All-In PodcastCan the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters
CHAPTERS
- 0:00 – 1:32
Besties riff and set the agenda for a packed episode
The hosts banter, welcome Friedberg back, and tee up an AI-heavy docket. The tone is playful, but they quickly pivot to regulation, Big Tech/legal drama, privacy, infrastructure, and science.
- •Besties’ comedic cold open and intros
- •Friedberg returns after a week off
- •Jason previews a multi-topic “full docket”
- •Transition into AI regulation proposal from Demis Hassabis
- 1:32 – 4:18
Demis Hassabis proposes a FINRA-style self-regulatory body for AI
Jason outlines Demis Hassabis’ proposal for an industry-funded, federally overseen standards body modeled after FINRA. The idea: pre-release model submissions, evolving benchmarks, and the ability to coordinate slowdowns in high-risk circumstances.
- •US-led international AI standards body modeled on FINRA
- •Frontier labs submit models ~30 days pre-release (voluntary at first)
- •Risk evaluation focus: cyber, national security, bio threats, other high-risk domains
- •Quarterly benchmark updates and potential coordinated “slowdown” mechanism
- •Broad support signaled by major AI and tech leaders
- 4:18 – 7:16
Why SROs could work: faster, expert-led oversight vs. slow-moving government rules
Friedberg explains what an SRO is and why AI’s pace makes traditional legislation brittle. He argues an SRO can adapt tests rapidly, use real experts, and maintain government oversight without ceding operational control to a new agency.
- •SRO definition and finance analogies (FINRA/NFA)
- •Government-written rules quickly become outdated in fast-moving tech
- •Independent experts can test for cyber/bio/weapons/social manipulation risks
- •Federal oversight without building a new “AI agency” bureaucracy
- •Industry alignment: oversight with less chance of innovation bottlenecks
- 7:16 – 15:39
Sacks’ five conditions to prevent regulatory capture (and avoid an 'FAA for AI')
Sacks says he could support an SRO if it meets strict guardrails: broad representation, only frontier coverage, catastrophic-risk scope, voluntary proof-of-work, and substituting (not adding) bureaucracy. He warns that an FAA-like approval regime would impose multi-year delays and effectively surrender the AI race.
- •Condition 1: broad representation incl. startups/open source to reduce capture
- •Condition 2: only true frontier models; don’t block incremental/lesser models
- •Condition 3: focus on catastrophic risks (cyber + CBRN), not speech moderation
- •Condition 4: voluntary first; prove efficacy before mandating
- •Condition 5: must replace (not add to) new agencies; FAA analogy implies 5–9 year delays
- 15:39 – 19:58
Anthropic, state-by-state rulemaking, and fears of a patchwork 'ratchet' strategy
The discussion shifts to state-level regulation and accusations of regulatory capture. Sacks cites reporting that Anthropic encourages increasingly strict state AI rules rather than a single national framework, and argues tech concessions invite escalating government control unless preemption and clear lines are set.
- •Claim: Anthropic pursuing tougher state guardrails via one-upmanship
- •Patchwork regulations vs. a single national framework
- •Argument that 'please regulate us' invites more government power
- •Need for preemption if an SRO becomes the chosen line in the sand
- •Concerns about political pressure and where an SRO would report inside government
- 19:58 – 24:17
Stripe + Block + Advent bid for PayPal: stablecoins, scale, and a Visa/Mastercard challenger
The hosts unpack a reported $53B offer structure and debate why PayPal is attractive despite aging product UX. The core thesis: combining merchant rails (Stripe), consumer accounts (PayPal/Venmo), and POS + Cash App (Block) could enable a vertically integrated payments network with stablecoin capabilities.
- •Deal rumor details and confusion over participants (Stripe/Advent + Block)
- •PayPal’s scale: ~439M consumer accounts and major brands (Venmo, Braintree)
- •Stablecoin angle: Stripe’s Bridge acquisition vs PayPal’s PYUSD
- •Sacks: combined entity could compete with Visa/Mastercard via new rails
- •Antitrust discussion hinges on market definition (merchant APIs vs card networks)
- 24:17 – 37:51
A new M&A playbook: 'AI-ify' mature digital businesses and unlock value
Friedberg predicts a wave of take-private/roll-up style deals where modern operators use AI and cost discipline to revive stagnating internet-era companies. They cite examples like Bending Spoons and discuss how capital + elite operators can restructure spend, product, and marketing.
- •Trend: mature, non-founder-led digital businesses become acquisition targets
- •AI and operational rigor as the key synergy (automation, efficiency, better UX)
- •Capital structure mechanics and who operates post-close
- •Bending Spoons as an operational model for revitalizing legacy assets
- •Macro environment: renewed corporate M&A appetite and liquidity dynamics
- 37:51 – 42:44
Apple sues OpenAI over alleged trade-secret theft tied to consumer hardware
Jason summarizes Apple’s lawsuit alleging OpenAI improperly obtained Apple IP via hiring and interview processes. The group emphasizes the key boundary: employees can bring what’s in their head, but not documents, parts, or data from prior employers.
- •Apple’s 41-page complaint and alleged interview 'show-and-tell' behavior
- •High-profile hiring (e.g., former Apple design leadership at OpenAI)
- •Reported scale of Apple employee poaching
- •Industry norm: no non-competes in CA, but trade-secret law still applies
- •Rule of thumb: take only knowledge/memory, never files/parts/credentials
- 42:44 – 47:41
Grok Build leak and the brittle reality of 'zero data retention' promises
They discuss reports that Grok Build uploaded more than necessary—potentially entire codebases—contrary to user expectations. The takeaway is broader: AI privacy guarantees are hard to validate, and enterprises may need third-party layers, strict boundaries, and tenant-controlled evaluation loops.
- •Reported behavior: broader-than-needed uploads during coding sessions
- •Rapid mitigation steps and claims of deletion; harness open-sourced afterward
- •Chamath: privacy in AI is fragile; 'ZDR' isn’t a guarantee
- •Sacks references enterprise playbook: trust boundaries, private evals, decoupled orchestration
- •Growing ecosystem pushback against monolithic closed-model stacks
- 47:41 – 59:53
Token economics shock: CFOs react as AI spend explodes (Ramp 'token maxing')
The conversation turns to cost: token spend growth, big price dispersion across models, and the new need for governance. A Ramp clip illustrates CFO pain—AI is a 'tab' employees can run up—driving new tooling to monitor, budget, and rate-limit model usage.
- •Ramp: token spend among customers up 21x
- •Model cost dispersion: premium frontier vs cheap open/Chinese models
- •CFO vs engineer incentives: performance vs ROI and budget discipline
- •Prediction: uncontrolled token spend could cause earnings misses
- •Mira Murati/Inkling positioning: 'just-below-frontier' + fine-tuning cheaper models
- 59:53 – 1:11:25
NY moratorium on hyperscale data centers: land, water, power, and politics
They criticize New York’s proposed pause, disputing claims about water usage, noise, and pollution, and argue data centers can be efficient and economically beneficial. The panel frames data center buildout as a national competitiveness issue tied directly to AI capability and energy strategy.
- •Debunking common critiques: closed-loop water systems, noise distance, land efficiency
- •Energy strategy: behind-the-meter generation vs grid competition
- •Economic upside: tax revenue, construction jobs, ongoing operations roles
- •Argument that moratoriums slow innovation and push investment to other states/countries
- •Compute will 'chase energy' if jurisdictions block buildouts
- 1:11:25 – 1:21:33
Influence campaigns and moral panic: from anti-GMO playbooks to anti-data-center sentiment
Friedberg draws a parallel between anti-GMO sentiment and media-driven influence operations, suggesting similar dynamics may be shaping public opposition to AI infrastructure. Sacks cites OpenAI’s claims about PRC-linked operations targeting US AI debates, tying infrastructure delays to AI-race geopolitics.
- •Friedberg’s Google Trends/RT example: sentiment spikes tied to media propagation
- •Concept of 'directed measures' and activism amplification loops
- •Polls: many Americans believe data centers raise utility/water costs regardless of facts
- •Sacks: references OpenAI post on PRC-linked influence targeting AI debates
- •Broader warning: moral panic may trigger self-sabotaging regulation
- 1:21:33 – 1:29:54
Science Corner: reversing aging by clearing extracellular 'gunk' with AI-designed enzymes
Friedberg explains a new paper (Calico + Revales) targeting extracellular aging via advanced glycation end products (CML). Using AlphaFold and directed evolution, researchers engineered an enzyme that degrades CML and demonstrated large reductions in human tissue—hinting at future therapies and huge cosmetic markets.
- •Aging focus shifts from cells to extracellular matrix and glycation buildup
- •Target molecule: CML (an advanced glycation end product)
- •Workflow: AlphaFold candidate → iterative directed evolution → high-throughput testing
- •Results: large CML degradation across proteins; human skin tests show major reduction
- •Commercial path: topical creams first, then systemic delivery (shots/RNA/protein)
