Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California’s Broken Elections

Source: Transcript of an episode of the All-In Podcast. The apparent principal speakers are Jason Calacanis, Chamath Palihapitiya, David Sacks and David Friedberg. The transcript contains phonetic or inconsistent renderings including “Chimath,” “Saxs,” “Freeberg,” “Daario/Daria,” and several other names. Speaker names are normalized here where the identity is clear from context; uncertain names remain flagged for verification.

Source context: The discussion covers five broad areas:

1. Anthropic's new “Fable 5” model, model restrictions, surveillance, AI safety and open-source competition.
2. Bernie Sanders' proposed public ownership of AI companies, sovereign wealth funds and the economic impact of AI.
3. The hosts' recent Liquidity event and observations about venture-capital returns.
4. Inflation, interest rates and the economic effects of the Iran war.
5. The Los Angeles mayoral primary, California voting rules and allegations or suspicions about election integrity.

The transcript internally refers to 2026, a June 1 Sanders proposal, May inflation data and an upcoming September 13–15 All-In Summit. These dates and all related current-event claims require independent verification before publication.

The source has severe duplicated-caption noise: most phrases appear two or three times consecutively. These duplicates have been removed in this record.

There are also numerous probable transcription errors or uncertain names/terms, including:

• “Fable 5”
• “Mythos”
• “Opus 4.8”
• “Nesh” of Palo Alto Networks
• “boweapons”
• “Daario/Daria”
• “Carlson Brothers” / “Collisons”
• “Saro Milthorn”
• “Nipia/Nitia Ramen”
• “Sarah Frier/Fry/Prior”
• “Thomas Leafant”
• “Chris Hoe”
• “Athena Wow”
• several garbled phrases and figures.

No questionable claim has been silently corrected. Material uncertainties are listed again under D. Verification list.

Jason introduces what the transcript calls Anthropic's “Fable 5”, saying it was released on Tuesday and tops nearly every benchmark. He says its tokens cost approximately twice as much as “Opus 4.8,” although the model might require fewer tokens overall because it is more capable.

He connects the release to an earlier model called “Mythos,” which he says Anthropic did not publicly release in April because of unusually strong hacking capabilities. According to Jason's account, “Nesh,” identified in the transcript as the CEO of Palo Alto Networks, had access to Mythos and said it was “the real deal”; Jason says Palo Alto Networks used it to identify or close vulnerabilities internally.

Jason says Fable 5 restricts subjects including bioweapons and hacking. He describes two policies that generated developer backlash:

• Anthropic reportedly retains prompts entered into the model for at least 30 days.
• If the system concludes that a user is doing frontier AI research—using Anthropic's model to improve or create another competitive model—it may route or downgrade that user's access without initially telling them.

Jason says these policies were described deep inside a 319-page document and prompted substantial criticism on X. He says Anthropic subsequently changed the policy so the safeguards or downgrades would be more visible.

Chamath praises the model's technical quality and says Anthropic deserves credit for continuing to push the frontier of closed models.

His concern is that Anthropic has revealed a broader operating model: it can inspect prompts, classify what a user is doing and decide what capabilities or responses the user should receive before generating an answer.

For an individual, Chamath says this creates a censorship risk.

For a company, he regards the problem as more serious. An ordinary employee, scientist or executive might unknowingly trigger a safeguard. He gives examples including:

• a scientist using cloud APIs;
• a business executive;
• someone conducting molecular or scientific research.

If that causes a company to lose access to an important model, the company could suddenly lose a source of competitive differentiation.

Chamath argues that enterprises therefore need an explicit AI-governance strategy addressing:

• who controls their AI infrastructure;
• who may learn from their information;
• whether a single model vendor represents a single point of failure;
• whether they have sufficient model and provider diversity.

He argues for broad diversity rather than dependence on one provider.

One speaker adds that Anthropic's distinguishing characteristic is that it tells users what it is doing, but says that the truth itself is uncomfortable: both censorship risk and enterprise-governance risk remain undesirable.

Jason asks Friedberg whether, as a business owner doing advanced research, he worries about being “rugpulled” by model companies or having proprietary work incorporated into their models.

Friedberg says the terms of service state that customer work cannot simply be taken and trained into the provider's model, but when asked whether he trusts that promise, he says he is not sure.

Despite those concerns, his company has chosen to use the systems because they are highly valuable in proprietary genomic research. He describes work involving:

• identifying genes and gene variants;
• estimating their effect on living organisms;
• agricultural and plant genomics;
• RNA-guide design for gene-editing tools;
• predicting whether a gene will produce a particular phenotype;
• designing genetic constructs used to produce specific proteins.

He says these tasks traditionally consume substantial scientific time but had become much faster with advanced models.

Friedberg says this changed over the preceding weeks. Providers began restricting work because of theoretical bioweapon risks. In his telling, legitimate scientific work now trips the same safeguards intended to stop dangerous biological applications.

His likely response is to use open-source models and run them locally.

He then makes a geopolitical argument: the strongest open-source models available, in his assessment, are Chinese, while American open-source offerings are weaker. He says restrictions imposed by American frontier labs are therefore driving startups and large enterprises toward Chinese open-source models.

Friedberg predicts that companies will increasingly take a foundational open model, combine it with proprietary data and create internal specialized models. His company, for example, expects to create its own genome language/prediction model.

His wider argument is that political pressure and restrictive safety policies in the United States could unintentionally advantage Chinese AI providers. AI capability cannot simply be abolished. If one country restricts itself while another continues advancing, the latter gains the advantage.

Sacks says he had argued about eight months earlier that Anthropic was pursuing a sophisticated regulatory-capture campaign based on fear. He says this once sounded extreme but now believes it is moving toward consensus.

He argues that developer outrage over Fable 5 is not simply about safety restrictions. In his account, it reflects a violation of trust.

Sacks claims Anthropic now retains every prompt and output sent to its “Mythos-class” models for 30 days, including for enterprise customers that previously negotiated zero-data-retention agreements.

He stresses that modern agent systems send far more than a short prompt. They may send:

• files;
• memories;
• stored user information;
• large context windows.

Therefore, he argues, a 30-day retention policy can encompass a substantial amount of business or personal information.

Sacks claims this information is then used to build a profile of the user, classify that user and determine what model capabilities the user should receive.

He says the original design could silently:

• move the user to a weaker model;
• rewrite the user's prompt behind the scenes;
• return a weaker or “nerfed” answer;
• continue charging for the product the user believed they were receiving;
• fail to disclose the downgrade.

He says Anthropic's narrow walk-back was to disclose when a downgrade occurs, rather than abolishing the downgrading itself.

Sacks cites reported examples where apparently ordinary research questions triggered restrictions:

• a question involving mitochondria;
• Ben Thompson of Stratechery asking about cancer risk and GLP-1 drugs.

He sees these examples as evidence that Anthropic's interpretation of risky biological or scientific content is very broad.

His broader prediction is that powerful model companies could decide who becomes an AI “have” and “have-not,” controlling access to information based on profiles and proprietary criteria.

Another speaker describes that prospect as “very Orwellian.”

Chamath identifies a separate risk: model providers may eventually have commercial incentives to favor some companies over competitors.

He gives hypothetical examples:

• if Anthropic had a strategic relationship with Novartis rather than Eli Lilly, information relating to competing GLP-1 products could conceivably be shaped;
• if a model company had a relationship with JPMorgan rather than Citibank, financial questions could theoretically be answered in ways that favor one institution.

His concern is not that the transcript establishes that this is happening. His argument is that the model architecture and lack of transparency could make such influence hard for users to detect.

Chamath then raises an auditability problem. If a user believes a response was intentionally altered, a regulator should theoretically be able to reconstruct the exact model run and determine:

• what model was used;
• what prompt actually reached it;
• whether the system altered the prompt;
• whether a weaker model was substituted;
• why a particular output resulted.

He doubts whether model providers will preserve a sufficient “fingerprint” or trace to make such investigations reliable.

He imagines an emerging company offering Anthropic substantial equity in exchange for strategic preference or routing. He says the model provider might then evaluate not only economics but a company's philosophy, morality or social positions. His fear is that companies aligned with the provider's preferred worldview could receive hidden advantages over others.

Jason offers a free-market counterargument. If many competing models remain available, he asks why Anthropic's conduct necessarily requires intervention.

He compares it with early search engines that accepted paid inclusion. Those products became worse for users, while Google won by offering better search results. His proposed logic is that consumers can similarly abandon an AI provider that manipulates or degrades answers.

Sacks responds that competition only solves the problem if alternatives remain legal and available.

He says Dario Amodei had simultaneously called for a new model-approval regulator comparable to the FAA or FDA. In Sacks' interpretation, Anthropic is not merely making restrictive choices for itself; it wants government rules that could restrict competitors as well.

He describes this as potentially anti-competitive regulatory capture: Anthropic could impose expensive compliance obligations that closed frontier labs can satisfy but which open-source models cannot.

The hosts characterize regulation of open source as a possible way of “scuttling” or “sandbagging” local open-source models and preventing users from escaping closed-provider restrictions.

Chamath says he and a partner bought approximately 2,000 acres in Arizona two years earlier, secured zoning and approval for a 2-gigawatt data center, and initially expected to sell or transfer the project to large infrastructure or technology companies such as Blackstone, Brookfield or Google.

He says he has reconsidered because open-source AI may require independent compute infrastructure.

He says he has also made an offer involving another gigawatt elsewhere, envisioning as much as 3 GW dedicated to deep, liquid access to open-source models.

The obstacle is capital. Chamath says that when the project began he thought development might cost roughly $4–5 billion, whereas he now estimates a gigawatt at roughly $100 billion, implying approximately $300 billion for 3 GW.

His argument is that open-source software alone does not ensure meaningful competition. If almost all power and compute remain concentrated behind the largest proprietary labs, open-source models may exist technically but lack sufficient infrastructure to compete.

The discussion connects this to a broader concern that the rest of the world, particularly China, may continue using unrestricted models even if US companies cannot. Participants specifically mention perceived Chinese advantages in biotech, materials science and industrial systems.

Later, Chamath reiterates that without access to hundreds of megawatts and ultimately gigawatts of power, creating an open-source frontier laboratory is largely academic.

Friedberg describes an open-source genomics model associated in the transcript with the Ark Institute and the “Collisons” or “Carlson Brothers” [transcript unclear].

He says the project ingested large quantities of genomic data and learned patterns in DNA in a way analogous to a language model learning valid English.

For plant breeding, his company can provide a gene variant and ask the model to estimate whether its sequence is likely to be functionally useful. The model evaluates the probability of DNA letters occurring together across organisms.

He says this model is already being used as an input in his company's plant-breeding program and cites it as evidence that philanthropy/community funding can create useful open alternatives to closed commercial models.

Jason deliberately attempts to present the strongest version of Anthropic's safety case.

He says that if Dario genuinely believes he has built an unusually powerful and potentially dangerous system, a cautious rollout is rational:

• initially restrict the model;
• provide it to trusted security partners;
• use those partners to close vulnerabilities;
• release it more broadly only after evaluation;
• monitor dangerous applications;
• allow users who dislike the restrictions to remain on older models or use competitors.

Jason says Anthropic may simply have communicated this approach badly.

He simultaneously criticizes Amodei's public rhetoric as hyperbolic and suggests that Amodei's worldview nevertheless attracts elite AI researchers. Jason estimates that perhaps 80–90% of talented AI researchers share similar beliefs about the profound or quasi-divine significance and risks of advanced AI. He contrasts them with libertarian researchers attracted to Elon Musk/xAI and people motivated more by compensation at firms such as OpenAI. These percentages and characterizations are opinions requiring verification if used as factual claims.

Jason also criticizes Meta's execution with Llama. He argues Meta could have deliberately commoditized frontier AI by releasing a top-tier open model, thereby eliminating competitors' margins. The hosts describe Meta's failure to establish a neck-and-neck open model as a strategic “fumble.”

Friedberg says safety concerns are not imaginary and compares AI with nuclear physics.

Splitting the atom made possible both:

• cheap nuclear energy;
• atomic weapons.

He says AI similarly creates capabilities with both beneficial and dangerous applications.

He divides weaponization risk into three broad groups:

1. cyber weapons;
2. physical weapons;
3. biological weapons.

The same underlying capabilities that may help design weapons can also:

• support cancer research;
• increase food production;
• build more powerful software;
• give workers greater leverage;
• expand entrepreneurial opportunity and income.

His central argument is therefore not that safeguards are unnecessary, but that safety policy should target dangerous outcomes rather than denying access to general-purpose intelligence.

Friedberg argues that society already possesses:

• product-liability law;
• laws against weapons;
• laws against hacking and cyberattacks;
• laws against bioweapon development.

He supports tracking, detection, guardrails and stage gates, but says restricting the underlying model itself sacrifices economic and scientific benefits.

He states a broader technological principle: once a capability is technically possible and publicly available, somebody will eventually attempt it. Open models are already distributed, copied and modifiable. He compares trying to eliminate them with trying to eliminate the internet, the typewriter or Gutenberg's printing press.

Jason offers fertilizer regulation as an analogy.

After the Oklahoma City bombing, he says governments placed stronger controls around fertilizer, including identification requirements and tracking. His steelman is that Anthropic, as the seller of a potentially dual-use capability, has a legitimate private-company interest in reducing liability and misuse.

He emphasizes that this is an argument he is testing, not necessarily his own settled view.

Chamath takes the analogy in another direction. If fertilizer requires identification, he asks why Anthropic does not use true KYC—Know Your Customer procedures.

He suggests a legitimate scientific enterprise should theoretically be able to identify itself, disclose its personnel or provide financial/security assurances and then receive unrestricted scientific access.

The hosts dispute whether an account plus a credit card constitutes even basic KYC.

Jason tests the system while the discussion is taking place.

He says he asks about fertilizer-bomb regulation without asking how to manufacture a bomb. The system reportedly analyzes his identity and legitimate-policy context but still moves him away from Fable to a weaker model.

He later asks about historical attempts by civilians to construct nuclear weapons and then asks which components are required and how they are regulated. He again reports being switched to “Opus 4.8.”

The displayed message reportedly tells him Fable 5 applies special safeguards to cybersecurity and biology, acknowledges that normal safe content may be flagged, and says the restrictions allow Anthropic to release Mythos-level capability in other areas sooner.

Friedberg says experiences like this will cause his company to reduce its use of Anthropic and switch to alternative models. Another host replies that this market solution fails if regulators subsequently ban those alternatives.

The hosts discuss a letter supporting mandatory nucleic-acid synthesis screening and recordkeeping.

The transcript says laboratories producing synthetic DNA or RNA already compare orders against databases intended to identify dangerous sequences such as those associated with pathogens.

A speaker says this process derives from a 2009 International Gene Synthesis Consortium agreement under which major laboratories voluntarily adopted safeguards.

The proposal discussed would codify such screening into mandatory rules after approximately 15 years of voluntary practice.

This is offered as an example of Friedberg's preferred principle: intervention can occur when digital instructions are translated into a physical biological product rather than preventing researchers from accessing general scientific reasoning.

The speakers say such screening is already fast enough not to materially obstruct legitimate laboratory research and could provide a stronger model for AI safeguards: intervene downstream, where dangerous outputs become physical, rather than broadly blocking questions about biology or medicine.

The section ends with the question: “Who gets to watch the watchman? Who put Dario in charge of all this?”

Jason introduces what he describes as a June 1 New York Times op-ed by Senator Bernie Sanders titled “AI is a public resource. You should own half of it.”

According to Jason's summary, Sanders also proposed an American AI Sovereign Wealth Fund Act involving a one-time 50% levy on the stock—not merely profits—of large AI companies including OpenAI, Anthropic and xAI.

Jason says those shares would be placed in a government sovereign wealth fund and would give the public voting rights and equal board representation at each company.

Jason quotes Sanders' stated rationale: AI is based on the collective intelligence embodied in humanity's books, songs, journalism, science and code, which Sanders characterizes as having been taken by extremely wealthy technology owners.

The hosts observe that the idea could appeal across unconventional political lines: Sanders on the left, populist figures such as Steve Bannon, and supporters of government sovereign-wealth-fund investments associated with President Trump.

Sacks initially rejects Sanders' proposal because he considers forced seizure of equity a straightforward confiscation of private property and a dangerous precedent.

However, he says he understands the argument for some voluntary mechanism through which ordinary citizens could participate in the upside of AI.

His reasoning is political.

AI CEOs themselves have repeatedly warned that AI may displace enormous numbers of workers. Sacks specifically attributes a 50% job-loss warning to Dario Amodei.

At the same time, those companies trained models using knowledge created by generations of humans who published work online without expecting that the resulting systems would later compete with their own labor.

Sacks argues that if technology leaders tell the public both:

• “we trained on humanity's collective work,” and
• “our technology may eliminate your job,”

then ordinary voters will naturally ask what they receive in exchange.

He therefore treats Sanders' proposal as a predictable political reaction to the AI industry's own rhetoric.

Sacks says he personally does not believe an AI job apocalypse is occurring.

He cites a May employment report which he says showed:

• 172,000 new jobs;
• more than twice economists' expectations;
• 4.3% unemployment;
• strong construction hiring;
• software-development employment at a three-year high.

The transcript contains a garbled phrase around “15% of Trump,” which should not be used without checking the source.

Sacks distinguishes the public positions of Elon Musk, Dario Amodei and Sam Altman.

• Musk, in Sacks' description, predicts an eventual abundance economy in which AI and robots produce so much that people no longer need to work and may receive “universal high income.” Sacks treats this as a potentially distant, utopian end state.
• Amodei, according to Sacks, made the more immediate claim that 50% of entry-level knowledge-worker jobs could disappear within one to five years.
• Altman, according to the discussion, previously warned about job losses but more recently acknowledged that employment data had not developed as he expected.

The hosts stress that these are materially different forecasts even if the public often hears only the simplified message that “AI means job loss.”

Sacks argues that AI companies have therefore created political risk for themselves by repeatedly telling the public that their own technology is dangerous and economically disruptive.

His frustration culminates in the argument that if executives genuinely believe they are building a “Frankenstein” capable of destroying society, they could simply stop building it rather than demanding that society regulate them while they continue racing ahead.

He also alleges a contradiction at Anthropic: the company reportedly warned that recursive self-improvement could threaten humanity while also hiring Andrej Karpathy [transcript attribution/name to verify] to work on recursive self-improvement. The hosts characterize this as hypocrisy.

Sacks cites Ben Thompson's interpretation that Anthropic's warning about recursive self-improvement may have helped justify its decision to restrict Fable users from researching machine learning, AI or chip design. This is presented as Thompson's analysis, not established fact.

The hosts note that OpenAI and Anthropic are described in the discussion as public-benefit corporations.

They explain a public-benefit corporation as one whose board must balance shareholder interests against an expressly stated public mission rather than maximizing shareholder value alone.

That leads to a provocative argument: if such companies say their purpose is to benefit humanity, perhaps public ownership or use of profits to reduce the national debt could be consistent with their claimed mission.

Sacks nevertheless maintains that compulsory confiscation is problematic, although the discussion becomes more open to public participation where:

• a company trained on humanity's knowledge;
• claims it will cause mass unemployment;
• presents itself as serving a public benefit;
• yet restricts public access to the resulting intelligence.

Later Sacks goes further, saying he might accept Sanders' idea in the specific case of a public-benefit corporation that claims it will cause massive job loss, trained for free on humanity's knowledge and then gatekeeps the output.

The hosts also propose another remedy: force such models to become open source rather than seize their equity. They connect this to the original rationale they attribute to Elon Musk in co-founding OpenAI—to prevent one company such as Google from monopolizing advanced AI.

Friedberg rejects conventional wealth taxes and compulsory equity seizure but proposes a different route for widespread ownership.

He says the United States already has a social safety-net system through Social Security, but describes it as a defined-benefit structure rather than an individual investment-account system.

He contrasts it with defined-contribution plans, where money is invested in an account and the individual owns the accumulated assets.

Friedberg says the Social Security trust fund effectively holds a special US Treasury obligation and claims the federal government owes it approximately $4 trillion.

His proposed reform would:

1. allow Social Security to own equities rather than only Treasuries;
2. convert the system toward individual accounts;
3. invest part of the capital in successful American companies;
4. include AI companies;
5. place resulting ownership interests economically in the accounts of American citizens.

He therefore supports a US sovereign wealth fund created through investment rather than confiscation.

Chamath says the broad concept resembles systems used successfully in Canada and Australia and calls Friedberg's idea theoretically excellent, while joking that its political difficulty makes implementation unlikely.

Friedberg forcefully rejects the idea that AI is currently destroying employment.

He divides a business into two sides:

• costs;
• revenues.

AI can reduce some costs by automating human work, but Friedberg argues that this effect is comparatively small.

The much larger opportunity is on the revenue side.

If one engineer can do 100 or 1,000 times more, a company can build vastly more products. He gives examples spanning:

• agricultural seeds;
• boats and ships;
• software;
• clothing.

In Friedberg's view, greater productive capacity allows companies to expand rather than simply cut workers.

He says his own product and engineering team had just requested approximately 15 additional engineers because AI had opened opportunities that previously could not be pursued.

His conclusion is that AI-driven productivity is a boom that creates demand for labor rather than a bust that eliminates it.

He summarizes: “everyone's gonna realize soon that this is a boom, not a bust.”

Friedberg cites Anthropic's own growth—described in the transcript as a “$40 billion revenue business”—as evidence that customers are using AI to create new products rather than merely firing support staff. This figure requires verification.

Chamath compares AI with the internet.

The internet generated extraordinary margins for companies such as Google and Facebook because adding one more search user or social-network user had close to zero marginal cost while network effects and advertising created significant additional revenue.

AI is different.

Every additional AI workload requires:

• GPU capacity;
• electricity;
• memory;
• infrastructure.

Therefore the marginal cost of serving additional AI usage is substantial rather than nearly zero.

Chamath says this creates a stronger argument for public participation because AI companies depend heavily on national infrastructure and energy systems.

He compares the situation to private transportation companies operating on federally financed interstate highways: if public infrastructure makes a private oligopoly possible, government may have greater negotiating leverage over the resulting economic value.

He says that if he ran a sovereign wealth fund, he would negotiate aggressively and jokes that he might end up owning 75% of AI companies.

His key economic observation is:

the incremental cost of performing AI is large, whereas the incremental cost of the previous internet model approached zero.

The hosts see that distinction as important to any future policy regarding infrastructure subsidies, regulation or public ownership.

The hosts claim that the US government under President Trump has already accumulated equity or equity-like interests in companies or sectors involving:

• critical metals and materials;
• Intel;
• possibly Micron;
• AMD;
• potentially SpaceX.

They speculate about using those holdings as seed assets for a sovereign wealth fund and potentially adding Anthropic or OpenAI.

These assertions and the exact ownership percentages are unclear in the transcript and require verification.

The discussion becomes jokingly expansive, with the hosts calculating hypothetical ownership of future “trillicorns”—trillion-dollar companies—and joking that portions of enough such firms could pay off the national debt or finance purchases of foreign territory. These remarks are humorous rather than serious policy proposals.

The hosts discuss recent interviews and presentations from their Liquidity conference.

They praise OpenAI CFO Sarah Friar [transcript spelling varies], describing her appearance as particularly impressive. One host jokes that people at the event were discussing her as a potential future OpenAI CEO, with Sam Altman as chairman; others push back on the speculation but agree that the two appear to work well together.

They also praise a presentation by Thomas [Laffont/Leafant — transcript unclear] on venture-capital outcomes.

The central data they recall:

• approximately 8% of unicorns reached decacorn status;
• approximately 13% of decacorns reached $100 billion;
• approximately 31% of $100 billion companies reached $1 trillion.

They interpret this as a striking inversion: once a company reaches extreme scale, its probability of another 10× increase may actually rise.

One host speculates, without data, that perhaps 60% of trillion-dollar firms could reach $10 trillion.

The hosts thank event partners and discuss their experience around Liquidity.

They mention:

• Thomas Keller and the French Laundry;
• EY sponsorship of a meeting hub;
• the New York Stock Exchange hosting a speaker dinner at the French Laundry;
• Niagen hosting a wellness lounge including recovery IVs and acting as an NAD-related partner;
• Jake Paul attending, whom several hosts praise as thoughtful and a strong investor;
• Athena as another event partner.

They discuss Thomas Keller's watch and favorite dishes at the French Laundry, including duck, Miyazaki beef, tapioca pearls and “oysters and pearls” with caviar.

They also describe a nine-hole golf scramble with approximately $100 wagered per hole against “Chris Hoe,” identified in the transcript as the chief business officer of Athena. One host says they lost roughly three units, or $300.

The hosts say Liquidity sold out and received roughly 3,000 applications.

They promote the All-In Summit for September 13, 14 and 15, saying Friedberg is developing unusually ambitious production elements.

Chamath distinguishes the two events:

• Liquidity is deliberately curated for major capital allocators to convene, debate, learn and form relationships.
• Summit is broader and more festival-like—a party or “vibe” combining learning and entertainment.

He says capital shapes what happens in the world, so Liquidity's participants and program should be selected carefully rather than simply sold to anyone able to buy a ticket.

Jason introduces May inflation data, describing both CPI and PPI as unexpectedly hot.

According to the transcript:

• CPI was 4.2% year over year, the highest since April 2023;
• PPI was 6.5% year over year, the highest since late 2022;
• Polymarket showed a 21% probability of inflation reaching 5% in 2026;
• Polymarket showed a 49% probability of a Federal Reserve hike that year, up from below 10% before the Iran war;
• the European Central Bank had just raised rates by 0.25 percentage point, its first hike since September 2023.

Every figure requires verification.

Friedberg says the Iran war created an energy shock that contributed to the inflation numbers.

However, he sees a deeper problem in:

• excessive government spending;
• persistent inflation;
• the eventual need for higher interest rates.

He suggests that under a Kevin Warsh Federal Reserve, overnight rates above 5.5% or even around 6% would not be inconceivable.

In further remarks, Friedberg says the basic cause of US inflation and even wealth inequality is excessive government spending. Politicians win elections by promising voters more benefits, causing spending commitments to accumulate over generations.

He expects higher interest rates to be one late-stage consequence.

Chamath says the inflation report is concerning and hotter than he expected.

He argues that Chinese behavior has so far helped smooth global energy demand, keeping oil below the extreme levels it might otherwise have reached.

He also says the situation has accelerated adoption of practical energy sources, particularly solar, because solar is comparatively simple, passive and subject to fewer permitting and clean-air constraints.

His downside scenario is that if China exhausts available reserves and must buy an additional roughly 3 million barrels per day on the spot market, oil could move well beyond $100 and potentially into a $150–$200 range.

Such a move would increase inflation not only through direct US energy prices but through oil-dependent inputs embedded throughout the global supply chain.

Chamath therefore treats the PPI number as a warning to find an off-ramp from the Iran conflict sooner rather than later.

He leaves open the possibility that President Trump may have information from discussions with Xi Jinping about Chinese inventories or supply slack that would permit a longer military strategy, but stresses that the public does not know whether such an understanding exists.

Sacks says he does not have inside information about Iran.

He agrees broadly with Chamath's economic analysis but notes that the PPI release was reportedly close to market expectations. He points to the market rising rather than falling as evidence that investors were not treating the inflation number as a major surprise.

Normally, he says, unexpectedly hot inflation should hurt stocks by increasing expectations of higher rates.

Jason calls starting the Iran war a “huge colossal error” and describes the downstream economic effects as disastrous. He hopes for a quick resolution rather than a conflict lasting one, two or three years, saying the United States cannot become trapped in another “forever war.”

Another speaker notes the Nasdaq was reportedly up roughly 2.5% despite the inflation print and interprets that as possible evidence that markets expect some resolution of the conflict. The speakers acknowledge this is only an interpretation.

The final major discussion concerns the Los Angeles mayoral primary.

Jason asks whether the election system is corrupt and whether candidate Spencer Pratt was denied a legitimate place in the runoff.

Friedberg responds provocatively that “there is no election,” arguing that the system has become one of appointment rather than genuine election.

The transcript presents the following alleged voting percentages:

• Spencer Pratt: 35%
• Karen Bass: 29%
• “Nipia/Nitia Ramen” [name unclear]: 26%

• Bass: 38%
• Pratt: 28%
• Ramen: 20%

• Ramen: 37%
• Bass: 35%
• Pratt: 19%

The speakers emphasize that Pratt's share falls dramatically in the late-ballot population while Ramen's rises sharply. They treat that divergence as statistically suspicious and say a map showed a large concentration of incremental Ramen votes around Skid Row.

These figures and the candidate identity must be independently checked before any publication.

Friedberg initially distinguishes fraud from what he sees as flawed but legal election rules.

He cites a US House Committee on House Administration report from May 2020 discussing California's 2018 midterm elections.

He says California Assembly Bill 1921 legalized broad ballot harvesting.

His description of the resulting system includes claims that:

• third parties may collect ballots from voters without a family relationship;
• California sends mail ballots to every registered voter;
• voter-registration rules do not require sufficient proof of citizenship;
• a gym-membership card can allegedly be used in registration;
• voters do not face adequate ID requirements;
• third parties can collect large numbers of ballots.

Some of these claims may be incomplete or inaccurate legally and require close verification.

Friedberg's key distinction is that he does not initially claim the system necessarily operates through illegal conduct.

Instead, he says the laws themselves create mechanisms that can convert elections into de facto appointments while remaining within the rules.

He regards this as an erosion of the democratic principle of one eligible person casting one vote.

When Jason says this sounds like fraud, Friedberg replies:

“This is not fraud. This is the way the laws are set up.”

He argues that “fraud” means breaking the law, whereas his concern is that problematic conduct has been legalized.

Friedberg acknowledges that each voting-access reform can be defended individually:

• increase participation;
• make voting easier;
• provide more channels;
• improve access.

His concern is cumulative.

He borrows the insurance concept of adverse selection: whenever a system contains an exploitable asymmetry, people motivated to abuse it will exploit it more aggressively than ordinary good-faith participants.

He compares this to hackers identifying vulnerabilities in a network.

His argument is that California's voting rules may have created so much “lubrication” that organized political operations can use the rules far more effectively than ordinary voters, transforming access-enhancing reforms into a machine for political control.

Sacks takes a much stronger position and repeatedly alleges that the outcome is crooked or fraudulent.

He says California mails ballots to roughly 23–24 million registered voters, while only around 9.5 million vote, leaving approximately 14 million unused ballots.

He recounts an anecdote from a friend who manages apartment buildings in Los Angeles and says unused ballots accumulate in common areas, including ballots addressed to former residents.

From this he alleges dirty or outdated voter rolls.

Sacks further claims:

• no adequate voter ID is required;
• registration itself lacks adequate ID;
• signature verification accepts a “40%” match [meaning unclear];
• a voter can use an X rather than a signature if witnessed;
• witnesses are not meaningfully registered or tracked;
• there is no adequate chain of custody;
• a ballot harvester could theoretically complete ballots, add marks for voters and witnesses, and submit them.

These are allegations or descriptions by Sacks and must not be published as established facts without legal and documentary verification.

He also claims Governor Gavin Newsom recently signed legislation making election audits more difficult.

Sacks argues that California already experiences substantial proven fraud in programs involving Medicaid, unemployment benefits, EBT, hospice services and homelessness programs, and asks why political actors would not similarly exploit weaknesses in election law.

That analogy is argumentative and does not itself establish election fraud.

Sacks ultimately says conduct might be “legalized” but still fraudulent in a broader moral sense.

Jason pushes back.

He argues that swinging 5,000–10,000 votes through an illegal operation would require coordination among dozens or hundreds of people and create many opportunities for detection.

The hosts discuss videos reportedly filmed by Nick Shirley on Skid Row and allegations that homeless or drug-addicted residents were paid a few dollars to vote.

A speaker cites Brenda Lee Brown Armstrong [name and case to verify] as having pleaded guilty in 2026 in a vote-payment case.

The transcript becomes garbled around this point, including an unexplained reference to an election being won by 43,000 votes.

These claims are among the most publication-sensitive portions of the transcript and require independent primary-source verification.

Chamath argues that California effectively operates under a political monopoly.

He says one political machine has shaped state law so extensively that elections for mayor, legislature, senate and governor increasingly function as a one-way door favoring one party.

He argues the state's governing institutions have become less capable while changing rules to reinforce their political dominance.

His long-term prediction is that the consequences of this institutional “rot” may take decades to become fully visible.

He says California voters should at least have had the opportunity to see Spencer Pratt challenge the existing political machine. Instead, he argues, the final candidates would offer little substantive choice.

Chamath ultimately reconciles Friedberg's and Sacks' positions:

• Friedberg is right, in his view, that the law may have been shaped to make previously impermissible practices legal.
• Sacks is right, in his view, that a rational observer could nevertheless consider the resulting system fraudulent in substance.

He says the political answer is to vote current leaders out.

When challenged that the same system makes political change difficult, Chamath proposes a “break the glass” route.

He argues that a reform candidate must win at the very top of the statewide ticket and names Steve Hilton.

If Hilton became governor, Chamath says he could declare a state of emergency on day one and use emergency authorities to begin changing the way California operates.

Chamath deliberately declines to elaborate further.

The legal scope of such emergency powers requires verification before publication.

Jason supports tighter safeguards, including:

• voter identification;
• ending automatic mailing of ballots to everyone;
• clearer registration requirements.

However, he says he does not believe election fraud has been demonstrated to have changed the outcome of a major modern US election.

He cites the Heritage Foundation's work collecting election-fraud cases and says lawsuits are routinely brought after elections by candidates of both parties.

He maintains that fraud exists at the margins but distinguishes that from evidence that it changed major election outcomes.

Other hosts invoke historical allegations involving:

• John F. Kennedy versus Richard Nixon;
• organized crime's alleged role in Kennedy's election;
• Lyndon Johnson's early Senate election.

Jason says those historical claims can be debated separately and confines his point to modern elections.

Jason argues that both political sides attempt to shape election rules to their advantage:

• one may restrict access;
• another may expand access excessively.

His concern is partly the appearance of impropriety: even if fraud is not outcome-determinative, systems that inspire widespread suspicion destroy trust.

The hosts broadly converge on the idea that Congress should establish stronger national protections ensuring that:

• voters are citizens;
• voters are properly registered;
• voters present valid identification;
• each eligible citizen gets one vote.

One speaker argues it is inconsistent that identification is necessary for many ordinary interactions with government but not uniformly required for voting.

The stated fear is not merely fraudulent ballots. It is loss of legitimacy.

If large numbers of people cease believing elections are real, the speakers warn that political conflict could become much uglier.

They therefore frame election-security reform as necessary to preserve confidence in democratic institutions.

They also mention a California voter-ID ballot measure and describe it as highly consequential, with one speaker calling it a potential “last chance” for election reform in the state.

Jason tells Friedberg that his argument sounds like election denial.

Friedberg rejects the relevance of the label and counters that Democrats challenged the legitimacy of Trump's 2016 victory through what he calls the “Russiagate hoax.”

He says his concern is based on the difference between in-person, early-mail and late-mail vote distributions, not simply dislike of the winning candidate.

Friedberg says he is willing to let socialist candidates win and govern places such as Los Angeles, New York, Seattle or Minnesota because he believes voters may then experience and reject socialism.

He argues socialism reduces economic mobility and individual liberty.

His election objection is therefore framed as procedural: he says he would accept a socialist winner if he believed the data clearly reflected the electorate's choice.

Sacks takes the most categorical position.

He invokes the Orwell line:

“The party told you to reject the evidence of your eyes and ears. It was their final most essential command.”

He says the charts themselves demonstrate that something crooked occurred.

He characterizes the vote pattern as “statistically impossible” and says the only unresolved question is whether the mechanism was illegal fraud or conduct made legal by election loopholes.

He explicitly says Spencer Pratt should have been in the runoff and denies that “Ramen” legitimately earned the position.

This is an allegation, not an established finding.

Another speaker argues that the California and federal attorneys general should investigate rather than rely on speculation.

Jason specifically urges the Trump administration and Department of Justice to audit or investigate the election if the statistics genuinely indicate wrongdoing.

Sacks accuses mainstream media organizations of refusing to investigate the anomalous voting pattern.

He says journalists should be asking how candidate shares changed so dramatically among ballots counted after election day.

Instead, he alleges that media institutions dismiss the controversy because investigating it might appear to support Trump or MAGA politics.

Chamath expands this into a polarization argument.

He says questioning a disputed election process can immediately place someone into a perceived pro-Trump camp, regardless of the substance of the question.

That social pressure, he argues, discourages institutions from investigating politically inconvenient claims.

He frames the issue as one of statistical literacy rather than partisan identity.

Jason repeatedly asks the other hosts for a legitimate explanation of the late-ballot discrepancy.

One speaker claims the probability of the observed change was “one in a trillion.” No methodology is provided in the transcript, so the number should not be used as a statistical finding without obtaining the underlying calculation.

Jason emphasizes that he is asking for a counterargument for intellectual rigor, not asserting that California elections are corruption-free.

The principal non-fraud explanation offered is ground-game sophistication:

• Ramen's campaign or allied political organizations may have had a much stronger ballot-collection operation;
• Spencer Pratt was a first-time, unconventional candidate without comparable organization;
• collecting ballots itself may be legal.

The hosts note that describing this as a stronger “ground game” may explain the result without requiring fabricated ballots.

Another theory mentioned from social-media commentary is strategic late voting.

Under this hypothesis, voters strongly opposed to Pratt waited to see which Democratic candidate needed support. Once polling suggested Bass was secure, late voters shifted toward Ramen to ensure that two Democrats—not Pratt—advanced.

The transcript says Pratt had alleged that Ramen registered as a candidate roughly one hour before the deadline, was friends with Bass and entered specifically to keep him out of the general election.

Those are Pratt's alleged claims and require verification.

Several hosts question how such mass coordination could occur organically. Others answer that modern group chats and organized ballot-harvesting operations could coordinate it.

The discussion ends without resolving whether the vote distribution resulted from legitimate strategic organizing, aggressive but legal ballot harvesting or actual fraud.

Jason closes with the most cautious formulation:

“No evidence of fraud right now. Statistics don't look good.”

He again urges the Department of Justice to investigate rather than treating suspicion as proof.

Quote: “If you were a person, you should generally now think there's a risk of censorship.”

Context: His assessment of Anthropic evaluating prompts before determining model access.

Topic: AI censorship and provider control.

Quote: “And what are the best open source models today? They're Chinese.”

Context: Friedberg argues that US model restrictions push businesses toward Chinese open-source alternatives.

Topic: Open-source AI and geopolitical competition.

Quote: “You can't just stop AI.”

Context: Argument that restrictions in one jurisdiction merely transfer technological advantage elsewhere.

Topic: AI regulation and international competition.

Quote: “They would nerf their models if it decided in anthropic soul discretion that you are not worthy of having access to that level of information.”

Context: Sacks' criticism of user profiling and model downgrades.

Topic: AI gatekeeping.

Quote: “They're going to decide whether you're worthy.”

Context: Prediction that AI companies could decide who receives frontier capabilities.

Topic: AI access and power.

Quote: “That's what's scary.”

Context: After describing the possibility of AI firms favoring companies based on commercial or ideological alignment.

Topic: Hidden AI influence.

Quote: “Technology is fundamentally deterministic. Whatever is possible will be tried at least once.”

Context: Friedberg's argument that AI capability cannot realistically be put back in the box.

Topic: Dual-use technology and regulation.

Quote: “The models have been put out in the world. It's like publishing a book.”

Context: Open-source models can be copied and modified after release.

Topic: Irreversibility of open-source AI.

Quote: “Who gets to watch the watchman? Who put Dario in charge of all this?”

Context: Closing question after discussing model gatekeeping and downstream safety mechanisms.

Topic: Accountability of AI companies.

Quote: “I'm not in favor of Bernie's proposal because it's a straightup confiscation of property.”

Context: Initial response to Sanders' proposed 50% public stake in major AI companies.

Topic: Public ownership of AI.

Quote: “What's in this deal for me?”

Context: Sacks' description of how ordinary Americans may react after hearing that AI could replace their jobs.

Topic: Political legitimacy of AI wealth.

Quote: “There is no job loss with AI. I will say it again.”

Context: Friedberg rejects the AI unemployment thesis.

Topic: AI and employment.

Quote: “Everyone's gonna realize soon that this is a boom, not a bust.”

Context: Conclusion of his productivity-and-hiring argument.

Topic: AI economic growth.

Quote: “The real opportunity with AI is on the revenue side.”

Context: He argues AI enables companies to create more products rather than simply cut headcount.

Topic: AI productivity.

Quote: “The incremental cost of performing AI is excessive and large.”

Context: Contrast between AI economics and near-zero marginal-cost internet economics.

Topic: Economics of AI infrastructure.

Quote: “The core problem with wealth inequality in this country, the core problem with inflation in this country, the core problem, all of it roots back to excess government spending.”

Context: Friedberg's fiscal explanation for inflation.

Topic: Inflation and fiscal policy.

Quote: “This is not fraud. This is the way the laws are set up.”

Context: Friedberg distinguishes legal election rules he dislikes from criminal fraud.

Topic: California election law.

Quote: “It may be legalized, but it's fraud.”

Context: Sacks rejects Friedberg's legal/moral distinction.

Topic: Election-integrity allegations.

Quote: “The party told you to reject the evidence of your eyes and ears. It was their final most essential command.”

Context: Sacks invokes Orwell while arguing that the vote-distribution charts themselves are suspicious.

Topic: Election legitimacy and media response.

Quote: “Spencer Pratt should be in the runoff.”

Context: Sacks explicitly rejects the legitimacy of the reported runoff result.

Topic: Los Angeles mayoral primary.

Quote: “No evidence of fraud right now. Statistics don't look good.”

Context: Jason's final attempt to distinguish suspicious data from proven fraud.

Topic: Evidentiary standard for election claims.