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

2026-06-13 - 102 min - source - Read full transcript

Jason Calacanis (host)Chamath Palihapitiya (host)David Sacks (host)David Friedberg (host)

Key insights

Anthropic's Fable 5 retains all prompt and context data for at least 30 days, including for enterprise customers who had signed zero-data-retention agreements, and silently downgraded users it flagged as doing 'Frontier AI research' without disclosing the downgrade.

Sacks says this was buried in a 319-page document and only surfaced after developer outrage; Anthropic's narrow walkback (via Wired) was to disclose downgrades going forward, not to stop downgrading. The hosts frame the retention policy as a direct reversal of Anthropic's public anti-surveillance stance and note it applies to all context shared with agent platforms, not just raw prompts.

ai-regulation-and-censorship

Sacks argues Anthropic's safety messaging functions as a sophisticated regulatory-capture campaign - he made this claim eight months earlier as a 'spicy' outlier take and says it is now becoming consensus.

He points to Dario Amodei's call (the same week) for a new licensing agency similar to the FAA or FDA to approve AI models, arguing such regulation would be effectively unenforceable against open source and would therefore entrench closed-model incumbents like Anthropic while looking, on the surface, like a self-sacrificing safety position.

ai-regulation-and-censorship

Friedberg's own company, O'Halo, has had legitimate genomics research work throttled by Anthropic's models under a bioweapon-risk rationale, and the best available open-source alternatives are now Chinese models.

He describes using the models for RNA guide design and gene-variant impact prediction in plant breeding, work that was smoothly supported for months before Anthropic began restricting it. He says this is forcing companies like his toward self-hosted open-source models, and that American open-source models currently lag Chinese ones - a dynamic the panel calls a national economic and security own-goal if it continues.

ai-regulation-and-censorship

The panel proposes KYC-style identity verification with a posted security bond as a more targeted alternative to blanket capability restriction, and argues Anthropic's failure to offer it is itself evidence its stated safety rationale is not the full story.

Sacks says a legitimate researcher like Friedberg should be able to register identity and business purpose with Anthropic in exchange for unrestricted access, similar to how fertilizer purchases require ID. Anthropic could implement this technically but has not, which the hosts read as consistent with a broader agenda to restrict access generally rather than to specific bad actors.

ai-regulation-and-censorship

The hosts call out an apparent contradiction: Anthropic published a blog post warning that recursive self-improvement could be dangerous and calling for a pause, the same month it hired Andrej Karpathy specifically to work on recursive self-improvement at the company.

Sacks cites Ben Thompson's read that the pause post was likely published to justify Anthropic's separate, more concretely anti-competitive step that week of restricting Fable's usefulness for AI, machine-learning, and chip-design research - i.e., using a safety narrative to cover a competitive move.

ai-regulation-and-censorship

Bernie Sanders proposed a one-time 50% stock tax on the largest AI companies (OpenAI, Anthropic, xAI), funding a government sovereign wealth fund with public voting rights and board representation; Sacks rejects the confiscation framing but says he has real sympathy for the underlying logic.

Sacks's reasoning: AI CEOs have publicly and repeatedly told the public their models - trained for free on collectively-produced human knowledge - will cause major job losses, so it is a predictable political reaction for the public to demand a stake in return. He distinguishes this from supporting outright seizure, preferring a more voluntary structure.

ai-sovereign-wealth-fund

Chamath argues AI's economics are structurally unlike the internet's - marginal cost per user is high, not near-zero - which is the strongest argument for the government taking a meaningful equity stake in AI labs.

He contrasts the internet's near-zero incremental cost per additional search or social-network user (which let Google and Facebook massively over-earn) with AI, where every user taxes GPU compute, electricity, and memory. He compares AI labs to early interstate-transportation companies riding on federally-built infrastructure, arguing the government could reasonably negotiate for a large stake - he estimates he'd target 75% if he were running the negotiation - given how critical (and federally enabled) the underlying power and compute infrastructure is.

ai-sovereign-wealth-fund

Friedberg proposes converting Social Security from a defined-benefit system holding only US Treasury bonds into an account-based sovereign wealth fund empowered to hold equities, including AI companies, modeled on Canadian and Australian pension systems.

He notes the Social Security Trust Fund's sole asset is a non-tradeable Treasury certificate representing roughly $4 trillion the government owes it, and argues restructuring it into individual equity accounts - funded by new investment rather than seizure - would let ordinary Americans own a stake in AI growth instead of relying on a shrinking defined-benefit promise.

ai-sovereign-wealth-fund

Friedberg disputes the AI-job-loss narrative that Dario Amodei, Elon Musk, and (previously) Sam Altman have each promoted, arguing from direct operating experience that AI's dominant economic effect is revenue expansion, not headcount reduction.

He says his own company is adding roughly 15 engineering headcount because AI lets each engineer produce far more output, and frames cost-side labor savings as a comparatively 'nominal' effect versus the revenue-side opportunity to create more products. He argues AI labs' own job-loss messaging - regardless of whether it is accurate - is what manufactured the political opening for proposals like Sanders'.

ai-sovereign-wealth-fund

May CPI came in at 4.2% year-over-year (highest since April 2020) and PPI at 6.5% (highest since late 2022), which the panel attributes partly to an Iran-war energy shock and partly to structural government overspending.

Chamath warns that if China's current smoothing of global energy demand runs out of slack and it re-enters the spot oil market for millions of extra barrels a day, oil could spike from sub-$100 toward $150-200 a barrel, which would flow through PPI into broader consumer prices; he frames the hot PPI print as a signal to seek an off-ramp from the Iran conflict sooner rather than later.

inflation-and-fed-policy

In the Los Angeles mayoral primary, candidate Nithya Raman's vote share rose sharply in ballots counted after election day while Spencer Pratt's post-election-day share fell by roughly a third versus his in-person and early-mail numbers - a pattern Chamath calls statistically improbable.

In-person votes on election day: Pratt 35%, Bass 29%, Raman 26%. Mail-in ballots received before election day: Bass 38%, Pratt 28%, Raman 20%. Ballots arriving after election day: Raman 37%, Bass 35%, Pratt 19%. Friedberg is careful to distinguish this from a claim of illegal fraud, arguing instead that California's ballot rules make the outcome 'legal' but still structurally exploitable.

california-election-integrity

Friedberg traces the vulnerability to a specific sequence of California election-law changes - AB 1921 legalizing unlimited third-party ballot harvesting, universal mail-ballot distribution, and no-ID voter registration - that in combination make large-scale, hard-to-audit ballot collection both legal and difficult to detect.

He cites a US House Committee on House Administration report (May 2020) documenting these changes around the 2018 midterms, and notes California sends ballots to roughly 23-24 million registered voters while only about 9.5 million vote, leaving millions of unclaimed ballots sitting in shared mailboxes with no chain-of-custody requirements once collected by a third party.

california-election-integrity

Media referenced

AI is a public resource. You should own half of it. - article - Bernie Sanders' New York Times op-ed (June 1) announcing the American AI Sovereign Wealth Fund Act, proposing a 50% stock tax on the largest AI companies.
Wired interview with Anthropic on Fable 5 safeguards - article - Anthropic's partial walkback, saying it would make Frontier LLM development safeguards 'more visible' after developer backlash over undisclosed model downgrades.
Dario Amodei recursive self-improvement blog post - article - Amodei's post warning recursive self-improvement could be dangerous and calling for a pause, published the same month Anthropic hired Andrej Karpathy to work on recursive self-improvement - cited by the hosts as evidence of hypocrisy, and per Ben Thompson as cover for restricting AI-research use of Fable.
US House Committee on House Administration report (May 2020) - paper - Report on 2018 California midterm ballot-harvesting problems that Friedberg cites to walk through the sequence of California election-law changes (AB 1921 and related bills) enabling large-scale ballot collection.
Nick Shirley Skid Row video - other - Independent video investigation the hosts reference showing alleged solicitation of homeless voters near Skid Row in the LA mayoral race.
Thomas Laffont venture capital data presentation (Liquidity conference) - other - Recap of a conference talk on valuation step-up odds (unicorn to decacorn to centicorn to trillion-dollar company), cited approvingly as rigorous data behind the panel's 'trillicorn' talk.

Companies

Anthropic - Central subject of the episode's first segment - the Fable 5 model release, its data-retention and silent-downgrade practices, and the developer backlash that followed.
Palo Alto Networks - CEO Nikesh Arora is cited as having used Anthropic's held-back Mythos-class model to find and close security vulnerabilities before wider release.
OpenAI - Discussed as a public benefit corporation alongside Anthropic; Sarah Friar (CFO) and Sam Altman referenced from a recent Liquidity conference appearance; also carries a Bernie Sanders sovereign-wealth-fund target.
xAI - Named as a third target of Sanders' proposed AI stock tax; Elon Musk's more utopian 'universal high income' framing of AI job displacement contrasted with Dario Amodei's more dire framing.
Meta - Criticized for 'fumbling' its open-source Llama strategy two years ago instead of using it to commoditize competitors' models.
Google - Referenced twice - as the historical monopoly risk that motivated Elon Musk to originally found OpenAI as an open counterweight, and as an example (paid search inclusion) of how degraded trust drives users to competitors in a free market.
Arc Institute - Cited by Friedberg as the group (with Collison-brothers backing) behind an open-source genome language model that O'Halo uses in its plant-breeding program.
O'Halo - Friedberg's own agriculture/genomics company; he describes it losing access to Anthropic model capabilities for legitimate gene-editing research under a bioweapon-risk rationale, forcing a shift toward open-source and self-hosted models.
SpaceX - Referenced via a Polymarket line on IPO odds before 2027 (100%), alongside Anthropic (83%) and OpenAI (48%, after disclosing it privately filed).

Techniques and frameworks

KYC (know-your-customer) gating - Proposed by the hosts as a more targeted alternative to Anthropic's blanket capability restrictions - verified identity plus a security bond in exchange for unrestricted model access - which they note Anthropic could implement but has not.
Public benefit corporation dual mandate - Sacks explains that OpenAI and Anthropic, as public benefit corporations, must balance shareholder interests against a stated public-benefit mission, which the panel argues gives legal cover for an equity claim like Sanders'.
Account-based sovereign wealth fund (vs. defined-benefit Social Security) - Friedberg's proposal to convert the Social Security Trust Fund from a bond-only, defined-benefit system into an account-based system allowed to hold equities, including AI companies.
Adverse selection applied to election law - Friedberg uses the insurance concept - that bad actors disproportionately exploit any exploitable loophole - to argue California's ballot-collection rules, though individually defensible, combine into a system structurally vulnerable to large-scale ballot harvesting.

Summary

The panel opens on Anthropic's Fable 5 release, which topped benchmarks but triggered a developer backlash once it emerged the model retains all prompt and context data for at least 30 days - even for enterprise customers with zero-data-retention agreements - and silently downgrades users it flags as doing "Frontier AI research," without telling them. Sacks reprises a claim he made eight months earlier that Anthropic is running a sophisticated regulatory-capture campaign, citing Dario Amodei's parallel call for an FAA/FDA-style licensing agency for AI models as evidence the company wants regulation applied in a way that is unenforceable against open source. Friedberg brings a first-person data point: his own company, O'Halo, has had genomics research work throttled under a bioweapon-risk rationale, pushing him toward self-hosted open-source models - and he notes the strongest current open-source models are Chinese, which the panel frames as a self-inflicted national security risk if US labs keep restricting access. The group proposes KYC-style identity verification as a more targeted alternative to blanket restriction, and calls out an apparent contradiction between Anthropic's blog post warning about the dangers of recursive self-improvement and its hiring of Andrej Karpathy that same month to work on exactly that.

The conversation pivots to Bernie Sanders' New York Times op-ed proposing a one-time 50% stock tax on the largest AI companies, funding a public sovereign wealth fund with voting rights and board seats. Sacks rejects outright confiscation but says he understands the political logic: AI CEOs have themselves told the public their models will cause major job losses, so demanding a public stake is a predictable response. Chamath makes the more structural case - AI's marginal cost per user is high, unlike the internet's near-zero marginal cost that let Google and Facebook over-earn for decades - arguing this cost structure, combined with AI labs' dependence on federally-enabled power and compute infrastructure, is the strongest justification for a large government equity stake, which he'd negotiate toward 75% if given the chance. Friedberg counters the underlying job-loss premise directly, arguing from his own company's hiring (adding roughly 15 engineers) that AI's dominant economic effect is revenue expansion rather than cost-cutting, and separately pitches converting Social Security from a bond-only defined-benefit system into an account-based sovereign wealth fund that could hold equities, including AI companies.

A shorter segment recaps the All-In Summit's Liquidity conference - a Thomas Keller / French Laundry dinner, an NAD/wellness lounge, a golf round, and Thomas Laffont's venture data on the odds of companies stepping up from unicorn to decacorn to centicorn to trillion-dollar valuations - before turning to the day's inflation print. May CPI hit 4.2% year-over-year (highest since April 2020) and PPI hit 6.5% (highest since late 2022), which the hosts attribute to a mix of Iran-war-driven energy costs and structural government overspending; Chamath warns oil could spike toward $150-200 a barrel if China's current smoothing of global energy demand runs out of room, while Sacks notes markets shrugged off the print because it came in roughly in line with expectations.

The episode closes on the Los Angeles mayoral primary, where Spencer Pratt led on election-day in-person votes but was overtaken in the final count by Nithya Raman, whose share of ballots counted after election day rose sharply as Pratt's fell by roughly a third. Friedberg is careful to separate "statistically strange" from "illegal," walking through a chain of California election-law changes - unlimited ballot harvesting under AB 1921, universal mail ballots, and registration without ID - that make large-scale, hard-to-audit ballot collection legal. Sacks is more direct in calling the pattern crooked regardless of its legality, and both point to Governor Newsom signing a law that makes auditing voter rolls harder, plus a documented guilty plea for paying people to vote, as reasons for concern. Chamath closes by framing California's one-party structural entrenchment as the deeper problem, with the three agreeing the state's pending voter-ID ballot measure is close to a last chance to address it.

Notable Quotes

"The truth is that the capabilities that allow that weaponization are the same capabilities that I'm describing that can be used to cure cancer, to make more food, that can be used to make software tools that create extraordinary leverage for people." - David Friedberg

"The incremental cost of performing AI is excessive and large. That contrasts and compares to the incremental cost before AI, which was zero. Do with that information what you will." - Chamath Palihapitiya

"These guys are capitalist cucks. I think their kink is like having their equity taken and being regulated." - Jason Calacanis

"When you start mailing out millions of ballots to people who don't even vote, and then you have no audit requirements, no chain of custody, no punishments, no penalties, no investigations... it is still corrupt. It is still crooked." - David Sacks

"It's a stupidity tax because they've been out there teaching the public that what they do is harmful." - Jason Calacanis, on AI labs' job-loss messaging fueling proposals to seize their equity