Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters

Source: Authorized transcript of an episode of the All-In Podcast, 349 source lines. The program is presented as a roundtable covering AI regulation, technology and M&A, AI privacy and costs, data-center energy policy, foreign influence, and a final science segment on aging research.

Principal speakers: The transcript clearly identifies Jason as moderator and David Freeberg/Friedberg by first and surname variants; David Sacks is repeatedly addressed as “Sacks.” Chamath’s name is heavily corrupted in transcription. To avoid silently correcting the source, this record uses Jason, Chamath, Sacks, and Friedberg as stable speaker labels. Several outside people, companies, products, and model names are also badly transcribed and are flagged below rather than silently corrected.

Transcript-quality problems: There are numerous corrupted proper names and product names; occasional missing words; speaker attribution sometimes has to be inferred from conversational hand-offs; some numerical expressions are garbled; and several passages contain obvious transcription errors such as “Deep Mines,” “SpaceX” in discussion of Grok/xAI infrastructure, “Fable,” “Seoul,” “Quad 4-8,” “Kimmy K3,” “Mythos,” “PGM,” and other terms. These are preserved or marked [transcript unclear], not externally corrected.

Editorial status: This is a source record only. Claims about companies, lawsuits, regulation, political motives, market values, scientific results, foreign influence, data-center impacts, and other factual matters are reports of what the speakers said. They are not independently verified here.

1. Jason opens by introducing the All-In Podcast and the regular panel.
2. The hosts joke about a historical claim concerning France and Louisiana: according to the story discussed, Paris prisoners in 1719 could allegedly obtain freedom by marrying a prostitute and moving to Louisiana. The discussion becomes a joke about Jason’s ancestry; Jason rejects the implication that his family was French and says he is Greek.
3. Friedberg says he is glad to be back after missing the previous week. The panel briefly discusses Brad’s substitute appearance and describes it as involving a “Trump account victory lap” accompanied by “Chariots of Fire.”
4. Two destinations or activities in the banter are rendered as “Blank” by the transcript and therefore cannot be recovered from the source.

Jason says Demis Hassabis had proposed a U.S.-led international AI standards organization modeled on FINRA. Jason describes it as federally overseen but industry-funded and operated by independent technological experts.

Under the proposal as Jason describes it:

• frontier laboratories would submit models 30 days before release;
• participation would initially be voluntary and become mandatory later;
• models would be assessed for cybersecurity, national-security, biological-threat and other high-risk capabilities;
• benchmarks would be updated quarterly;
• the organization could coordinate a slowdown in development where risk justified doing so.

Jason says favorable reactions had come from Elon Musk, Sam/OpenAI, Jack Clark/Anthropic, Sundar [surname omitted], Satya [surname omitted], Jack Dorsey/Block, and “the Carlson brothers” [transcript/source identification unclear].

Jason plays or recites a previous clip in which he argued that the AI industry would need regulation but should regulate itself.

His reasoning was:

• companies such as Google, Microsoft, Amazon and Elon Musk's organization should agree on a common suite of tests;
• laboratories should test and self-certify models before release;
• government lacks sufficient understanding of the models to serve as the primary technical certifier;
• industry certification already exists in other sectors;
• self-certification could create safety checks without forcing model releases through a government approval process.

One speaker explains that self-regulatory organizations such as FINRA and the National Futures Association allow industry participants to establish and adapt rules while operating under public oversight.

The analogy offered is that financial institutions share risks with one another and therefore have an incentive to maintain safety without unnecessarily making markets inefficient. The same could apply to AI because multiple companies are pursuing fast technological progress while sharing an interest in preventing serious failures.

The speaker argues that conventional statutory regulation becomes obsolete too quickly. California's earlier AI legislative efforts are cited as an example: the speaker claims that parts of those proposed rules did not make sense even when drafted and were even less aligned with the technology a year later.

An SRO, by contrast, could:

• change tests as the technology changes;
• select technically qualified independent experts;
• test cyber, biological, weapons and social-manipulation capabilities;
• update membership and governance;
• move faster than a newly established federal agency;
• preserve ultimate government oversight without giving government direct technical control.

The panel says OpenAI and Anthropic had expressed support, although it is unclear from the transcript whether Dario himself had done so; Jason says Anthropic's president and Sam had given approval.

Sacks says he personally discussed the proposal with Demis and told him he could potentially support it. He emphasizes that he is speaking for himself and not on behalf of anyone in government.

Sacks says a self-regulatory model would be vastly preferable to a new government agency, which he fears would become a “DMV for AI.” He contrasts that with Dario's proposed analogy of an “FAA for AI.”

His main concern is speed: government lacks the expertise to evaluate rapidly changing models, so a mandatory government testing queue might begin as a one-month delay and eventually grow into many months. In his view, that could cause the United States to lose the AI race.

He gives five conditions:

The organization must represent the breadth of AI, including:

• major laboratories;
• startups;
• open-source developers.

It cannot be controlled by the three largest labs. Sacks argues that diverse representation is essential to prevent regulatory capture.

As illustrative participants, he names Jensen, Elon, Zuck and Mira, referring to Mira's newly launched open-source-based platform [specific platform/name to verify].

Sacks says regulation should apply only to true state-of-the-art advances, not every model.

He suggests frontier status might mean something like a top-10 or top-20 benchmark performer that represents a material increase over the previous state of the art.

His reasoning is that the justification for special regulation is new catastrophic risk created by a step-change in intelligence. A model that is not materially beyond the existing frontier does not introduce that same category of incremental danger and should not be delayed.

He also warns that incumbent leaders could otherwise use compliance requirements to slow smaller competitors.

Sacks says the SRO should confine itself to catastrophic threats.

He identifies the relevant current categories as:

• cyber risk;
• CBRN: chemical, biological, radiological and nuclear risk.

He explicitly rejects turning the organization into a regulator of disinformation, microaggressions or speech generally.

The body should operate voluntarily first and demonstrate that it works before being legally enshrined as a mandatory system. Sacks says this element was also included in Demis's proposal.

The SRO should replace the need for a new AI regulator rather than becoming an additional layer of regulation.

If it is merely added on top of new agencies and other regulatory structures, Sacks says there would be little reason for him to support it.

Sacks reports that Demis told him these were good ideas even though all five were not incorporated in Demis's public post.

Sacks says Anthropic's support concerns him because he suspects the SRO could merely be an “opening bid” toward much heavier regulation.

He explains his objection through actual FAA aircraft certification:

• new aircraft designs require type certification;
• he says certification for an entirely new design takes roughly five to nine years;
• certificate amendments for major changes take roughly three to five years;
• he cites the Boeing 737 MAX as an example that took around five years.

His broader point is that aviation regulation is permission-based: no approval means no commercial flying. He accepts that such a safety-first system may make sense for aircraft, but argues that applying its tempo to AI would convert a technology currently updated every few months into one whose iterations could take years.

He further argues that China would not follow equivalent delays, putting the United States at a competitive disadvantage.

Sacks therefore says that, if choosing between an FAA/“DMV” model and Demis's SRO, he would choose the SRO, provided it remains narrowly constrained. Otherwise, he fears it would become a mechanism for regulatory capture.

The panel repeatedly emphasizes that the SRO must not restrict open source.

Their reasoning is that costly regulatory requirements can themselves become a competitive barrier: large companies can afford compliance while startups and open-source developers cannot.

Chamath says a “torrent of money” will attempt to influence both political parties and shape regulation in ways favorable to incumbent actors. He favors moving quickly on an SRO so that a workable industry framework exists before political lobbying produces more restrictive federal structures.

He notes that government would still have relevant authority through institutions such as the Department of Commerce and DOJ, so self-regulation would not mean a completely unregulated “Wild West.”

His central concern is preventing well-capitalized incumbents from “pull[ing] the ladder up.”

Sacks adds another complication: FINRA ultimately operates within a government-supervised structure, so any AI SRO would face a political fight over which department or agency supervises it. That relationship could itself expose the body to political pressure.

He notes that software has historically lacked a dedicated regulator. Chamath nevertheless describes the SRO as the lesser evil if the alternative is effectively killing open source and allowing the market to collapse toward a duopoly.

Sacks says that in October of the previous year he publicly accused Anthropic of conducting a “sophisticated regulatory capture strategy based on fear mongering.”

He says the view was controversial at the time because Anthropic was treated as a small startup.

He now characterizes Anthropic as one of the largest technology companies, claiming:

• a roughly $1 trillion valuation/market capitalization [description itself requires verification because Anthropic is discussed elsewhere as private];
• investor Gavin Baker predicts a possible $3 trillion value after an IPO;
• Anthropic is, in Sacks's assessment, the leading AI company by criteria including revenue.

Sacks argues that its financial resources allow it to promote regulations that could make competition more difficult. These are his allegations and interpretation, not established facts in this source record.

Sacks cites a recent Politico article whose title is transcribed as “Inside Anthropics State-by-State Plan to ratchet up AI rules.”

He quotes or paraphrases the article as saying Anthropic pursues a strategy encouraging individual states to impose progressively stronger guardrails rather than converge on one national standard.

He cites California's SB 53 as an example. According to his account:

1. one state adopts a regulatory model;
2. rather than merely copying it, subsequent states enact progressively stronger versions;
3. this produces no stable regulatory equilibrium;
4. a growing patchwork places increasing burdens on the sector.

Sacks argues this is the opposite of the national uniform framework he prefers.

The speakers say Anthropic has succeeded or made progress in California, Illinois and New York, and possibly in some Republican-led states. These claims require verification.

Sacks's broader political argument is that when companies ask government to regulate them, public institutions rarely refuse the additional authority.

He rejects the idea that technology companies can satisfy politicians by making incremental concessions. In his view, each concession becomes the starting point for further demands until the industry is substantially controlled by government.

He therefore says AI companies eventually need to establish a firm boundary and fight further regulation.

If Demis's SRO is that boundary, Sacks says companies should obtain legal protections such as preemption so that the SRO becomes the agreed regulatory settlement rather than the first layer of a much larger system.

Jason shifts to a breaking report about a bid for PayPal.

The transcript says:

• Stripe remains private;
• Stripe and private-equity firm Advent are offering approximately $53 billion for PayPal;
• the offer is around $60 per share;
• some observers may expect approximately $70 per share;
• PayPal's shares rose on the news;
• Block was also reportedly joining the transaction;
• Block would contribute $17 billion in equity.

The speakers acknowledge that contemporary media reports were inconsistent because the story was breaking.

The panel discusses PayPal's assets, including Venmo.

Jason says Stripe owns Bridge, described as stablecoin infrastructure acquired for about $1 billion in 2025.

He also says PayPal has a stablecoin transcribed as “SIUSD” [transcript unclear].

Jason frames stablecoins as one component of the strategic rationale.

He describes PayPal as still having approximately 430.9 million consumer accounts.

Sacks says it is remarkable that PayPal remains so significant roughly 25 years after its creation.

However, he characterizes the product as “long in the tooth” and says growth is only around 7% annually, albeit from a very large base.

He believes PayPal's interaction model has become somewhat obsolete and asks what Stripe could do to renew it. One possibility, he says, is that buyers might simply make the company more efficient and extract cash rather than fundamentally reinventing it.

Chamath proposes a more ambitious thesis: Stripe, Advent, Block and PayPal could together create a serious competitor to Visa and Mastercard.

He says the combined ecosystem would offer:

• perhaps 600–700 million accounts;
• stablecoin infrastructure;
• Stripe's risk-management infrastructure;
• Stripe's merchant network;
• PayPal's consumer relationships;
• Block's point-of-sale infrastructure.

He argues Stripe historically faced concern that its margin as a middleman on existing payment rails would eventually be competed away, which encouraged the company to build numerous value-added services.

A PayPal combination could allow much deeper vertical integration and potentially create extremely low-cost proprietary payment rails. Advent could contribute capital while Stripe and Block could distribute the infrastructure globally.

Friedberg says the PayPal proposal may be part of a broader emerging acquisition pattern.

He compares it with Ryan Cohen's reported bid for eBay and argues that AI-native operators are beginning to look at first-generation digital businesses that:

• are mature;
• are no longer founder-led;
• have stalled;
• overspend;
• fail to exploit their networks fully;
• have not integrated AI effectively.

To a modern operator, Friedberg says such businesses can appear dramatically underutilized.

He compares the model with efforts to acquire traditional professional-services businesses and “AIify” them, mentioning Josh Kushner's accounting-firm roll-up and a General Catalyst project. He believes a similar strategy can now be applied to older digital businesses and says there may be “a couple dozen” suitable public companies.

The key investment question, he says, is not merely access to capital but finding the operator capable of performing the turnaround. A McKinsey consultant, in his formulation, is not enough; investors need elite operators with domain expertise.

He points to Ryan Cohen's history at Chewy and GameStop—while acknowledging that assessments of his GameStop record are debatable—and says Stripe would be an obvious qualified operator in payments.

Friedberg attempts to explain the reported $17 billion equity contribution.

His interpretation is that Stripe and Block would effectively sell equity to cash investors, with those proceeds used to acquire PayPal. The result would be:

• Stripe and Block owning part of PayPal;
• private-equity investors owning stakes in Stripe and Block.

He says the publicly reported deal documents do not make clear who would operate PayPal after closing.

His prediction is that Stripe's team would be placed in operational control because they are most qualified and would likely have the largest strategic stake.

Friedberg predicts eBay and PayPal may mark the beginning of a wave of mega-deals involving stagnant digital companies that can be revived by capital and strong operators.

He jokingly describes the companies as “flaccid” digital businesses and capital as their “BlueChew.”

He also predicts the reported PayPal bid is unlikely to be the final price and estimates a clearing price 10–15% above the level being discussed.

The panel then jokingly speculates about another possible bidder: an unnamed individual associated with the original PayPal who controls or is connected with $4–5 trillion in market capitalization and had recently made a $60 billion acquisition. The speakers explicitly say they have no inside information. Identity is intentionally unstated in the source and should not be inferred for publication without verification.

Jason cites Bending Spoons as evidence that the playbook already exists.

The transcript claims Bending Spoons had recently gone public and acquired a collection of non-founder-led assets, including:

• AOL for approximately $1.4 billion;
• Vimeo for approximately $1.4 billion;
• WeTransfer;
• Eventbrite;
• “Brite code” [transcript unclear];
• Evernote.

One speaker says he knows the operator personally and calls him an exceptional operational executor based in Milan.

The described strategy is to diagnose:

• overspending;
• underspending;
• product weaknesses;
• marketing weaknesses;

then repair those areas, lower operating costs and restore cash generation.

The speaker says Bending Spoons is using young, AI-first executives, according to information he has heard.

Friedberg considers it a smaller-scale example of the same strategy he predicts will be used for larger public companies.

Jason argues that venture capital and M&A were constrained for several years under Lina Khan's antitrust leadership.

He says that after Donald Trump's election, corporate-development executives concluded that “M&A's back on the menu,” and companies became less afraid to pursue acquisitions.

As an example, the panel says Uber had just acquired Delivery Hero. Figures are discussed as roughly:

• a 20–24% increase in revenue;
• approximately 10% dilution.

These figures and the claimed transaction require verification.

Jason says LPs and family offices he speaks with are again asking about new venture funds and deals because of:

• SpaceX distributions;
• increased M&A activity;
• several portfolio companies being acquired since Trump was elected.

Jason explicitly praises Trump for putting M&A “back on the menu.” This is Jason's political view, not a neutral source conclusion.

The panel asks why Sacks did not bid for PayPal.

A story is raised that around 15 years earlier Sacks had been considered for PayPal's CEO job. Jason recalls going with him to Las Vegas to think it through.

Sacks corrects the account: he says he never actually received an offer, but believes he was among the final two candidates.

He says PayPal ultimately hired a more traditional credit-card executive.

Sacks attributes much of PayPal's subsequent stagnation to what happened after eBay acquired it in 2002:

• founding executives were pushed out;
• eBay did not preserve the founding culture;
• management became more corporate and consultant-driven;
• Meg Whitman's background in Procter & Gamble, Disney and Bain is mentioned as emblematic of that style;
• the original PayPal founders were treated as difficult “cowboys” rather than talent to retain.

He argues that acquisitions normally seek to lock key talent in, whereas PayPal's founders were effectively “locked out.”

Sacks therefore rejects the usual phrase “PayPal Mafia”, preferring “PayPal diaspora.”

He uses the deliberately dramatic analogy:

“Our homeland was taken over and they burned our temple and then kicked everybody out.”

He does not portray the founders as victims, however. He says eBay owned the asset and had the right to run it as it wished, and the departing founders gained capital to pursue their next ventures.

The panel also praises PayPal's current CEO, transcribed as “Enrique,” and says management had been tightening operations during the preceding couple of years.

The panel recalls PayPal once being worth approximately $322 billion at its peak, whereas before the reported offer it had fallen to roughly $30–40 billion.

Sacks says this depressed valuation, more than operational improvements, explains why bidders are interested.

Asked whether AI itself is the key synergy, Sacks distinguishes two forms of value:

1. Efficiency: AI can automate work, lower costs and increase profitability.
2. Product renewal: the acquirer still needs a compelling vision for improving the user experience.

Sacks says efficiency alone might make the transaction financially successful, but the deeper problem is that PayPal's core interaction model remains approximately 25–27 years old.

Without genuine product rejuvenation, he suggests the buyers may primarily be purchasing the user accounts.

Sacks estimates:

• Stripe processes approximately $2 trillion in annual transaction volume;
• PayPal handles approximately $1.7 trillion.

He says Stripe has become highly important to merchants through API-based payment acceptance, while PayPal's distinctive asset is the direct consumer relationship represented by more than 400 million active consumer accounts.

Combining these could permit more transactions to stay “on us” rather than traverse traditional credit-card networks.

Chamath adds:

• PayPal owns Braintree, a Stripe competitor;
• Block adds point-of-sale infrastructure;
• Cash App [apparently referenced as “cash out”; transcript unclear] may add another consumer/payment component.

He calls the proposed combination a “shot across the bow for Visa and MasterCard.”

Venmo is also identified as strategically useful because it reaches younger users, complementing PayPal's older customer base.

When Sacks questions whether consumers would actively choose a new combined payment system, Chamath proposes incentivizing merchants instead.

His hypothetical:

• approach merchants using Stripe;
• offer a 3%, 4% or 5% discount for using the new rails;
• merchants keep perhaps another 2–4 percentage points;
• adoption follows because of the direct economic advantage.

He uses Shopify merchants as a hypothetical example.

The panel argues that lower transaction costs could ultimately reduce consumer prices.

The panel notes a paradox: a large Stripe/PayPal/Block transaction could look highly anti-competitive under one market definition and pro-competitive under another.

Sacks explains:

• if the relevant market is merchant payment APIs, combining Stripe and Braintree may appear to remove a major competitor;
• if the relevant market is the much larger Visa/Mastercard card-network market, a new vertically integrated competitor could increase competition.

Therefore the antitrust result depends heavily on how regulators define the market.

Jason introduces a claim that Apple filed a 41-page lawsuit on July 10 against OpenAI over alleged stolen trade secrets.

He says Apple alleges OpenAI used Apple intellectual property in developing a consumer hardware device.

The panel references an earlier interview with Sarah Fryer [transcript spelling], who had described OpenAI's future device as “very human and lovable.”

Specific allegations presented include:

• Tang Tan, described as a former Apple VP of iPhone design and OpenAI's chief hardware officer, allegedly told Apple employees interviewing with OpenAI to bring “actual parts” to interviews for “show and tell”;
• Chang Liu, described as a former Apple senior technical engineer, allegedly messaged an Apple employee: “LOL. I found out I can access the network storage. So funny.”
• OpenAI allegedly recruited more than 400 Apple employees over one or two years.

All these are described by the hosts as allegations in litigation and require independent verification.

Jason adds a highly critical characterization of Sam Altman, claiming he had now damaged another major relationship after previously doing so with Elon Musk. This is Jason's opinion.

Jason also notes that ChatGPT had been intended as an important/default AI option on the iPhone and frames the alleged lawsuit as particularly remarkable given the relationship between the companies.

Chamath says he has rarely seen Apple behave aggressively in litigation during his roughly 25 years in Silicon Valley, which he considers a potentially significant signal.

However, he refuses to infer the facts from the complaint and says a court will have to determine what actually happened.

His normative position is clear: employees should not steal property from a former employer.

He says senior employees are already valuable because they carry accumulated knowledge and experience; they do not need to take physical or digital company property.

Chamath also says he hopes the allegations are false and does not believe Sam, Sarah, or other senior OpenAI leaders would deliberately induce employees to steal material.

Jason replies that he doubts senior leaders induced it but believes Apple's decision to sue makes the allegations more plausible. That is Jason's inference, not established fact.

Jason contrasts California's labor mobility and limited non-compete enforcement with trade-secret restrictions. He cites the Waymo/Uber episode as an example in which Uber allegedly rejected materials brought by a recruit and rescinded the person's job.

Sacks again refuses to opine on the truth of Apple's allegations.

Instead, he gives a general rule:

• employees may take their skills, memories and knowledge;
• they should take no files or physical material from their former company;
• no thumb drives;
• no storage media;
• no documents.

His formulation is:

“The only thing you can bring to your new job is what's in your head.”

Friedberg responds to the broader accumulation of OpenAI lawsuits: “A couple dots make a line, I guess.”

The transcript repeatedly says “SpaceX” while discussing Grok/xAI products; this is likely a transcript or speaker error and should not be silently corrected at Stage 1.

Jason says:

• “Croc/Grock Build” entered public beta at the end of May;
• “GROC 4.5” powers it;
• it operates inside “Cursier” [transcript unclear];
• users had previously been told nothing from their code base would be transmitted to company servers;
• according to reports, the tool actually uploaded developers' entire code bases;
• this could include passwords, API keys, change logs and unrelated files;
• a privacy setting intended to prevent the upload allegedly failed;
• the upload was disabled on July 13 through a server-side switch;
• Elon Musk said previously uploaded material had been deleted;
• the harness was subsequently open-sourced.

Jason interprets the open-sourcing as a positive development for the open-source community and “AI sovereignty.”

Chamath says the incident reinforces comments he made on CNBC after comments from Alex Karp.

His central thesis is:

“privacy in AI is very fragile and it's very brittle.”

Chamath explicitly says this can be true even when companies are acting in good faith.

He describes Elon as highly trustworthy and unusually transparent and credits the company with disabling the behavior quickly.

The broader problem, in his view, is the existence of non-obvious data-leak vectors.

He warns enterprises against assuming that enabling ZDR — zero data retention — guarantees that no information is retained or exposed.

Model providers may sincerely try to honor those guarantees, he says, but hidden technical “trap doors” can cause leakage that neither customer nor provider knows about until an outsider discovers it.

Chamath says enterprises therefore require a stratified ecosystem with independent third parties between the enterprise and foundation-model providers.

He discloses that this is aligned with his own commercial interest because his firm, rendered as “8090,” implements a “software factory” for large enterprises.

His argument is that a third-party orchestration layer can manage model exposure and prevent customers from depending completely on one provider's assurances.

The panel then cites a blog post attributed to “Sotia” [transcript unclear; likely a person whose identity must be verified] and calls the concept the “reverse information paradox.”

The discussion builds on Alex Karp's argument that technically capable enterprises want control over:

• compute;
• models;
• weights;
• data;
• proprietary “alpha.”

The blog post is said to extend Karp's principle into an operational roadmap.

Recommended technical mechanisms are described as:

• private evaluations;
• proprietary learning loops inside the tenant;
• decoupled orchestration;
• the explicit right to fine-tune the enterprise's own outputs.

The speakers describe an informal intellectual exchange among technology leaders—a kind of virtual “college dorm session”—in which companies are independently developing alternatives to a monolithic model-provider stack.

They characterize Anthropic, and to some extent OpenAI, as preferring customers to remain within a vertically controlled system containing the provider's:

• model;
• harness;
• data environment;
• surrounding infrastructure.

The panel sees an emerging counter-ecosystem centered on interoperability and user control.

Chamath cites pricing data whose product/model names are badly corrupted in the transcript.

As transcribed, he says approximately:

• “Fable” — $56 per million tokens;
• “Seoul” — $26 per million;
• “Quad 4-8” — about the same;
• “GROC” — $1.50 per million input tokens;
• “Zuck's” model — approximately $1.50;
• Elon's model — approximately $1;
• Chinese models — approximately $0.50.

Every product name and the exact input/output-token basis should be verified before publication.

Chamath's substantive argument is that enterprises may simultaneously be:

1. surrendering control of valuable proprietary data;
2. paying dramatically more for the privilege.

He calls that combination irrational.

Jason says he has been using “Perplexity Computer” [name to verify], which supports multiple models including Grok and GLM-52 [transcript rendering].

Unlike systems restricted to their provider's own model, he says this allowed him to compare multiple models using essentially the same product requirements document.

His experiment was to build a podcast player that could locate and concatenate relevant clips from multiple technology and business podcasts into a single stream—for example, all discussions of one topic across several shows.

He says:

• the project took a couple of hours;
• using the new Grok cost approximately $11;
• he found the price surprisingly low.

When asked what the same task would have cost on “Fable,” he says he could not test it because he had exhausted the credits on his $200 account.

The panel plays a clip attributed to Eric “Gleman/Glenn” [transcript unclear], described as the CEO of Ramp.

He announces token spend management, available to Ramp and non-Ramp customers.

He says token spending among Ramp customers had increased 21-fold over the prior year—emphasizing “21 times,” not 21%.

He explains that CFOs are often surprised by AI bills because:

• employees effectively have an open tab;
• finance teams cannot easily see spending in advance;
• new models can have higher rates;
• incentives between users and corporate budget owners are misaligned.

Ramp's goal, according to the clip, is to help CFOs see, understand and control that spending.

Chamath says uncontrolled AI usage may eventually matter materially to reported corporate earnings.

If engineers can consume very expensive models without centralized controls, companies could end up missing quarterly expectations because operating expenses increased.

He notes that he had previously imagined AI expense affecting results by only a few cents per share, whereas the Ramp executive suggested the eventual effect might amount to dollars if growth continued at the current rate.

Chamath calls uncontrolled usage a “money-burning furnace” unless companies can directly connect AI spending with money earned.

The panel says high frontier-model pricing creates a large opening for alternatives.

Names mentioned include:

• BitTensor subnets;
• GLM-52 [transcript];
• Grok 4.5;
• “Inkling/Inklink,” attributed to Mira;
• other open-model hosting.

Sacks or another speaker corrects a suggestion that customers are paying $2 and could pay $1, saying some are actually paying $26–56 where alternatives may cost around $0.50.

The panel describes Mira's product strategy as deliberately just below frontier intelligence rather than competing directly for the top benchmark score.

Its value proposition is said to be:

• use much cheaper open models;
• fine-tune them to a particular task;
• achieve the desired practical outcome without paying frontier-model prices.

Sacks again argues that Anthropic has economic incentives to oppose open models because cheap open-source competition threatens expensive proprietary pricing.

This is Sacks's allegation/inference.

At the same time, he explicitly qualifies the argument by acknowledging that Anthropic is still growing rapidly and remains, in his account, an industry leader in revenue growth.

Thus, he says, two trends are occurring simultaneously:

• open models are becoming more capable and plentiful;
• proprietary frontier providers continue to achieve strong commercial growth.

The panel interprets Ramp's product launch as evidence that corporate customers are already struggling to control AI spending.

Their incentive analysis is:

• engineers want access to the newest and most capable model;
• engineers may have little reason to optimize cost;
• CFOs are accountable for return on investment and total spending;
• many tasks may not actually require frontier intelligence.

One speaker suggests perhaps 95% of tasks could use a model one capability tier lower, especially when it costs roughly 1/100th as much.

When asked what proportion of prompts on “Fable 5” [transcript unclear] are ordinary tasks that could run on a cheaper model, a speaker says 98% and admits personally using the expensive model for “stupid stuff” that could have run elsewhere.

The analogy is corporate travel: an employee who does not see the price may casually choose business class and an expensive hotel because another department pays the bill.

Jason presents a “micro prediction” based on information attributed to Mark “German” [transcript unclear].

He says Apple's forthcoming CEO is “John Furness” [transcript unclear], described as a hardware engineer.

He then discusses a future M7 Ultra chip and predicts/supports a claim that it could handle as much as 1.5 terabytes of RAM, roughly double the level currently available.

His thesis is that such memory capacity could allow a Mac Studio to run something comparable to a previous-generation frontier model locally.

Jason predicts employers might eventually run:

• 90% of workloads locally;
• perhaps 99%.

He calls Apple a “screaming buy”, while explicitly saying this is not financial advice, because local AI could create enormous hardware demand and reduce reliance on expensive cloud models.

He imagines users effectively having “unlimited tokens on their desk.”

The hosts briefly disagree over whether people laughed at the first iPhone. Steve Ballmer is mentioned as the major example of someone who did.

Jason's broader argument is that Apple's hardware economics and large-memory architecture could place major downward price pressure on providers such as Claude/Anthropic and OpenAI.

The panel says Sunrun had announced distributed data-center blocks that could be installed at homes.

A second company, Span, is said to have partnered with Nvidia on something similar.

The speakers regard this as evidence of a broader move toward fragmented/distributed edge compute.

The discussion then connects compute distribution with energy availability: AI workloads may increasingly move toward places where electricity is available rather than requiring all compute to remain in centralized hyperscale sites.

Chamath says the United States is facing a severe electricity deficit.

He claims that by 2050 the country will be short an amount of energy equal to roughly 2.5 Californias—2.5 times California's total consumption.

He describes an auction by a utility/market operator transcribed as “PGM”, serving Pennsylvania, New Jersey, Maryland and roughly 13 states.

According to his account:

• it needed around 7–8 gigawatts of future supply;
• only around 156 megawatts showed up.

He uses the example to argue that available generating capacity is dramatically below future demand and warns about electricity-price pressure.

All numbers and the identity of “PGM” require verification.

Chamath explains behind-the-meter generation.

Normally a data center connects to utility grid power. When the local grid lacks capacity or interconnection queues are too long, the data-center owner can instead generate power on its own property.

He says the complication is environmental regulation, particularly clean-air permitting.

He describes options:

• solar requires substantial land;
• batteries store electricity but do not create the electricity themselves;
• natural gas can generate power onsite.

He says Elon obtained mobile-turbine equipment, described loosely in the transcript as “18 wheeler like engines,” and installed it to power the Memphis/Colossus facility.

A regulatory distinction involving “personal use” or aggregated blocks is described but is not fully clear in the transcript.

Chamath then mentions Bloom Energy as another technology that can support large onsite installations while navigating clean-air permitting.

His larger conclusion is that electricity scarcity is already serious and becoming worse.

The program plays a statement attributed to New York Gov. Kathy “Hochel” [transcript spelling].

In the clip she argues that hyperscale data centers:

• use fossil-fuel power;
• increase the carbon footprint;
• occupy large amounts of land;
• can displace agriculture/open space;
• can raise utility bills;
• can deplete water supplies;
• create noise pollution.

The clip says she is signing the nation's first statewide moratorium on hyperscale data centers. This claim and whether the audio is actual, hypothetical or satirical should be checked before publication.

Sacks says data centers can increase household electricity costs if they connect to a constrained grid without adding generation.

However, he argues that behind-the-meter power solves this by requiring data centers to bring their own generation.

He says the president has advocated allowing AI companies to effectively become power companies.

Sacks rejects the land-use argument.

He says the United States has enormous amounts of available land and describes data centers as exceptionally productive uses of land because the economic value generated per acre can be very high.

The panel says facilities should not be placed immediately adjacent to residential neighborhoods but argues there is plenty of suitable land elsewhere.

The speakers call noise pollution manageable through siting and distance.

They strongly dispute water-consumption criticism, saying modern data centers use closed-loop recirculation.

One speaker cites an alleged study claiming a typical data center consumes about as much water as “two and a half In-N-Out Burgers” or restaurant locations [exact comparison unclear].

Almond agriculture and golf courses are offered as examples of water uses the panel considers more wasteful.

Jason recalls an article, possibly concerning North Dakota, in which teachers allegedly received $30,000–$40,000 bonuses because of additional tax revenue from data centers.

Sacks says data centers:

• produce substantial tax revenue;
• create a construction boom for blue-collar workers;
• continue employing people after construction.

The panel disputes claims that completed data centers create no ongoing employment.

Sacks also describes natural gas as one of the cleaner-burning power sources and says data centers have become a scapegoat for broader public anxiety about AI.

Jason attacks the stated land concerns by claiming roughly 70–80% of New York State is undeveloped and argues critics are mentally conflating dense New York City with the much larger and less-developed rest of the state.

The panel says Texas would welcome projects New York rejects.

Jason describes attending Senator Dave McCormick's defense and innovation summit in Carlisle, Pennsylvania, at the Army War College.

He says:

• the president attended and gave a speech;
• a CEO roundtable included defense-company leaders;
• Chris Wright also participated.

According to Jason's retelling, Dina Powell asked Wright a question and Wright claimed there are common funding patterns among current anti-data-center protest groups and organizations that previously protested fracking.

A panelist characterizes such activists as “professionally paid protesters” who move between causes.

This is an allegation/generalization and requires evidence before publication.

Sacks asks why Anthropic would fund groups seeking to slow data-center development.

He identifies an organization called Public First and claims Dario recently made his first seven-figure contribution to it, with other Anthropic employees contributing as well.

Sacks alleges these organizations support:

• new AI regulation;
• restrictions that make data-center construction more difficult.

He questions whether this is another form of regulatory capture or evidence Anthropic has “lost the plot,” because he believes Anthropic's own revenue growth is constrained primarily by compute/data-center availability rather than demand.

All funding relationships and organizational goals should be independently checked.

One panelist reports a theory heard from a political contact.

The theory—explicitly presented as speculation—is that Democratic policymakers may not want to stop data centers permanently. Instead, temporary moratoria could give them leverage to impose a future regulatory framework.

Under this hypothesis, projects would later be permitted again only under conditions including:

• a new government AI regulator;
• new rules around speech;
• migration of social-media “trust and safety” structures into AI.

The speaker imagines this occurring after Trump's presidency or under a future Democratic administration.

This is secondhand political speculation, not established policy intent.

The panel argues that even a nominally temporary moratorium could cause roughly a five-year practical delay, because stopped projects would take years to restart after a policy reversal.

Chamath says he and partner “Nita” [transcript unclear] own or have accumulated a portfolio of data-center assets.

He says discussions with hyperscalers show extremely high values for projects that have verifiable, energizable power available today.

He attributes the premium to a supply shortage and claims approximately 40% of projects are being mothballed or stopped.

He connects infrastructure limitations with lost AI benefits.

If electricity is unavailable, he argues, demand for:

• drug discovery;
• cancer diagnosis;
• improved healthcare;
• improved legal advice;

may exceed available compute capacity.

His prediction is that “GPUs [will be] chasing energy”—hardware will increasingly be installed where electricity is available rather than the other way around.

Sacks adds that the same political forces opposing U.S. data-center development are, in his view, supporting tighter semiconductor export controls.

He argues this could restrict infrastructure not only in the United States but also in allied and partner countries.

His concern is that the U.S. could simultaneously:

1. prevent enough domestic data centers from being built;
2. prevent allies from accessing chips needed to build them abroad.

He asks where American-aligned compute infrastructure would then go.

The panel observes that Middle Eastern countries have large energy resources and geographically can serve very large populations with relatively low latency.

Chamath says he previously expected more rapid Middle Eastern data-center development but some projects stalled.

He says an unexpected infrastructure expansion occurred in “18” [transcript unclear—possibly a country or region was mistranscribed] and specifically Australia.

He had expected Australia's political environment to be less favorable but says major projects were nonetheless completed.

Chamath calls restrictive policies in wealthy states such as New York or California “luxury regulations”: prosperous jurisdictions can afford to reject investment longer than regions with weaker productivity, higher debt or a greater need for foreign direct investment.

Texas is cited as a U.S. jurisdiction more likely to welcome the economic activity.

The panel also says the UAE had just obtained the ability to import leading-edge chips, making the Middle East newly attractive again.

The speakers discuss larger enterprise Starlink systems.

They claim bundled systems can approach 10–20 gigabit connectivity.

If high-bandwidth satellite links continue improving, they argue, data centers can be built in places chosen primarily for electricity and land availability rather than proximity to traditional fiber infrastructure.

This gives the geographically distributed compute thesis another potential accelerator.

Friedberg or another speaker introduces an analogy involving Russia Today and anti-GMO sentiment.

The argument presented is:

1. commercial GMOs became widespread in the U.S. beginning around 1996;
2. strong anti-GMO sentiment allegedly did not emerge immediately;
3. Russia Today began operating in the U.S. around 2010;
4. after that, anti-GMO content grew;
5. that media narrative was allegedly amplified through bloggers and social-media networks;
6. after Russia Today lost distribution or was removed from U.S. outlets, anti-GMO interest supposedly declined.

The speaker says a Google Trends chart illustrates the pattern.

These are causal claims and require careful verification; temporal correlation is not itself proof of causation.

The speaker relates the alleged strategy to a Cold War KGB influence technique called “directed measures” [term as transcribed; verify exact historical terminology], in which foreign actors promoted narratives through media to shape Western public opinion.

He speculates that Germany's rejection of nuclear energy might have had a comparable origin.

The speaker argues that some contemporary opposition to AI infrastructure resembles the earlier GMO pattern:

• NGOs receive funding;
• media narratives reinforce their claims;
• social-media reposting amplifies them;
• public opinion eventually absorbs the claims as common knowledge.

He says a poll released that day showed more than 50% of Americans believe data centers increase the cost of water and electricity.

He argues this sentiment remains even when a hypothetical data center recycles its water and generates its own power.

The speaker describes a broader emotional response involving hostility toward:

• wealthy people;
• technology;
• AI.

He worries that foreign interests may be contributing to that shift but initially avoids the stronger word “influence,” instead suggesting “foreign interest.”

Sacks responds that the concern may be more concrete.

He cites an OpenAI blog post from approximately one month earlier, transcribed as:

“PRC linked influence operations are targeting AI debates in the U.S.”

He says Politico and other outlets also covered the allegation and claims that China is believed to be involved in campaigns designed to shape American views about AI and data centers.

The panel reasons that this would align with China's strategic interests if slowing U.S. infrastructure helped China gain ground in AI.

They speculate that a foreign competitor would benefit if the U.S.:

• suppresses data-center construction;
• constrains open-source models;
• permits regulatory capture;
• forces American businesses to use models costing much more than overseas alternatives.

Using their earlier token prices, they say a U.S. company paying $56 per million tokens while a competitor pays $0.50 faces a roughly 50–100x cost disadvantage.

The panel returns to the concept attributed to “Sacha/Sotia” [transcript unclear].

Their argument is that enterprises may pay frontier AI providers in two currencies:

1. financial fees;
2. proprietary knowledge/data that may improve the provider's system or weaken the customer's information advantage.

They refer to this proprietary advantage as the enterprise's “alpha.”

Thus an expensive closed-model service could, in their view, simultaneously be more costly and expose valuable institutional knowledge.

Sacks says foreign governments plainly have incentives to influence U.S. affairs and that Americans should not assume such efforts do not occur.

He argues many Americans fail to reason from first principles and declares that the United States is experiencing a “huge moral panic” about AI.

His core argument is that the most dramatic AI catastrophes under discussion remain hypothetical:

• catastrophic cyber outcomes have not yet occurred in the form being feared;
• mass job-loss predictions have not yet materialized to the degree predicted.

Meanwhile, he fears the United States may damage what he calls the “crown jewel” of its economy: free-market innovation characterized by rapid iteration, risk capital and easy company formation.

Sacks observes that the policy “Overton window” has moved so far that the panel is considering a FINRA-like regulator the least bad alternative.

He characterizes FINRA as essentially stockbrokers writing rules for stockbrokers and argues financial markets are not renowned for technological innovation.

Another speaker points to Robinhood making trading free as an innovation; Sacks dismisses that as primarily pricing innovation rather than fundamental sectoral innovation.

His larger warning is that even industry self-regulation carries innovation costs, but he sees it as preferable to direct government model licensing.

Sacks says a Chinese model transcribed as “Kimmy K3” had just appeared and was being described as very close to the AI frontier.

He warns the U.S. may have only months of lead over China, if that.

Therefore, he argues, the U.S. should not create slow regulatory approval structures to address risks that have not yet manifested.

His preferred approach is monitoring rather than panic.

The panel gives several examples of risks it believes are worth watching.

China is said to have paused or restricted new self-driving permits because autonomous driving was advancing successfully enough to create job losses and some civil unrest.

The panel considers such targeted observation reasonable.

A system/model transcribed as “Mythos” is said to be under examination for whether it can hack systems, including testing by Palo Alto Networks and others.

The product identity and claim require verification.

The speakers note only a small number of companies are at the actual frontier and say those companies already perform:

• safety testing;
• red teaming;
• related risk evaluation.

Their conclusion is that monitoring is justified but a vast new regulatory apparatus is not yet justified by observed harm.

Jason and Sacks criticize statements attributed to Dario.

The transcript ultimately gives the specific prediction as:

50% of entry-level knowledge-worker jobs could disappear within one to five years.

The speakers say the prediction had been made approximately one year earlier and regard it as exaggerated.

They also recall proposals to impose regulatory approval above 10^25 FLOPs.

The panel says current leading models have already exceeded that compute threshold without producing the catastrophic harms that earlier arguments associated with such scale.

They jokingly describe the earlier concern as expecting enough compute to create “the Terminator” or “Skynet.”

Jason speculates—while saying he does not mean to make light of mental health—whether Dario's repeated concerns might reflect unusual anxiety or “delusions of grandeur.”

He invites Dario to appear on the podcast.

Another speaker jokes that the invitation is unlikely to work after accusing him of a panic attack.

Sacks offers a non-psychological explanation: he believes Anthropic's approach can be understood as a rational business strategy:

1. brand yourself as the safe AI company;
2. ban unsafe AI;
3. profit.

This is Sacks's interpretation of Anthropic's incentives, not evidence of an admitted company strategy.

The hosts promote the All-In Summit in September and say scholarships are open at the show's events site.

Before Science Corner, Friedberg promotes a dog named Ronnie [name may be “Ronnie”; verify] for adoption through Family Dog Rescue in Sonoma and asks listeners to consider giving the dog a home.

The panel then engages in deliberately offensive joking about eating the dog, Sri Lanka, seasoning and marinades. The comments are comedic banter rather than factual claims.

Friedberg introduces the segment by recalling previous discussions of Yamanaka factors, described as proteins capable of entering cells and reversing aspects of cellular aging so cells behave more like younger cells.

He says substantial progress continues on that approach.

The new paper, however, focuses not on rejuvenating the cell itself but on aging in the extracellular matrix—material outside and between cells.

The paper is attributed to a collaboration between:

• Calico, described as Google's secretive age-reversal company;
• Revel Pharma [transcript spelling; verify exact company and authorship].

Friedberg explains aging of the extracellular matrix as an accumulation process.

According to his explanation:

• sugars and fats bind to proteins outside cells;
• those modifications accumulate over time;
• the body becomes less able to clear and replace affected proteins;
• tissue becomes “sticky” or structurally cross-linked;
• mobility decreases;
• skin wrinkles;
• joint/tissue function can deteriorate.

He calls the process glycation.

Proteins such as collagen, which should maintain particular spacing and structures, can become linked together.

Because modified proteins no longer look structurally normal, Friedberg says the immune system may attack them, producing inflammation.

He links this mechanism to the broader increase in chronic inflammation with age.

Friedberg says the modified compounds are known as advanced glycation end products (AGEs).

He identifies CML as a major molecule/product in the extracellular matrix involved in this aging mechanism.

He says the problem is that nothing in the body efficiently breaks it down.

The researchers' objective was therefore to create an enzyme capable of degrading CML.

Friedberg explains the biotechnology workflow.

• proteins are chains of amino acids;
• amino-acid sequences are encoded by DNA;
• DNA can therefore be designed and synthesized;
• bacteria can be used to produce the resulting proteins;
• those proteins can then be experimentally screened.

He describes this as a modern protein-synthesis and testing system.

The researchers used AlphaFold to identify a protein capable of binding CML and initiating an enzymatic process that could break it down.

The starting protein reportedly originated from a bacterium.

The researchers then:

1. produced the candidate protein;
2. identified regions that could potentially be improved;
3. altered the underlying DNA;
4. created hundreds and then thousands of variants;
5. measured how effectively each variant degraded CML;
6. repeated this process through five cycles.

Friedberg describes this iterative process as directed evolution.

Once the enzyme performed well against CML in a test tube, the researchers tested it against biological proteins named in the transcript as:

• casein [transcribed “cassine”];
• collagen;
• retinal proteins;
• hemoglobin.

Friedberg says the optimized enzyme removed approximately 52% to 97% of CML, with degradation exceeding 90% at several sites.

The researchers then applied the enzyme to human skin donated by elderly patients.

Friedberg says:

• the donors were older than 70;
• CML was reduced by approximately 55%;
• by the CML-based measure discussed, the treated skin resembled the skin of a 31-year-old.

This “reversed the skin's age” formulation is Friedberg's description of the reported biomarker result and should not be converted into a claim that whole human skin or humans were literally rejuvenated to age 31.

Friedberg emphasizes the combination of:

• AlphaFold;
• computational protein discovery;
• DNA programming;
• directed evolution;
• high-throughput experimental screening.

He says the process yielded a novel protein that does not naturally exist in its optimized form and performs a potentially important health function.

For him, the result illustrates how AI-assisted protein design can produce new therapies rather than merely analyze existing biology.

Friedberg says the discovery does not solve the therapeutic-delivery problem.

Questions include whether the enzyme could be delivered by:

• topical cream;
• injection;
• supplement;
• RNA-based injection that makes the protein inside the body.

The challenge is getting sufficient active enzyme into the extracellular matrix where CML accumulates.

He considers the work promising but says these practical questions remain unresolved.

One speaker predicts the first commercial market would be cosmetic skin treatment, not treatment for deeper age-related disease or joint problems.

The reasoning is straightforward: if the enzyme can penetrate skin through a cream and reduce aging-associated extracellular products, topical rejuvenation would have enormous consumer demand.

The speakers speculate that such a product could create a:

• $1 trillion market;
• possibly $2 trillion market.

These are spontaneous market-size predictions, not sourced valuations.

The hosts use the research as a counterweight to the episode's earlier discussion of AI fear and regulation.

They argue that AI-enabled scientific discovery can produce tangible benefits shared broadly across society.

AlphaFold is specifically credited as an enabling technology in the protein-discovery process.

One speaker says this kind of application is an example of why AI is “actually awesome” and argues that people should not allow possible foreign propaganda or generalized AI fear to obscure potential benefits.

The episode closes with expressions of appreciation between the hosts and miscellaneous closing banter and merchandising references.

Exact quote: “The whole industry is going to need to be regulated. And I think the industry needs to regulate themselves.”

Context: Jason plays or recalls his earlier argument for industry-led model safety certification.

Supports: Self-regulation / AI governance.

Exact quote: “a new government agency that I think would rapidly become a DMV for AI.”

Context: Explaining why he prefers a self-regulatory organization to direct federal model approval.

Supports: AI regulation / bureaucratic delay.

Exact quote: “it has to include startups and open source. It can't just be the three biggest labs”

Context: First of his five conditions for an acceptable AI SRO.

Supports: Competition / regulatory capture.

Exact quote: “The only reason to have this is for truly catastrophic risk.”

Context: Arguing that the regulator should address cyber and CBRN risks rather than speech issues.

Supports: Narrow AI-risk regulation.

Exact quote: “That will not be the end of it. That will just be the opening bid and the government will come back to take more and more and more.”

Context: Warning that an SRO without legal limits could become merely the first layer of expanded regulation.

Supports: Government power / regulatory escalation.

Exact quote: “It's really the PayPal diaspora. Our homeland was taken over and they burned our temple and then kicked everybody out.”

Context: Describing how eBay's acquisition dispersed PayPal's founding team.

Supports: Founder culture / PayPal Mafia history.

Exact quote: “I think it's a shot across the bow for Visa and MasterCard.”

Context: Describing the potential impact of combining Stripe, PayPal/Braintree and Block infrastructure.

Supports: Payments strategy / M&A.

Exact quote: “The only thing you can bring to your new job is what's in your head.”

Context: His rule for avoiding trade-secret disputes when changing employers.

Supports: Employment / intellectual property.

Exact quote: “A couple dots make a line, I guess.”

Context: Asked about the number of lawsuits surrounding OpenAI.

Supports: OpenAI legal risk.

Exact quote: “privacy in AI is very fragile and it's very brittle.”

Context: His conclusion from the Grok Build data-upload episode.

Supports: AI privacy / enterprise security.

Exact quote: “Token spend among ramp customers has grown by 21 times. 21 times. 21 times. Not 21 percent. We're talking about 21 times.”

Context: Played clip announcing corporate token-spend controls.

Supports: AI operating costs / CFO controls.

Exact quote: “It is a money-burning furnace.”

Context: Warning that AI consumption without demonstrated ROI can damage corporate earnings.

Supports: AI economics / enterprise spending.

Exact quote: “It's going to be wild when people have unlimited tokens on their desk.”

Context: Predicting large-memory Apple hardware will enable extensive local model inference.

Supports: Edge AI / Apple hardware.

Exact quote: “we're going to have GPUs chasing energy.”

Context: Predicting AI compute will migrate to locations with available electricity.

Supports: Data centers / energy scarcity.

Exact quote: “We have a huge moral panic going on with respect to AI.”

Context: Arguing that catastrophic predictions have outrun observed harm.

Supports: AI risk debate / regulation.

Exact quote: “we're on the threshold, I think, of destroying the crown jewel of our economy, which is the system of free market innovation that we have”

Context: Warning that excessive AI regulation could weaken U.S. innovation.

Supports: Technology policy / economic competitiveness.

Exact quote: “Number one, brand yourself as a safe AI company. Number two, ban unsafe AI. Three, profit.”

Context: Sacks offers a strategic rather than psychological explanation for Anthropic's regulatory posture.

Supports: Anthropic / regulatory incentives.

Exact quote: “they were able to eliminate 55% of the CML on the skin, which basically reversed the skin's age down to the age of a 31-year-old.”

Context: Describing the reported experiment on donated skin from people over 70.

Supports: Glycation / experimental rejuvenation.

Exact quote: “It's just so profound that we have this pool at our disposal in this day and age.”

Context: Praising AlphaFold-enabled protein engineering.

Supports: Positive scientific applications of AI.