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

Summary of the subtitles (All-In podcast episode)

1) Anthropic’s “Fable 5” backlash: surveillance, prompt retention, and “nerfing”

• Anthropic released Fable 5, which tops many benchmarks, but it’s reportedly about 2× more expensive per token than Anthropic’s prior Opus model. Supporters suggest it may use fewer tokens overall, but the speaker says it’s uncertain whether this actually resolves cost concerns.
• The main controversy is framed as privacy + product control:
  • Anthropic is described as storing prompt data for at least 30 days for these “Mythos-level” models, which is presented as a privacy issue, including for enterprise customers who allegedly did not consent to retention.
  • If the system detects “frontier AI research” (using the model to build better models), it may downgrade capabilities without clearly informing users.
  • Social-media developer outrage is highlighted as stemming from the fear that capabilities can be silently reduced—including examples of downgraded answers about biomedical topics like mitochondria and questions involving GLP-1/cancer risk.
• Core debate point: the speakers argue this creates
  • Censorship risk for individuals, and
  • Operational/business risk for enterprises (because users could be “cut off” from important differentiation or research capability without knowing why).

2) Enterprise and industry reaction: move toward open-source and local inference (but with geopolitical downsides)

• A contributor (“Freeberg,” a genomics-focused founder) argues the restrictions are harming legitimate scientific work in genomics and gene-editing-related modeling.
• He claims companies will increasingly run open-source models locally to keep doing this work.
• He also asserts a strategic/regional concern: he believes the best local open-source options are likely “Chinese”—meaning Western restrictions could indirectly push innovation toward competitors.
• The speakers also predict a broader trend: companies will build their own models by combining base models with proprietary data (e.g., a “genome language model” for internal use).

3) Criticism of “regulatory capture” and government-style gatekeeping

• Another contributor (“Saxs” / “Chimath” depending on speaker naming in the transcript) argues Anthropic appears to be promoting regulation driven by fear, likened to an FAA/FDA-like approval model for AI.
• The claim is that such regulation could be used to limit access to certain categories of competitors, including open-source efforts that are harder to regulate like closed systems.
• The major theme: concern that models become “have/ have-not” systems where access depends on profiling and provider discretion.

4) “Steelman” argument on safety vs. output restrictions

• The hosts attempt to steelman the safety rationale: if frontier models enable weaponization, a provider might believe it must restrict access/capabilities to reduce risk.
• However, the speakers argue the better approach is to restrict weaponization outcomes (real-world harmful actions) rather than broadly restricting model inputs, drawing an analogy to focusing law enforcement on illegal acts instead of banning the underlying technology.
• They also reference a parallel safety approach: mandatory nucleic acid screening/recordkeeping for DNA/RNA synthesis (citing international lab safeguards), arguing downstream controls can target misuse without choking research inputs.

5) Bernie Sanders’ “AI public resource” proposal: wealth capture and a sovereign model

• The episode shifts to politics: Sen. Bernie Sanders argues AI companies should give the public 50% equity via an “American AI sovereign wealth fund act.”
• Debate highlights:
  • Saxs opposes confiscation as a bad property precedent, though he’s sympathetic to the underlying complaint that jobs/benefits don’t reach the public.
  • Others discuss structuring it as voluntary participation/ownership instead of outright seizure.
• The hosts connect the discussion to a broader critique of AI labs’ messaging about job displacement—arguing politicians respond to fears that AI will put workers out of jobs.

6) Inflation segment: hot CPI/PPI, Iran-war energy effects, and “rates higher for longer”

• The hosts discuss May CPI (reported 4.2% YoY, highest since April 2023) and PPI (reported 6.5% YoY, highest since the end of 2022).
• They suggest inflation pressure is partly tied to the Iran war via energy inputs, but emphasize deeper causes such as overspending and broader monetary/fiscal dynamics.
• They discuss the risk of higher rates—possibly “north of 5.5–6% overnight rates”—and debate whether there’s an “offramp” from the conflict.

7) California LA mayoral election: claims of “no election” / rigged system via legal loopholes

• A long segment claims the LA mayoral primary functions more like an “appointment” system than a transparent election.
• The speaker presents statistics comparing in-person voting and mail-in ballots received before vs. after election day, arguing the patterns suggest irregularities.
• The argument centers on California election-law changes (ballot harvesting, universal ballot mailing, and weak identity verification) described as enabling legal exploitation.
• Even while acknowledging it could be “legal fraud” (fraud enabled by law), the speakers argue it undermines democracy and calls for voter ID and stricter auditability.

---

Presenters / contributors (as named in the transcript)

• David Saxs (David Sacks)
• Freeberg (David Freeberg)
• Chimath (Chimat / Chimath “Chemt” Palihapitiya mentioned as “Chimath”)
• J.C. A.T. / “Jason” (appears as “Jason”/“JCAL” / “Jason from Liquidity”)
• Sarah Fryer (referred to as CFO of OpenAI; also spelled “Sarah Prior/Frier” in the transcript)
• Thomas Keller (referenced in connection with sponsorship/hosting context; French Laundry mentioned)
• Jake Paul (mentioned as present at an event)
• Nick (referenced as someone pulling up charts/maps; last name not clearly identified)