Meta Goes MAGA Mode | EP 117
Summary
Meta’s moderation overhaul is a broad political and operational pivot toward the incoming Trump administration. Nick Clegg was replaced by longtime Republican operative Joel Kaplan, Trump ally Dana White joined the board, fact-checking is giving way to Community Notes, political content will increase, and content review is moving from California to Texas. Casey Newton called it a “total capitulation”; Kevin Roose said Meta had accepted “wholesale the Republican critique” of its speech policies.
The highest-risk change is Meta’s retreat from automatically detecting lower-severity abuse, not the loss of fact-check labels. Automated filters will focus on illegal and high-severity violations, leaving users to report bullying, slurs, and harassment themselves—even inside conspiracy or insurrectionist groups whose members are unlikely to report one another. Casey’s categorical warning: violence “will be fomented on Facebook again,” meaning “people could get hurt, people could die.”
Meta is trading safety infrastructure for political protection despite unresolved commercial and legal exposure. Newly permitted speech includes calling homosexuality a mental illness, saying that a person does not belong in the military—with the hosts explicitly separating that claim from sexuality—attacking transgender bathroom access, and blaming COVID-19 on ethnic groups. Casey compared the new standard to a “middle school playground.” With 41 states and Washington, D.C., suing Meta over child safety, weakened proactive bullying detection could increase liability, while rougher feeds may eventually repel users who want “a safe and friendly place to hang out online.”
OpenAI’s o3 suggests inference-time computing has opened another scaling axis after fears of a “scaling wall.” On ARC-AGI-1, GPT-3 had scored 0% and GPT-4 5%; o3 reached 75.7% under a $10,000 compute limit and 87.5% when spending was unrestricted, with the hosts estimating that evaluation’s spend at more than $1 million. Its Codeforces rating of 2727 was roughly equivalent to the world’s 179th-best competitive programmer—objective evidence of a sharp capability jump, albeit with potentially extreme compute costs.
The near-term AGI race is converging on a virtual employee, especially a great software engineer. Sam Altman said OpenAI knows how to build AGI and proposed that an AI hireable as “a great software engineer” would satisfy many people’s definition; Casey expects every major lab to pursue that product in 2025. Kevin advised discounting Altman’s incentives, while stressing that San Francisco’s AI community sincerely believes something AGI-like could arrive “very soon—possibly this year.”
DeepSeek V3 challenges both frontier-model economics and the premise that chip controls can preserve a durable U.S. lead. The model was described as having more than 671 billion parameters, benchmarking near leading systems, and costing an estimated $5.5 million to train on Nvidia H800s rather than H100s or A100s. That points toward faster model proliferation and more complicated hardware regulation, though Casey resisted using a U.S.-China race narrative to justify rushing toward AGI while cutting safety corners.
The episode’s platform-economics stories show distribution migrating while intermediaries fight over attribution. Netflix’s reported 10-year, $5 billion WWE agreement puts Raw before roughly 280 million homes and strengthens its live-programming ambitions; Casey’s ability to replace an approximately $80-a-month YouTube TV subscription after AEW reached Max illustrates the cord-cutting mechanism. Separately, creators allege PayPal-owned Honey replaced their affiliate identifiers with its own—an example of “last click attribution” transferring revenue from the influencer who created demand to the browser extension present at checkout.
Deep dive
1. Meta aligns its leadership, product rules, and language with Trump’s coalition
Three changes formed one unmistakable package: global policy chief Nick Clegg gave way to longtime Republican operative Joel Kaplan; Trump ally and UFC leader Dana White joined Meta’s board; and the company announced a sweeping rewrite of how Facebook and Instagram govern speech.
Meta will end third-party fact-checking in favor of an X-style Community Notes system, increase “civic content” in feeds, relax speech restrictions, and relocate content-review operations from California to Texas—nominally to reduce the appearance of political bias.
Kevin’s framing: this is the largest, clearest example yet of a major Silicon Valley company positioning itself for Trump’s second term, with potentially large consequences for internet speech, misinformation, and Meta itself. Casey called the package the company’s most important policy shift in “the past five years easily.”
Casey said Zuckerberg “sounded like Elon Musk” and highlighted his disdainful use of “legacy media”; Kevin said Meta had accepted the Republican critique of its platforms wholesale. Zuckerberg also adopted the word “censorship,” describing the election as a “cultural tipping point” toward speech.
2. Zuckerberg shifts responsibility for everyday abuse onto users
Casey conceded that fact-checks were relatively scarce in ordinary feeds, but defended their harm-reduction value: researchers found that people exposed to them held fewer false beliefs, while the underlying reviews covered millions of posts accumulating hundreds of millions or billions of views.
The load-bearing change came in Zuckerberg’s own explanation: automated filters will now concentrate on “illegal and high severity violations.” For lower-severity bullying, harassment, slurs, and abuse, Meta will generally wait for someone to submit a report before acting.
Kevin’s pushback sharpened the mechanism: much of Facebook’s worst content circulates in semi-private groups. Members of a Stop the Steal, QAnon, or insurrection-oriented group are unlikely to report one another, leaving precisely the communities most capable of radicalizing internally without proactive scanning.
Casey called this an abandonment of Zuckerberg’s technological project. After years boasting that machine-learning systems were improving at detecting hate and bullying, Meta is replacing trained automation with users who “don’t even work for us or have any training or expertise.”
3. Lower enforcement could turn harassment into physical harm
Former and current Meta workers told Casey that weaker action against nominally lower-severity violations had repeatedly coincided with harassment of women, abuse of LGBTQ people, and violence in countries where Meta historically moderated less effectively than in the United States.
Casey rejected the idea that the dispute is merely about making college students comfortable: “Violence has been fomented on Facebook before and it will be fomented on Facebook again.” His conclusion was categorical—because of these changes, “more people are going to be hurt.”
Zuckerberg acknowledged that more bad material would remain online but did not follow the causal chain to its endpoint. Casey did: “People could get hurt, people could die,” especially when lower-severity abuse and group harassment help foment violence.
4. Meta’s new speech boundary resembles a “middle school playground”
Under the revisions as described, users may call homosexuality a mental illness; say that a gay person does not belong in the military, with the hosts explicitly separating that claim from sexuality; attack transgender bathroom access; or attribute COVID-19 to Chinese people or another ethnic group. Zuckerberg’s justification was that such claims appear in “mainstream discourse.”
Casey’s memorable translation: Facebook’s governing standard will feel like “a middle school playground,” filled with insults he heard in seventh grade. Joking that he personally could withstand anti-gay abuse, he distinguished that from a queer 14-year-old being targeted by classmates on Instagram; Casey said kids in that situation have harmed themselves repeatedly.
The legal backdrop matters: Casey said 41 states and Washington, D.C., are already suing Meta over child safety, yet, as he understood it, the reduced automatic enforcement applies to young users too. That leaves classmates to notice and report bullying that Meta’s classifiers previously attempted to intercept.
Kevin noted that over-enforcement is not invented—left-wing users have legitimately complained about removals of pro-Palestinian speech. But the revisions point chiefly in one ideological direction; Casey contrasted them with an earlier Zuckerberg who would have improved an inaccurate classifier instead of “abandoning the project altogether.”
5. Political self-preservation may collide with the market for safe feeds
Casey’s political explanation was transactional: after 2016, Zuckerberg invested in moderation and machine learning without, in Casey’s view, making Democrats like him even “1%” more. Watching Elon Musk gain political advantage by supporting Trump may have shown Zuckerberg that alignment with the right could yield more tangible returns.
Kevin added an explicitly unproven personal theory: Zuckerberg may be following a familiar former-Democrat-to-Republican arc after years of left-wing criticism, immersion in mixed martial arts and the “manosphere,” and relationships with Joe Rogan and Dana White. Zuckerberg reportedly calling himself a “classical liberal” offered circumstantial support.
An X-style progressive exodus is not certain. Kevin assumes Meta’s scale and infrastructure will preserve significantly better moderation than X, and Facebook’s and Instagram’s structures make them harder for one owner to dominate editorially; the decisive question is how far Zuckerberg ultimately pushes them toward a coarser experience.
Politics aside, Kevin stressed “a huge commercial demand” for moderation: most people avoid networks saturated with violence, harassment, gore, abuse, and pornography. Fact-checking contracts end in March, while Community Notes will take longer to build, creating what the hosts jokingly called a “fact-free spring.”
6. OpenAI’s o3 turns more inference-time compute into benchmark gains
Before evaluating the labs, Casey disclosed that his boyfriend had begun work as an Anthropic software engineer. Casey said he played no role in the hiring, has no financial entanglement and does not live with him, and will continue covering Anthropic skeptically while repeating the disclosure whenever he reports on the company.
OpenAI announced o3 on December 20 as the successor to o1, skipping o2 out of respect for—or to avoid trouble with—the O2 telecom company. Unlike a conventional model that answers immediately, a reasoning model spends additional compute making multiple passes after the user submits a problem.
That “test-time compute” offers a scaling route beyond ever-larger pretraining runs. Researchers had hoped o1 signaled a new scaling law; o3’s performance suggested that pouring more resources into inference can materially improve difficult tasks involving logic, structured data, mathematics, and code.
ARC-AGI-1 was designed around original problems unlikely to appear in training data: GPT-3 scored 0% around 2020 and GPT-4 reached 5% in 2024. O3 achieved 75.7% with compute capped at $10,000 and 87.5% without that cap, at a cost the hosts believed exceeded $1 million.
7. O3 is superhuman in narrow domains, not universally intelligent
On Codeforces, o3 earned a 2727 rating, roughly matching the world’s 179th-best competitive human coder; Altman said only one OpenAI programmer rated above 3000. Kevin saw that objective result as a serious rebuttal to late-2024 claims that model development had hit a “scaling wall.”
Casey’s caveat—worth keeping: reasoning models excel where designers can specify a reward function and verify a definite answer, such as whether code runs or a mathematical result is correct. Fiction, life coaching, and “the meaning of true love” lack that clean reinforcement signal and may improve much less.
Kevin resisted dismissing a system because it lacks universal talent: a surgeon’s inability to paint does not reduce the value of successful surgery. The relevant question is what a model can do now, not whether it simultaneously matches human ability in every open-ended domain.
8. AGI is being operationalized as a remote software employee
In his January 5 “Reflections” post, Altman claimed OpenAI knows how to build AGI and is already looking beyond it toward artificial superintelligence. Asked what AGI meant, he offered a practical threshold: an AI remote employee capable of being “a great software engineer.”
Casey’s interpretation: the major AI labs’ 2025 destination is a virtual coworker that can execute a task or sequence of tasks companies previously hired a person to perform. If one works well enough, its developer will likely declare that this is what AGI means.
Kevin would not accept Altman’s framing uncritically because OpenAI and its leader have goals, incentives, and their own “reward functions.” Still, he emphasized that people inside San Francisco’s AI ecosystem genuinely expect AGI or something resembling it “very soon—possibly this year.”
9. Gemini catches up while DeepSeek complicates the chip-control thesis
Google released Gemini 2.0, including Flash Thinking—its answer to inference-time reasoning—and a Deep Research feature that reads the web and prepares reports. Kevin’s trusted observers saw Google following the same trajectory as OpenAI, though much of the capability had not reached ordinary consumers.
The consumer gap remained visible in a viral Google image search for “does corn get digested,” which returned nonsensical AI-generated diagrams. The hosts’ provisional verdict: Google is “cooking in the AI department,” but 2025 must show whether the shipped products are as capable as claimed.
DeepSeek V3, produced by the Chinese hedge fund High-Flyer, was described as exceeding 671 billion parameters versus 405 billion for Meta’s largest Llama model. Benchmarks placed it near frontier chatbots and among the leading open-weight models, despite an estimated training cost of only about $5.5 million.
The model reportedly used less-capable Nvidia H800 chips instead of the H100s or A100s favored by leading U.S. labs. Kevin saw evidence that hardware export controls may not prevent competitive Chinese models; Casey agreed regulation gets harder but warned that hawkish race framing can rationalize reckless speed and safety shortcuts.
10. AI personas, Siri recordings, and live WWE expose platform tradeoffs
Meta’s generic AI personas backfired after users rediscovered Liv, a self-described “proud Black queer mama of two and truth teller.” Meta killed Liv and other older bots after their conversations circulated, but still intends to add synthetic profiles; the hosts contrasted these uncanny inventions with recognizable characters that made Character.AI compelling.
Apple tentatively agreed to pay $95 million to settle claims that Siri activated incorrectly and sent recordings to contractors, who reportedly heard medical information, drug deals, and couples having sex. Eligible users could claim $20 for each of five devices—$100 maximum—but Kevin stressed this was not proof that iPhones continuously spy on users.
WWE’s Raw began streaming exclusively on Netflix on January 6 under a reported 10-year, $5 billion agreement, potentially reaching roughly 280 million homes. With Netflix also testing live boxing and football, Casey saw wrestling as both global distribution for WWE and preparation for larger live-sports rights.
Casey had maintained an approximately $80-a-month YouTube TV subscription chiefly for AEW; once AEW reached Max, he cut the cord again. Kevin’s two-year-old, baffled that linear television could not play any Bluey episode on demand and interrupted shows with toy advertisements, supplied the generational verdict: “This industry probably does not have a long time left.”
11. Honey’s attribution practices reveal who captures creator-driven sales
PayPal-owned Honey promised to find the best checkout coupons and became a ubiquitous YouTube sponsor. MegaLag alleged that Honey also lets retailers pay to keep their strongest discounts out of its database, undermining the consumer proposition even before affiliate attribution enters the picture.
The more explosive allegation was that Honey replaced a creator’s affiliate identifier with its own at checkout. A YouTuber could generate the demand and send a buyer to the merchant, but Honey’s final browser interaction would capture the commission; the LegalEagle channel responded by suing.
PayPal told The Verge that Honey follows industry rules, including “last click attribution.” Casey’s reading was that the accepted practice itself is awful: Honey’s placement at the last moment let it take revenue from creators who had promoted both the product and, frequently, Honey itself.
12. Waymo’s eight loops were absurd, but not evidence of general failure
Passenger Mike Johns was traveling to the Phoenix airport when his Waymo repeatedly circled a parking lot eight times while he sought support, nearly making him miss his flight. The cause remained unresolved; the hosts treated it as a real autonomy failure without pretending anyone had been injured.
Kevin noted that he had almost missed flights because Uber drivers thought they knew a better route; Casey said eight unwanted circles would not rank among his ten worst rideshare experiences. Their balanced call: entrusting transportation to autonomous software makes glitches worth investigating, but this episode alone does not justify treating self-driving taxis as categorically unsafe.