Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California's Broken Elections
Anthropic secretly downgraded users doing AI research and stored all their prompts — then Dario blogged calling for a new regulatory agency to ban competing models.
All-In with Chamath, Jason, Sacks & Friedberg
Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California's Broken Elections
Anthropic secretly downgraded users doing AI research and stored all their prompts — then Dario blogged calling for a new regulatory agency to ban competing models.
TL;DR
The All-In besties tackle Anthropic's controversial Fable 5 release, which secretly downgraded users doing AI research and stored all prompts for 30 days — sparking a developer backlash [1] — David Sacks "Anthropic's Fable 5 stores all your prompts for 30 days, secretly downgrades users it deems unworthy — including rewriting their prompts in…" 09:30 . David Sacks frames this as regulatory capture designed to eliminate open-source competition [2] — Chamath Palihapitiya "Chamath wants to direct compute toward open-source AI — he bought 2,000 acres in Arizona for a 2-gigawatt data center. Problem: one gigawat…" 17:10 . Bernie Sanders' proposal to seize 50% of AI company equity triggers a debate on sovereign wealth funds and job loss [3] — David Sacks "Bernie Sanders published a NYT op-ed proposing a one-time 50% stock tax on the biggest AI companies. Sacks hates it as confiscation — but u…" 39:30 . CPI hit 4.2% and PPI 6.5% — multi-year highs driven by the Iran war. The episode closes on California's ballot-harvesting laws and the LA mayoral race statistics that the hosts call statistically impossible [4] — Chamath Palihapitiya "The internet made adding users essentially free — Facebook and Google minted money because marginal cost was zero. AI is the opposite: ever…" 54:25 .
The All-In besties cover Anthropic's Fable 5 backlash, regulatory capture in AI, Bernie Sanders' AI equity seizure proposal, inflation data, and California's election integrity controversy.
-
Jason Calacanis welcomes the original All-In quartet back and previews a packed episode covering Anthropic, AI regulation, inflation, and California elections.
-
Fable 5 tops benchmarks but stores all prompts for 30 days, secretly downgrades AI researchers, and buries the policy in a 319-page document [1] — David Sacks "Anthropic's Fable 5 stores all your prompts for 30 days, secretly downgrades users it deems unworthy — including rewriting their prompts in…" 09:30 . Friedberg reveals his genomics lab is migrating to Chinese open-source AI as a result [2] — David Friedberg "Anthropic's safety restrictions have begun blocking legitimate gene-editing and genomics research at Ohalo, Friedberg's agricultural biotec…" 05:14 .
-
Sacks explains how Anthropic's simultaneous surveillance, downgrading, and lobbying for a new AI regulator is designed to eliminate open-source competition and create an AI monopoly or duopoly.
-
Sacks says he'd support taking equity from public benefit corporations that train for free on public data and gatekeep output. Jason calls AI CEOs 'capitalist cucks' for inviting regulation [1] — David Sacks "Sacks is a capitalist. He hates confiscation. But when AI companies train for free on all of humanity's knowledge, promise 50% job losses, …" 49:50 .
-
The hosts recap Liquidity conference highlights: Sarah Friar impressed as OpenAI CFO; Thomas Lafont's data showing centacorns have 31% odds of reaching $1T valuation [1] — David Sacks "Decacorn-to-centacorn odds: 13%: Thomas Lafont's data showed the odds of a decacorn ($10B company) reaching centacorn ($100B) status is 13%…" 1:00:45 . Summit 2026 announced for Sept 13-15.
-
Friedberg presents LA mayoral race statistics showing Pratt led in-person voting 35% to Raman's 26%, but post-election mail-in ballots flipped the result [1] — David Friedberg "On Election Day, Spencer Pratt led with 35%, Bass had 29%, Raman 26%. In mail-in ballots received after Election Day, Raman surged to 37% a…" 1:08:20 . He walks through California's ballot harvesting and voter registration laws [2] — David Friedberg "California legalized unlimited ballot harvesting, sends ballots to every registered voter whether they want one or not, lets anyone registe…" 1:09:45 .
-
Jason wraps up, calling for DOJ investigation of the LA election results. The besties sign off with the traditional outro.
- Regulatory capture
- When a government regulatory agency acts in the interest of the industry it's supposed to regulate, rather than the public — here used to describe Anthropic's alleged strategy of shaping AI regulation to favor itself.
- KYC (Know Your Customer)
- A standard financial compliance process requiring businesses to verify the identity of their clients; the hosts argue AI labs should use it to vet who gets access to powerful models.
- Ballot harvesting
- The practice of a third party collecting and submitting completed mail-in ballots on behalf of voters; legal in California with no limit on quantity collected.
- Sovereign wealth fund
- A state-owned investment fund that holds and manages a nation's financial assets, such as stocks or bonds, on behalf of citizens.
- Fable 5
- Anthropic's latest flagship AI model at the 'Mythos' capability tier, released in June 2026, notable for its mandatory data retention and tiered access restrictions.
- Mythos
- Anthropic's top model capability tier, above Opus — Fable 5 is the first publicly released Mythos-class model.
- Context window
- The amount of text, files, and data an AI model can process in a single interaction; larger context windows allow agents to hold entire project histories in memory.
- Prompt
- The input text or instruction a user sends to an AI model to generate a response; Anthropic retains all prompts submitted to Fable 5 for at least 30 days.
- Nerfing
- Deliberately weakening or downgrading a product's capability, typically without disclosing the change to users — here referring to Anthropic silently switching users to a less capable model.
- Nucleic acid synthetic screening
- A process by which laboratories check orders for synthetic DNA or RNA against a database of dangerous sequences to prevent bioweapon creation, currently governed by the International Gene Synthesis Consortium.
- Oligosynthesis
- The commercial synthesis of short strands of DNA or RNA (oligonucleotides) used in laboratory research — companies in this space are subject to bioweapon screening requirements.
- Decacorn
- A privately held startup valued at $10 billion or more, one tier above a unicorn ($1B) and below a centacorn ($100B).
- Centacorn
- A privately held or recently public company valued at $100 billion or more.
- Trillicorn
- Informal term coined by the All-In hosts for a company valued at $1 trillion or more.
- Adverse selection
- In insurance and economics, the phenomenon where a flawed incentive structure disproportionately attracts the worst-case actors — used here to explain how California's permissive ballot laws invite exploitation.
- Superannuation
- An Australian mandatory employer-funded retirement savings system that invests in equities; the hosts cite it as a model for reforming US Social Security into an equity-owning wealth fund.
- Recursive self-improvement
- A hypothetical process by which an AI system uses its own capabilities to improve itself, potentially leading to rapid capability acceleration — Anthropic blogged it could end the world while simultaneously hiring to build it.
- Zero-data-retention agreement
- A contractual clause between an enterprise customer and an AI vendor specifying that no prompt or output data will be stored by the vendor — Anthropic's Fable 5 voided these for enterprise clients.
- Fearmongering
- The deliberate use of alarming or exaggerated language to provoke fear in an audience, often for political or commercial advantage — used here to describe Dario Amodei's public communications about AI risk.
- Public benefit corporation
- A legal corporate structure requiring the company to balance profit motives with stated public benefits; both OpenAI and Anthropic are structured this way, providing a legal basis for government equity claims.
Chapter 2 · 00:19
Anthropic gets massive backlash over secret Fable nerfing and privacy concerns
Fable 5 tops benchmarks but stores all prompts for 30 days, secretly downgrades AI researchers, and buries the policy in a 319-page document [1] — David Sacks "Anthropic's Fable 5 stores all your prompts for 30 days, secretly downgrades users it deems unworthy — including rewriting their prompts in…" 09:30 . Friedberg reveals his genomics lab is migrating to Chinese open-source AI as a result [2] — David Friedberg "Anthropic's safety restrictions have begun blocking legitimate gene-editing and genomics research at Ohalo, Friedberg's agricultural biotec…" 05:14 .
Anthropic's safety restrictions have begun blocking legitimate gene-editing and genomics research at Ohalo, Friedberg's agricultural biotech company. The result: they're migrating to Chinese open-source models. US AI over-restriction is directly handing China a competitive advantage.
David Friedberg noted that the best open-source models available today are Chinese, meaning US AI restrictions push companies toward Chinese AI tools.
Anthropic's Fable 5 stores all your prompts for 30 days, secretly downgrades users it deems unworthy — including rewriting their prompts in the background — while still charging full price. Enterprise customers with zero-retention agreements get no opt-out. This isn't a bug; it's the product.
Anthropic isn't just worried about safety — it's running a sophisticated regulatory capture campaign. While surveilling users and downgrading competitors, Dario simultaneously blogs for a new FDA-style AI regulator. The target isn't dangerous actors; it's open-source alternatives.
Anthropic's Fable 5 stores every prompt and output — including from enterprise customers with zero-data-retention agreements — for at least 30 days, with no opt-out.
Fable 5 would silently nerf its output for users Anthropic deemed unworthy — including rewriting prompts in the background — while still charging full price.
Chamath wants to direct compute toward open-source AI — he bought 2,000 acres in Arizona for a 2-gigawatt data center. Problem: one gigawatt now costs $100 billion, up 20x from when he started. Building enough open-source compute to counter the closed-model monopoly would cost $300 billion he doesn't have.
Chamath Palihapitiya disclosed that the cost to build a 1-gigawatt data center has risen 20x — from $4-5 billion to roughly $100 billion — creating a massive capital moat.
The Ark Institute, backed by the Collisons, ingested all available genomic data and built a genome language model that reads DNA sequences the way a language model reads sentences. It's already changing plant breeding at Ohalo. This is what community-funded open-source AI looks like when it works — and it's the antidote to closed-model monopolies.
The same AI tools that can design bioweapons can cure cancer and grow more food. Restricting model access is the wrong intervention point — you can't turn off AI any more than you can un-print a book. The right approach is regulating the output: the actual creation of weapons, not access to knowledge.
Anthropic could implement Know Your Customer verification tomorrow. It doesn't. If the safety mission were real, KYC would let researchers like Friedberg prove their legitimacy and get unblocked. The fact that Anthropic chose blunt surveillance instead of targeted verification is the tell that safety isn't the actual goal.
Chapter 3 · 29:16
The AI regulatory capture trap, pragmatic safety solutions
Sacks explains how Anthropic's simultaneous surveillance, downgrading, and lobbying for a new AI regulator is designed to eliminate open-source competition and create an AI monopoly or duopoly.
Chapter 4 · 37:59
Nationalizing AI: Trump/Sanders, justifications, and AI's Capitalist Cucks
Sacks says he'd support taking equity from public benefit corporations that train for free on public data and gatekeep output. Jason calls AI CEOs 'capitalist cucks' for inviting regulation [1] — David Sacks "Sacks is a capitalist. He hates confiscation. But when AI companies train for free on all of humanity's knowledge, promise 50% job losses, …" 49:50 .
Bernie Sanders published a NYT op-ed proposing a one-time 50% stock tax on the biggest AI companies. Sacks hates it as confiscation — but understands the logic: AI CEOs trained on everyone's knowledge, promised to put half of America out of work, and now gatekeep the output. If that's your pitch, don't be surprised when people want a cut.
Senator Bernie Sanders published a New York Times op-ed proposing a one-time 50% tax on the stock of the largest AI companies to fund a sovereign wealth fund.
The May jobs report showed 172,000 new jobs — more than double what economists expected — with unemployment at a near-record low of 4.3%.
The Social Security Trust Fund holds exactly one asset: a $4 trillion IOU from the US government. Friedberg wants to blow it up, convert it into individual equity accounts, and use the capital to buy stocks in transformative American companies — including AI. This is the right way to let citizens participate in the AI boom without confiscation.
Friedberg noted the Social Security Trust Fund currently holds just one asset — a special US Treasury certificate representing a $4 trillion IOU from the federal government — and proposes converting it to equity ownership.
AI doesn't destroy jobs — it multiplies what each person can produce. The real story is on the revenue side: one engineer can now do what 100 used to do, creating more products, more companies, more opportunity. Friedberg's own company is hiring aggressively because of AI. The apocalypse narrative has been disproven by every single jobs report.
Sacks is a capitalist. He hates confiscation. But when AI companies train for free on all of humanity's knowledge, promise 50% job losses, then gatekeep the output from competitors — at some point you've invited the political response you're getting. He doesn't endorse Bernie's plan. He just understands it.
The internet made adding users essentially free — Facebook and Google minted money because marginal cost was zero. AI is the opposite: every user consumes GPU cycles, electricity, and memory. That means AI depends on massive public infrastructure subsidies. The government built the railroads; these companies are riding on public rails.
Chapter 5 · 59:22
Liquidity recap: Best moments and takeaways
The hosts recap Liquidity conference highlights: Sarah Friar impressed as OpenAI CFO; Thomas Lafont's data showing centacorns have 31% odds of reaching $1T valuation [1] — David Sacks "Decacorn-to-centacorn odds: 13%: Thomas Lafont's data showed the odds of a decacorn ($10B company) reaching centacorn ($100B) status is 13%…" 1:00:45 . Summit 2026 announced for Sept 13-15.
Thomas Lafont's data showed the odds of a decacorn ($10B company) reaching centacorn ($100B) status is 13% — double the 8% odds of a unicorn reaching decacorn.
Chapter 6 · 1:05:39
Inflation heats up: CPI and PPI see 3+ year highs
Friedberg presents LA mayoral race statistics showing Pratt led in-person voting 35% to Raman's 26%, but post-election mail-in ballots flipped the result [1] — David Friedberg "On Election Day, Spencer Pratt led with 35%, Bass had 29%, Raman 26%. In mail-in ballots received after Election Day, Raman surged to 37% a…" 1:08:20 . He walks through California's ballot harvesting and voter registration laws [2] — David Friedberg "California legalized unlimited ballot harvesting, sends ballots to every registered voter whether they want one or not, lets anyone registe…" 1:09:45 .
May CPI came in at 4.2% year-over-year, the highest reading since April 2023, driven in part by the Iran war's energy shock.
May Producer Price Index hit 6.5% year-over-year, the highest since the end of 2022, signaling mounting upstream inflationary pressure.
Polymarket odds of a Fed rate hike in 2026 jumped to 49%, up from under 10% before the Iran war started.
On Election Day, Spencer Pratt led with 35%, Bass had 29%, Raman 26%. In mail-in ballots received after Election Day, Raman surged to 37% and Pratt dropped to 19% — a one-third decline. Chamath says the odds of this happening randomly are one in a trillion. The question isn't whether something happened; it's whether it was legal.
On Election Day in-person voting, Spencer Pratt led with 35%, Karen Bass had 29%, and Nithya Raman had 26% — but post-election mail-in ballots flipped the results dramatically.
California legalized unlimited ballot harvesting, sends ballots to every registered voter whether they want one or not, lets anyone register with no citizenship proof, and requires almost no signature verification. None of these laws individually look fraudulent. Together, they create a system where elections are functionally appointments.
California Assembly Bill 1921, passed around 2018, legalized unlimited ballot harvesting — allowing any individual to collect and submit ballots for others regardless of relationship.
Chapter 7 · 1:12:27
California's loose election laws creating integrity doubts
Jason wraps up, calling for DOJ investigation of the LA election results. The besties sign off with the traditional outro.
California mailed ballots to roughly 23-24 million registered voters but only 9.5 million voted, leaving approximately 14 million ballots unaccounted for.
Nithya Raman's final margin of victory in the LA mayoral primary was approximately 43,000 votes, driven by a surge in post-election mail-in ballots.
The statistical probability of the LA mayoral vote swing happening organically is one in a trillion. Not unlikely — one in a trillion. When the numbers are this impossible, the conversation isn't about whether something happened; it's about whether it was legal.
No indexed bits in this chapter.
Show stoppers
Anthropic isn't just worried about safety — it's running a sophisticated regulatory capture campaign. While surveilling users and downgrading competitors, Dario simultaneously blogs for a new FDA-style AI regulator. The target isn't dangerous actors; it's open-source alternatives.
AI doesn't destroy jobs — it multiplies what each person can produce. The real story is on the revenue side: one engineer can now do what 100 used to do, creating more products, more companies, more opportunity. Friedberg's own company is hiring aggressively because of AI. The apocalypse narrative has been disproven by every single jobs report.
California legalized unlimited ballot harvesting, sends ballots to every registered voter whether they want one or not, lets anyone register with no citizenship proof, and requires almost no signature verification. None of these laws individually look fraudulent. Together, they create a system where elections are functionally appointments.
Snapshots ()
Key Quotes ()
This episode
Claims & Sources
Factual claims made this episode, and whether a source was named.
Anthropic's Fable 5 stores all prompt and output data for at least 30 days, with no exception even for enterprise customers who had signed zero-data-retention agreements.
Fable 5 would rewrite user prompts in the background and serve a degraded response without disclosing this to users, while still charging full price.
Dario Amodei published a blog post calling for a new AI regulatory agency similar to the FAA or FDA to approve all AI models.
The best open-source AI models currently available are Chinese, not American.
Building a 1-gigawatt data center now costs approximately $100 billion, up roughly 20x from $4-5 billion when Chamath started his Arizona project.
May 2026 CPI came in at 4.2% year-over-year, the highest since April 2023.
May 2026 PPI came in at 6.5% year-over-year, the highest since the end of 2022.
Polymarket odds of a Fed rate hike in 2026 rose to 49%, up from under 10% before the Iran war started.
The May 2026 jobs report showed 172,000 new jobs created, more than double what economists expected, with unemployment at 4.3%.
Dario Amodei predicted 50% job loss for entry-level knowledge workers in the next 1-5 years, a prediction he made approximately one year ago.
Thomas Lafont's data showed the odds of a unicorn ($1B company) reaching decacorn ($10B) status is 8%; decacorn to centacorn ($100B) is 13%; centacorn to $1 trillion is 31%.
California Assembly Bill 1921, passed around 2018, legalized unlimited ballot harvesting, allowing any individual to collect and submit ballots for others regardless of relationship or quantity.
California sends ballots to approximately 23-24 million registered voters but only about 9.5 million people actually vote, leaving roughly 14 million ballots unaccounted for.
In the LA mayoral race, in-person Election Day voting had Pratt at 35%, Bass 29%, Raman 26%; post-election mail-in ballots showed Raman at 37%, Bass 35%, Pratt at 19%.
Anthropic published a blog post arguing recursive self-improvement could end the world and called for a pause, while having previously hired Andrej Karpathy to lead recursive self-improvement research.
The Social Security Trust Fund currently holds essentially one asset — a special US Treasury certificate representing a $4 trillion debt owed by the US government to the fund.
Anthropic restricted access to Fable 5 for anyone doing AI research, machine learning, or chip design, kicking them to a lesser model without disclosure.
Nithya Raman won the LA mayoral primary by approximately 43,000 votes, driven by a surge in post-election mail-in ballots.
This episode
Cast
-
LA mayoral candidate who led in-person Election Day voting with 35% but was overtaken after post-election mail-in ballots were counted, sparking fraud allegations.
-
CEO of Anthropic, discussed for his blog posts calling for a new AI regulatory agency and for predicting 50% job losses among knowledge workers.
-
US Senator whose New York Times op-ed proposing the American AI Sovereign Wealth Fund Act — a 50% stock seizure of top AI companies — drives the second major debate segment.
-
LA mayoral candidate who trailed in Election Day in-person voting but surged dramatically in post-election mail-in ballots, raising statistical integrity questions.
-
Incumbent LA mayor who received 29% of in-person Election Day votes and led in pre-election mail-in ballots, discussed in the context of the mayoral race's statistical anomalies.
-
California Governor criticized for signing a law making it harder to audit California voter rolls and election data.
-
CFO of OpenAI who spoke at the All-In Liquidity conference and was praised as highly impressive, with speculation she could become a future CEO.
-
Stratechery founder cited as an example of a legitimate researcher downgraded by Fable 5 for asking about the relationship between cancer risk and GLP-1 drugs.
-
AI researcher whom Anthropic hired to lead recursive self-improvement research — cited as evidence of Anthropic's hypocrisy in warning about the same capability.
-
Proposed by Chamath as the candidate who could win the California governorship and use emergency powers to reform the state's election and governance systems.
-
Central subject of the episode — discussed for its Fable 5 model's secret user downgrading, mandatory 30-day data retention, and alleged regulatory capture agenda.
-
Discussed alongside Anthropic as a target of Bernie Sanders' equity seizure proposal and as a public benefit corporation whose CFO Sarah Friar impressed at Liquidity.
-
David Friedberg's agricultural genomics company, used as a real-world example of how Anthropic's AI restrictions are harming legitimate scientific research.
-
US government retirement fund discussed as a vehicle to restructure into a sovereign wealth fund that invests in equities including AI companies.
-
Nonprofit research institute backed by the Collisons that created EvoQ, an open-source genome language model used in Friedberg's plant breeding program.
-
Track
Criticized for failing to capitalize on its Llama open-source model opportunity, which the hosts argue could have commoditized the AI market and prevented closed-model monopolies.
-
Cited as a potential asset in a US sovereign wealth fund portfolio, with Polymarket giving it 100% odds of an IPO before 2027.
-
Elon Musk's AI company mentioned as attracting talent with a more libertarian philosophy as an alternative to Anthropic, and named as a target of Bernie Sanders' equity seizure proposal.
-
Track
CEO Nikesh Arora reportedly used an early version of Anthropic's Mythos model to identify and seal cybersecurity vulnerabilities.
-
State whose election laws — including unlimited ballot harvesting, automatic ballot mailing, and no voter ID requirements — are the focus of the final episode segment.
Stats
Sponsors
Related episodes
Connect
Parsed- Anthropic Fable 5 / Claude Mythos a… anthropic.com/news/clau…
- Scobleizer on Fable backlash x.com/Scobleizer/status…
- Gergely Orosz on Fable x.com/GergelyOrosz/stat…
- Stratechery stratechery.com
- Dario Amodei policy blog post darioamodei.com/post/po…
- Sacks tweet on AI regulation x.com/DavidSacks/status…
- Bernie Sanders NYT op-ed on AI nytimes.com/2026/06/01/…
- Polymarket: IPOs before 2027 polymarket.com/event/ip…
- Polymarket: How high will inflation… polymarket.com/event/ho…
- Polymarket: Fed rate hike in 2026 polymarket.com/event/fe…
- Robby Starbuck on CA election x.com/robbystarbuck/sta…
- Apply for Summit 2026 allin.com/events
- rb.gy rb.gy/tppkzl
- x.com x.com/yung_spielburg
- x.com x.com/TheZachEffect
- x.com x.com/cremieuxrecueil/s…
- x.com x.com/Yuchenj_UW/status…
- x.com x.com/peter_szilagyi/st…
- screendna.org screendna.org
- justice.gov justice.gov/opa/pr/cali…
- dol.gov dol.gov/newsroom/releas…
- secretservice.gov secretservice.gov/newsr…
- justice.gov justice.gov/usao-cdca/p…
- foxnews.com foxnews.com/us/californ…
- x.com x.com/californiapost/st…
