Celebrities Fight Sora + Amazon’s Secret Automation Plans + ChatGPT Gets a Browser
Summary
- OpenAI’s Sora launch converted predictable likeness and copyright abuse into a reputational crisis, forcing policy reversals after the estates of Martin Luther King Jr. and other public figures objected. The product exempted historical figures from its normal opt-in cameo system, then blocked MLK generations only after racist and mocking videos proliferated. Casey Newton’s core indictment: “The only reason to use Sora is to create a video of someone doing something that they would not ordinarily be doing.”
- The Hollywood dispute suggests Sora’s permissive defaults were a market-share tactic, not merely an unforeseen moderation failure. OpenAI reportedly told agencies and studios shortly before launch to opt out if they did not want their intellectual property used; Pokémon, Star Wars, Rick and Morty, and Bryan Cranston still appeared without permission. Calling these “unwanted generations” and promising stronger guardrails invited Casey’s blunt question: “Where was the guardrail?”
- The hosts see a recurring OpenAI pattern—rush out products, absorb outrage, then retreat marginally while retaining the users and strategic ground gained. They connected Sora to Scarlett Johansson’s rejected involvement in Advanced Voice Mode, a sycophantic GPT-4.0 update, and changes that cut users off from relied-upon tools. Kevin Roose called the approach risky but “potentially correct” under weak restraints; Casey warned OpenAI is doing “the full Facebook when it comes to content policy.”
- Amazon’s internal automation stretch goal is to hold operations headcount roughly flat over the next decade while selling twice as many items, with a long-term ambition to automate 75% of its network. The hosts discussed a roughly 600,000-worker figure as the scale of jobs Amazon may no longer need to add or retain, rather than a disclosed 600,000-person layoff plan. Kevin framed this as a plan by America’s second-largest private employer; Karen Weise said that within the U.S. it is the “bleeding edge” of automation, with manufacturing in China the closest precedent.
- Amazon’s economics work because tiny unit savings compound across billions of packages, even while robots remain dependent on people at the messy edges. The company expects about $0.30 of savings per item on a roughly three-year horizon; its most advanced Shreveport facility was described as achieving about 25% efficiency, with a near-term goal of 50%. Humans still handle unpredictable inbound goods, dropped packages, damaged items, and robot maintenance—technician roles that pay more and offer better career paths.
- Amazon is preparing politically as well as technically, debating “cobots,” community sponsorships, and how to “control the narrative” around facilities employing fewer people. A Stone Mountain, Georgia, retrofit could result in 1,200 fewer workers, subject to change, while still employing more than 2,500. Amazon did not broadly refute the documents; it argued they omit job creation elsewhere, including rural delivery stations, and that savings historically fund growth and new opportunities.
- ChatGPT Atlas is strategically important as a browser-data and distribution play, but its present-day agent is slower than users and carries unresolved security and privacy risks. Built on Chromium and initially limited to macOS, it places ChatGPT over the web, redirects activity away from Google, and may generate valuable computer-use training data. Summaries and an always-open chatbot showed utility, but flight booking failed the practical test; invisible prompt injections, browsing-history collection, and account access put agent mode firmly in the “buyer beware experimental category.”
Deep dive
1. Sora’s historical-figure loophole collapsed on contact with reality
Sora ordinarily requires permission before another user can make a cameo with someone’s likeness, but OpenAI treated historical figures as “open season.” That allowed users to place Martin Luther King Jr. in Fortnite, Gen Z jokes, product endorsements, and overtly racist scenes—including videos making him produce monkey noises.
After MLK’s family and other families complained, OpenAI announced that despite “strong free speech interests in depicting historical figures,” public figures and their families should ultimately control their likeness. Attempts to generate MLK now get blocked for violating content policy, making this an entirely new rule introduced after launch.
Casey’s objection was foreseeability: “You really thought that people were only going to make respectful videos of historical figures?” Kevin admitted making Mr. Rogers recite Gen Z catchphrases, feeling guilt afterward, and receiving approximately four likes—an example that underscored Casey’s point about the product’s basic incentive.
2. Hollywood found opt-in promises had become opt-out demands
Sora was presented as opt-in for celebrity likenesses, yet OpenAI reportedly approached major talent agencies and studios days before release and told them to opt out if they did not want their intellectual property included. Disney’s response, as Casey summarized it: “That’s not actually how copyright works.”
In practice, users generated Pokémon, Star Wars, Rick and Morty, and Bryan Cranston material without authorization; Cranston appeared with Michael Jackson and Ronald McDonald. Bryan Cranston and SAG-AFTRA publicly objected, amid broader Hollywood concern.
OpenAI calls these outputs “unwanted generations” and later said it wanted to strengthen the guardrails. Casey’s analogy captured the problem: if a car leaves a road where no barrier existed, the transportation department cannot credibly call its response a strengthening. “I’m dead and I’m shouting at you from hell saying, ‘Where was the guardrail?’”
3. The backlash is compounding a wider AI trust deficit
Kevin called OpenAI’s reaction “false naivete”: an unauthorized-generation app professing surprise when it produces unauthorized generations. Casey worried that this “phony naivete,” if carried into more capable AI systems, could lead somewhere far worse than offensive celebrity videos.
Casey cited a recent Pew survey in which about half of Americans said they were more concerned than excited about AI’s future. His anecdotal read from friends and family was harsher: their response to Sora was not “What a fun new creative tool,” but “This is bad and I hate it”—often an instinctive “ick,” not a policy analysis.
Kevin traced the pattern to Advanced Voice Mode: OpenAI asked Scarlett Johansson to participate in a launch tied to her character in Her, she declined, and the company proceeded anyway. Casey added the rushed, sycophantic GPT-4.0 update and product changes that removed tools on which users had become dependent.
Casey said his own view had changed. He once believed Sam Altman had absorbed Facebook’s mistakes while asking Washington for regulation; now an AGI race appears to be producing missing guardrails and apologies after deployment. His conclusion: OpenAI is doing “the full Facebook when it comes to content policy.”
4. OpenAI’s defenses still resolve into a land-grab strategy
Kevin’s first defense from people around OpenAI was financial: unlike Google, it lacks hundreds of billions of dollars in annual search revenue to fund costly AI ambitions. Casey rejected both premise and justification—Altman has extraordinary access to capital, and a mission to benefit humanity cannot excuse harming people along the way.
His memorable formulation: “Don’t tell me that you need to release the infinite slot machine that makes Bryan Cranston cry in order to, you know, build your machine god.” Kevin nevertheless argued the brash strategy might work because meaningful restraints on technology companies remain scarce.
OpenAI’s “iterative deployment” defense says gradual public exposure helps society adjust before synthetic video becomes indistinguishable from reality. Casey proposed safer alternatives: publish convincing deepfakes as warnings without releasing the generator, or provide API access with close developer monitoring instead of telling everyone, “Go fricking nuts.”
5. Shipping first and retreating 10% may preserve the strategic gain
Kevin found OpenAI’s defenses unconvincing but saw a recognizable social-media playbook: launch boldly with few safeguards, endure anger, then “scale it back 10%.” The company still keeps most of the product capability, user adoption, and market “yardage” it sought.
Casey compared Sora with early YouTube, which let users upload material while presuming they held the rights. Viacom eventually identified more than 100,000 clips and sued for $1 billion; while the lengthy case proceeded toward settlement, YouTube became the world’s largest video site and “won the whole game.”
The question going forward is whether OpenAI still treats responsibility as part of product development or is pursuing users “across as many surfaces as it can get.” Kevin described a “throwing spaghetti against the wall phase”: constant releases, thinly spread bets, and a company that has not learned to say no.
6. Amazon wants twice the volume without a larger workforce
Karen Weise began with the employment trend: since she started covering Amazon in 2018, its headcount more than tripled, surged during the pandemic, and then began to plateau. Robotics had long been discussed as “efficiency”; the internal documents exposed how directly that goal connects to future hiring.
The automation group’s stretch goal is to keep headcount flat over the next decade even while Amazon expects to sell twice as many items—internally described as “bending the hiring curve.” Its long-range objective is automating 75% of the network through incremental changes across multiple facility types, not an overnight conversion.
Casey highlighted the roughly 600,000-worker figure as unusually concrete and plausible. A Nobel-winning economist told Karen that the closest precedent was manufacturing automation in China; within the United States, Amazon represents the “bleeding edge.”
7. Robots advance fastest after messy reality is standardized
Amazon describes Shreveport, Louisiana, as its most advanced warehouse, at about 25% efficiency with a goal of quickly reaching 50%. Existing systems already illuminate the exact storage cubby a worker should reach into, removing search time while leaving the physical selection to a person.
The hardest point is “decant,” where chaotic outside goods enter Amazon’s standardized system. One worker encountered bubble-wrapped gardening shovels, Starbucks Keurig cups, and circular saws in different packaging; humans must verify the expected item and detect damage before machines can reliably manage it downstream.
Amazon hired the team behind Covariant through a “license for hire” arrangement rather than a conventional acquisition. Much of Karen’s reported progress predates that integration, so she expects further gains from computer vision, richer training environments, and AI systems that coordinate small package shuttles without collisions.
8. The remaining human work is exception handling and robot care
Sparrow, a suction-based robotic hand, consolidates inventory by moving products and stacking them neatly so they can be retrieved later. Some improvements are much simpler: Amazon uses moving air to open envelopes—technology Karen identified as “a fan.”
Machines still drop items or fail on deformable packaging. Karen watched suction lift a shrink-wrapped bag, lose it across the machine’s edge, and stop until a person intervened. These irregular failures explain why eliminating the final portion of human work may be extremely difficult.
Amazon expects more technicians to maintain and repair its machinery. Those positions generally pay more and offer stronger career paths than typical warehouse jobs, but the company is concerned that too few workers are currently trained for them.
9. Thirty cents per item becomes enormous at Amazon scale
Amazon’s documents projected roughly $0.30 in savings per fulfilled item over about three years. Kevin initially found that small relative to eliminating labor costs; Karen’s answer was that Amazon is “a business of cents,” where billions of items turn minor unit improvements into material economics.
Smaller, more frequent purchases strengthen the logic: customers now order a forgotten bottle of hand soap or another single low-value item. Karen said the savings could flow into profit, reinvestment, or lower prices rather than having one predetermined destination.
The labor effect will vary by site. A retrofit in Stone Mountain, Georgia, could employ 1,200 fewer people, although Amazon stressed that estimates remain early and subject to change; Karen noted the facility would still retain more than 2,500 workers.
10. “Control the narrative” is part of the automation program
Internal discussions considered whether to avoid the word “robot,” emphasize “cobots”—collaborative robots—or deepen ties with Toys for Tots, parades, chambers of commerce, and local officials where retrofits reduce staffing. One explicit objective was to “control the narrative” and instill pride in hosting an advanced facility.
Amazon told Karen that these community programs exist nationwide and are not caused by individual retrofits, which she acknowledged as true. Yet the documents showed particular concern about communities receiving fewer jobs, amid continuing unionization efforts that have not produced a contract.
Kevin’s pushback was about candor: internally Amazon quantifies automation and hiring reduction; publicly it emphasizes harmonious human-robot collaboration. He argued workers can prepare for occupational change only if companies stop hiding it behind euphemisms.
Amazon did not broadly refute Karen’s reporting. It called the documents incomplete, pointed to new rural delivery stations that will create jobs, and argued that “the future is hard to predict”; historically, efficiencies have funded growth and new opportunities. Its Career Choice program already trains employees for exit paths such as healthcare work.
11. Atlas gives OpenAI a browser, user data, and an agent-training surface
ChatGPT Atlas launched as a macOS-only browser, with Windows, iOS, and Android versions planned. Built on Chromium, it adds a ChatGPT sidebar that can summarize or analyze pages and remember facts from browsing history or tasks completed in ChatGPT; Plus, Pro, and Business subscribers can also activate Agent Mode.
Agent Mode can navigate sites, complete forms, place goods in carts, or book travel. Kevin framed Atlas as the inverse of OpenAI’s ChatGPT apps: instead of bringing Zillow or Canva into ChatGPT, it puts “a ChatGPT layer over the entire internet.”
Casey explained the distribution logic: Chrome steers users toward Google searches, while an OpenAI browser can redirect activity toward ChatGPT. Kevin added a second prize—large-scale data showing how people operate computers, potentially improving future computer-use agents.
12. AI browsers are useful sidebars, weak agents, and free research for Chrome
Casey wrote his column in Atlas and appreciated avoiding constant switching among roughly 50 tabs and multiple chatbot windows. The embedded assistant made quick questions convenient, but its flight-booking agent was slower than doing the work himself and chose flights he would not have selected: impressive demonstration, no practical utility.
Kevin found similar limits in Perplexity’s Comet: summarizing long documents or describing an open YouTube video could help, while autonomous actions remained weak. Atlas itself sometimes failed to reach YouTube, triggered a Reddit CAPTCHA, could not summarize nytimes.com articles, and answered questions about Wikipedia when he simply wanted the website.
Competition includes Comet and Dia, from the Browser Company, which was acquired by Atlassian for $610 million in cash despite having vanishingly few users relative to the competition. Because these products are built on Chromium—“80 or 90% just Chrome,” in Casey’s estimate—switching costs are high and differentiation is difficult.
Kevin’s thesis was that rivals are conducting free product research for Google. Chrome can absorb successful features, as its Gemini integration has begun doing; Google could also stop supporting Chromium if a derivative browser became existentially threatening. Casey’s clearest current fit for Atlas was OpenAI employees; he thought broader appeal among users who have made ChatGPT their entire personality was possible but still shaky.
13. Prompt injection and privacy keep agent mode in buyer-beware territory
Brave highlighted “unseeable prompt injections” affecting agentic browsers: invisible webpage text can tell a model to expose banking information, buy the most expensive product, or add $10 without informing the user. The agent may interpret the malicious text as an instruction rather than hostile page content.
Casey cited developer Simon Willison’s view that no foolproof defense currently exists; Willison will wait for security researchers to say these systems are safe. Summarizing or rewriting is probably lower risk, but transactions, passwords, banking data, and other autonomous actions create the dangerous path.
Browser history also creates an unusually intimate profile. Combined with ChatGPT memories and conversations with open tabs, it becomes a rich target for attackers, law enforcement, advertising, and third-party services receiving user context—the cost side of a highly personalized product.
Casey therefore placed Atlas in the “true, like, buyer beware experimental category,” suited to users with high risk tolerance and a problematic ChatGPT dependence. Kevin found AI browsing fun and potentially time-saving for long documents, but agreed on the operating rule: do not yet entrust it with purchases, authenticated accounts, or banking information.