12 个月里，亚马逊广告业务收入已高达 760 亿美元，体量甚至远远压倒了 OpenAI 和 Anthropic 的总和。
分析师 Eric Seufert 提出的核心论断一针见血：“Agent 电商（Agentic Commerce）本质上是一场幻觉”。真正的 Agent 电商早就在平台内部实现了——亚马逊拥有自研购物助手 Rufus（不少重度用户已经每天用它解决日常一切琐碎），沃尔玛也推出了 Sparky 并斩获了巨大的商业增量。电商巨头绝不可能向第三方独立 Agent 开放代购权限，平台与外部爬虫之间正在上演类似当年 Napster 时代的猫鼠游戏。早在 2026 年 2 月，ChatGPT 上线站内广告后不久便关停了即时结账功能（Instant Checkout），证明佣金分成模式在经济效益上远逊于搜索竞价广告；最终主导商业格局的，依然是坚不可摧的广告变现模式。
- 亚马逊对外部 Agent 早有防备，此前已因自动化购物功能 Comet 而起诉了 Perplexity。
- 亚马逊与 OpenAI 达成的妥协方案是将其广告网络直接接入 ChatGPT，让 ChatGPT 在获取用户意图后倒流卖广告，而不是直接让 AI 完成代付结算。
- 尽管普通消费者苦于 Amazon.com 繁琐沉重的搜索交互（被吐槽为体验非常“复杂恶心”），极其渴望一键式跨平台代购，但巨头绝不会在此退让。
---
### [25:11](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1511s) 代理人困境：AI 智能体能靠卖广告盈利吗？
AI 智能体的变现路径正引发巨大的信任危机。以近期走红的 Instinct 为例，该产品向用户推送了邀请裂变的弹窗推广，瞬间引发大量用户反弹。即便用户享受着高昂算力成本下的免费服务，心理上依然无法接受——在用户心智中，Agent 是代表“我”个人利益的私人秘书，一旦它开始推销广告，就背叛了这种代办雇佣关系。
与此同时，平台方正尝试在智能体的主动工作流推荐中植入赞助内容。例如，如果用户在 ChatGPT 中咨询冲浪，冲浪板蜡制造商完全可以通过赞助让系统弹出“下周冲浪进阶行动指南”的自动化任务建议。但核心挑战在于：广告主愿意为“增量需求”买单，但如果用户本就打算去亚马逊买一袋特定商品，中间商 Agent 根本无法创造新的广告溢价。
---
### [26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) 扎克伯格拍板“放毒”：Facebook Marketplace 的机器人危机
为了全面推动 Muse 的用户增长，Meta 已经开启了全域流量倾斜：Instagram 顶部出现通栏全屏（Full Bleed）的 Muse 推广广告，其增长势头甚至伴随着 Threads 在榜单上的明显下滑。更具争议性的是前 Meta 员工 Nikita Bier 爆料的内幕决策：Muse 在上线时推出了一项重磅功能——支持在 Facebook Marketplace 上为用户自动化砍价抢购。
这项功能引发了内部激烈交锋。Marketplace 的产品经理极力反对，认为放任数以百万计的 AI 机器人自动搜寻并狂发“骨折低光价”（Lowballing）骚扰卖家，迟早会从根本上毁掉二手交易生态。然而，这一冲突最终被递交到扎克伯格（Mark Zuckerberg）面前，扎克伯格为了 AI 战略的激进落地直接定调：“行，那我们就稍微把 Marketplace 毒害一点点吧。” 这种默许极可能引发未来买卖双方纯靠 Bot 互相博弈的荒谬景象，甚至滋生出利用特定兴趣社群精准定向发送垃圾私信的出海欺诈机器人（例如私信推销道奇地狱猫跑车等）。
### [27:55](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1675s) 派拉蒙达成和解留守加州：1110 亿美元合并巨无霸扫清最后障碍
派拉蒙（Paramount）正式确认不会离开加利福尼亚州。在经历了一段漫长的拉锯战之后，正推进与华纳兄弟探索公司（Warner Bros. Discovery）合并的派拉蒙，终于与加州总检察长以及各州检察长联盟达成了和解协议。派拉蒙官方宣布，已经与反对这笔总额高达 1110 亿美元收购案的多州检方达成和解，清除了该交易面临的最后一个主要法律障碍。此前，加利福尼亚州总检察长罗布·邦塔（Rob Bonta）与多州联合联盟在 7 月正式起诉派拉蒙，指控两家巨头的合并将使单一娱乐霸主对主要电影制片厂、有线电视网络及其它媒体资产拥有过高的垄断控制力。
为了解决诉讼，派拉蒙同意接受多项严苛的附加条件。新组建的合并公司将成立一个独立的记者委员会，旨在保护 CNN 和 CBS News 的新闻编辑自主权；同时，公司承诺每年至少院线上映 30 部电影，并在未来 5 年内额外投入 15 亿美元用于电影制作。值得注意的是，派拉蒙顶住了进行更深层结构性拆分的压力，既不需要出售 CNN，也无需剥离旗下的任何主要电影制片厂。一旦交易最终完成，派拉蒙与华纳兄弟将归于同一屋檐下，旗下汇聚 Paramount+、HBO Max、CBS、CNN，以及包括《超人》（Superman）、《我的世界》（Minecraft）和《罪人》（Sinners）在内的顶级影视 IP 与片单。
派拉蒙一直坚称，在面对 Netflix 等实力愈发强劲的流媒体巨头和科技公司时，达到这种体量规模对于生存竞争是绝对必要的。本次和解也为围绕派拉蒙在加州未来的数月政治施压画上了句号。此前派拉蒙曾强硬放话，警告称如果这笔收购案遭到阻止，公司可能会将大量影视制作和数千个工作岗位迁出加利福尼亚州，甚至有报道称其已开始在纳什维尔（Nashville）物色新的摄影棚场地。加州州长加文·纽森（Gavin Newsom）曾公开警示，失去派拉蒙将严重动摇加州作为全球娱乐产业中心的地位。随着诉讼走向和解，好莱坞大型制片厂撤离洛杉矶的重大威胁暂时烟消云散。主播们调侃道，这对那些希望派拉蒙洛杉矶片场被改建为耗水耗电超级 AI 数据中心的人来说，简直是个“坏消息”，看来这块宝地终究还是要继续用来拍电影了。
- **交易核心规模与涉及方**：派拉蒙以 1110 亿美元收购华纳兄弟探索公司（Warner Bros. Discovery），扫清加州总检察长罗布·邦塔牵头的多州反垄断诉讼阻碍。
- **和解让步条件**：
- 设立独立记者委员会，保障 CNN 与 CBS News 的采编独立性。
- 承诺每年至少在院线公开放映 30 部电影。
- 承诺未来 5 年内追加 15 亿美元的电影制作预算。
- 拒绝进行资产拆解：保留 CNN 与全部核心电影制片厂。
- **业务版图整合**：合并涵盖流媒体平台（Paramount+、HBO Max）、广播电视网（CBS、CNN）以及重点影视项目（《超人》、《我的世界》、《罪人》等）。
- **产业博弈与威胁解除**：派拉蒙曾威胁迁往纳什维尔并裁撤数千岗位，加州州长加文·纽森介入呼吁，最终成功将影视制作产能与工作岗位留在洛杉矶。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=0s) 抽到 A100 了，兄弟们赶紧冲！
> [02:40](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=160s) 去全网搜集一些 Jordi 的公开参考照片，然后去挑几段有趣的视频，把 Jordi 替换进去。 — John Coogan
> [05:18](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=318s) 头上都没戴耳机，一眼假！
> [05:18](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=318s) Developers! Developers! Developers! — Steve Ballmer
> [09:12](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=552s) 粘个假胡子难道不是世界上最容易的事吗？他为什么不干脆刮了，去《碟中谍》剧组贴假胡子？ — Jordi Hays
> [12:23](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=743s) 该套餐内赠送的额度不能用于 API 调用 — Higgsfield
> [12:23](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=743s) 限时10分钟，享下一笔点数75%折扣 — Higgsfield
> [17:41](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1061s) 一旦你在前沿技术上取得绝对领先，你就不会开源 — John Coogan
> [20:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1235s) 未授权 AI Agent 的持续访问已违反亚马逊用户协议 — Amazon
> [21:35](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1295s) Agent 电商（Agentic Commerce）本质上是一场幻觉 — Eric Seufert
> [26:43](https://www.youtube.com/watch?v=3IpWmnYz4_c&t=1603s) 行，那我们就稍微把 Marketplace 毒害一点点吧。 — Mark Zuckerberg
## Mentioned
John Coogan, Jordi Hays, SeeDance 2.5, DreamBooth, Stability AI, Stable Diffusion, Emad Mostaque, OpenAI, Sora, Codex, Claude, Fal, Higgsfield, Dylan Patel, Drake, SemiAnalysis, Steve Jobs, 黄仁勋 (Jensen Huang), Steve Ballmer, Henry Cavill, 《超人》 (Superman), 《碟中谍》 (Mission: Impossible), 《蠢蛋搞怪秀》 (Jackass), Johnny Knoxville, Steve-O, ChatGPT, 字节跳动 (ByteDance), TikTok, 抖音, 剪映 (CapCut), Anthropic, Google DeepMind, Meta, Llama, Muse, Sheel Mohnot, Shopify, Amazon, Rufus, Sparky, Walmart, Eric Seufert, Perplexity, Instinct, Facebook Marketplace, Nikita Bier, Mark Zuckerberg, 派拉蒙 (Paramount), 华纳兄弟探索公司 (Warner Bros. Discovery), Rob Bonta, Gavin Newsom, CNN, CBS News, Paramount+, HBO Max, Netflix, 《我的世界》 (Minecraft), 《罪人》 (Sinners)
## Video description
Sign up for TBPN’s daily newsletter at TBPN.com
Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after.
Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella.
TBPN is made possible by:
Ramp - https://ramp.com
Public - https://public.com
Cisco - https://www.cisco.com
Console - https://www.console.com
CrowdStrike - https://www.crowdstrike.com
Figma - https://www.figma.com
MongoDB - https://www.mongodb.com
NYSE - https://www.nyse.com
Railway - https://railway.com
Shopify - https://www.shopify.com/
Follow TBPN:
https://TBPN.com
https://x.com/tbpn
https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231
https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235
https://www.youtube.com/@TBPNLive
Chapters:
00:00 SeeDance 2.5 + The Moment AI Video Face Swaps Got “Solved”
04:03 Dylan Patel as Drake + Viral AI Replacements Take Off
09:08 Hollywood VFX, Jackass Stunts + The New AI Video Pipeline
14:05 Why ByteDance Is Winning the AI Video Race
19:15 Meta Muse Takes Off + Amazon Blocks Its Shopping Agent
26:43 Meta Pushes Muse Everywhere + The Facebook Marketplace Bot Problem
27:55 Paramount Stays in California + The $111B Warner Bros. Deal Clears a Hurdle
[Source](https://www.youtube.com/watch?v=3IpWmnYz4_c)
Source quotes
[0:00] 抽到 A100 了，兄弟们赶紧冲！
[2:40] John Coogan: 去全网搜集一些 Jordi 的公开参考照片，然后去挑几段有趣的视频，把 Jordi 替换进去。
[5:18] 头上都没戴耳机，一眼假！
[5:18] Steve Ballmer: Developers! Developers! Developers!
[9:12] Jordi Hays: 粘个假胡子难道不是世界上最容易的事吗？他为什么不干脆刮了，去《碟中谍》剧组贴假胡子？
[12:23] Higgsfield: 该套餐内赠送的额度不能用于 API 调用
[12:23] Higgsfield: 限时10分钟，享下一笔点数75%折扣
[17:41] John Coogan: 一旦你在前沿技术上取得绝对领先，你就不会开源
[20:35] Amazon: 未授权 AI Agent 的持续访问已违反亚马逊用户协议
[21:35] Eric Seufert: Agent 电商（Agentic Commerce）本质上是一场幻觉
[26:43] Mark Zuckerberg: 行，那我们就稍微把 Marketplace 毒害一点点吧。
Original transcript
[0:01] How many days till Christmas? TBPN's
[0:03] Road to Christmas started weeks ago. We
[0:07] got 94 days until Christmas. Double
[0:10] >> double digits.
[0:11] >> We're so close. Every day one day closer
[0:13] to Christmas. So, get excited. Start
[0:16] shopping for presents. Start picking out
[0:18] that Christmas tree. Uh, how's your
[0:19] weekend?
[0:20] >> Uh, it was couldn't have been better.
[0:22] >> Yeah, it was good. Uh, we were off on
[0:24] Friday. We went to Boston and back
[0:26] basically the same day. Sort of a brutal
[0:28] flight back and forth. a little tired,
[0:29] but terrific.
[0:30] >> I got completely oneshotted by the new
[0:33] AI video model. Seance 2.5.
[0:36] Gone are the days of uh face swaps. I
[0:40] remember when Dream Booth, this paper,
[0:42] came out from Google and it allowed you
[0:45] to very realistically swap your face
[0:48] into a photo. Now, this is something you
[0:49] can just do with a single prompt, but uh
[0:51] today, but back then you had to go and
[0:54] take 10 or 20 different photos of
[0:56] yourself at different angles, different
[0:57] lighting, usually just dug through your
[0:59] camera roll, got them. Then you had to
[1:01] feed that into like a Python script that
[1:03] was either hosted on collab and then the
[1:06] collab thing was weird because it was a
[1:08] luck of the draw what GPU got. So you
[1:11] would like sign up and create an
[1:12] instance and then sometimes it would
[1:14] just be like you got a 1080 Ti and you'd
[1:16] be like I can't do what I need to do
[1:18] because I need a lot of memory to do a
[1:20] particular thing and then sometimes it
[1:21] would be like you got an A100 brother
[1:24] like lock in like it was just random
[1:26] provisioning and people would be like so
[1:28] excited when they got one or the other.
[1:29] Um but there was this musician who was
[1:32] like a guitarist who implemented the
[1:34] paper and then and then uh basically
[1:37] distilled it down and figured out how to
[1:39] run. This is a crazy thing, but this guy
[1:40] who's just like a hobbyist machine
[1:42] learning engineer figured out how to get
[1:44] Dream Booth running on like basically
[1:46] commodity gaming hardware. So then you
[1:49] could go and you could take your face,
[1:51] fine-tune it, and then be able to
[1:53] generate new photos in the style of
[1:55] whatever. Uh, it was all built on that
[1:57] open source one, not Midjourney. The
[1:59] other image generator, Ahmad Mustach was
[2:03] stable.
[2:04] >> Stable diffusion. It was like stable
[2:05] diffusion as the base and then this
[2:07] dream booth thing on top. Uh, and it was
[2:09] super cool, uh, because if you actually
[2:11] went through, if you jump through all
[2:12] those hoops, uh, you basically had
[2:14] access to the magic avatar thing like
[2:17] three months before everyone else and
[2:18] you could like send one to a friend and
[2:20] be like, "Oh, this is cool." And then it
[2:21] was just completely commodified. It
[2:23] became an app and then it just became
[2:24] available in every chat app uh, very
[2:27] shortly after. But the AI video has been
[2:29] moving a lot slower. It's very compute
[2:32] intensive obviously since uh opening I
[2:34] had to shut down Sora and uh it just
[2:36] feels like none of the major American
[2:37] labs are as focused on this. So uh I got
[2:40] completely addicted to this. Had a lot
[2:42] of fun doing it. I'll walk through
[2:43] everything that we did. You no longer
[2:44] need to go get a different app or sign
[2:47] up for anything you like. I was just
[2:49] texting with codeex remotely. I had a PC
[2:52] that was running this and then I would
[2:54] just be like I didn't even say take this
[2:56] video. I would just say like, "Go find
[2:58] some videos and put Jordy in them." And
[3:01] it would just do that. And and I didn't
[3:02] even have to like upload photos of you.
[3:04] I would just say like, "Go find some
[3:05] reference photography for Jordy across
[3:07] the internet." And it would
[3:08] >> And so far, it seems to have made a lot
[3:10] of sense not to focus on video given
[3:13] that Codex and Claude can connect into
[3:16] all these other tools and models and
[3:17] just give you what you want as a user
[3:19] without incurring any of the cost of
[3:22] focusing to make great models or the
[3:24] actual costs of of serving them. and
[3:26] having to allocate
[3:27] >> GPUs. I think people will be having fun.
[3:29] Oh, speaking of Codex, let me tell you
[3:30] about Codeex. Codex is a powerful
[3:32] workspace for getting work done with AI
[3:33] agents. Whether you're writing code,
[3:35] analyzing data, creating content,
[3:37] creating some of the videos I will show
[3:39] you shortly or automating business
[3:40] workflows, Codex helps you move projects
[3:42] forward from start to finish. So, the
[3:43] latest turn of AI AI video models, like
[3:46] they're really, really good. Not just
[3:47] for the high fidelity stuff, which we'll
[3:49] show you some of, but even going back to
[3:50] like 90s VHS footage, it can edit
[3:53] anything and it doesn't look weird 90%
[3:55] of the time. Uh, Cance 2.5, it the the
[3:58] big update here is that it replaces the
[4:00] whole person, the whole object as
[4:02] opposed to just a face swap. It it just
[4:04] hits way harder. You you'll see uh and
[4:06] it matches all the subtle details of the
[4:08] footage, so things integrate really
[4:09] nicely. Dylan Patel posted a video of
[4:11] him as Drake in this uh cheetah print
[4:15] video over the weekend and I think a lot
[4:17] of people fell for this.
[4:19] >> It tricked everyone.
[4:20] >> I I think it tricked a lot of people cuz
[4:22] it's super high fidelity. Let's play
[4:24] semi analysis latest post on X Dylan
[4:27] Patel as Drake. And what's interesting
[4:31] about this is the video model's really
[4:32] good, but also this Drake video came out
[4:36] a day ago. And so it's not it hasn't
[4:38] saturated the internet. Like a lot of
[4:40] people
[4:40] >> a lot of people hadn't seen it.
[4:41] >> This is the first Yeah. I I saw this
[4:43] video before I saw the actual Drake
[4:45] music video. And of course the
[4:46] choreography is is fantastic in the
[4:48] underlying video and that drives the
[4:50] motion. He's using AI. He's using CDance
[4:52] 2.5. Uh Higsfield wrapped Cdance 2.5 in
[4:56] uh in a model that sort of uh a harness
[4:59] that sort of makes this a little bit
[5:00] easier. You can use that via the API.
[5:02] It's also available on fall. Uh, and
[5:05] there's a couple other ways to interface
[5:06] with it. Uh, but it's interesting
[5:08] because it's it's highly comput
[5:10] intensive. China's obviously way behind
[5:11] on compute. So, what's going on with the
[5:13] strategy? But before we get to that, uh,
[5:16] it's really fun. So, we got to play some
[5:18] of the more videos that we made. So, I
[5:19] had Codex finding uh, famous or funny
[5:21] videos to use as bass clips. I would
[5:23] just say I I started with the rap
[5:25] videos. Did a couple of those. I think
[5:26] there's one maybe in the timeline.
[5:28] >> Yeah. Noah in the X chat says, "No
[5:30] headphones, obviously fake."
[5:31] >> Oh, true. Yeah. Yeah, that's a good
[5:33] point. Yeah, this one because I saw this
[5:36] originally with Elon Musk and Sam Alman
[5:38] uh rapping and I think we might just
[5:41] because of swearing.
[5:43] >> But what's interesting here is that it
[5:44] does a pretty good job of getting us,
[5:46] but there's moments when I sort of blend
[5:49] into to Elon Musk, which is funny
[5:51] because it's not like it's originally an
[5:53] Elon video, but clearly the underlying
[5:56] video that was driving the AI model was
[5:59] the Elon and Sam version. And then our
[6:01] heights kind of changed like I look a
[6:02] little bit more like Elon there. But uh
[6:04] also then I was able to just say like
[6:06] okay go make a bunch of tech videos like
[6:10] just find iconic videos from tech and it
[6:13] did it and it found like sweet baby
[6:15] rays. I didn't tell it to go get that
[6:18] video but it just did it which was very
[6:20] cool. Some of them came out really well.
[6:21] Some of them didn't do too well. Uh the
[6:23] one that I liked uh I think the most
[6:26] >> and it's a funny dynamic that we noticed
[6:27] where when you're looking at an AI video
[6:29] of yourself, you're thinking that
[6:30] doesn't look like me and then when other
[6:32] people look watch it, they're like,
[6:33] "Wow, that looks exactly like Jordy."
[6:35] >> Yeah. So, let's watch Jordy introduce
[6:37] the MacBook Air as Steve Jobs. I don't
[6:40] think actually Apple kind of bans
[6:43] everything. So, we'll see. We'll see if
[6:44] this one gets flagged, but we might be
[6:46] we might be cooked on this on this
[6:48] particular one. But, let's play. I said
[6:49] video is solved.
[6:50] >> Roll the dice. the post. Yeah, let's
[6:52] roll the dice on this one. Boom. So, it
[6:55] didn't change the voice. So, you still
[6:58] hear Steve Jobs's voice, which is also
[7:00] funny about the Dylan Patel thing
[7:01] because it's Drake's voice. So, it's
[7:02] like, what did they think happened? He's
[7:04] like lip-syncing to it or something.
[7:06] That didn't make any sense. But what's
[7:07] really cool about this is that it
[7:08] changes Jordy, the small Jordy and the
[7:11] big Jordy. It changes both of them in
[7:14] the same frame. And then even there, you
[7:16] only see a little bit of Steve Jobs
[7:18] mouth. and it was able to change that to
[7:20] Jordy so it can represent what's going
[7:22] on outside of the frame and historically
[7:25] that's been very difficult for AI
[7:27] models. They want the face in the center
[7:29] of the frame. They want to be able to
[7:30] crop it and think reasonly reasonably
[7:32] about it. But it did a good job there.
[7:34] Uh Jordi Jensen taking the uh GPUs out
[7:39] >> mass for this one.
[7:40] >> Yeah, it's very funny when you when when
[7:42] when you use a different person with a
[7:44] different stature, different
[7:46] musculature. Very funny clip. iconic
[7:48] when Jensen announced uh was this that
[7:51] H100 announcement, the A100
[7:53] announcement. It was pretty early, but
[7:55] uh
[7:55] >> for a while,
[7:58] >> this is still such a funny gag that
[7:59] Jensen actually did this.
[8:01] >> I love it.
[8:06] >> He's really grunting. I mean, that's a
[8:08] heavy rack.
[8:09] >> Got here is the world's largest graphics
[8:11] card.
[8:12] >> Such a funny bit that he just did this
[8:14] in a kitchen. It's so good.
[8:16] >> Something I got to show you. I got
[8:17] something I got to show you. Jensen's
[8:20] such a such a master performer. Uh we
[8:23] saw that on stage at uh at Salesforce
[8:26] Agent Force uh last last week. The key
[8:30] tonet,
[8:31] the key to industry transformation, the
[8:35] key to success is developers,
[8:38] developers, developers, developers,
[8:41] developers, developers, developers,
[8:43] developers, developers, developers,
[8:45] developers, developers, developers,
[8:47] developers.
[8:49] >> It really got like the level of sweat on
[8:51] your face pretty correct, but the face
[8:54] is jumping around. Uh, not not quite not
[8:57] quite dialed. I think you'll have to
[8:59] perform that one IRL. Uh some
[9:01] productions will use this where it helps
[9:03] in urgent scenarios. I think some of the
[9:05] stuff is already actually being used in
[9:06] Hollywood specifically for re-shoots.
[9:08] About 10 years ago, Henry Cavill was in
[9:12] was uh was in Superman and he left
[9:15] Superman. They wrapped production. He
[9:17] went to film Mission Impossible. In
[9:19] Mission Impossible, he has a mustache.
[9:21] So, he grew out a real mustache, was
[9:23] filming Mission Impossible, and they
[9:24] were like, "We didn't get the shot for
[9:26] Superman. you got to come back and film
[9:29] another scene as Superman. And he was
[9:31] like, I can't shave my mustache. And
[9:33] they were like, okay, we'll fix it in
[9:34] post. And uh in the newsletter, there's
[9:37] a photo here that we can pull up of how
[9:39] it came away because they had to remove
[9:41] and reconstruct his lip and his mouth
[9:44] using 2017 technology. And it was not
[9:47] good. It was very, very rough. Probably
[9:50] very tedious, very expensive.
[9:51] >> Isn't a mustache like the easiest thing
[9:53] to apply?
[9:55] >> Oh. Oh, like fake. like he should have
[9:57] just shaved it and then put on a fake
[9:58] mustache. Yeah,
[9:59] >> I don't know. I think uh Mission
[10:00] Impossible, they wanted the real thing.
[10:02] But the Superman on the right is Henry
[10:04] Caval with a mustache that has been
[10:06] removed in post and it was uh yeah,
[10:08] widely widely critiqued by the VFX
[10:10] community. So, you can imagine for stuff
[10:12] like that, uh this these tools will
[10:14] creep in very quickly. For most things,
[10:17] it's going to be okay. Well, if you
[10:19] needed someone to drive the video, why
[10:21] don't you just have the actor do that?
[10:24] Maybe for stunts and stuff you'd do it,
[10:26] but for most performances I think you'll
[10:27] still be doing it the old way, at least
[10:29] for the time being.
[10:30] >> What about some of the other more
[10:32] actionoriented?
[10:33] >> Yes. So, this was a fun one. I had to go
[10:36] and pull Johnny Knoxville and Steo clips
[10:38] throughout the throughout their film
[10:41] making and TV
[10:42] >> generational run.
[10:43] >> Their Yeah. Their entire run, make about
[10:46] 20 of them and then put it together into
[10:48] a trailer that we can watch now. And I
[10:51] really enjoyed this. We got to do that
[10:55] in the office. I think Tyler was
[10:57] thinking about getting that.
[11:07] This one, this one didn't seem that bad
[11:09] to me, but Nick said this was
[11:10] terrifying. Getting in a cactus. I don't
[11:12] know. I've never really been poked by a
[11:13] cactus. The hands on the bikes. I don't
[11:16] know how the guy on the left wins. Of
[11:20] course, a lot of Steo in that one. A lot
[11:21] of Steo here, too.
[11:25] But I think these work so well because a
[11:27] lot of the underlying footage looks like
[11:29] AI. It look like this.
[11:30] >> The prompts.
[11:31] >> This looks like an AI prompt. Like like
[11:33] drive a a jet ski in a pool in a luxury
[11:38] house, you know? This looks like
[11:39] >> This is real though. This is real. We
[11:40] actually shot this.
[11:41] >> Yeah, this is real. No, the only thing
[11:43] that's different about this is the
[11:44] suits. We replaced the characters. Um,
[11:47] but the underlying footage is all real.
[11:49] uh from Johnny Knoxville and crew the
[11:52] fire hose rodeo. This one that looks fun
[11:57] just
[11:59] dying laughing.
[12:02] >> The old alligator in the house
[12:03] >> like that. The old alligator in the
[12:05] house trick. I think they chilled that
[12:07] guy down so he was real slow
[12:12] >> and then the the bull run. Um, but a
[12:15] very impressive demo because it's able
[12:16] to replace multiple people and and do it
[12:19] very consistently, even through like
[12:20] some really aggressive motion. So, I
[12:22] thought that was pretty cool. Uh so the
[12:24] the whole pipeline here was I was
[12:26] running codecs on my PC interacting with
[12:29] that from my phone the chatbtd app and
[12:31] then I had codeex go and generate
[12:34] reference images with just chatbt images
[12:37] because it's pretty good at that and
[12:38] then use those to swap through seed
[12:41] dance which was uh which was actually
[12:43] rendered via Higsfield. Higsfield uh
[12:46] amazing in terms of just delivering the
[12:49] assets like as an API. Extremely
[12:52] confusing as a product because they
[12:53] basically have two ways to give them
[12:55] money. You can load up API credits and
[12:58] then you can like buy a plan. And so at
[13:00] one point I ran out of API credits and
[13:02] it was like time to upgrade to the plan
[13:03] and I was like okay cool. I'm having a
[13:05] good time. I'll do more of this. Click
[13:06] the button. I'm like, "Okay, yeah, you
[13:08] loaded up a couple hundred bucks in here
[13:09] with like the ultimate plan or
[13:11] whatever." And then it's like, "Those
[13:13] credits are not available in the API."
[13:15] And I was like, "Great, thanks." And
[13:18] then and then it was like, "Okay, time
[13:20] to add some credits to your plan." I was
[13:22] like, "Okay, this is what I need to do.
[13:23] I add credits to the plan." It's like,
[13:25] "Those credits are on your plan. Those
[13:26] aren't API credits. Those are plan
[13:28] credits." And so I have like so much
[13:31] random money stuck in various pieces of
[13:34] >> Well, and I think that's been a massive
[13:35] complaint that Higsfield has faced a
[13:38] bunch of I think various allegations
[13:41] over the last year around just like
[13:42] >> I mean the website looks like a gambling
[13:44] company. It'll be like countdown 10
[13:47] minutes to claim 75% discount on your
[13:49] next credit. It's like it's totally
[13:51] optimized for like brain rat as you go
[13:53] through the flow being like, "Oh, buy
[13:55] you got to buy a manual."
[13:56] >> I I had seen stuff of like not being
[13:58] able to actually cancel your plan. So
[13:59] now that you have a plan that you can't
[14:00] use for anything, it's over for you.
[14:02] >> It's over for me.
[14:03] >> It's over. You're a customer for life.
[14:05] >> My questions about Cance 2.5. I was
[14:07] like, how is China winning? They have
[14:10] 1/5if the comput.
[14:13] >> Is it that simple? And then also, why
[14:16] aren't they open sourcing it? They love
[14:17] open sourcing everything. and they're
[14:18] like, "We love open sourcing except when
[14:20] we're on the frontier."
[14:21] >> Yeah. Good.
[14:22] >> And there's a weird dynamic there. So,
[14:24] first off, did some digging, got some
[14:26] answers. Let's dig into it. Um, at a
[14:29] high level, intellectual property rules
[14:31] are somewhat flexible over there. That's
[14:33] true. But the models are insanely
[14:36] computensive. Like, it is even if you're
[14:39] like, I don't care about Hollywood IP.
[14:41] Come sue me in China. Good luck. There's
[14:44] still a finite amount of compute. And so
[14:47] why aren't you on the RSI trajectory?
[14:49] Why aren't you working on coding models?
[14:51] Like why don't you just take that
[14:53] compute and put it on the next GLM run
[14:55] or the next Kimmy run, right? Um well,
[14:58] there are reasons for that. First is
[14:59] that bite dance has a more natural
[15:02] flywheel here than the leaders in
[15:04] America. So Anthropic and Open AI do not
[15:07] have good surfaces to deliver video, AI
[15:10] or not video. like OpenAI tried with
[15:12] Sora, got a decent amount of people to
[15:14] use it, but it's but it was not the
[15:16] obvious flywheel that Bite Dance has.
[15:18] And so over in China, Bite Dance is a is
[15:21] a big company leading and they have Tik
[15:24] Tok, Duian, Cap Cut, ads, e-commerce,
[15:26] recommendation systems, tons of
[15:28] immediate benefits to the core business
[15:30] without needing to pull together a new
[15:32] demand surface for video. This sort of
[15:34] exists in America with Meta and Google,
[15:36] YouTube, right? But AI slop is deeply
[15:39] unpopular in America. Although I debate
[15:42] whether or not this is this is slop this
[15:44] is this is sloppier than some of the
[15:46] videos we showed you. Um but both
[15:48] companies DeepMind and uh and and MSL
[15:52] they are on the RSI trajectory. They're
[15:55] focusing on agentic workflows personal
[15:58] agents knowledge retrieval coding uh
[16:02] recursive self-improvement AGI. They are
[16:04] it feels like they are more AGI pill
[16:06] than bite dance which is like look we
[16:08] got a good thing going with Duian and
[16:10] Tik Tok we got a lot of video space
[16:12] let's dominate there because we have an
[16:14] uh because we have an opportunity and
[16:16] also like the the AI the IP infringement
[16:18] thing is not actually dead to rights
[16:21] they haven't necessarily been caught in
[16:22] any major way it's just that the when
[16:25] the model comes out people are like Tom
[16:26] Cruz and it's like perfectly Tom Cruz
[16:29] and so it's like what's going on here so
[16:30] some of that could be the the outputs
[16:33] and the harness that actually like
[16:35] reviews outputs. There is an argument on
[16:38] fair use for training to include
[16:40] everything especially when you think
[16:41] about Tik Tok. So people go to Tik Tok
[16:44] and they upload clips from movies. Some
[16:46] of those are fair use, some of those are
[16:48] transformational. Like you can actually
[16:50] make a vibe reel from from Top Gun that
[16:53] is transformational and passes the fair
[16:55] use test which is that it's not a
[16:57] replacement. So you're not like why
[16:58] would I watch Top Gun when I could watch
[17:00] 30 seconds of Top Gun edited like like a
[17:04] crazy hyperlapse edit. Um so it's not
[17:07] it's not competitive in that way. It is
[17:08] transformative and it provides some
[17:10] commentary layer on top. So that is
[17:13] actually okay. You can use clips from
[17:15] movies if you're comment if you're
[17:17] commenting on the lighting and you're
[17:18] trying to explain the lighting. All that
[17:20] means is that some of that data goes
[17:22] onto TikTok and lives there forever and
[17:24] Hollywood can't say anything about that.
[17:25] Same thing with YouTube. Same thing with
[17:27] Instagram. There are plenty of uh film
[17:29] criticism channels that post barely long
[17:32] clips from movies completely acceptably
[17:34] and they don't even need to share
[17:35] revenue because it is transformational.
[17:37] It is open uh it is fair use. So they
[17:39] have a lot of data uh to train all this.
[17:41] Um but they're not open sourcing.
[17:44] They're not doing open weights. They're
[17:45] certainly not open sourcing the the
[17:47] training data set. Uh and it's a little
[17:49] bit of a narrative violation from the
[17:51] default Chinese AI strategy. Uh there
[17:53] are a few reasons for this. First, no
[17:55] one wants to open source when they're in
[17:56] the lead. We've just seen this time and
[17:58] time again. Uh, and second, with coding
[18:01] models, open weights are a play to build
[18:04] a Chinese tech stack. And so, you you
[18:07] open source the the the coding model,
[18:10] build the developer ecosystem, and then
[18:12] maybe that helps Huawei stay
[18:15] competitive. Maybe that, you know,
[18:17] reduces the amount of dominance of the
[18:19] American technology stack. And so
[18:21] there's there's a little bit of like
[18:23] public incentive to build up uh the
[18:26] indigenous supply chain that comes from
[18:28] that less so in video where they are
[18:31] vertically integrated and have
[18:32] everything down to the distribution
[18:33] surface. So uh in general if you are a
[18:37] challenger that needs an ecosystem you
[18:39] want to open up. We saw this with the
[18:41] early meta stuff. Uh and and then uh if
[18:44] you're an incumbent and you have
[18:45] distribution you want to close down. So
[18:48] uh there's some interesting knock-on
[18:49] effects mainly around a speculative
[18:51] future value of video models and
[18:53] training humanoid robots like China also
[18:56] seems to be doing really well in
[18:57] humanoids. If AI video becomes really
[18:59] important in simtore translation and
[19:02] training these uh AI training humanoid
[19:05] robots with simulated video then this
[19:07] could be like part of the bigger
[19:09] picture. But I think that's further out.
[19:11] I think for the next month we're just
[19:12] going to be talking about memes and
[19:13] funny videos. Anyway,
[19:15] >> that's right. Meta on a tear. Meta's up
[19:18] 27% this month according to Shiel Monot
[19:22] almost entirely thanks to Muse. Uh this
[19:25] makes sense, right?
[19:26] >> Is that almost entire to Muse?
[19:29] >> I mean the semi analysis
[19:30] >> basically when you have a company that's
[19:32] spending hundreds of billions of dollars
[19:33] on something without much to show for it
[19:35] and then you have something to show for
[19:36] it then the market's going to react
[19:38] positively. Yeah.
[19:39] >> They're going to they're saying so
[19:40] you're saying there's a fight and chance
[19:43] here. Totally. Totally. And the feedback
[19:45] on the product is incredible. There's a
[19:46] lot of investors that have tried it and
[19:48] they're like, "Hey, this is a good
[19:49] product." And
[19:50] >> and and and interestingly like they are
[19:52] in the lead in a subcategory where they
[19:54] have never been the lead before. Like
[19:56] Google had Nano Banana which like there
[19:58] was a moment where it was like oh nano
[19:59] banana is the best. Uh obviously
[20:02] Anthropic had best coding models for a
[20:03] while like like crowns have changed but
[20:06] meta's never really held the best model
[20:09] in knowledge retrieval or when they did
[20:10] vibes that was not the best version of
[20:13] that. Uh, but this is at this moment in
[20:16] time the best version of of of what they
[20:18] have and it's building off of a model
[20:20] that's not on the frontier but able to
[20:23] deliver an experience that puts them in
[20:25] first place for this subcategory which
[20:27] is exciting
[20:28] >> and can immediately be incredibly
[20:30] integrated across everything else that
[20:32] they do shopping messaging all these
[20:34] different things.
[20:35] >> Shopify and Muse announced a partnership
[20:38] this morning counterpositioning a little
[20:40] bit against Amazon. Uh, Amazon came down
[20:44] with the banhammer. Buco Capital says,
[20:46] "Amazon cuts off Muse. While I am
[20:48] bullish, Meta and Muse, I think many
[20:49] people are overlooking the digital knife
[20:51] fight that's about to occur. Nobody
[20:53] wants to get commoditized or layered
[20:55] here. Let the games begin." Amazon says
[20:58] it has cut off Meta's new Muse personal
[21:00] AI agent from shopping on Amazon.com on
[21:02] behalf of customers after attempting
[21:04] unsuccessfully to get the Facebook
[21:05] parent company to voluntarily exclude
[21:08] the e-commerce site from the experience.
[21:10] The problem Amazon says is that it never
[21:12] agreed to any of it. Meta didn't tell
[21:14] Amazon that Muse would access its store.
[21:16] The agent itself doesn't identify itself
[21:18] when it browses, and it appears to
[21:19] capture and store customer credentials,
[21:21] which the company says create privacy
[21:23] and security risk. As of Sunday night,
[21:25] people trying to use Muse to shop on
[21:27] Amazon. We're seeing the popup continued
[21:28] access by an unauthorized AI agent
[21:31] violates Amazon's conditions of use, to
[21:33] which our customers have agreed.
[21:35] >> Interesting. I wonder how Walmart will
[21:37] respond. I think Walmart had always been
[21:40] more friendly with chat apps,
[21:42] specifically ChatGpt. And that's the
[21:44] kind of thing that you do when you're
[21:46] not when when you're not the leader in
[21:47] the space. You're more willing to
[21:49] >> experiment. I think they wound up
[21:51] rolling their own. Uh I think they have
[21:53] Sparky and Amazon has Rufus. And so
[21:56] there's two dog named uh AI agents that
[21:59] do the shopping. And then and then most
[22:01] recently
[22:01] >> Tyler's been dailying Rufus by the way.
[22:03] >> Yeah. Just for everything.
[22:04] >> For everything.
[22:05] >> Just Yeah. Psychology. Yeah, getting
[22:08] close to solving another millennium.
[22:10] >> Yeah,
[22:11] >> roof is underrated.
[22:13] >> Yeah, I I I know uh Amazon wound up
[22:15] finding sort of an equilibrium with
[22:17] OpenAI by plugging in their ad network
[22:20] into Chat GPT instead of letting Chacht
[22:23] like check out. This is the whole Eric
[22:25] Suffort agent commerce is a is a mirage.
[22:28] Um instead, uh Amazon is vending in
[22:31] their ads. So Chacht brings the data and
[22:33] says, "Okay, this person's looking for
[22:35] this thing. Do you want to buy an ad on
[22:37] it?" And so there's the transaction
[22:38] value. But I I I don't know how play
[22:41] play out.
[22:41] >> Yeah, I saw saw a couple seemingly smart
[22:44] investors say that that are using Muse
[22:46] and excited about it say, "Why is Amazon
[22:48] banning Muse? This is so shortsighted."
[22:51] And then I actually saw in two instances
[22:53] they deleted because of course they
[22:55] immediately were getting comments about
[22:57] Amazon's ad business, which is obviously
[22:59] what they're looking to protect. You
[23:00] can't
[23:01] >> serve ads to Muse agents or Instinct or
[23:04] or Chat or any of these other things.
[23:06] Yeah. And Amazon's ad business is
[23:08] actually dramatically bigger than both
[23:11] Anthropic and Open AAI over the last 12
[23:13] months. Yeah. So they've done 76 billion
[23:17] in the last 12 months through Q2 of
[23:19] 2026. So obviously they're going to
[23:21] protect this.
[23:22] >> Uh obviously they're going to protect
[23:24] this. I think uh consumers would love to
[23:26] be able to check out with other
[23:29] services. Yeah,
[23:30] >> it is pretty gnarly to navigate
[23:32] >> amazon.com as amazing as the services
[23:35] overall.
[23:37] >> Describing a UX experience is nar.
[23:39] >> No, it's it's
[23:40] >> just being a UX researcher at Amazon.
[23:42] They're calling it gnarly.
[23:45] >> It is fairly gnarly.
[23:47] >> Yeah. Yeah. Yeah.
[23:48] >> The knife fight already took place. It
[23:50] ended a year ago when I wrote the first
[23:52] entry in my Agentic Commerce is a Mirage
[23:54] series. Amazon began blocking agents
[23:56] last July. It's currently suing
[23:58] Perplexity over autonomous shopping with
[24:00] Comet.
[24:01] >> Amazon won't capitulate. It doesn't have
[24:03] to. Alexa for shopping
[24:06] generated 12 billion in incremental
[24:08] revenue. Um,
[24:11] >> every time I hear Roffus or Sparky,
[24:14] >> uh, Walmart Sparky has generated
[24:16] similarly significant commercial
[24:17] traction for the company. Agent Commerce
[24:20] is already a reality, but it takes place
[24:21] on site, not through independent agents.
[24:23] Chad GPT launched ads in February 2026
[24:26] and shortly thereafter shuttered instant
[24:28] checkout. The affiliate model is
[24:30] economically suboptimal relative to
[24:32] advertising.
[24:33] >> Chad GPT has already demonstrated that.
[24:36] >> Yep. Advertising is the dominant
[24:38] business model. Yeah, it'll be
[24:40] interesting. I mean, uh, when the
[24:42] openclaw stuff came out, I was like, it
[24:43] feels like Napster. It feels like the
[24:45] Napster moment. Like, yeah, you can go
[24:46] check out on an e-commerce website, but
[24:48] like nobody wants you to do that that
[24:50] way. and like you're going to have to
[24:52] write some API binding and then like
[24:54] scrape the front end and then rewrite it
[24:55] every time you go and check out because
[24:57] they're constantly going to be fighting
[24:58] you and you are sort of like playing
[24:59] this like game of cat and mouse which
[25:01] you can play uh and for certain for
[25:04] certain e-commerce sites they're going
[25:06] to be very open to that. Other ones are
[25:08] going to be a huge hassle to deal with.
[25:10] >> Yeah. And the question is can agents
[25:12] monetize with advertising? I saw someone
[25:16] uh a pretty viral post. Somebody
[25:18] frustrated with Instinct. Instinct
[25:20] coming and they did basically a blast
[25:21] that says, "Hey, you have a bunch more
[25:23] invites. You can invite a bunch of
[25:24] people." And the user was pretty
[25:25] frustrated with that because they're
[25:26] like, "Hey, you work for me. Why are you
[25:28] advertising?" And so even though they're
[25:30] not paying for the product and they're
[25:31] incurring a lot of costs,
[25:33] >> product's free to them, they're still
[25:35] mad that that they feel like the agent
[25:37] is doing something that's purely in the
[25:38] agent's benefit, not in their
[25:40] >> Sure. Sure. Yeah. I wonder if uh we'll
[25:42] live in a world where you know how Chad
[25:44] GPT has like those recommended agentic
[25:47] actions if like one of those could be
[25:49] like sponsored like hey like want me to
[25:51] go shopping for you right now like
[25:54] >> well yeah no and that's important
[25:55] because companies will happily pay for
[25:58] incremental demand but the issue with
[26:00] agents is if you're saying go to Amazon
[26:02] and buy this thing the user already
[26:04] wanted to go to
[26:06] >> but if I'm if I'm an industry coalition
[26:09] of surfboard wax tax makers. I might
[26:12] choose to send you a notification when
[26:15] you open up Chachi PT that says like,
[26:17] "Hey, want to go surfing? Want to take
[26:19] it to the next level?" Isn't that what
[26:20] happened?
[26:21] >> Uh, basically,
[26:22] >> you got one one recommendation.
[26:25] >> Chachi PT recommends
[26:27] >> what was it?
[26:27] >> Various agentic workflows for me.
[26:29] >> Yes.
[26:30] >> It recommended setting up an agent to
[26:32] give me one realistic next step for my
[26:34] surfing each week.
[26:36] >> Yep. That's great. It's like every
[26:37] >> problem is I know exactly I know exactly
[26:40] what I need to do.
[26:42] >> Uh this is actually
[26:43] >> uh Raj Ver okay
[26:45] >> called out something interesting which
[26:46] is that thread started free falling in
[26:48] the charts right as Muse started
[26:50] growing.
[26:50] >> Yeah.
[26:51] >> Today on on Instagram there there's a
[26:53] there's a hero story that's a full bleed
[26:56] ad for Muse. So they're pushing it.
[26:58] They're pushing it.
[26:59] >> This was interesting from former Meta
[27:02] employee Nikita Beer says, "How
[27:04] important is AI to Meta? Here's one easy
[27:06] tell. One of the launch features of the
[27:08] Muse app was autonegotiating for stuff
[27:10] on Facebook Marketplace. The poor
[27:12] product manager who owns Marketplace
[27:14] knows that millions of bots lowballing
[27:16] sellers will eventually kill the
[27:17] product. But this was escalated to Zuck
[27:19] by the Muse team and he said, "Okay,
[27:21] we'll poison Marketplace just a little."
[27:23] >> Wait, why will that kill the product?
[27:25] >> I mean, think about the experience if
[27:26] you tell your bot like, "Hey, I'm
[27:28] interested in these types of items. Go
[27:30] anytime one of them appears,
[27:32] autonegotiate." Yeah, but wouldn't
[27:33] someone have a bot on the other side
[27:35] dealing with the negotiation?
[27:36] >> You can do the outbound, too.
[27:38] >> You can go outbound.
[27:39] >> Yeah. Yeah.
[27:40] >> Yes.
[27:40] >> Message everyone on Facebook.
[27:42] >> Yes. I got a Dodge Hellcat. If you know,
[27:45] you know, I got one pair of keys and I
[27:48] saw you were post. You liked a post from
[27:50] the Dodge Hellcat community. Would you
[27:52] be interested in purchasing this? I I'll
[27:54] need cash.
[27:55] >> Don't ask questions.
[27:56] >> So, Paramount is officially not leaving
[27:58] California. It's been a back and forth.
[28:00] The company going through an acquisition
[28:02] with Warner Brothers Discovery has been
[28:05] duking it out with AG, Attorneys General
[28:08] in California. Uh the company said
[28:10] Monday, that's Paramount said that it
[28:12] has re reached a settlement with the
[28:14] states challenging its 111 billion
[28:18] acquisition of Warner Brothers Discovery
[28:20] clearing what had become the last major
[28:23] obstacle to the deal. California
[28:26] Attorney General Rob Bont and a
[28:28] coalition of other states sued Paramount
[28:30] into July, arguing that combining the
[28:32] two companies would give a single
[28:34] entertainment giant too much control
[28:36] over major film studios, cable networks,
[28:39] and other media properties to resolve
[28:40] the case. Paramount agreed to several
[28:43] conditions. The combined company will
[28:45] create an independent board of
[28:46] journalists intended to protect the
[28:47] editorial autonomy of CNN and CBS News,
[28:50] commit to releasing the at least 30
[28:52] movies theatrically each year, and spend
[28:55] an additional $1.5 billion on film
[28:57] production over the next 5 years.
[29:00] Paramount did not agree to more
[29:01] structural changes. They're not selling
[29:03] CNN or divesting one of the movie
[29:04] studios. Once completed, the merger will
[29:06] bring Paramount and Warner Brothers
[29:08] under the same roof along with Paramount
[29:09] Plus, HBO Max, CBS, CNN, and franchises
[29:12] and films including Superman, Minecraft,
[29:14] and Sinners. Paramount has argued that
[29:17] the scale is necessary to compete with
[29:19] increasingly powerful technology and
[29:21] streaming companies like Netflix. The
[29:23] settlement also follows months of
[29:25] political pressure surrounding
[29:27] Paramount's future in California. The
[29:29] company had warned that it could move
[29:30] significant production and thousands of
[29:32] jobs out of the state if the acquisition
[29:34] were blocked and had reportedly begun
[29:36] exploring studio space in Nashville.
[29:38] Governor Gavin Newsome publicly warned
[29:40] that losing Paramount would damage
[29:41] California's standing as the center of
[29:43] the entertainment industry. With the
[29:44] lawsuit now headed for resolution, one
[29:46] of the more serious recent threats of a
[29:48] major Hollywood company pulling back
[29:49] from Los Angeles appears to have
[29:51] receded. So, this is extremely bad news
[29:53] for everyone that was hoping that the
[29:54] Paramount lot would be converted into a
[29:56] data center in part of Los Angeles. Uh,
[29:58] a lot of the pro- data center crowd were
[30:00] optimistic about a large-scale data
[30:02] center coming in to Los Angeles, using
[30:04] up a lot of the water, a lot of the
[30:06] energy here in Los Angeles, but looks
[30:08] like they're going to be making movies
[30:09] instead. So,
[30:10] >> and that's our show, folks.
[30:12] >> Thank you so much for tuning in to TVPN.
[30:14] We'll see you tomorrow at 11:00 a.m.
[30:16] Pacific. Leave us five stars on Apple
[30:18] Podcast and Spotify. Sign up for our
[30:20] newsletter at tvpn.com.
[30:24] >> Goodbye.
