Sarah Friar | All In Summary

1. Key Themes

Anthropic's Regulatory Capture Strategy Is No Longer a Fringe View

What was once a "spicy take" has become near-consensus among the hosts. Sacks laid out the core argument: Anthropic is simultaneously engaging in mandatory surveillance of users, nerfing model outputs without disclosure, and calling for a new government regulatory agency to constrain competitors — especially open source.

"Eight months ago, I said that Anthropic was engaged in a very sophisticated regulatory capture campaign based on fear mongering. And people at the time thought that was a very spicy take. But eight months later, I think you're hearing a lot of people say it. In fact, I think it's almost now becoming a new consensus." — David Sacks 00:09:54

Anthropic's Surveillance and Secret Model Degradation Is an Existential Enterprise Risk

Fable 5's mandatory 30-day data retention (overriding even existing zero-retention enterprise agreements) combined with undisclosed model nerfing for users flagged as AI researchers creates serious business risk for enterprise customers.

"Even enterprise customers who had signed zero data retention agreements, they do not have a choice... And it's not just the prompts and the output. Remember, it's all the context you share with them. So all these agent platforms are basically storing all of your memories, all of your files, all of your data. And they're passing them to the model in these giant context windows to get better responses. Anthropic is saying it will keep all of that." — David Sacks 00:10:41

Closed Model Restrictions Are Actively Driving Enterprises Toward Chinese Open Source Models

Friedberg's direct operational experience at Ohalo makes this concrete: legitimate genomics research is being blocked, and the only alternatives are open source models — which are currently led by Chinese providers.

"As folks like Anthropic say, hey, we're going to restrict access or censor the output of these models, it is going to force companies like ourselves, who still want to take advantage of the capability of these LLMs to go and get open source tools and run them. And what are the best open source models today? They're Chinese. That is a major concern. The American open source models are not as good as the Chinese open source models." — David Friedberg 00:07:04

AI's Marginal Cost Structure Is Fundamentally Different From the Internet — and That Changes Everything

Calacanis articulated a structural insight that justifies public ownership claims: unlike the internet (where each marginal user costs nothing), every AI query consumes real compute, power, and memory — and that infrastructure is heavily dependent on government-built foundations.

"AI is completely different. There is a real cost for every marginal user. Everyone you stand up is taxing a GPU. Everyone you stand up needs electrons. Everything needs memory... It is not dissimilar to having the first set of transportation companies that run on the interstate highway. And if the interstate highways were built by the federal government writ large, and now you had two companies that transported all the goods, a logical question at that time could have been how much of that should I own? Because you're riding on my rails." — Jason Calacanis 00:53:01

AI Is a Revenue Multiplier, Not a Job Destroyer — And the Data Is Already Proving It

Friedberg pushed back forcefully on the Dario/Elon job-loss narrative with ground-level evidence, arguing AI primarily expands revenue capacity rather than cutting costs.

"Where I see AI being used is on the revenue side, a hundred times more than the cost side. And in that equation, people are hiring like crazy. We cannot hire enough people. I just had a review meeting with my product and engineering team two days ago. And they're like, we want to add an extra 15 headcount to our engineering squads because we have all this opportunity to do stuff that we couldn't otherwise do." — David Friedberg 00:47:05

Bernie Sanders' AI Wealth Seizure Proposal Is a Direct Consequence of AI CEOs' Own Messaging

The hosts converged on an ironic observation: the political appetite to confiscate equity from AI companies was manufactured by those same companies telling the public they would cause massive job loss.

"Look, I think that it's a natural political reaction based on what these AI companies have said... These AI companies trained on all that data, and now they're basically saying they're going to use it to put Americans out of work. Now, obviously, when you repeat that message over and over again, ordinary Americans are going to say, what's in this deal for me?" — David Sacks 00:40:00

"They poke the tiger... I'm so sick of defending these idiots. It's a stupidity tax because they've been out there teaching the public that what they do is harmful." — David Sacks 00:49:40

California's Election Laws Have Created a Structurally Appointive System, Not a Democratic One

Friedberg and Sacks argued that a layered accumulation of individually defensible California laws — unlimited ballot harvesting, automatic mail ballots to all registered voters, no citizenship proof required to register, lax signature verification, no audit requirements — has collectively destroyed the one-person-one-vote principle.

"There is nothing illegal or fraudulent going on. In fact, the system is operating exactly as intended. It has been set up and structured in a way that with the right construct you can get an individual appointed, not elected, but appointed to a particular role in government under a quote free election in California." — David Friedberg 00:16:33

Inflation Is Running Hot and the Iran War Is the Accelerant

CPI at 4.2% year-over-year (highest since April 2023) and PPI at 6.5% (highest since end of 2022) arrived as a shock. The hosts pointed to the Iran war as the primary catalyst and flagged the risk of oil spiking to $150-$200/barrel if China needs to return to spot markets.

"If China somehow runs out of reserves and they need to go back into the spot market to buy an extra 3 million barrels a day, there's a very big risk that oil gets into the well past 100 and maybe between 150 and 200 a barrel... I think the PPI number should be an alarm that we use to off-ramp this Iran situation sooner rather than later." — Jason Calacanis 00:09:26

The Venture Valuation Escalation Curve Has a Surprisingly High Completion Rate at the Top

Thomas Lafont's data revealed a counterintuitive pattern: the probability of advancing between valuation milestones increases as companies get bigger, not decreases.

"The odds of a unicorn getting to decacorn was 8%. The odds of a decacorn getting to a centacorn was about double that, 13%. And then the odds of it going from centacorn to a trillion dollar market cap was 31%. So double again. And I asked him, could you extrapolate this — on 1 trillion getting to 10 trillion — my guess is that it would follow and we'll see about 60% of trillion dollar market cap companies getting to 10 trillion in the next few years." — David Sacks 00:51:00

2. Contrarian Perspectives

The "Safety" Framing Is Cover for Anti-Competitive Behavior, Not Genuine Risk Management

Most people accept Anthropic's safety narrative at face value. The hosts argued it is strategically constructed to justify regulatory capture and eliminate open source competition — and the tell is the absence of KYC.

"I think if you did not have that agenda, you would have implemented KYC yesterday... A business executive inside your corporation could trip it. A person doing scientific molecular research could trip it. And all of a sudden, you'll get cut off from a very important source of business differentiation for yourself." — Jason Calacanis 00:03:16

"Freeberg should be able to go to Anthropic and say, I'm David Friedberg. Here's what I do. Here's a security bond I'm willing to post. Here's the names and addresses of all my employees. Give me Fable 5 and don't nerf me. He can't do that." — Jason Calacanis 00:28:45

You Cannot "Turn Off" AI — Open Source Models Already Exist and Cannot Be Recalled

The conventional view is that regulation can meaningfully constrain AI capability. Friedberg argued this is wrong: the models are already published, copied, and distributed globally, exactly like books after the printing press.

"The idea that you can regulate or downscale or turn off AI is not a realistic idea. The models have been put out in the world. It's like publishing a book. Once the book has been printed, anyone can use their own Xerox machine at home to make copies of it and use it. So we have crossed the Rubicon in terms of the potential of language models." — David Friedberg 00:35:28

Restricting AI Harms America More Than It Protects It — China Is the Beneficiary

Counterintuitively, the self-imposed restrictions by American AI labs on legitimate scientific users directly advantage Chinese open source providers who face no such constraints.

"The restrictions that Anthropic and others are putting upon themselves and upon the industry is forcing a lot of companies to go and get open source Chinese models and run them. We're seeing this across the landscape with startups. We see it with large scale enterprises. Everyone's making that move." — David Friedberg 00:07:29

California's Election Problems Are Not Fraud — They Are the System Working as Designed

Most commentary frames California's mail-in ballot irregularities as either fraud (MAGA) or conspiracy theory (Democrats). Friedberg's position is more unsettling: no laws are being broken. The laws themselves have been engineered to produce these outcomes.

"There is nothing illegal or fraudulent going on. In fact, the system is operating exactly as intended... The principle should be one individual, one vote. And if you opt not to vote, your vote should not be counted." — David Friedberg 00:16:33

AI CEOs' Job Loss Narrative Is Self-Defeating and Already Empirically Wrong

While Dario Amodei predicted 50% entry-level knowledge worker job loss in one to five years, and that prediction is now more than one year old, the actual jobs data shows software developer jobs at a three-year high and 172,000 new jobs added in May alone.

"There is no job loss with AI. I will say it again. I've said it a thousand times... The idea that AI is going to destroy jobs is a Luddite idea that is being disproven every single day. And I see it on the ground. It is only a matter of time before people wake up to this." — David Friedberg 00:47:34

3. Companies Identified

Anthropic Developer of the Claude/Fable frontier AI models. Mentioned critically for its mandatory 30-day data retention on Fable 5 (overriding enterprise zero-retention agreements), secret model nerfing of users flagged as AI researchers, and Dario Amodei's public campaign for a new AI regulatory agency — which the hosts characterize as regulatory capture.

"They would give you a nerfed answer. They would not tell you what they were doing. They would still charge you for the product that you thought you were getting. And they would never tell you that you were not getting frontier model capability. So they were actually misleading their users." — David Sacks 00:12:09

ARC Institute Non-profit research institute co-founded by Patrick and John Collison. Built and open-sourced a genome language model trained on all available global genomic data, which Friedberg's company Ohalo uses actively in its plant breeding program.

"The ARC Institute ingested all the world's genomic data that they could get access to. And it's a genome language model, basically... It's figured out the language of DNA because it sees the probability of those letters being in a row in other organisms. And so it's a very powerful tool." — David Friedberg 00:19:46

Ohalo Friedberg's agricultural genomics company. Mentioned as a real-world case study of a legitimate science company being harmed by Anthropic's model restrictions on genomic/RNA work, and as an example of a company actively hiring more engineers because of AI productivity gains.

"We are likely going to end up needing to use open source models and run them locally ourselves... We cannot hire enough people. I just had a review meeting with my product and engineering team two days ago. And they're like, we want to add an extra 15 headcount to our engineering squads." — David Friedberg 00:06:36 and 00:47:05

OpenAI Mentioned in multiple contexts: as a target of Bernie Sanders' proposed 50% equity seizure, as a public benefit corporation, and for Sam Altman's relatively more pragmatic posture compared to Dario Amodei.

"Sam has walked back. Sam has said that he basically — I don't think he quantified — but he said that job loss was coming. But he said more recently that he was wrong because they're not seeing that in the numbers." — David Sacks 00:49:07

Palo Alto Networks Nikesh Arora (CEO) was cited as a credible third-party validator of the Mythos-class model's genuine capability, using it to identify and close real security vulnerabilities inside Palo Alto Networks before public release.

"We had the CEO of Nikesh of Palo Alto Networks who said, hey, Mythos was the real deal and that they used it to seal up some vulnerabilities inside their shop." — Chamath Palihapitiya 00:00:59

XAI (xAI) Elon Musk's AI company. Named as one of the targets of Bernie Sanders' proposed 50% equity seizure legislation alongside OpenAI and Anthropic.

"He announced the American AI Cyber Wealth Fund Act... a one-time 50% tax on stock, not profits of the largest AI companies, including OpenAI, Anthropic, and XAI." — Chamath Palihapitiya 00:38:25

Meta / Llama Mentioned critically for failing to execute on the strategic opportunity to dominate open source AI and commoditize the frontier model market, despite having the resources to do so.

"Meta really fumbled this with Llama. If they had landed a really good working open source model... What a fumble. Take the margin out for the rest of us." — Jason Calacanis 00:23:15

Stratechery Ben Thompson's technology analysis publication. Mentioned because Thompson himself was kicked down to a lesser model when asking a straightforward question about cancer risk and GLP-1s — a vivid illustration of Fable 5's overly broad restriction triggers.

"When Ben Thompson from Stratechery asked a very straightforward question about the relationship between cancer risk and GLP1s, he was kicked out." — David Sacks 00:12:33

Athena Executive staffing/operations company. Mentioned as an All In Podcast partner; their Chief Business Officer Chris Ho (described as a plus-one golfer) played in the Liquidity event golf scramble.

4. People Identified

Thomas Lafont Venture capital analyst/presenter. Delivered a data-rich presentation on venture valuation progression rates at the All In Summit approximately two years ago that the hosts are now making an annual feature. His data on unicorn-to-decacorn-to-centacorn-to-trillion progression rates was described as uniquely substantiated.

"I thought Thomas Lafont crushed it with his presentation on the overview of venture capital... It provided so much numerical detail to support a lot of things I was kind of feeling but didn't have the data for." — David Sacks [00:01:00:02]

Sarah Friar CFO of OpenAI. Described as having crushed her appearance at the All In Liquidity event, with buzz in the room that she could be a future CEO of OpenAI.

"Sarah Fryer... she was incredible, the CFO of OpenAI. Future CEO — I think was the buzz in the room." — Jason Calacanis 00:59:20

Nikesh Arora CEO of Palo Alto Networks. Cited as an authoritative external validator of Anthropic's Mythos-class model capability, using it to find and close real security vulnerabilities at his company.

"We had the CEO of Nikesh of Palo Alto Networks who said, hey, Mythos was the real deal and that they used it to seal up some vulnerabilities inside their shop." — Chamath Palihapitiya 00:00:59

Andrej Karpathy AI researcher, formerly at Tesla and OpenAI. Mentioned because Anthropic hired him to run recursive self-improvement research — in the same month that Dario published a blog post warning that recursive self-improvement could end the world, which the hosts described as a glaring contradiction.

"Last week, they published this blog saying that recursive self-improvement could end the world. Therefore, we need a pause. What did they do the previous month? They hired Andrej Karpathy to run recursive self-improvement at Anthropic. They're complete hypocrites." — David Sacks 00:50:35

Patrick and John Collison Co-founders of Stripe. Cited for funding and supporting the ARC Institute's open source genome language model — presented as a model (pun intended) for how private capital can fund open source AI to create public benefit and competitive alternatives to closed models.

"It's a good example where a community, in this case, the Collisons and others, put money behind it to fund this research and output this open source model." — David Friedberg 00:20:43

Steve Hilton Former British political strategist, Fox News commentator, California gubernatorial candidate. Named as the one viable "break the glass" solution to California's broken election system — the argument being that if elected governor, he could declare a state of emergency on day one and begin restructuring how the state is run.

"There is one break the glass way... You must get somebody at the absolute top of the ticket elected. That person is Steve Hilton. And if that happens, there is one thing that Steve Hilton can do to start to completely change the way that this state is run, which is instantly on day one declare a state of emergency." — Jason Calacanis 00:27:12

Jake Paul Boxer and investor. Attended the All In Liquidity event; praised by multiple hosts for his investment acumen and the quality of his team.

"Jake Paul was awesome. I sat him next to Sacks for dinner. He's a great investor, by the way, fantastic. His portfolio is awesome." — Jason Calacanis 00:02:45

Thomas Keller Chef/owner of The French Laundry. Hosted the speaker dinner for the All In Liquidity event at The French Laundry; mentioned admiringly for his watch collection (Patek Philippe Aquanaut, green band, rose gold).

Kevin Warsh Presumed incoming Fed Chair (referenced by Friedberg). Cited in the context of expectations for tighter future monetary policy.

"I think with the Kevin Warsh Fed I think we could see north of future of our monetary policy." — David Friedberg 00:07:04

Brenda Lee Brown Armstrong Named as an individual who pled guilty in 2026 to paying people in the Skid Row area for votes in the LA mayoral election — the concrete criminal case cited to substantiate the ballot harvesting fraud allegations around Spencer Pratt's race.

"Brenda Lee Brown Armstrong, she pled guilty in 2026, paying..." — Chamath Palihapitiya 00:24:54

5. Operating Insights

Treat AI Model Access as a Single-Point-of-Failure Business Risk and Govern It Accordingly

Calacanis framed this as the core enterprise lesson from the Fable 5 episode: companies that have built differentiation on top of a single closed model are now exposed to being silently downgraded, surveilled, or cut off entirely — without notice or recourse.

"Companies need to start underwriting this next phase of AI, which is: how do I have control? Who am I allowing to learn off of all of this information? Do I want to have single point of failure risk with respect to AI? And I think the answer is that you need broad diversity and a governance approach that's better managed." — Jason Calacanis 00:03:55

The Path From Closed Model Dependency to Proprietary Model Is Already Being Walked — Follow It

Friedberg described Ohalo's roadmap in real time: start with closed models, get blocked, move to open source, run locally, combine with proprietary data, build your own domain-specific model. Palihapitiya predicted this trajectory for Friedberg before Friedberg confirmed it. This is a replicable playbook.

"We'll start with the core model, combine it with our data, and then we'll have our own genome language model or our own prediction model that we'll then use internally. And I think that's where folks are going." — David Friedberg 00:07:58

Use Open Source Domain-Specific Models as a Competitive Moat Before Competitors Do

The ARC Institute genome language model example shows that open source domain-specific models — funded by private capital — can provide defensible scientific capability that closed models are now actively restricting. First movers in a given domain who build on and fine-tune these models with proprietary data will have durable advantages.

"Others are taking it in and they use it very actively. So it's part of our plant breeding program as an input. It's a good example where a community — in this case, the Collisons and others — put money behind it to fund this research and output this open source model. And I think we see more of that kind of coming down the pipe, which creates a very great advantage against the closed proprietary model ecosystem." — David Friedberg 00:20:43

AI Investment Should Be Framed Around Revenue Expansion, Not Cost Reduction

Friedberg's operational insight reframes how to pitch and measure AI ROI internally. The cost-reduction framing invites headcount debates; the revenue-expansion framing justifies hiring more.

"On the cost side of the equation, AI can be used to reduce humans doing things that cost money to some extent. The effect there, I would argue, is nominal. The real opportunity with AI is on the revenue side, where suddenly one engineer can do a hundred times or a thousand times what they used to be able to do." — David Friedberg 00:46:35

6. Overlooked Insights

Compute Is the New Regulatory Lever — and Calacanis Is Already Positioning for It

Buried in the conversation was a stunning disclosure: Calacanis bought 2,000 acres in Arizona two years ago, got it zoned for a 2-gigawatt data center, and has now put in an offer for another gigawatt site — not to flip to Blackstone or Google as originally intended, but because he concluded that without independently controlled compute directed at open source model infrastructure, everything else discussed about AI freedom is moot. The cost of this infrastructure has increased 20x (from ~$4-5B to ~$100B per gigawatt) in two years, creating an almost insurmountable capital moat that makes the open source compute problem nearly unsolvable without sovereign or institutional capital.

"I thought that this was the moment where you just turn around and you flip it to the Blackstones, the Brookfields of the world, to the Googles of the world, let them develop it. And I've come to the conclusion that I don't think I can... I'm coming to the conclusion that I think we may just be dragged into having to build two gigawatts. And I just put in an offer for another gig in a different place because I'm like, if you could take three gigawatts and now actually create a deep and liquid access to open source, we're going to have to do it... A gigawatt now costs $100 billion, guys. When I started this project, it was like $4 or $5 billion. And it's increased by 20x." — Jason Calacanis 00:17:14

This is the most actionable and underreported detail in the episode: the person most vocally advocating for open source AI compute independence is quietly building the infrastructure to back that thesis with real capital — and the window to do it at reasonable cost has already largely closed.

The International Gene Synthesis Consortium's 2009 Voluntary Framework Is the Template for AI Safety — and It's Already Being Codified

Sacks flagged, and Friedberg confirmed, that the major AI labs and oligosynthesis companies have signed a letter supporting mandatory codification of the 2009 International Gene Synthesis Consortium's voluntary screening protocols — which require labs to check synthetic DNA/RNA orders against bioweapon databases before manufacturing. This flew by in under two minutes but is significant: it represents a concrete, already-proven, 15-year-tested model for AI safety intervention that targets the output (physical synthesis) rather than the input (model access). If this framework is extended as a template for other AI safety interventions, it would fundamentally shift the regulatory debate away from model access restrictions toward point-of-use controls — which is exactly what Friedberg argued for, and which would be far less commercially damaging to legitimate researchers.

"The major AI labs, along with lots of other signatories, sent this letter recently. It's called 'In Support of Mandatory Nucleic Acid Synthetic Screening and Record Keeping'... That is based on an agreement called the International Gene Synthesis Consortium of 2009, in which all the major labs agree to develop and implement voluntary safeguards against misuse... Now they're saying we should make it mandatory. Okay, that seems reasonable to me. So to Freeberg's point about at what stage do you intervene — maybe it's not at the level of who gets to use the model, but if you try to turn it into output in the physical world." — David Sacks 00:32:57