Anthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California’s Broken Elections
Summary
Anthropic’s Fable 5 may top nearly every benchmark, but its 30-day data retention and user-specific capability downgrades turned model quality into an enterprise-governance risk. Chamath warned that an employee or scientist could unknowingly trip a safeguard and lose a critical capability, while Sacks emphasized that Anthropic may retain prompts, outputs, files, memories, and context even for customers with zero-data-retention agreements. The narrow walkback—telling users when they are downgraded—does not remove the underlying profiling or gatekeeping: “They’re creating a new level of AI haves and have-nots.”
The panel expects restrictive closed models to accelerate local deployment, custom model-building, and demand for open-source compute—currently with an advantage for Chinese providers. Friedberg said Anthropic’s biology restrictions were already interfering with Ohalo’s genomic modeling, genetic-construct design, and RNA-guide work, making locally run open-source models the likely alternative. Chamath framed power as the binding constraint: his planned capacity expanded from 2 GW toward 3 GW, but estimated costs had risen from roughly $4 billion–$5 billion per gigawatt to $100 billion—“in the absence of power, this is all a moot conversation.”
The strongest safety compromise discussed was to regulate dangerous physical outputs rather than silently restrict access to general-purpose intelligence. Friedberg accepted that models can enable cyber, physical, and biological weapons, but argued that the same capabilities can cure cancer, improve food production, and multiply software output; “the same tools are hand in hand.” The discussion contrasted access controls such as KYC with downstream controls such as nucleic-acid synthesis screening, where laboratories have operated voluntary safeguards since 2009 and now support mandatory recordkeeping.
Bernie Sanders’s proposed 50% equity tax exposed the political cost of AI executives repeatedly telling the public that their products will eliminate work. Sacks rejected compulsory confiscation but understood the reaction to companies training on humanity’s accumulated knowledge, predicting 50% job loss, and then gatekeeping the resulting capability: ordinary Americans will ask, “What’s in this deal for me?” Friedberg’s alternative was investment rather than seizure—convert the $4 trillion Social Security trust-fund claim into citizen accounts holding equities in productive American companies.
Friedberg sees AI as a revenue and hiring boom, not a labor-displacement event, and the panel carefully separated the CEOs’ competing forecasts. He said AI is used “a hundred times more” to create products than to reduce costs, with one engineer potentially doing 100× or 1,000× more work; his own team requested 15 additional engineers. Sacks contrasted Elon’s distant abundance and “universal high income” end state, Dario’s specific forecast of 50% entry-level knowledge-worker losses within one to five years, and Sam Altman’s more recent statement that observed job losses had not matched expectations.
May inflation raised the episode’s clearest macro risk: CPI reached 4.2% year over year and PPI 6.5%, with an Iran-driven energy shock potentially feeding much higher goods prices. Friedberg contemplated 5.5%–6% overnight rates under a Kevin Warsh Fed, while Chamath’s tail case had China returning for another 3 million barrels per day and oil moving from below $100 to $150–$200. Sacks said the print was largely expected and the market was up; Jason and Chamath read the Nasdaq’s 2.5% rise as potentially pricing a resolution rather than an uncontrolled inflation spiral.
The Los Angeles mayoral count produced sharp but unresolved disagreement over whether permissive election rules merely look vulnerable or enabled a corrupt result. The panel highlighted Pratt’s 35% in-person share falling to 19% among post-election mail ballots, while Raman rose from 26% in person to 37%; Friedberg called the structure “a system of appointment,” and Sacks called it crooked while also saying there was no evidence of fraud “right now.” Jason pressed for investigation rather than treating the discrepancy as proof, while Friedberg and Sacks argued for voter ID, cleaner rolls, tighter custody, and fewer unsolicited ballots.
Deep dive
1. Fable 5 wins benchmarks but breaks the trust layer
Jason introduced Fable 5 as a Mythos-level Anthropic model topping nearly every benchmark, though tokens cost twice as much as Opus 4.8. Anthropic’s economic claim was that stronger reasoning should consume fewer tokens overall; whether that actually relieves token-cost pressure remained unclear.
The release followed Anthropic’s April decision not to publish Mythos broadly because of hacking capability. Palo Alto Networks’ CEO Nikesh Arora said Opus was “the real deal” and that his company used it to close vulnerabilities, while Fable 5 arrived with bioweapon and hacking topics blocked and frontier-AI research subject to additional scrutiny.
The operational surprise was mandatory retention: Anthropic would store prompts for at least 30 days and classify whether someone was conducting frontier AI research. Its original policy could silently move such users to a weaker model, a mechanism disclosed deep inside a 319-page document.
Anthropic’s partial walkback was visibility, not restored access: it would disclose when safeguards downgraded frontier LLM developers. Chamath’s split verdict captured the episode’s tone—“an incredible model,” with a company “firing on all cylinders,” attached to censorship and governance risks that could make it a nonstarter for serious enterprises.
2. Enterprise adoption now requires model diversity and exit options
Chamath’s enterprise concern was accidental dependency failure. A scientist, executive, or downstream API user could trip Anthropic’s classifier without knowing why, abruptly losing “a very important source of business differentiation”; companies therefore need control, provider diversity, and governance that eliminates a single point of failure.
His blunt assessment of Anthropic was that “they tell the truth”—but “the truth sucks when you actually take it and you eat it.” Individuals face censorship risk; corporations face a broader nonstarter involving continuity, proprietary-data control, unpredictable permissions, and uncertainty over who learns from their information.
Friedberg said Ohalo’s terms of service ostensibly prohibit training on customer work, though his honest answer on whether he trusts that promise was, “I’m not sure.” The company nevertheless uses LLMs for proprietary genomics: predicting gene phenotypes, designing RNA guides, and designing genetic constructs that produce specific proteins.
Those workflows had been very valuable and fast for scientists, but biology restrictions introduced over the preceding weeks were removing useful capabilities. His expected response was direct: use an open model locally, combine it with Ohalo’s data, and build an internal genomic language and prediction model.
3. Safety restrictions are exporting demand to Chinese open models
Friedberg’s causal chain was stark: American closed labs restrict scientific uses; companies still need the capability; American open-source models lag; therefore firms adopt Chinese open-source models and run them locally. He said this migration was already visible among startups and large enterprises.
The competitive risk is not simply model revenue. If US labs and politicians impose constraints that other countries do not, China can compound advantages in biotechnology, materials science, and industrial systems while American companies lose productivity and the workforce eventually bears the cost.
His objection was categorical: “You can’t just stop AI.” Political or social attempts to suppress it do not remove the technology; they “force the hand of someone else” by granting foreign providers the advantage that American firms denied themselves.
Jason saw a second-order destination beyond local inference: companies will start with an open base model, add proprietary information, and own the specialized model. Friedberg confirmed that this was precisely Ohalo’s intended path.
4. Profiling plus regulation could turn safety into market control
Sacks reiterated his eight-month-old accusation that Anthropic was pursuing “a very sophisticated regulatory capture campaign based on fear-mongering.” The Fable controversy moved that view, in his telling, from a spicy minority position toward a developer consensus.
His version of mandatory surveillance extended well beyond prompts and answers. Agent systems pass memories, files, and extensive context to the model; Anthropic would retain that material for 30 days, construct a profile, and use it to determine which capabilities a customer deserves—even where enterprise customers previously negotiated zero retention.
The most deceptive original mechanism, Sacks argued, could rewrite a prompt or reroute it to a lesser model without notice while charging for frontier capability. Disclosure now accompanies the downgrade, but the company still decides who is “worthy”; examples cited included a mitochondria query and Ben Thompson’s question about GLP-1s and cancer risk.
Sacks illustrated a subtler commercial danger with hypothetical deals favoring Novartis over Eli Lilly or JPMorgan over Citibank; Jason supplied the examples. Without a preserved fingerprint of the exact model run, regulators and customers might never prove whether information was deliberately nerfed or manipulated.
5. The free-market defense fails if alternatives are regulated away
Jason’s free-market pushback was that biased models should lose, just as paid-inclusion search engines lost users to better products. He was willing to let Anthropic impose private restrictions because customers could choose another provider.
Sacks’s rebuttal was that Dario simultaneously advocated an FAA- or FDA-like agency to approve models. If regulation blocks rivals—especially open source—the market cannot discipline Anthropic; users could be left with one to three approved companies operating a monopoly or duopoly alongside a revolving-door regulator.
The panel viewed open-source regulation as the intended choke point: decentralized weights cannot easily be supervised like a centralized API. Calling for model regulation while imposing surveillance and anti-competitive degradation therefore looked to them like “a preemptive strike against token maxing on your local machine.”
Jason called the absence of real KYC “the tell.” Sacks said Anthropic could identify users, take a security bond, and allow a disclosed enterprise to request unnerfed access. Friedberg countered that an account and credit card provide some basic KYC; Sacks rejected that as merely an email address and payment method.
6. Open-source AI is becoming an infrastructure trade
Chamath had bought 2,000 acres in Arizona with a partner, obtained approval for a 2 GW data center, and initially expected to flip the project to Blackstone, Brookfield, Google, or another developer. He was reconsidering because most megawatts still serve large proprietary labs, leaving the open-model ecosystem with relatively little compute.
He had also offered on roughly another gigawatt elsewhere, contemplating 3 GW of open-source capacity. The obstacle was capital: what began as a roughly $4 billion–$5 billion project had, by his estimate, become $100 billion per gigawatt—$300 billion for the full buildout.
Friedberg’s working counterexample was the Arc Institute’s open genome language model, which ingested the genomic data it could access. Like judging “good English or bad English,” it infers whether a DNA-letter sequence looks plausible across organisms; Ohalo already uses it as an input to plant breeding.
Once open weights exist, Friedberg said, copies proliferate like a printed book through Xerox machines. Meta’s failure to produce a neck-and-neck open-source model was therefore a strategic fumble: a strong American open model could have commoditized capability and removed margins from closed competitors.
7. Serious safety controls belong closer to dangerous output
Jason’s steelman was that Dario might sincerely have seen frightening capability: withhold Mythos, give it to selected cybersecurity partners, then release Fable gradually while monitoring misuse. False positives might catch Friedberg “trying to make better potatoes,” but customers remain free not to use that model.
Friedberg accepted the dual-use premise through the Manhattan Project analogy. Models can help create cyber, physical, or biological weapons, but the same intelligence can cure cancer, grow more food, increase incomes, and let individuals build businesses: “The same tools are hand in hand.”
His proposed boundary was manifestation, not knowledge. Product-liability rules and existing prohibitions on weapons, hacking, espionage, and bioweapons can be reinforced with tracking and stage gates; restricting the general system upfront sacrifices scientific and economic upside without recovering technology already “out of the box.”
Fertilizer provided the practical compromise: dangerous supply can require identification without prohibiting general knowledge. Sacks pointed to nucleic-acid synthesis screening, developed voluntarily by major laboratories under a 2009 consortium agreement; after more than 15 years of operation, the signatories supported mandatory screening and recordkeeping at the physical-order stage.
8. Anthropic’s own product demonstration made the criticism tangible
While discussing fertilizer regulation, Jason queried Fable 5 and watched it assess him as “a VC and podcaster” with a potentially legitimate policy interest, then downgrade him. A later nuclear-weapons-components question triggered another switch to Opus 4.8 despite being framed as journalistic research.
Anthropic’s interface explained that Fable’s measures may flag “safe normal content” involving cybersecurity or biology so Mythos-level capability can be released sooner elsewhere. Jason said the blockages would chase users away; Friedberg had already described Ohalo’s shift toward open models.
Sacks warned that “safety” could expand as it did during early-2020s social-media battles—from physical harm to microaggressions, psychological safety, debanking, and exclusion from payments. AI profiling would be more consequential because it mediates not merely speech distribution but access to productive intelligence.
The irreversibility point closed the debate: “You can’t turn off the internet and you can’t just turn off the typewriter and you can’t turn off the Gutenberg press.” The open models already published can be copied, fine-tuned, and evolved regardless of what a US regulator later demands.
9. Sanders’s 50% proposal is the backlash AI leaders invited
Senator Bernie Sanders’s proposed American AI Sovereign Wealth Fund Act would impose a one-time 50% tax in stock—not profits—on major AI companies including OpenAI, Anthropic, and xAI. The public fund would receive voting rights and equal board representation.
Sanders’s moral claim was that AI rests on “our collective and human intelligence”: books, songs, journalism, science, and code contributed by society and then effectively taken by a small number of wealthy owners. Jason framed the proposal as a horseshoe coalition with sovereign-wealth ideas attractive across populist factions.
Sacks rejected compulsory seizure as an intolerable property precedent, yet sympathized with the politics. AI CEOs trained on accumulated human knowledge, warned that 50% of people could lose work, and offered no obvious participation; the predictable public response is, “What’s in this deal for me?”
Anthropic and OpenAI’s public-benefit-corporation status sharpened the contradiction. If their boards balance shareholder value with public benefit, Sacks suggested that paying down the national debt might qualify. Jason argued that their “Frankenstein” messaging also makes confiscation easier: if the monster is truly existential, “stop making Frankenstein.”
10. A citizen-owned equity fund is the capitalist alternative
Friedberg proposed restructuring Social Security rather than confiscating property. Its trust fund, he said, effectively contains one special Treasury certificate worth roughly $4 trillion because it may own only government debt; that defined-benefit structure should become individually attributable, investment-backed accounts.
The fund could buy equity in productive American businesses, including AI companies, with new capital entering the enterprises in exchange for shares. Those certificates would sit in citizens’ accounts, making broad ownership an investment rather than “a seizure.”
Chamath called the proposal genius—and therefore politically unlikely—but endorsed the Canadian and Australian model. A US sovereign fund could be seeded with the government’s existing portfolio of critical metals and materials; panelists also floated Intel, Micron, AMD, SpaceX, and frontier AI companies.
His negotiation stance was more aggressive than Sanders’s: because AI companies depend on indispensable infrastructure and national resilience, he said he might emerge owning “75% of these companies” if he were running the fund and negotiating directly.
11. AI’s labor and cost curves point in opposite political directions
Sacks rejected the immediate jobs apocalypse using the episode’s May figures: 172,000 new jobs, more than twice economists’ expectation; unemployment at 4.3%; hundreds of thousands of construction jobs; and software-development employment at a three-year high.
Friedberg divided a business into cost and revenue. AI can remove some cost-side labor, but he called that effect nominal beside revenue creation: one engineer can produce 100× or 1,000× more products, and his own engineering leaders requested 15 additional hires to pursue newly feasible work.
The forecasts were not interchangeable. Elon described a potentially distant abundance state with robots, optional work, and “universal high income”; Dario forecast 50% losses among entry-level knowledge workers within one to five years; Sacks said Sam Altman had more recently acknowledged he was wrong because the numbers were not showing the expected job losses.
Chamath contrasted AI with the internet’s zero marginal cost: each new AI user consumes GPUs, electricity, and memory. That infrastructure dependence supports a public-equity argument; Jason countered that locally run open models can spread fixed costs differently and weaken centralized tollbooths.
12. Scale statistics favor companies that have already crossed major thresholds
Thomas Laffont’s Liquidity presentation supplied an investor counterintuition: only 8% of unicorns reached decacorn status, but 13% of decacorns reached $100 billion and 31% of $100 billion companies reached $1 trillion.
Sacks asked whether the pattern might continue from $1 trillion to $10 trillion and guessed that roughly 60% of trillion-dollar companies could reach that level in the next few years. He presented this as an extrapolation, not as a quoted figure from the data.
The figures were aimed at capital allocators: once a company has crossed a major valuation threshold, its odds of another large step may be better than the odds of reaching that threshold from below. The discussion connected that scale advantage to the capital, power, and infrastructure barriers facing open-source competitors.
13. Hot inflation and disputed elections close on the same trust problem
May CPI printed at 4.2% year over year, its highest since April 2023, while PPI reached 6.5%, the highest since late 2022. A prediction market assigned a 21% chance to 5% inflation during 2026 and 49% to a Fed hike that year, up from below 10% before the Iran war; the ECB had raised rates by a quarter point.
Friedberg attributed part of the jump to wartime energy but returned to excess government spending as the root cause of inflation and inequality. Under a Kevin Warsh Fed, he considered 5.5%–6% overnight rates plausible; Chamath’s tail risk was China buying another 3 million barrels daily and oil jumping from below $100 to $150–$200.
Sacks said the PPI print was largely in line with expectations and that the market was up. Jason and Chamath interpreted the Nasdaq’s 2.5% rise as potentially pricing a resolution; Jason nevertheless called the Iran war a “huge colossal error” whose inflationary effects made a quick off-ramp urgent.
In Los Angeles, the cited mayoral data showed Pratt at 35% in person, Bass at 29%, and Nithya Raman at 26%; pre-election mail favored Bass 38%, Pratt 28%, Raman 20%, while later ballots shifted to Raman 37%, Bass 35%, Pratt 19%. Friedberg called that pattern statistically alarming.
Friedberg blamed the legal architecture—universal mailed ballots, unlimited harvesting under AB 1921, weak identification, and broad registration—calling it “a system of appointment.” Sacks went further, alleging dirty rolls, millions of unused ballots, weak signature matching, missing custody, and either “illegal fraud or legal fraud.”
The discussion noted that swinging thousands of votes would require coordination. Jason resisted treating the discrepancy as proof and pressed for investigation; his proposed legitimate explanation was a stronger Democratic ground game or late voters coordinating to place both Bass and Raman ahead of Pratt. Sacks replied that legal ballot harvesters, rather than individual voters, could execute such a strategy.
Proposed remedies included voter ID, cleaner rolls, less unsolicited mailing, verifiable custody, and audits. Friedberg warned that public faith collapses when basic anomalies cannot be explained. Sacks said there was “no evidence of fraud right now” while maintaining that the statistics looked bad and that Spencer Pratt should be in the runoff; Jason left the issue as one requiring investigation rather than a proven result.