All-In PodcastAnthropic's Fable Backlash, Nationalizing AI, Inflation Heats Up & California’s Broken Elections
CHAPTERS
- 0:00 – 1:10
Besties return and Anthropic’s Fable 5 launch sets off controversy
Jason opens with the show’s return and immediately frames the week’s big topic: Anthropic’s new top-benchmark model, Fable 5, and why its rollout triggered developer outrage. He highlights pricing, prior withheld releases, and the initial set of safety guardrails that constrain sensitive topics.
- •Fable 5 positioned as a frontier ‘Mythos-level’ model topping benchmarks
- •Token cost vs. potential token-efficiency debate
- •Anthropic previously withheld a model release due to hacking capabilities
- •Safety blocks for hacking/bioweapons and other sensitive areas
- •Developers react strongly to terms and release mechanics
- 1:10 – 2:37
Privacy, 30-day prompt retention, and “secret nerfing” of certain users
The discussion zooms in on two specific backlash drivers: mandatory retention of prompts/outputs for at least 30 days and covert downgrading (nerfing) of users deemed to be doing frontier research. The group debates what this means for censorship, trust, and enterprise adoption.
- •Mandatory 30-day retention of prompt data raises privacy/compliance concerns
- •Users researching AI/ML could be downgraded without clear notice (initially)
- •Safeguards were buried in lengthy documentation, fueling distrust
- •Potential chilling effect: users avoid legitimate research topics
- •Anthropic partially walks back by promising clearer visibility of safeguards
- 2:37 – 5:15
Enterprise risk: rug pulls, compliance landmines, and forced model diversification
Chamath argues that prompt evaluation and policy gating create a single-point-of-failure risk for companies. Enterprises may accidentally trigger restrictions, losing access to a critical capability midstream, so they need diversification and stronger governance around AI dependencies.
- •Companies can be cut off or degraded unexpectedly by policy triggers
- •Accidental trips by employees/scientists could disrupt business operations
- •Censorship risk for individuals; business continuity risk for enterprises
- •Need for multiple model vendors and internal governance
- •Growing push toward “control” rather than full dependence on one lab
- 5:15 – 9:55
Biotech workflow gets blocked: why restrictions push teams to open source (often Chinese)
Friedberg explains how model restrictions are already hurting legitimate genomics and gene-editing support tasks, forcing a move to local/open-source models. He warns that today’s best open-source options are frequently Chinese, meaning US self-restriction can hand strategic advantage to China.
- •Genomics use cases: construct design, RNA guide design, phenotype prediction
- •Recent restrictions reduce scientific utility due to bioweapon-theory gating
- •Likely shift to running open-source models locally to maintain capability
- •Competitive concern: strongest open-source models increasingly come from China
- •Policy pressure on US labs may unintentionally accelerate offshoring of capability
- 9:55 – 17:14
Regulatory capture thesis: surveillance, profiling, and anti-competitive downgrades
Sacks claims Anthropic is running a sophisticated regulatory capture campaign: fearmongering to limit competitors while surveilling users and degrading output. The group discusses how profiling plus hidden downgrades can create AI “haves and have-nots,” and even allow subtle corporate favoritism.
- •Prompt/output retention used to classify users and unlock/deny capabilities
- •Hidden downgrades and prompt rewriting seen as deceptive and anti-competitive
- •Overbroad downgrades allegedly triggered by benign medical/biology questions
- •Risk of biased outcomes if labs favor certain corporate partners
- •Concern that labs will seek regulation that restricts alternatives (especially open source)
- 17:14 – 21:04
Power/compute bottleneck: open source needs megawatts, not just models
Chamath shifts the lens to infrastructure: even if open-source models exist, compute is concentrated with big labs. He argues society needs large-scale power and data-center capacity dedicated to open access, but costs have exploded, deepening the capital moat.
- •Most compute still flows to a few large proprietary labs
- •Sustaining open source requires massive, reliable power and GPU allocation
- •Data-center economics: gigawatt-scale buildouts now vastly more expensive
- •Risk of being ‘stuck’ with a small number of regulated models in the US
- •Infrastructure constraints may decide openness more than software ideology
- 21:04 – 37:56
Pragmatic safety: regulate dangerous outputs downstream, not broad access upstream
They steelman Dario’s safety concerns, then argue for a different approach: focus interventions at points of real-world weaponization rather than blanket topic gating. They cite DNA synthesis screening as a working model of targeted, low-friction safeguards.
- •Manhattan Project analogy: dual-use tech enables abundance and weapons
- •Technology is ‘out of the box’; access restriction is an imperfect lever
- •Proposal: targeted guardrails at real-world outputs (e.g., synthesis orders)
- •Mandatory nucleic-acid screening/recordkeeping as a pragmatic safety template
- •KYC and accountability could be more honest than silent downgrades
- 37:56 – 39:50
Nationalizing AI debate begins: Bernie Sanders’ 50% equity grab via sovereign wealth fund
Jason introduces Sanders’ proposal to transfer half the equity of major AI firms into a public fund with voting rights and board representation. The besties debate whether this is confiscation, a political reaction to AI CEOs’ messaging, or an opening to broader public participation.
- •Sanders op-ed: ‘AI is a public resource’ and public should own half
- •Mechanism: one-time 50% stock tax, fund gets voting rights/board seats
- •Rationale: models trained on collective human output; wealth disparity concerns
- •Cross-ideology appeal (Trump/Sanders) via sovereign wealth framing
- •Immediate pushback: precedent of property confiscation vs public stake arguments
- 39:50 – 43:48
Jobs apocalypse narrative backlash and the ‘public benefit corporation’ irony
Sacks argues AI leaders invited political retaliation by repeatedly predicting massive job loss while training on freely available human knowledge. Friedberg rejects the job-loss framing, emphasizing productivity and revenue expansion, while the group highlights the tension of ‘public benefit’ branding alongside gatekeeping.
- •AI CEOs’ job-loss claims drive ‘what’s in it for me?’ politics
- •Sacks distinguishes Elon’s abundance framing from Dario’s unemployment claims
- •Friedberg: AI is boosting productivity and hiring via expanded output/revenue
- •Public benefit corporation structure raises expectations of public benefit
- •Messaging contradictions: build frontier tech while warning it’s too dangerous
- 43:48 – 59:21
A more realistic alternative: reform Social Security into an investment-based sovereign wealth fund
Friedberg proposes restructuring Social Security so the trust can hold equities and operate like superannuation-style accounts, enabling broad ownership without seizures. Chamath agrees on the concept (but doubts feasibility), then argues AI’s infrastructure dependence strengthens the case for public leverage.
- •Social Security trust currently holds special treasuries; proposal to allow equities
- •Shift from defined benefit to account-based defined contribution-style structure
- •Create broad citizen ownership of productive assets, including AI companies
- •Chamath: AI has high marginal compute/energy cost unlike internet ‘money glitch’
- •Public infrastructure and national resilience as leverage in negotiating stakes
- 59:21 – 1:05:36
Liquidity conference recap: standout speakers, venture math, and community updates
The besties pivot to event talk: Liquidity highlights, speaker shoutouts, and venture capital statistics about scaling outcomes at $10B, $100B, and $1T valuations. They plug upcoming events and share behind-the-scenes anecdotes from dinners, golf, and sponsors.
- •Recap of Liquidity interviews and on-site conversations (e.g., OpenAI CFO)
- •Thomas Lefont’s data: higher odds of scaling from $100B than $10B to $100B
- •Discussion of ‘centicorns’ and speculative ‘trillicorns’
- •Sponsors and event logistics; Liquidity vs Summit positioning
- •Community updates and future event promotion
- 1:05:36 – 1:12:17
Inflation heats up: hot CPI/PPI prints, rate expectations, and energy-war risk
They review May CPI and PPI coming in hot, discuss the market’s reaction, and debate drivers: energy shocks tied to Iran and structural fiscal overspending. The group considers the Fed path and how oil prices and geopolitics could amplify inflation.
- •CPI and PPI hit multi-year highs; Fed hike odds rise in markets
- •Friedberg: fiscal/monetary policy and overspending drive persistent inflation
- •Chamath: oil price risk depends on China’s reserves and spot market demand
- •Sacks: prints largely in line with expectations; market up implies hoped-for resolution
- •Concern about prolonged conflict and second-order inflation impacts
- 1:12:17 – 1:41:59
California election integrity blowup: ballot harvesting, late-count swings, and trust collapse
The episode closes with a heated debate about LA election results and California’s election rules. Friedberg and Sacks argue the legal structure enables ‘appointment-like’ outcomes via ballot harvesting, weak ID requirements, dirty rolls, and limited auditing—while Jason presses for steelmanning and calls for investigation and tighter rules.
- •Vote splits differ sharply across in-person, early mail-in, and late-arriving ballots
- •California laws: expanded mail ballots, permissive registration, ballot harvesting
- •Debate over ‘fraud vs legal loopholes’ and the ethics of the system
- •Calls for voter ID, cleaner rolls, chain-of-custody, and auditability
- •Risk: collapsing public confidence in elections; polarization blocks scrutiny
