TBPN Sep 21, 2026 44m 30m saved
Amazon has cut Meta's new Muse agent off from Amazon.com, and the number underneath the decision is an advertising business that did $76 billion in the twelve months through the second quarter of 2026.
The reading circulating on Monday morning was that Amazon was being short-sighted about a product customers like. TBPN's hosts made the opposite case: agents that check out for you are a threat to the ad slot, and an ad slot is the thing Amazon actually sells.
"Nobody wants to get commoditized or layered here. Let the games begin."
TBPN runs a daily live tech show out of Los Angeles, and Monday's news block ran from ByteDance's Seedance 2.5 video model through Meta's share price, Amazon's block, the return of model hallucination and the settlement that keeps Paramount in California.
The full segment is covered here so you can skip it. 44 minutes of audio, 14 minutes of reading.
Here are the 10 takeaways that matter.
Key Takeaways
Seedance 2.5 swaps the whole person, not the face, which is why a video of an analyst dressed as Drake fooled a large part of X over the weekend The give-away that still shows up is blending: a subject's features drift back toward the original performer in close shots
ByteDance is the one Chinese lab not open-sourcing its best model, and the hosts' explanation is distribution Open weights are how a challenger buys an ecosystem; ByteDance already owns the surface the video lands on
Meta is up 27% this month, which the hosts put almost entirely down to Muse rather than to anything in the underlying model
Amazon's advertising revenue was "76 billion in the last 12 months through Q2 of 2026", which is the business an agent checkout would route around
Agentic commerce is happening on-site, not through third-party agents — Rufus and Sparky, not an outside shopper with your password
Hallucination is back at the consumer layer because free products at scale run on cheaper, older models
Meta shipped an agent that auto-negotiates on Facebook Marketplace, which a former employee reads as a deliberate decision to degrade Marketplace for Muse's benefit
Paramount settled with California and the other states for no divestitures, agreeing to 30 theatrical releases a year and $1.5 billion of extra production spending
1. Seedance Replaces People
The show opened on what one host had spent the weekend doing: feeding real video clips to ByteDance's Seedance 2.5 and replacing the people in them.
The change is that the model swaps the whole subject, not the face
The big update here is that it replaces the whole person, the whole object, as opposed to just a face swap
TBPN
The reference point was Dream Booth, the Google paper that made face-swapping possible in 2022 and required ten or twenty photographs, a Python script and a Colab notebook with whatever GPU the lottery handed you. A hobbyist guitarist got it running on gaming hardware, which is how it reached everyone else.
That entire pipeline is now one prompt
Gone are the days of face swaps.
TBPN
The example that carried the point was a video of the semiconductor analyst Dylan Patel appearing as Drake, posted the day after the real music video dropped. The hosts said it tricked a large number of people, in part because the underlying footage had not yet saturated the internet.
What they said the model does well is hold the subject when the camera is not cooperating: small in frame, behind a chandelier, looking away, two versions of the same person in one shot. Where it still slips is close-up facial structure, where the original performer starts bleeding back through.
The hosts' own attempts included one host as Steve Jobs introducing the MacBook Air, as Jensen Huang pulling GPUs out of an oven, and as Steve Ballmer at the developers speech. They also cut a trailer of Johnny Knoxville and Steve-O clips with themselves substituted in, and noted that the footage worked partly because the real stunts already look like AI prompts.
2. The Higgsfield Credit Maze
The production path was Codex running on a PC, driven from a phone, generating reference images through ChatGPT and rendering through Higgsfield, which wraps Seedance in a harness and exposes it as an API.
The complaint was billing. Higgsfield sells API credits and it sells plans, and the two do not convert into each other, so money loaded in one place cannot be spent in the other.
Money gets stranded between two kinds of credit
And so I have like so much random money stuck in various pieces of Higgs.
TBPN
One host compared the checkout flow to a gambling site, with countdown timers offering a discount on the next credit purchase. The other raised complaints he had seen about being unable to cancel a plan. The verdict on the product was that the assets themselves were delivered well and that a cleaner API was worth trying next.
3. Reshoots Get Cheaper
Asked where the technology lands commercially, the hosts separated the "it's over for traditional filmmakers" reaction from what studios will actually use it for.
Their example was Henry Cavill, who grew a mustache for a Mission: Impossible shoot and was then called back for Superman reshoots. The mustache was removed in post with 2017 tools, at high cost, and the result was widely criticized by visual-effects people. That job is now trivial.
The likely use is repair work, not performance
For most things, it's going to be, okay, well, if you needed someone to drive the video, why don't you just have the actor do that?
TBPN
Stunts were the one exception they allowed. For ordinary performances, they expect the old method to hold.
4. Why ByteDance Stays Closed
The section the hosts flagged as the real question: China has roughly a fifth of the compute and is winning at video, and is not open-sourcing the model.
Their first answer was that ByteDance has somewhere to put the output. Video models are compute-intensive, and a Western lab spending that compute has to go and build a demand surface for it.
The American labs have no place to put a video model
So Anthropic and OpenAI do not have good surfaces to deliver video, AI or not video.
TBPN
ByteDance has TikTok, Douyin, CapCut, an ad business, e-commerce and recommendation systems already running. Meta and Google have the same shape in the United States, but the hosts noted that AI-generated video is unpopular with American audiences in a way it is not in China.
On training data, they walked through the fair-use argument rather than asserting the outcome. A short, recut, commented-on clip of a film is transformative and legal, which means a very large amount of Hollywood footage sits permanently on the platforms.
A film-criticism clip is fair use, and it stays on the server
You can use clips from movies if you're commenting on the lighting, and you're trying to explain the lighting.
TBPN
The exposure, they said, is on the output side rather than the input side. Likeness rights survive training. Their test case was a music-generation service that will refuse to produce a named artist's song and offer a generic one instead.
Then the open-source question, which they answered twice.
Nobody releases weights from in front
First, no one wants to open source when they're in the lead.
TBPN
Open weights are a challenger's tool, not an incumbent's
So in general, if you are a challenger that needs an ecosystem, you want to open up. We saw this with the early meta stuff. And then if you're an incumbent and you have distribution, you want to close down.
TBPN
Applied to coding models, open weights build a Chinese developer stack and give Huawei something to be competitive on. Applied to video, where ByteDance is vertically integrated down to distribution, they buy nothing. The hosts' longer-dated argument was that video models may matter for training humanoid robots in simulation, and that China is doing well in humanoids too, but they put that well beyond the current cycle.
5. Meta Up 27% on Muse
Meta's share price came up next, with a chart the hosts read as almost entirely a Muse story.
A company that spends heavily and then ships gets paid for shipping
Basically, when you have a company that's spending hundreds of billions of dollars on something without much to show for it, and then you have something to show for it, then the market's going to react positively.
TBPN
What they found unusual is the category. Meta has not held the best model in knowledge retrieval, and its earlier video product was not the best version of that either. Muse puts it in first place in a subcategory for the first time, built on a model that is not at the frontier, and the surface connects immediately to shopping, messaging and the small-business base. Shopify and Muse announced a partnership the same morning, which the hosts read as positioning against Amazon.
6. Amazon's Ban Hammer
Amazon then cut Muse off. The hosts read the reporting: Amazon asked Meta to exclude the store voluntarily, Meta did not, and Amazon blocked it.
Amazon's stated objections were that it never agreed to any of this, that Meta did not tell it Muse would access the store, that the agent does not identify itself when browsing, and that it appears to capture and store customer credentials. Users trying to shop through Muse on Sunday night got a pop-up telling them continued access by an unauthorized AI agent violates Amazon's conditions of use.
The framing the hosts took came from the investor account Bucco Capital.
The knife fight is between agents and the stores they shop at
While I am bullish meta and Muse, I think many people are overlooking the digital knife fight that's about to occur. Nobody wants to get commoditized or layered here. Let the games begin.
TBPN, reading Bucco Capital
They noted that several investors publicly called the block short-sighted and then deleted the posts once the replies pointed at Amazon's advertising revenue.
The business being defended is much larger than either frontier lab
And Amazon's ad business is actually dramatically bigger than both Anthropic and OpenAI over the last 12 months.
TBPN
And the number
So they've done 76 billion in the last 12 months through Q2 of 2026.
TBPN
The hosts expect Walmart to behave differently, because Walmart is not the leader and has historically been more willing to work with chat products. Both retailers now run their own shopping assistants — Rufus at Amazon, Sparky at Walmart.
The legal precedent they raised is the Perplexity ruling: Amazon cannot block an agent that is running on the customer's own computer, because the customer opened the tab. One host expects that to be where this lands, and expects the store to keep fighting anything running as a third-party service.
7. Ads Beat Agent Checkout
The hosts leaned on Eric Seufert's argument that independent agentic commerce is not the business it looks like.
Amazon has already found an equilibrium with OpenAI, they said, by plugging its ad network into ChatGPT rather than letting ChatGPT check out: ChatGPT supplies the intent, Amazon bids for the placement.
The knife fight Seufert says already happened
Agentic Commerce is already a reality, but it takes place on site, not through independent agents.
TBPN, reading Eric Seufert
And the reason it settles this way
The affiliate model is economically suboptimal relative to advertising.
TBPN, reading Eric Seufert
The supporting figures they read out were Rufus generating $12 billion in incremental revenue for Amazon and Sparky producing comparable traction for Walmart. They also noted that ChatGPT launched advertising in February 2026 and shut its instant checkout shortly afterwards.
One host's structural point was that agent advertising has an incrementality problem. A user who tells an agent to buy something on Amazon was already going to Amazon, so nobody will pay for that impression. The slot worth paying for is the one that creates demand the user did not arrive with — his example was a surfboard-wax trade body pushing a suggestion into the assistant. He noted that ChatGPT already recommends agentic workflows unprompted, having offered to set up a weekly surfing agent for him.
They also flagged a second problem: users resent it. A widely shared complaint came from someone whose agent used its own turn to push invite codes, and the objection was not the cost — the product was free — but that the agent was working for itself.
8. Hallucinations Come Back
The hosts' view is that model hallucination had stopped being a practical problem for ordinary retrieval this year, and has returned with the consumer agents.
Retrieval got reliable and then stopped being reliable
But for knowledge retrieval, the hallucination problem kind of went away. But it's coming back in full force.
TBPN
The instance they cited was a user whose agent invented a middle name while booking a flight, leaving him with a boarding pass that did not match his identification. The hosts' explanation was economic: a free product served at scale has to be served cheaply, which means older and smaller models behind it.
They were not alarmed by the specific failure. One host had been through misspelled flight details before and said it is resolved at the desk, at the cost of being sent back through the security line.
9. Poisoning Marketplace
The sharpest item of the block came from the former Meta employee Nikita Beer, who used one Muse launch feature as a read on the company's priorities.
Marketplace was knowingly degraded to give Muse a launch feature
The poor product manager who owns Marketplace knows that millions of bots, low-balling sellers will eventually kill the product.
TBPN, reading Nikita Beer
Beer's account is that the feature was escalated to Mark Zuckerberg, who approved poisoning Marketplace a little. One host pushed back on the premise — if buyers run bots, sellers will run bots too — and the other pointed out that the bots also go outbound, cold-messaging everyone who has ever engaged with a listing category.
10. Paramount Settles
Between guests the hosts read the Paramount news. Paramount said Monday it had settled with the states challenging its $111 billion acquisition of Warner Bros. Discovery, removing the last obstacle to the deal.
California's attorney general Rob Bonta and a coalition of states had sued in July, arguing the combination would give one company too much control over film studios, cable networks and other media properties.
What Paramount gave up, and what it did not
The combined company will create an independent board of journalists intended to protect the editorial autonomy of CNN and CBS News, commit to releasing through at least 30 movies theatrically each year and spend an additional $1.5 billion on film production over the next five years.
TBPN
Nothing is being sold. CNN stays, both studios stay. The company had warned it could move production and thousands of jobs out of California if the deal were blocked and had reportedly looked at studio space in Nashville, and Governor Gavin Newsom had said publicly that losing Paramount would damage the state's standing in the industry.
One host's closing observation was that the subscription stack this produces — Paramount+, HBO Max, CBS, CNN — already costs a household about $60 a month.
Bonus Insights
The FAA cleared 10,000 feet of airspace over SpaceX's McGregor site
The restriction covers a 1.5 nautical mile radius ahead of Tesla's October 1 Roadster reveal, and the hosts said that altitude is well outside the norm for the site.
That altitude is unusually high for McGregor, where restrictions tied to static engine fires and component testing have often topped out closer to 2,000 feet.
TBPN
The hosts have a running bet on how high the Roadster gets off the ground
Their own guesses ran from half a foot to 10,000 feet. Tesla has confirmed nothing beyond launch-themed imagery and the phrase "go for launch."
The deposit math on the Roadster is a ten-day option
One host worked through it out loud: the deposit is refundable, the reveal is on October 1, and so the money is committed for about ten days against the chance of an early car. He put a $250,000 price as competitive and $800,000 as a different market entirely.
A clip of a government adviser claiming humans answer ChatGPT queries got a split verdict
The hosts said the claim is wrong about inference and roughly right about training data, given the scale of the labeling industry, and used it to note that several governments now have a podcast-style figure who says striking things in public.
David Sacks co-chairs something abbreviated P-CAST
The President's Council of Advisors on Science and Technology shortens to something that sounds like "podcast," which the hosts enjoyed more than the substance.
Taste as the remaining differentiator came up as a joke and then as a thesis
The fedoras the hosts wore all episode were the pretext. The claim underneath it was that as generation gets cheap, the scarce input is judgment about what to generate.
The hosts' bottom line is that the agent era's first real fights are commercial rather than technical: the models can already replace a person in a video or check out a shopping cart, and what decides the outcome is who owns the surface the output lands on.
Products, Companies & Tools Mentioned
Seedance (ByteDance's video model, version 2.5, which replaces the whole subject rather than the face and which the hosts say China is deliberately not open-sourcing)
Higgsfield (Wraps Seedance as an API; praised for asset delivery, criticized for a plan-versus-API credit system that strands money)
Codex and ChatGPT (The pipeline: Codex driving the render from a phone, ChatGPT generating the reference images)
ByteDance, TikTok, Douyin and CapCut (The distribution the hosts say removes ByteDance's reason to open the weights)
Meta and Muse (Up 27% on the month, which the hosts attribute almost entirely to the new personal agent)
Amazon and Rufus (Blocked Muse from the store; its assistant is credited with $12 billion of incremental revenue)
Walmart and Sparky (The retailer the hosts expect to take the opposite side, and its own shopping assistant)
Shopify (Announced a Muse partnership the same morning, which the hosts read as positioning against Amazon)
Perplexity (The Comet ruling the hosts cite: a store cannot block an agent running on the customer's own machine)
Anthropic and OpenAI (Named as having no natural surface to deliver video, which is the hosts' explanation for ceding the category)
Huawei (The beneficiary of an open Chinese coding stack, and the reason open weights make sense there and not in video)
Paramount and Warner Bros. Discovery (The $111 billion deal settled with the states, with no divestitures)
Tesla and SpaceX (The Roadster reveal on October 1 and the McGregor airspace restriction behind the speculation)
Books & Resources Mentioned
Agentic commerce is a mirage – Eric Seufert (The essay series the hosts lean on for the claim that agentic commerce happens on-site and that advertising beats affiliate economics)
SemiAnalysis (Dylan Patel's research outfit, whose Drake video became the weekend's demonstration of Seedance 2.5)
DreamBooth (The 2022 Google paper the hosts use as the before picture for how hard face replacement used to be)
Listen to the full episode
🔴 YouTube | 🔗 Episode page
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: