Featuring
Episode 170 bounced from candid CEOs and one-person AI companies to the main event: a 35-minute TikTok brawl over foreign control, secret evidence, free speech, and whether Washington had built Patriot Act 2.0. Jason had the best episode because his sell-it-don't-silence-it position survived both the argument and two years of hindsight. Sacks supplied the necessary civil-liberties stress test, while Chamath and Friedberg closed with a surprisingly useful fight over when “regulatory capture” explains too much.
Spice rack
Did TikTok's foreign-control risk justify forced divestiture, or did the bill create an overbroad power before the government proved a narrowly tailored case?
Original point: The bill risked becoming 'Patriot Act 2.0' because its foreign-control definitions could be stretched beyond TikTok and used against domestic political targets.
What everyone argued
Chamath Palihapitiya
Chamath argued that TikTok should be shut down, not merely sold, because hidden access paths and algorithmic influence could survive a nominal migration. He treated Congress's 50-0 committee vote as a strong signal that lawmakers had seen classified national-security evidence the public would never receive.
Jason Calacanis
Jason argued for divestiture: a foreign adversary should not control a platform with massive U.S. reach, detailed user data, and a recommendation algorithm capable of steering attention. He stressed that ByteDance personnel had already misused TikTok data and that a sale preserved the product while removing the control risk.
David Sacks
Sacks demanded public proof, warned that 'subject to the direction or control' was vague, and argued that a general foreign-adversary framework could invite executive abuse against domestic platforms. He remained open to a proven, narrowly tailored TikTok remedy but did not think the March 2024 case had met that burden.
David Friedberg
Friedberg opposed reciprocity for its own sake and favored consumer choice, transparency, audits, technical controls, or a sale if actual spying could be shown. He did not see a TikTok-specific harm that justified forcing a transaction without clearer evidence.
Winner circle
Jason wins the central policy question. His conditional-divestiture position best matched the demonstrated risk, the Supreme Court's narrow as-applied holding, and the eventual deal that kept TikTok online. Sacks raised the strongest caution and correctly identified genuinely broad language, but he overstated how easily the Attorney General alone could target unrelated domestic platforms. Chamath demanded more than the evidence justified, while Friedberg gave too little weight to ownership and enforcement risk.
Commentary
Chamath Palihapitiya
Assumptions and fact checks
Unanimous congressional support likely meant members had received decisive classified evidence.
Why it mattersClassified evidence existed, but the Supreme Court expressly resolved the case on the public record. Unanimity alone does not prove what any classified material showed.
No divestiture or audit could reliably separate TikTok's U.S. service from ByteDance control.
Why it mattersThe 2026 joint venture does not eliminate every verification concern, but it demonstrates that a legally accepted separation structure was feasible.
Jason Calacanis
Jason wins the policy question because his remedy matched the eventual workable outcome. He would have been stronger if he had dropped the microphone speculation and stayed with ownership, data access, and algorithm governance.
Assumptions and fact checks
Foreign control of a large recommendation platform creates material potential harm even before covert manipulation is publicly proved.
Why it mattersRisk regulation can reasonably consider capability, incentives, scale, and difficulty of detection. The Court accepted this predictive logic on the public record.
If ByteDance resisted divestiture, that resistance itself would prove malign intent.
Why it mattersA firm can resist losing a valuable asset for ordinary commercial, legal, or strategic reasons. Resistance adds concern but is not proof of espionage or manipulation.
The House passed the original TikTok measure 352-65.
CheckThe congressional record cited by the Supreme Court reports a 352-65 House vote; the later enacted package passed the Senate 79-18.
TikTok collects extensive personal and behavioral data, creating a plausible foreign-control risk even without proof that China had already demanded every category of U.S. data.
CheckThe Supreme Court record describes collection including location, device, contacts, messages, viewing behavior, keystroke patterns, and network data, and held Congress's predictive national-security judgment was supported by substantial evidence.
David Sacks
Sacks supplied the debate's most valuable objection, but his statutory parade of horribles omitted procedural gates that materially narrowed the broader designation power.
Assumptions and fact checks
National-security unanimity should increase, not reduce, skepticism about new executive power.
Why it mattersThe Patriot Act analogy is imperfect, but scrutiny of vague definitions and secret evidence is an important discipline when speech and platform access are at stake.
Public proof of actual completed espionage was necessary before Congress could act.
Why it mattersCongress may regulate a well-supported prospective national-security risk; it need not wait for the feared harm to be completed and publicly documented.
The bill defined foreign-adversary control to include a person 'subject to the direction or control' of a covered foreign person or entity.
CheckThat phrase appears in the statutory definition exactly as Sacks described.
An Attorney General could use that phrase alone to sweep a domestic platform into the law.
CheckFor companies outside the named ByteDance/TikTok group, the statute also required a presidential finding of a significant national-security threat, advance public notice, and a report to Congress. The Attorney General enforced the prohibition rather than unilaterally making the whole designation Sacks described.
David Friedberg
Friedberg's anti-paternalism is coherent, but he treats each technical control as if the controls can be cleanly separated from ownership and enforceability—the very point under dispute.
Assumptions and fact checks
Audits, U.S. hosting, and platform-level microphone controls could address the core risk without ownership separation.
Why it mattersThose tools address parts of the risk, but they do not necessarily resolve control over code, recommendation logic, data flows, or compelled cooperation by the parent company.
Reciprocity with China's platform restrictions is not, by itself, a sufficient American policy principle.
Why it mattersA liberal democracy should justify restrictions through its own security and rights framework, not merely mirror an authoritarian state's choices.
Can a video model train on publicly accessible internet content without licenses, or do copyright and platform rules require permission?
Original point: YouTube is public, and teaching a model from public material should not be treated as creating a derivative work.
What everyone argued
Jason Calacanis
Jason argued that OpenAI would need licenses to build a commercial model from creators' video and that OpenAI's licensing deals undercut its claim that public-web training needs no permission.
David Sacks
Sacks favored fair use for training on publicly available material while also saying the CTO's hesitation likely reflected litigation risk and that YouTube probably appeared somewhere in Sora's internet-scale data.
David Friedberg
Friedberg argued that models should be able to learn from anything openly viewable on the internet, just as people can, and rejected the claim that training itself creates a derivative work.
Winner circle
Sacks gives the best answer because he separates a plausible fair-use principle from OpenAI's litigation and disclosure risk. Friedberg is directionally supported by later transformative-use rulings, but his 'open internet means usable' rule is too broad. Jason correctly identifies access and creator-market concerns, yet his claim that licenses are simply required is also too absolute. The law that emerged is annoyingly lawyerly but more accurate: training, acquisition, and outputs must be analyzed separately.
Commentary
Jason Calacanis
Jason saw the creator-market problem clearly, but he turned a strong factor into a universal legal rule.
Assumptions and fact checks
A commercial video generator trained on expressive video will usually need licenses.
Why it mattersLicensing and output competition can weigh strongly against fair use, but current law does not impose Jason's categorical rule for every training act and every lawfully accessed work.
YouTube treats unauthorized scraping of creator content as a violation of its Terms of Service.
CheckYouTube's 2024 creator-tools statement explicitly says unauthorized third-party scraping violates its terms and that it uses technical measures to block scrapers.
David Sacks
Sacks wins by being the least categorical: training may be fair while acquisition, platform terms, or substitutive outputs can still create exposure.
Assumptions and fact checks
Training that learns patterns rather than reproducing source works should generally receive fair-use protection.
Why it mattersThat reasoning carried in the 2025 Anthropic and Meta rulings on their records, but it is not a universal safe harbor.
Murati's hesitation makes YouTube training likely.
Why it mattersHer answer and later internet-scale disclosure make the inference plausible, but no cited public source confirms YouTube specifically.
OpenAI later disclosed that Sora used a mix of selected public data, proprietary partnership data, and custom in-house data.
CheckOpenAI's Sora system card describes exactly that mixture, including public machine-learning datasets and web crawls, but it still does not specifically confirm YouTube as a source.
David Friedberg
Friedberg had the best technological intuition but flattened several legally distinct acts into the word 'watchable.'
Assumptions and fact checks
Anything publicly viewable should be legally available for model training.
Why it mattersPublic visibility does not erase copyright, access restrictions, contracts, or the four-factor fair-use inquiry.
At least one 2025 federal court held that training an AI model on lawfully obtained copyrighted books was fair use while treating a pirated central library as a separate infringement problem.
CheckThe Bartz v. Anthropic order found the training use transformative and fair but did not excuse the acquisition and retention of pirated copies.
Were Florida's cultivated-meat ban and the TikTok divestiture law two versions of regulatory capture, or materially different risk problems?
Original point: Florida's protection of ranchers and the push against TikTok both showed incumbents using government to block a competing product under an unproven-threat story.
What everyone argued
Chamath Palihapitiya
Chamath agreed that Florida's meat law was cronyism but insisted TikTok was categorically different because one dispute concerned domestic ranchers versus a federally regulated food product, while the other concerned a foreign-controlled surveillance and influence surface.
David Friedberg
Friedberg argued that both episodes featured incumbents invoking speculative harm to block innovation and consumer choice. He warned that Florida's law would encourage other ranching states to ban a product before it could compete and predicted affected companies would seek federal preemption.
Winner circle
Chamath wins the comparison. He grants the common lobbying incentive but correctly separates the risk models: domestic agricultural competition is not foreign control of a mass data and recommendation system. Friedberg wins the narrower Florida argument—its official rationale and categorical design look protectionist—but stretches that success too far when he treats TikTok as the same case. The shared word 'ban' was doing more work than the underlying facts.
Commentary
Chamath Palihapitiya
Chamath wins by refusing a label-driven analogy. 'Regulatory capture' can describe political incentives without proving the policies have the same evidence or proportionality.
Assumptions and fact checks
Industry lobbying can be present in both cases without making their underlying risks equivalent.
Why it mattersMotive and mechanism are separate. Competitors may lobby around both laws, but a federally reviewed food product and a foreign-controlled communications platform present different failure modes.
Florida enacted an outright ban on manufacturing for sale, selling, holding for sale, or distributing cultivated meat, backed by misdemeanor and business penalties.
CheckFlorida Statute 500.452 contains those prohibitions and penalties; the governor signed the measure on May 1, 2024.
David Friedberg
Friedberg is plainly right on Florida and plainly too eager to cash that win out against TikTok. His analogy attacks a shared political smell, not the strongest version of the national-security case.
Assumptions and fact checks
Florida's cultivated-meat ban was principally an incumbent-protection and culture-war measure rather than a tailored food-safety rule.
Why it mattersThe categorical structure and official signing statements focused on defending farmers, ranchers, 'real beef,' and resistance to global elites, while federal safety oversight already existed.
Because regulatory capture may influence both policies, the TikTok and cultivated-meat cases are intellectually similar enough to judge together.
Why it mattersThe analogy ignores different risks, actors, evidence, and remedies. Similar lobbying incentives do not make a food-market ban equivalent to conditional separation from foreign platform control.
Federal agencies already had a regulatory pathway for cultivated meat before Florida banned it.
CheckFDA completed premarket consultations for cultivated chicken, and USDA FSIS established inspection and labeling responsibilities for cell-cultured meat and poultry products.
Federal preemption would likely knock out Florida's ban.
CheckIn March 2026, the Eleventh Circuit held that the Poultry Products Inspection Act did not preempt Florida's product ban and affirmed denial of a preliminary injunction.
Chamath was right to focus on code and control rather than flags on a cap table, but he jumped from a difficult audit problem to an absolute claim that no auditable remedy could work.